Israeli-American rabbi ridicules Hindu beliefs over cow worship, faces backlash | Video

US midterms 2026: Michigan Democrats thought they had made peace. Then Haley Stevens renewed clash with Abdul El-Sayed

All votes counted, Netanyahu’s Likud emerges as biggest party with 27 seats; majority eludes both blocs

iPhone Air 2 may get 2 rear cameras, smaller Dynamic Island and A20 Pro chip: What to expect

Turkish ambassador Ali Murat Ersoy says India and Turkey ‘cannot remain strangers’

Air Marshal Ashutosh Dixit takes charge as India’s 29th Air Force Chief, succeeds AP Singh

'One little banana heist': UP Police shares elephant mom-calf clip to teach honesty, internet teaches it a lesson instead | Viral video

ChatGPT blamed for Flavio Bolsonaro's bizarre 'president-elect' claim: 5 Times AI use made a fool of global leaders | In Pics

Kagiso Rabada completes 350 Test wickets, joins elite club in Cape Town Test vs Australia

IND vs NZ: Shreyas Iyer breaks record for most T20I runs by an Indian captain in a year

Pakistan hikes diesel amid West Asia crisis; essential goods likely to cost more

Diwali at the Jonas'! Priyanka Chopra and Nick Host party, Winnie Harlow and others attend

Trump says US 'looking at' possible North Korean role in $6.3 billion Binance crypto heist

Prediction markets are betting big on the US midterms. But can they get it right?

ISRO's GSLV-F18 launch success revives NAVIC, but there's a catch...

Omar Abdullah marries journalist Nidhi Razdan, announces ‘new journey’ on social media

'Make s**t houses great again': Norway hits back at Trump's insult with genius level toilet humour

Did rebels shoot down helicopter carrying Lt Gen Sanjay Mitra's family in Uttarakhand? Indian Army fact-check reveals...

Director Sanjay Gupta, who made underworld look cool on screen, died after bathroom fall

Women hold just 17.4% of Trump's cabinet-level posts, down from 52% under Biden: Report

INDIA bloc holds symbolic ‘public impeachment’ of CEC Gyanesh Kumar, lists 6 charges | Watch

Arijit Singh and Karan Aujla to perform at ‘HOME’ in Goa: Read details here

Ilhan Omar's alleged brother and ex-husband calls himself a 'gay man' in academic paper - 'extremely feminine Somali guy'

Captured on CCTV video: Six-year-old special needs girl brutally attacked by Class III student at Panchkula school

Tehran may move enriched uranium stockpile to another country, open Hormuz under new proposal: Report

US midterms: Early voting crosses record 47.9 million, Democrats lead Republicans

Federal judge blocks Trump administration's election pressure tactics in sweeping emergency order

Physicist predicted Friday the 13th 'doomsday' in 1960. It is 13 days away

WPL 2027 auction: Top 5 most expensive signings

BJP to stage protest over anti-Sikh riots on November 1, same day as INDIA bloc’s ‘Chalo Dilli’ rally

Poland in political turmoil as Karol Nawrocki returns mid-flight amid Constitutional Court 'coup d'etat' claims. What's happening?

India slams energy supply disruptions, says oil trade cannot be held hostage

Doomsday warning? Why governments are warning people to stockpile emergency supplies for 72 hours

Putin's war warning shakes Europe: NATO builds war bunkers, Britain stockpiles food | Gravitas

'Shut up, shut up. Dope. Shut up. Loser': Trump lashes out while getting heckled again | WATCH

Enterprises are moving from paper-based and spreadsheet-heavy purchasing to intelligent procurement, empowered by growing technology, including artificial intelligence (AI), machine learning (ML), and robotic process automation (RPA), in order to transform the entire source-to-pay (S2P) lifecycle. Several firms have shifted from using AI merely to analyse and spend toward deploying it as an operator, as agents now create purchase orders, route multi-level approvals, and flag supplier anomalies before any involvement of humans. Interestingly, teams in the firms deploying AI typically cut cycle times by 40 to 60 per cent and improve productivity and efficiency of the result more precisely without consuming extra effort and time. While AI, automation and cloud platforms are becoming increasingly accessible, many organisations continue to operate with legacy systems, fragmented data and processes that were not designed for today’s pace of business. For large enterprises, the challenge is no longer simply about investing in newer technology. It is about connecting existing systems, improving the quality of data and redesigning workflows so that technology can deliver measurable improvements in productivity, efficiency and decision-making. This shift is particularly visible in functions such as procurement, where organisations are looking to move beyond manual processes and disconnected systems towards more intelligent, integrated and user-friendly experiences. At the same time, the growing use of AI is raising new questions around governance, security, human oversight and how businesses can translate experimentation into tangible operational value. In a conversation with WION, Trishna Patel, Head, Ewiz Procure, Powerweave, and Reena Ravindran, VP – Delivery, Powerweave, discuss the major challenges enterprises experience in modernising legacy environments and the changing expectations around procurement. They also focused on how AI can deliver measurable business outcomes, and the importance of balancing automation with human judgement and responsible technology adoption.&nbsp;1. Many enterprises continue to operate with legacy systems and fragmented data. What are the biggest challenges they face when modernising their technology infrastructure?&nbsp;Trishna Patel, Head- Ewiz Procureat Powerweave: One of the biggest misconceptions is that modernisation means replacing legacy technology. In reality, the bigger issue is often the fragmentation of data, processes and workflows around it. Large enterprises have invested heavily in ERP and other core platforms, but many still operate with disconnected systems, inconsistent master data and manual processes outside those platforms. Modernisation is therefore as much a data and operating-model challenge as a technology one.&nbsp;The starting point should be the business problem: where is the process breaking down, where is data unreliable, and where are manual interventions limiting visibility or efficiency?&nbsp;"We see this often in procurement. With ewiz procure, we work alongside systems such as SAP, Oracle and Coupa. We first understand the business objective and process gaps, then fix the underlying data and governance, automate the right workflows, and introduce AI where it can add meaningful value. That is where enterprise modernisation is heading: not replacement for replacement’s sake, but making existing investments work harder through cleaner data, better-connected processes and more intelligent automation," Trishna said.&nbsp;2. With AI becoming more accessible, how can businesses move beyond experimentation and translate AI investments into measurable operational outcomes?&nbsp;Reena Ravindran, VP-Delivery at Powerweave: AI will create real value when businesses stop treating it as a standalone technology experiment and start embedding it into the way work actually gets done. The right question is not how many AI pilots an organisation has launched, but which business metric has improved because of them. To get there, AI needs to sit inside a defined workflow where it can influence decisions, automate repetitive steps and improve cycle times. That also requires the right foundation: reliable data, clear process ownership, team training, change management and controlled evaluation. Without those elements, even a strong AI capability can struggle to deliver sustained operational value. We have seen this in practice. In one chemical manufacturing deployment, the results came not from AI alone, but from redesigning the procurement workflow around structured data, automation, system integration and digital decision support. PR-to-PO processing time was reduced by 87%, while supplier onboarding time dropped by 95%. I think that is the real shift enterprises need to make. AI should not sit beside the process as another tool. It needs to be embedded into the workflow, supported by good data, governance and adoption. That is what turns experimentation into measurable business outcomes. 3. How are customer expectations evolving in procurement, particularly among manufacturers, distributors and wholesalers?&nbsp;Trishna Patel, Head- Ewiz Procure at Powerweave: Procurement users are no longer comparing their experience only with other B2B systems. They are comparing it with the best digital experiences they use every day, raising expectations around speed, visibility, self-service and ease of use. They expect every interaction with a procurement tool to feel seamless and intuitive, whether they are searching for a product, raising a request, approving a purchase or onboarding a supplier. With tools like ChatGPT shaping user behaviour, that experience is also becoming more conversational. Users increasingly expect to describe what they need in natural language and be guided to the right action without navigating complex menus. For manufacturers, distributors and wholesalers, the challenge is to hide complexity rather than expose it. Data, workflows, approvals and intelligence need to work together behind the scenes while the user gets a simple, connected experience. The competitive advantage will increasingly come from removing friction across the procurement journey while improving control, visibility and compliance. 4. How does Powerweave measure the impact of its solutions for customers? Could you share examples of business challenges addressed and outcomes achieved?&nbsp;Reena Ravindran, VP-Delivery at Powerweave: Technology should never be the outcome. The outcome is the only reason to deploy technology in the first place. We therefore establish a baseline before measuring impact. In ewiz procure, customers have seen 8–12% savings on indirect procurement, sourcing cycles around 30% faster and 27% lower process costs. In one procurement data cleaning project, we helped uncover €40M vendor-consolidation opportunity for a global manufacturer. The opportunity was already there. We just helped them uncover it. In awspromostack, AI has reduced order-entry effort by 40% and art and booking effort by 60%, while storefront and product-listing creation is 80% faster. Our operations process more than 40,000 orders a day at over 98% accuracy. Across these examples, the principle is consistent: we measure what changed in the business, not how much technology was deployed 5. With customers across multiple geographies, how does Powerweave adapt its solutions to different markets, regulatory environments and business requirements?&nbsp;Trishna Patel, Head- Ewiz Procure, Powerweave: Global technology does not mean identical technology everywhere. The foundation can be common, but the context cannot. Regulations, workflows, data structures and customer expectations vary significantly by market. The real capability lies in knowing what should be standardised and what must remain local. Snowkap is a good example. An Indian business may operate under BRSR and BRSR Core while also responding to frameworks such as CSRD, GRI, TCFD or IFRS when serving global markets. The same principle applies to procurement. ewiz procure works with the customer’s existing ERP, S2P and supplier landscape, while adapting workflows, data structures, approvals and governance to local business requirements rather than forcing a single operating model everywhere. Our approach is therefore to build adaptable technology around the customer’s reality. Scale comes from a strong common foundation; relevance comes from understanding the market, regulations and operating model in which that foundation has to work. 6. As AI becomes more deeply integrated into business operations, what responsibilities do technology leaders have in ensuring its adoption is practical, secure and aligned with business objectives?&nbsp;Reena Ravindran, VP-Delivery, Powerweave: The responsibility of a technology leader is no longer simply to ask whether AI can be deployed. It is to decide where AI can create measurable business value, how it should fit into existing workflows, and where human judgement must remain. AI is highly effective at processing information, identifying patterns and automating repetitive work. But relationships, commercial judgement, risk and strategic decisions often depend on context that cannot be captured in data alone. That is why we believe in progressive adoption: automate administrative work first, augment human decision-making next, and introduce greater autonomy only where the controls, data and context support it. Security and governance are equally fundamental. That means clear access controls, data governance, auditability, defined decision rights and human oversight, alongside strong enterprise security standards. Powerweave's platforms and operations are ISO 27001 certified and SOC 2 Type II audited. Responsible AI ultimately comes down to four things: a clear business purpose, reliable data, appropriate controls and clear accountability for the outcome.&nbsp; &nbsp;

The shift is particularly visible in functions such as procurement, where organisations are looking to move beyond manual processes and disconnected systems towards more intelligent, integrated and user-friendly experiences. 

Enterprises are moving from paper-based and spreadsheet-heavy purchasing to intelligent procurement, empowered by growing technology, including artificial intelligence (AI), machine learning (ML), and robotic process automation (RPA), in order to transform the entire source-to-pay (S2P) lifecycle. Several firms have shifted from using AI merely to analyse and spend toward deploying it as an operator, as agents now create purchase orders, route multi-level approvals, and flag supplier anomalies before any involvement of humans. Interestingly, teams in the firms deploying AI typically cut cycle times by 40 to 60 per cent and improve productivity and efficiency of the result more precisely without consuming extra effort and time. While AI, automation and cloud platforms are becoming increasingly accessible, many organisations continue to operate with legacy systems, fragmented data and processes that were not designed for today’s pace of business. For large enterprises, the challenge is no longer simply about investing in newer technology. It is about connecting existing systems, improving the quality of data and redesigning workflows so that technology can deliver measurable improvements in productivity, efficiency and decision-making. This shift is particularly visible in functions such as procurement, where organisations are looking to move beyond manual processes and disconnected systems towards more intelligent, integrated and user-friendly experiences. At the same time, the growing use of AI is raising new questions around governance, security, human oversight and how businesses can translate experimentation into tangible operational value. In a conversation with WION, Trishna Patel, Head, Ewiz Procure, Powerweave, and Reena Ravindran, VP – Delivery, Powerweave, discuss the major challenges enterprises experience in modernising legacy environments and the changing expectations around procurement. They also focused on how AI can deliver measurable business outcomes, and the importance of balancing automation with human judgement and responsible technology adoption. 1. Many enterprises continue to operate with legacy systems and fragmented data. What are the biggest challenges they face when modernising their technology infrastructure? Trishna Patel, Head- Ewiz Procureat Powerweave: One of the biggest misconceptions is that modernisation means replacing legacy technology. In reality, the bigger issue is often the fragmentation of data, processes and workflows around it. Large enterprises have invested heavily in ERP and other core platforms, but many still operate with disconnected systems, inconsistent master data and manual processes outside those platforms. Modernisation is therefore as much a data and operating-model challenge as a technology one. The starting point should be the business problem: where is the process breaking down, where is data unreliable, and where are manual interventions limiting visibility or efficiency? "We see this often in procurement. With ewiz procure, we work alongside systems such as SAP, Oracle and Coupa. We first understand the business objective and process gaps, then fix the underlying data and governance, automate the right workflows, and introduce AI where it can add meaningful value. That is where enterprise modernisation is heading: not replacement for replacement’s sake, but making existing investments work harder through cleaner data, better-connected processes and more intelligent automation," Trishna said. 2. With AI becoming more accessible, how can businesses move beyond experimentation and translate AI investments into measurable operational outcomes? Reena Ravindran, VP-Delivery at Powerweave: AI will create real value when businesses stop treating it as a standalone technology experiment and start embedding it into the way work actually gets done. The right question is not how many AI pilots an organisation has launched, but which business metric has improved because of them. To get there, AI needs to sit inside a defined workflow where it can influence decisions, automate repetitive steps and improve cycle times. That also requires the right foundation: reliable data, clear process ownership, team training, change management and controlled evaluation. Without those elements, even a strong AI capability can struggle to deliver sustained operational value. We have seen this in practice. In one chemical manufacturing deployment, the results came not from AI alone, but from redesigning the procurement workflow around structured data, automation, system integration and digital decision support. PR-to-PO processing time was reduced by 87%, while supplier onboarding time dropped by 95%. I think that is the real shift enterprises need to make. AI should not sit beside the process as another tool. It needs to be embedded into the workflow, supported by good data, governance and adoption. That is what turns experimentation into measurable business outcomes. 3. How are customer expectations evolving in procurement, particularly among manufacturers, distributors and wholesalers? Trishna Patel, Head- Ewiz Procure at Powerweave: Procurement users are no longer comparing their experience only with other B2B systems. They are comparing it with the best digital experiences they use every day, raising expectations around speed, visibility, self-service and ease of use. They expect every interaction with a procurement tool to feel seamless and intuitive, whether they are searching for a product, raising a request, approving a purchase or onboarding a supplier. With tools like ChatGPT shaping user behaviour, that experience is also becoming more conversational. Users increasingly expect to describe what they need in natural language and be guided to the right action without navigating complex menus. For manufacturers, distributors and wholesalers, the challenge is to hide complexity rather than expose it. Data, workflows, approvals and intelligence need to work together behind the scenes while the user gets a simple, connected experience. The competitive advantage will increasingly come from removing friction across the procurement journey while improving control, visibility and compliance. 4. How does Powerweave measure the impact of its solutions for customers? Could you share examples of business challenges addressed and outcomes achieved? Reena Ravindran, VP-Delivery at Powerweave: Technology should never be the outcome. The outcome is the only reason to deploy technology in the first place. We therefore establish a baseline before measuring impact. In ewiz procure, customers have seen 8–12% savings on indirect procurement, sourcing cycles around 30% faster and 27% lower process costs. In one procurement data cleaning project, we helped uncover €40M vendor-consolidation opportunity for a global manufacturer. The opportunity was already there. We just helped them uncover it. In awspromostack, AI has reduced order-entry effort by 40% and art and booking effort by 60%, while storefront and product-listing creation is 80% faster. Our operations process more than 40,000 orders a day at over 98% accuracy. Across these examples, the principle is consistent: we measure what changed in the business, not how much technology was deployed 5. With customers across multiple geographies, how does Powerweave adapt its solutions to different markets, regulatory environments and business requirements? Trishna Patel, Head- Ewiz Procure, Powerweave: Global technology does not mean identical technology everywhere. The foundation can be common, but the context cannot. Regulations, workflows, data structures and customer expectations vary significantly by market. The real capability lies in knowing what should be standardised and what must remain local. Snowkap is a good example. An Indian business may operate under BRSR and BRSR Core while also responding to frameworks such as CSRD, GRI, TCFD or IFRS when serving global markets. The same principle applies to procurement. ewiz procure works with the customer’s existing ERP, S2P and supplier landscape, while adapting workflows, data structures, approvals and governance to local business requirements rather than forcing a single operating model everywhere. Our approach is therefore to build adaptable technology around the customer’s reality. Scale comes from a strong common foundation; relevance comes from understanding the market, regulations and operating model in which that foundation has to work. 6. As AI becomes more deeply integrated into business operations, what responsibilities do technology leaders have in ensuring its adoption is practical, secure and aligned with business objectives? Reena Ravindran, VP-Delivery, Powerweave: The responsibility of a technology leader is no longer simply to ask whether AI can be deployed. It is to decide where AI can create measurable business value, how it should fit into existing workflows, and where human judgement must remain. AI is highly effective at processing information, identifying patterns and automating repetitive work. But relationships, commercial judgement, risk and strategic decisions often depend on context that cannot be captured in data alone. That is why we believe in progressive adoption: automate administrative work first, augment human decision-making next, and introduce greater autonomy only where the controls, data and context support it. Security and governance are equally fundamental. That means clear access controls, data governance, auditability, defined decision rights and human oversight, alongside strong enterprise security standards. Powerweave's platforms and operations are ISO 27001 certified and SOC 2 Type II audited. Responsible AI ultimately comes down to four things: a clear business purpose, reliable data, appropriate controls and clear accountability for the outcome.

The shift is particularly visible in functions such as procurement, where organisations are looking to move beyond manual processes and disconnected systems towards more intelligent, integrated and user-friendly experiences.&nbsp;

From manual processes to intelligent procurement: How are enterprises embracing AI and automation

Apple could be preparing a significant upgrade for its next slim smartphone. The iPhone Air 2 is reportedly expected to launch in March 2027, alongside the standard iPhone 18 and more affordable iPhone 18e. The upcoming model could address some of the limitations of the first-generation iPhone Air with a second rear camera, a more powerful processor and better battery life.&nbsp;The expected launch is also part of a possible change in Apple's iPhone release strategy. Instead of launching its entire line-up together in September, the company could split launches between the second half and the first half of the following year.<h2>iPhone Air 2 Launch: When Is Apple Expected To Announce It?</h2>According to reports, Apple is planning to introduce the iPhone Air 2 in the first quarter of 2027, with March being a possible launch window.The smartphone could arrive alongside the iPhone 18 and iPhone 18e. Meanwhile, the iPhone 18 Pro models and Apple's foldable iPhone are expected to remain part of the September launch cycle.Apple has not officially confirmed these plans or announced a launch date.<h2>iPhone Air 2 May Get A Second 48MP Camera</h2>One of the biggest changes could be on the back of the phone. The first-generation iPhone Air features a single 48-megapixel rear camera. Its successor could add a second 48-megapixel ultra-wide camera, according to early reports. This would give users more flexibility when capturing landscapes, group photographs and wider scenes without moving further away from the subject. The phone is also expected to retain a 6.5-inch display, although a smaller Dynamic Island could free up more usable screen space.<h2>A20 Pro Chip And Better Battery Life</h2>Apple could equip the iPhone Air 2 with its upcoming A20 Pro processor, reportedly based on TSMC's 2nm manufacturing process. A move to a newer chip could deliver improvements in processing speed and power efficiency compared with the A19 Pro. Better efficiency could also help address battery-life concerns associated with the first iPhone Air. Apple is reportedly exploring battery improvements, but it remains unclear whether these will come from a larger battery, chip efficiency or a combination of both.For now, the iPhone Air 2 remains a rumoured product. If these upgrades materialise, Apple's next slim iPhone could offer a more balanced combination of design, camera capabilities and battery performance.

Apple's iPhone Air 2 could launch in March 2027 alongside the iPhone 18 and iPhone 18e. Leaks suggest a second 48MP rear camera, smaller Dynamic Island, A20 Pro chip and improved battery life, although Apple has not confirmed the details.

Apple could be preparing a significant upgrade for its next slim smartphone. The iPhone Air 2 is reportedly expected to launch in March 2027, alongside the standard iPhone 18 and more affordable iPhone 18e. The upcoming model could address some of the limitations of the first-generation iPhone Air with a second rear camera, a more powerful processor and better battery life. The expected launch is also part of a possible change in Apple's iPhone release strategy. Instead of launching its entire line-up together in September, the company could split launches between the second half and the first half of the following year. iPhone Air 2 Launch: When Is Apple Expected To Announce It? According to reports, Apple is planning to introduce the iPhone Air 2 in the first quarter of 2027, with March being a possible launch window. The smartphone could arrive alongside the iPhone 18 and iPhone 18e. Meanwhile, the iPhone 18 Pro models and Apple's foldable iPhone are expected to remain part of the September launch cycle. Apple has not officially confirmed these plans or announced a launch date. iPhone Air 2 May Get A Second 48MP Camera One of the biggest changes could be on the back of the phone. The first-generation iPhone Air features a single 48-megapixel rear camera. Its successor could add a second 48-megapixel ultra-wide camera, according to early reports. This would give users more flexibility when capturing landscapes, group photographs and wider scenes without moving further away from the subject. The phone is also expected to retain a 6.5-inch display, although a smaller Dynamic Island could free up more usable screen space. A20 Pro Chip And Better Battery Life Apple could equip the iPhone Air 2 with its upcoming A20 Pro processor, reportedly based on TSMC's 2nm manufacturing process. A move to a newer chip could deliver improvements in processing speed and power efficiency compared with the A19 Pro. Better efficiency could also help address battery-life concerns associated with the first iPhone Air. Apple is reportedly exploring battery improvements, but it remains unclear whether these will come from a larger battery, chip efficiency or a combination of both. For now, the iPhone Air 2 remains a rumoured product. If these upgrades materialise, Apple's next slim iPhone could offer a more balanced combination of design, camera capabilities and battery performance.

Apple's iPhone Air 2 could launch in March 2027 alongside the iPhone 18 and iPhone 18e. Leaks suggest a second 48MP rear camera, smaller Dynamic Island, A20 Pro chip and improved battery life, although Apple has not confirmed the details.

Chinese AI companies DeepSeek and Moonshot AI reportedly used outputs from OpenAI’s GPT-6 Astra to train their latest artificial intelligence models, according to a report by The Information. The development has brought the controversial practice of AI distillation back into focus as competition between US and Chinese technology companies intensifies. The report said Moonshot AI’s Kimi K3.1 Pro, released last week, and DeepSeek’s upgraded V4.1-Flash 1017 model, launched on October 17, benefited from distilling GPT-6 Astra. The companies are yet to publicly respond to the report. OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters’ requests for comment.<h2>What Is AI Distillation And Why Is It Controversial?</h2>AI distillation is a technique in which developers use the outputs of a more advanced AI model to help train a smaller or competing system.In simple terms, developers can ask an established AI model to solve problems, write code or respond to complex prompts. They can then use those responses as training examples for another model. This can help developers reproduce some capabilities of a powerful AI system without having to build everything from scratch. It can also reduce the time and cost involved in developing advanced models. However, the practice has become controversial. Leading AI companies argue that large-scale extraction of model outputs can allow competitors to benefit from capabilities developed through substantial research and investment.<h2>OpenAI Had Previously Flagged Moonshot-Linked Activity</h2>The latest report follows an earlier statement from OpenAI. In a September 30 blog post, the company said it had linked a core part of a distillation campaign, which began in July, to individuals associated with Moonshot AI. OpenAI has raised concerns about attempts to use its models to develop competing systems through large-scale distillation. The latest report adds DeepSeek to the discussion, although the claims about the two companies' use of GPT-6 Astra have not been independently confirmed in the information provided.<h2>Anthropic Says New Measures Cut Distillation Data Collection</h2>The dispute is not limited to OpenAI. Anthropic and Google have also raised concerns about competitors using their models to reproduce advanced AI capabilities. Earlier this month, Anthropic said new safeguards designed to prevent distillation attacks had reduced the amount of data harvested from its Claude models by around 90 to 99 per cent. The issue highlights a growing challenge for AI developers: protecting the capabilities of their models while making them available to users and developers. As companies compete to build more powerful AI systems, control over model outputs and training data is becoming an increasingly important part of the global AI race.

DeepSeek and Moonshot AI reportedly used OpenAI’s GPT-6 Astra outputs to train their latest models, according to The Information. The report has renewed concerns over AI distillation, a technique that can help competitors reproduce advanced AI capabilities.

Chinese AI companies DeepSeek and Moonshot AI reportedly used outputs from OpenAI’s GPT-6 Astra to train their latest artificial intelligence models, according to a report by The Information. The development has brought the controversial practice of AI distillation back into focus as competition between US and Chinese technology companies intensifies. The report said Moonshot AI’s Kimi K3.1 Pro, released last week, and DeepSeek’s upgraded V4.1-Flash 1017 model, launched on October 17, benefited from distilling GPT-6 Astra. The companies are yet to publicly respond to the report. OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters’ requests for comment. What Is AI Distillation And Why Is It Controversial? AI distillation is a technique in which developers use the outputs of a more advanced AI model to help train a smaller or competing system. In simple terms, developers can ask an established AI model to solve problems, write code or respond to complex prompts. They can then use those responses as training examples for another model. This can help developers reproduce some capabilities of a powerful AI system without having to build everything from scratch. It can also reduce the time and cost involved in developing advanced models. However, the practice has become controversial. Leading AI companies argue that large-scale extraction of model outputs can allow competitors to benefit from capabilities developed through substantial research and investment. OpenAI Had Previously Flagged Moonshot-Linked Activity The latest report follows an earlier statement from OpenAI. In a September 30 blog post, the company said it had linked a core part of a distillation campaign, which began in July, to individuals associated with Moonshot AI. OpenAI has raised concerns about attempts to use its models to develop competing systems through large-scale distillation. The latest report adds DeepSeek to the discussion, although the claims about the two companies' use of GPT-6 Astra have not been independently confirmed in the information provided. Anthropic Says New Measures Cut Distillation Data Collection The dispute is not limited to OpenAI. Anthropic and Google have also raised concerns about competitors using their models to reproduce advanced AI capabilities. Earlier this month, Anthropic said new safeguards designed to prevent distillation attacks had reduced the amount of data harvested from its Claude models by around 90 to 99 per cent. The issue highlights a growing challenge for AI developers: protecting the capabilities of their models while making them available to users and developers. As companies compete to build more powerful AI systems, control over model outputs and training data is becoming an increasingly important part of the global AI race.

DeepSeek and Moonshot AI reportedly used OpenAI’s GPT-6 Astra outputs to train their latest models, according to The Information. The report has renewed concerns over AI distillation, a technique that can help competitors reproduce advanced AI capabilities.

What is AI distillation? DeepSeek and Moonshot reportedly used OpenAI’s model to train AI

Google has unveiled Gemini 4 Carbon, a new AI model built to tackle difficult coding tasks and help cybersecurity teams find and fix software weaknesses. Announced on October 28, 2026, it follows Gemini 4 Argon by just two weeks, signalling Google’s push to develop more capable AI systems.&nbsp;&nbsp;The company says Carbon ranks first on the latest version of the Artificial Analysis Intelligence Index, with a score of 59. It also reports leading results across several cybersecurity benchmarks. However, the model is not yet available to everyone: Google plans to begin with a limited group of cyber defenders before expanding access.<h2>Gemini 4 Carbon: What Can It Do?</h2>Gemini 4 Carbon is designed for long, complicated tasks that may require an AI system to work through several steps, rather than simply answer a question. Google says the model can help identify subtle bugs across large codebases, trace their causes and develop fixes. It is also designed to support full-stack web development and scientific work.&nbsp;&nbsp;The cybersecurity focus is particularly significant. Software flaws can expose organisations to attacks, and finding them early can help developers strengthen their systems before vulnerabilities are exploited.<h2>Google Shares Three Key Cybersecurity Scores</h2>Google reported the following results for Gemini 4 Carbon:- 88.2% on SEC-bench Pro: identifying software vulnerabilities.- 54.8% on ExploitGym: assessing whether vulnerabilities can be exploited.- 71.0% on CWE-bench: developing fixes for security weaknesses.These tests cover different stages of cybersecurity work, from discovering flaws to assessing risk and patching the underlying problem. Google says the results demonstrate Carbon’s capabilities across the cybersecurity defence process.&nbsp;&nbsp;The company also reported internal improvements in secure code generation compared with Argon. These are Google’s published results; benchmark scores alone do not guarantee that an AI model will find every vulnerability or produce a correct fix in real-world systems.<h2>Gemini 4 Carbon vs GPT-6 Astra: What About Availability?</h2>The supplied comparison says Gemini 4 Carbon outperformed OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 on selected coding and cybersecurity tests. Those comparisons should be treated cautiously unless the specific benchmark results and testing conditions can be independently confirmed.&nbsp;For now, Google is taking a controlled approach to release. A limited group of cybersecurity defenders is due to receive access first, followed by additional trusted partners through its Fairwind Program. Wider access for developers, businesses and consumers is planned for a later stage.&nbsp;&nbsp;The practical impact will depend on how reliably Carbon works outside benchmark tests. If it can help security teams find and repair flaws faster, it could become a useful tool for software development and cyber defence. But its performance in everyday use will matter more than any single leaderboard position.

Google has launched Gemini 4 Carbon, an AI model focused on complex coding and cybersecurity. Google reports an 88.2% score on SEC-bench Pro for finding vulnerabilities. Access will initially be limited to trusted cyber defenders before a wider rollout.

Google has unveiled Gemini 4 Carbon, a new AI model built to tackle difficult coding tasks and help cybersecurity teams find and fix software weaknesses. Announced on October 28, 2026, it follows Gemini 4 Argon by just two weeks, signalling Google’s push to develop more capable AI systems. The company says Carbon ranks first on the latest version of the Artificial Analysis Intelligence Index, with a score of 59. It also reports leading results across several cybersecurity benchmarks. However, the model is not yet available to everyone: Google plans to begin with a limited group of cyber defenders before expanding access. Gemini 4 Carbon: What Can It Do? Gemini 4 Carbon is designed for long, complicated tasks that may require an AI system to work through several steps, rather than simply answer a question. Google says the model can help identify subtle bugs across large codebases, trace their causes and develop fixes. It is also designed to support full-stack web development and scientific work. The cybersecurity focus is particularly significant. Software flaws can expose organisations to attacks, and finding them early can help developers strengthen their systems before vulnerabilities are exploited. Google Shares Three Key Cybersecurity Scores Google reported the following results for Gemini 4 Carbon: - 88.2% on SEC-bench Pro: identifying software vulnerabilities. - 54.8% on ExploitGym: assessing whether vulnerabilities can be exploited. - 71.0% on CWE-bench: developing fixes for security weaknesses. These tests cover different stages of cybersecurity work, from discovering flaws to assessing risk and patching the underlying problem. Google says the results demonstrate Carbon’s capabilities across the cybersecurity defence process. The company also reported internal improvements in secure code generation compared with Argon. These are Google’s published results; benchmark scores alone do not guarantee that an AI model will find every vulnerability or produce a correct fix in real-world systems. Gemini 4 Carbon vs GPT-6 Astra: What About Availability? The supplied comparison says Gemini 4 Carbon outperformed OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 on selected coding and cybersecurity tests. Those comparisons should be treated cautiously unless the specific benchmark results and testing conditions can be independently confirmed. For now, Google is taking a controlled approach to release. A limited group of cybersecurity defenders is due to receive access first, followed by additional trusted partners through its Fairwind Program. Wider access for developers, businesses and consumers is planned for a later stage. The practical impact will depend on how reliably Carbon works outside benchmark tests. If it can help security teams find and repair flaws faster, it could become a useful tool for software development and cyber defence. But its performance in everyday use will matter more than any single leaderboard position.

Google has launched Gemini 4 Carbon, an AI model focused on complex coding and cybersecurity. Google reports an 88.2% score on SEC-bench Pro for finding vulnerabilities. Access will initially be limited to trusted cyber defenders before a wider rollout.

Google unveils Gemini 4 Carbon just two weeks after Argon: Here’s what’s new

India is heading to build defences to keep pace with attackers who now use artificial intelligence to discover vulnerabilities, craft convincing phishing campaigns, and automate attacks within minutes. The Ministry of Electronics and Information Technology (MeitY) has acknowledged that generative AI is reducing the barrier for cybercriminals. In response, it has urged ministries, state governments, and regulators to urgently review their security posture as India strengthens its cyber defence framework against threats that are becoming faster, more sophisticated and harder to detect. Between June and July 2026, CERT-In, the Indian Computer Emergency Response Team, ran ten customised cybersecurity exercises and drills, with the theme of building resilience against frontier AI-driven cyber threats. Earlier this year, it also released targeted technical guidance, a 'Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure,' published in May 2026 to help protect digital infrastructure from fast-evolving AI threats. In April, CERT-In went further, issuing a high-risk advisory concerning frontier AI systems capable of allowing autonomous cyberattacks, including hacking and multi-stage, sophisticated malicious operations. In November 2025, MeitY released AI Governance Guidelines, laying out a framework made to bolster safe, inclusive, and accountable AI adoption across industries, as India still needs dedicated AI legislation. Three institutions anchor this architecture: the AI Governance Group for high-level coordination, the Technology and Policy Expert Committee for technical standards, and a proposed AI Safety Institute tasked with audits and certification. Crucially, incidents tied to AI are now folded into existing reporting rules; AI-related cybersecurity incidents must be reported to CERT-In within six hours where applicable.&nbsp;Also read: <a href="https://www.wionews.com/trending/chatgpt-guided-woman-to-taper-epilepsy-medication-she-died-parents-sue-openai-1793251205711">ChatGPT guided woman to taper epilepsy medication; she died, parents sue OpenAI</a><h3>Experts weigh in</h3>Discussing the topic, Srinivasa Sundar Bandepalli, Practice Head, Cards &amp; Payments at Expleo, told WION, "AI agents are acting beyond the scope of their authorisation, raising new cybersecurity concerns worldwide. In payments, where transactions happen in real time, compromised AI decisions can quickly translate into fraud and loss of customer trust. While robust infrastructure and smarter threat detection matter, they're only part of the answer. As AI systems are vulnerable to threats such as prompt injection and data poisoning, AI Assurance must sit at the heart of a multi-layered defence, built in from design through deployment and operations. India is taking important steps to strengthen its defences with CERT-In guidelines, RBI's digital payment security controls, zero-trust security principles, and cyber skills. As organisations increasingly turn to AI to fight AI-powered threats, we must remember that technology cannot be left to guard itself. Human oversight, stronger guardrails, and continuous AI assurance are needed to make advanced technology trusted and secure so that innovation can create value for the people and businesses." Rahail Ali, Enterprise Head, End Compute &amp; Infra Business, Ample, said, "Cybersecurity Awareness Month asks whether employees recognise cyber risks. Resilience asks a more practical question: if an incident disrupts the business, how quickly can critical operations recover?Most enterprises can assess awareness. Fewer have tested their ability to recover. That is because resilience is built continuously, not in a one-month sprint. It is built into how the IT lifecycle runs. Every device provisioned, every access granted, every system refreshed, and every endpoint retired is a security decision, whether or not anyone calls it one. When one accountable partner owns those decisions across identity, endpoint, network, cloud, and response, protection keeps pace with the business instead of interrupting it. That is what year-round resilience looks like in practice: security that lets business flow, not security that stops it to check." Similarly, Debraj Dam, Chief of Supertron VAD, said, "India is strengthening its cybersecurity through better sharing of threat information, closer coordination with industry and greater use of AI to detect suspicious activity. These efforts need to translate into how businesses manage their systems every day. AI allows attackers to find and exploit weaknesses faster. An outdated server, an unsecured device or an account with unnecessary access can leave an entire business exposed. Organisations need to be disciplined about updating systems, reviewing access and helping employees recognise suspicious requests. They should also have clear answers to some practical questions: who takes charge during an incident, how it is contained and how quickly operations can resume. As businesses adopt new technology, these decisions need to be made at the planning stage. India’s digital growth depends on people trusting the services they use, and businesses being prepared to protect that trust when something goes wrong." On the rising concern over deepfakes and synthetic fraud, experts also caution that India’s own defensive AI capability still trails global leaders, leaving the country reliant on foreign technology even as it tries to build indigenous detection and governance systems, a gap that will likely shape how effectively these new rules and drills translate into real-world resilience.

MeitY has acknowledged that generative AI is reducing the barrier for cybercriminals. It has urged ministries, state governments, and regulators to urgently review their security posture as India strengthens its cyber defence framework against threats.

India is heading to build defences to keep pace with attackers who now use artificial intelligence to discover vulnerabilities, craft convincing phishing campaigns, and automate attacks within minutes. The Ministry of Electronics and Information Technology (MeitY) has acknowledged that generative AI is reducing the barrier for cybercriminals. In response, it has urged ministries, state governments, and regulators to urgently review their security posture as India strengthens its cyber defence framework against threats that are becoming faster, more sophisticated and harder to detect. Between June and July 2026, CERT-In, the Indian Computer Emergency Response Team, ran ten customised cybersecurity exercises and drills, with the theme of building resilience against frontier AI-driven cyber threats. Earlier this year, it also released targeted technical guidance, a 'Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure,' published in May 2026 to help protect digital infrastructure from fast-evolving AI threats. In April, CERT-In went further, issuing a high-risk advisory concerning frontier AI systems capable of allowing autonomous cyberattacks, including hacking and multi-stage, sophisticated malicious operations. In November 2025, MeitY released AI Governance Guidelines, laying out a framework made to bolster safe, inclusive, and accountable AI adoption across industries, as India still needs dedicated AI legislation. Three institutions anchor this architecture: the AI Governance Group for high-level coordination, the Technology and Policy Expert Committee for technical standards, and a proposed AI Safety Institute tasked with audits and certification. Crucially, incidents tied to AI are now folded into existing reporting rules; AI-related cybersecurity incidents must be reported to CERT-In within six hours where applicable. Also read: ChatGPT guided woman to taper epilepsy medication; she died, parents sue OpenAI Experts weigh in Discussing the topic, Srinivasa Sundar Bandepalli, Practice Head, Cards & Payments at Expleo, told WION, "AI agents are acting beyond the scope of their authorisation, raising new cybersecurity concerns worldwide. In payments, where transactions happen in real time, compromised AI decisions can quickly translate into fraud and loss of customer trust. While robust infrastructure and smarter threat detection matter, they're only part of the answer. As AI systems are vulnerable to threats such as prompt injection and data poisoning, AI Assurance must sit at the heart of a multi-layered defence, built in from design through deployment and operations. India is taking important steps to strengthen its defences with CERT-In guidelines, RBI's digital payment security controls, zero-trust security principles, and cyber skills. As organisations increasingly turn to AI to fight AI-powered threats, we must remember that technology cannot be left to guard itself. Human oversight, stronger guardrails, and continuous AI assurance are needed to make advanced technology trusted and secure so that innovation can create value for the people and businesses." Rahail Ali, Enterprise Head, End Compute & Infra Business, Ample, said, "Cybersecurity Awareness Month asks whether employees recognise cyber risks. Resilience asks a more practical question: if an incident disrupts the business, how quickly can critical operations recover?Most enterprises can assess awareness. Fewer have tested their ability to recover. That is because resilience is built continuously, not in a one-month sprint. It is built into how the IT lifecycle runs. Every device provisioned, every access granted, every system refreshed, and every endpoint retired is a security decision, whether or not anyone calls it one. When one accountable partner owns those decisions across identity, endpoint, network, cloud, and response, protection keeps pace with the business instead of interrupting it. That is what year-round resilience looks like in practice: security that lets business flow, not security that stops it to check." Similarly, Debraj Dam, Chief of Supertron VAD, said, "India is strengthening its cybersecurity through better sharing of threat information, closer coordination with industry and greater use of AI to detect suspicious activity. These efforts need to translate into how businesses manage their systems every day. AI allows attackers to find and exploit weaknesses faster. An outdated server, an unsecured device or an account with unnecessary access can leave an entire business exposed. Organisations need to be disciplined about updating systems, reviewing access and helping employees recognise suspicious requests. They should also have clear answers to some practical questions: who takes charge during an incident, how it is contained and how quickly operations can resume. As businesses adopt new technology, these decisions need to be made at the planning stage. India’s digital growth depends on people trusting the services they use, and businesses being prepared to protect that trust when something goes wrong." On the rising concern over deepfakes and synthetic fraud, experts also caution that India’s own defensive AI capability still trails global leaders, leaving the country reliant on foreign technology even as it tries to build indigenous detection and governance systems, a gap that will likely shape how effectively these new rules and drills translate into real-world resilience.

MeitY has acknowledged that generative AI is reducing the barrier for cybercriminals. It has urged ministries, state governments, and regulators to urgently review their security posture as India strengthens its cyber defence framework against threats.

Trust deficit: How is India strengthening cybersecurity in age of AI-powered threats?

Apple’s new MacBook Pro line-up is drawing attention for two reasons: a touchscreen display and a price tag that can reach Rs 17.99 lakh in India for a fully configured model. The launch also comes with an internet-friendly comparison: the touchscreen MacBook Pro has arrived before GTA 6.Apple’s two options target different users. The touchscreen model focuses on OLED display technology and premium hardware, while the 14-inch MacBook Pro with the M6 chip puts greater emphasis on graphics, AI performance and value within Apple’s high-end laptop range. &nbsp;<h2>Apple Touchscreen MacBook Pro: Price in India</h2>According to the supplied details, the touchscreen MacBook Pro is available in 14-inch and 16-inch sizes, with M5 Pro and M5 Max chip options.Model Starting price in India14-inch MacBook Pro with M5 Pro Rs 3,49,90016-inch MacBook Pro with M5 Pro Rs 4,09,90014-inch MacBook Pro with M5 Max Rs 5,49,90016-inch MacBook Pro with M5 Max Rs 5,89,900&nbsp;The most expensive configuration could cost Rs 17,99,700, with an 18-core CPU, 40-core GPU, 128GB of unified memory and 16TB of storage, alongside the stated software options.Pre-orders reportedly began on October 28, with availability scheduled for November 9.<h2>OLED Touchscreen vs Liquid Retina XDR: What Changes?</h2>The touchscreen model’s biggest difference is its display. It is described as featuring a tandem OLED panel with touch support, a variable refresh rate of up to 120Hz and peak brightness of up to 2,000 nits. It also brings Dynamic Island to the MacBook Pro line-up, according to the supplied information. The feature is familiar to iPhone users, but its reported arrival on a Mac would mark a notable change in the laptop experience. The M6 MacBook Pro, by comparison, retains the Liquid Retina XDR display without touchscreen controls. Users would continue to rely on the keyboard and trackpad.<h2>M6 MacBook Pro: Performance and AI Features</h2>The 14-inch MacBook Pro with the M6 chip starts at Rs 2,39,900, with 16GB of unified memory and 1TB of SSD storage. The stated education price is Rs 2,27,900. Apple claims the M6 delivers up to 30 per cent faster graphics performance and up to 25 per cent faster on-device AI performance than the M5. The chip has a 12-core CPU and 12-core GPU, alongside a Neural Engine for AI workloads. Both laptops are said to offer up to 24 hours of battery life. The M6 model also supports Wi-Fi 7 and Bluetooth 6.<h2>Which should you choose?&nbsp;</h2>

The M6 model appears to be the more accessible option for buyers prioritising performance and AI features. The touchscreen version is aimed at users who want touch controls, an OLED display and higher-end configurations. The right choice depends on whether those additions justify the premium.

Apple’s reported touchscreen MacBook Pro brings OLED, touch controls and Dynamic Island, with prices starting at Rs 3,49,900. The 14-inch M6 model starts at Rs 2,39,900 and focuses on graphics and AI, while a top touchscreen configuration could reach Rs 17.99 lakh.

Apple’s new MacBook Pro line-up is drawing attention for two reasons: a touchscreen display and a price tag that can reach Rs 17.99 lakh in India for a fully configured model. The launch also comes with an internet-friendly comparison: the touchscreen MacBook Pro has arrived before GTA 6. Apple’s two options target different users. The touchscreen model focuses on OLED display technology and premium hardware, while the 14-inch MacBook Pro with the M6 chip puts greater emphasis on graphics, AI performance and value within Apple’s high-end laptop range. Apple Touchscreen MacBook Pro: Price in India According to the supplied details, the touchscreen MacBook Pro is available in 14-inch and 16-inch sizes, with M5 Pro and M5 Max chip options. Model Starting price in India 14-inch MacBook Pro with M5 Pro Rs 3,49,900 16-inch MacBook Pro with M5 Pro Rs 4,09,900 14-inch MacBook Pro with M5 Max Rs 5,49,900 16-inch MacBook Pro with M5 Max Rs 5,89,900 The most expensive configuration could cost Rs 17,99,700, with an 18-core CPU, 40-core GPU, 128GB of unified memory and 16TB of storage, alongside the stated software options. Pre-orders reportedly began on October 28, with availability scheduled for November 9. OLED Touchscreen vs Liquid Retina XDR: What Changes? The touchscreen model’s biggest difference is its display. It is described as featuring a tandem OLED panel with touch support, a variable refresh rate of up to 120Hz and peak brightness of up to 2,000 nits. It also brings Dynamic Island to the MacBook Pro line-up, according to the supplied information. The feature is familiar to iPhone users, but its reported arrival on a Mac would mark a notable change in the laptop experience. The M6 MacBook Pro, by comparison, retains the Liquid Retina XDR display without touchscreen controls. Users would continue to rely on the keyboard and trackpad. M6 MacBook Pro: Performance and AI Features The 14-inch MacBook Pro with the M6 chip starts at Rs 2,39,900, with 16GB of unified memory and 1TB of SSD storage. The stated education price is Rs 2,27,900. Apple claims the M6 delivers up to 30 per cent faster graphics performance and up to 25 per cent faster on-device AI performance than the M5. The chip has a 12-core CPU and 12-core GPU, alongside a Neural Engine for AI workloads. Both laptops are said to offer up to 24 hours of battery life. The M6 model also supports Wi-Fi 7 and Bluetooth 6. Which should you choose?

Apple’s reported touchscreen MacBook Pro brings OLED, touch controls and Dynamic Island, with prices starting at Rs 3,49,900. The 14-inch M6 model starts at Rs 2,39,900 and focuses on graphics and AI, while a top touchscreen configuration could reach Rs 17.99 lakh.

Rs 18 Lakh Apple MacBook Pro? Apple launches new touchscreen models in India, check prices

&nbsp;Alphabet shares fell more than 10% after Google’s parent company reported weaker-than-expected quarterly earnings and a sharp drop in free cash flow, raising fresh questions about the cost of its artificial intelligence (AI) expansion. Revenue rose 25% year on year to $127.58 billion, but both revenue and adjusted earnings per share missed analyst estimates.&nbsp;The company’s results highlight a growing challenge for Big Tech: AI is driving demand for cloud computing and new products, but building the infrastructure to support it requires enormous investment.<h2>Alphabet earnings miss estimates as cash flow turns negative</h2>Alphabet reported adjusted earnings per share of $2.67, below the $3.02 expected by analysts polled by LSEG. Revenue of $127.58 billion also fell short of the $128.22 billion forecast. Net income declined to $26.17 billion from $34.98 billion a year earlier. More concerning for investors, the company recorded negative free cash flow of $9.07 billion, significantly worse than the $2.70 billion analysts had expected.&nbsp;Free cash flow measures the money left after operating expenses and capital investment. A negative figure means the company spent more cash than it generated under this measure during the period. Alphabet spent $51.36 billion on capital expenditure in the quarter, below analyst expectations of $55.09 billion. However, the company maintained its 2026 spending forecast of $195 billion to $205 billion.<h2>Google Cloud grows as AI investment accelerates</h2>Google Cloud was a bright spot. Revenue more than doubled to $30.7 billion, beating the $28.7 billion analysts expected. The division provides cloud infrastructure and AI services to businesses and developers. Google also said its Gemini app had reached one billion users, while chief executive Sundar Pichai pointed to strong growth in search activity supported by AI investments. Search revenue rose 12% to $63.35 billion, but still missed the $64.96 billion estimate. The results suggest that AI is creating opportunities for Google while also putting pressure on the company to demonstrate that its investments can generate sufficient returns.<h2>Alphabet signals bigger AI spending in 2027</h2>Chief financial officer Anat Ashkenazi said Alphabet expects capital expenditure to increase significantly in 2027. Spending could exceed $300 billion that year, according to the supplied report. The company’s long-term debt also climbed to $125.1 billion, compared with $21.6 billion a year earlier.&nbsp;For investors, the concern is not simply how much Alphabet spends, but how quickly those investments translate into sustainable revenue and profits. Strong cloud growth and Gemini adoption offer encouraging signs, yet the latest earnings miss shows that the market is demanding clearer financial returns from the AI boom.

Alphabet shares fell more than 10% after the Google parent missed earnings and revenue estimates and reported negative free cash flow of $9.07 billion. Strong Google Cloud growth and Gemini adoption failed to offset investor concerns about the cost of its AI expansion.

Alphabet shares fell more than 10% after Google’s parent company reported weaker-than-expected quarterly earnings and a sharp drop in free cash flow, raising fresh questions about the cost of its artificial intelligence (AI) expansion. Revenue rose 25% year on year to $127.58 billion, but both revenue and adjusted earnings per share missed analyst estimates. The company’s results highlight a growing challenge for Big Tech: AI is driving demand for cloud computing and new products, but building the infrastructure to support it requires enormous investment. Alphabet earnings miss estimates as cash flow turns negative Alphabet reported adjusted earnings per share of $2.67, below the $3.02 expected by analysts polled by LSEG. Revenue of $127.58 billion also fell short of the $128.22 billion forecast. Net income declined to $26.17 billion from $34.98 billion a year earlier. More concerning for investors, the company recorded negative free cash flow of $9.07 billion, significantly worse than the $2.70 billion analysts had expected. Free cash flow measures the money left after operating expenses and capital investment. A negative figure means the company spent more cash than it generated under this measure during the period. Alphabet spent $51.36 billion on capital expenditure in the quarter, below analyst expectations of $55.09 billion. However, the company maintained its 2026 spending forecast of $195 billion to $205 billion. Google Cloud grows as AI investment accelerates Google Cloud was a bright spot. Revenue more than doubled to $30.7 billion, beating the $28.7 billion analysts expected. The division provides cloud infrastructure and AI services to businesses and developers. Google also said its Gemini app had reached one billion users, while chief executive Sundar Pichai pointed to strong growth in search activity supported by AI investments. Search revenue rose 12% to $63.35 billion, but still missed the $64.96 billion estimate. The results suggest that AI is creating opportunities for Google while also putting pressure on the company to demonstrate that its investments can generate sufficient returns. Alphabet signals bigger AI spending in 2027 Chief financial officer Anat Ashkenazi said Alphabet expects capital expenditure to increase significantly in 2027. Spending could exceed $300 billion that year, according to the supplied report. The company’s long-term debt also climbed to $125.1 billion, compared with $21.6 billion a year earlier. For investors, the concern is not simply how much Alphabet spends, but how quickly those investments translate into sustainable revenue and profits. Strong cloud growth and Gemini adoption offer encouraging signs, yet the latest earnings miss shows that the market is demanding clearer financial returns from the AI boom.

Alphabet shares fell more than 10% after the Google parent missed earnings and revenue estimates and reported negative free cash flow of $9.07 billion. Strong Google Cloud growth and Gemini adoption failed to offset investor concerns about the cost of its AI expansion.

Why Alphabet stock fell 10% despite strong revenue growth and Google Cloud gains

There is a high probability that most enterprise AI projects will not fail on the demo. They falter somewhere between what is right and what is right for the business. A model can generate fluent and confident text when trained on a pilot, but it may not perform well if it is used as a component of a real decision, if it hasn't been trained on what constitutes "correct" for a particular company's orders, backlogs, and their rules about service levels. Now, for all of these years, Sandeep Nutakki has wandered in that void and he is blunt. The problem of moving AI from a trial run to a trusted production decision system is not, fundamentally, a modeling issue. It's a systems issue.&nbsp;&nbsp;&nbsp;Nutakki is a senior AI and data engineering practitioner with more than seven years across enterprise-scale organizations, including a major international airline and a global e-commerce and logistics company before his current work at an AI-native data-engineering firm. The throughline of his career is not a single product. It is the discipline of making machine output trustworthy enough that operations teams will act on it. That discipline lives at the intersection of three forces that constantly pull against each other: accuracy, speed, and business impact.&nbsp;&nbsp;Those three forces are why so many promising systems stall. A team can push accuracy higher by adding validation, retrieval, and review, but every layer adds latency. It can chase speed by simplifying the pipeline, but a faster wrong answer is worse than a slow one when a fulfillment decision is on the line. And neither accuracy nor speed counts as impact until the output changes what the business does. Balancing the three is the real engineering work, and Nutakki has built the pieces that make the balance hold.&nbsp;&nbsp;The first force is accuracy, and his answer to it is to treat evaluation as a built-in component rather than a final checkpoint. At the AI-native company, he contributed to an evaluation harness that scores agent-generated SQL and narrative answers against structured ground-truth answers.&nbsp;The harness does not ask whether the output reads well. It asks whether the output matches what the business already knows to be true. "Benchmark AI outputs against business ground truth from day one," Nutakki says. "Treat evaluation as a first-class component, not a final QA step." That shift, from human eyeball grading to a structured judging loop, is what lets a team tell the difference between a system that is improving and one that is merely sounding more confident.&nbsp;&nbsp;Accuracy of that kind depends on a less visible layer: the semantic alignment between plain-language questions and enterprise data. Concepts like order status, backlog, and service-level definition do not map cleanly onto a general model's reasoning, because each company encodes them differently. Nutakki's approach is to treat that mapping as infrastructure, an enforced contract between the business and the model, rather than something coaxed out through prompt wording. At his current company he has validated more than 100 AI and semantic functions against enterprise domain objects and expected outcomes, the unglamorous work that keeps a confident answer honest.&nbsp;&nbsp;The second force is speed, and it is where many evaluation-heavy systems quietly collapse. Checking output against ground truth at the scale modern models operate at can mean assembling enormous prompts and running thousands of comparisons. Nutakki worked on an evaluation harness that runs those comparisons in parallel and uses adaptive token-budget handling to keep very large prompts within a model's context limits, so that thoroughness does not become a bottleneck. He has applied the same instinct to execution itself. Instead of one monolithic prompt that does everything at once, he favors composable execution, where a business question is decomposed into discrete, governed steps. "Prefer composable execution over monolithic prompts," he says. The payoff is practical: behavior becomes inspectable, individual steps can be swapped, and the system stays safe to change while it is running in production.&nbsp;&nbsp;The third force, business impact, is the one that decides whether any of it mattered. Impact, in Nutakki's framing, is governed correctness rather than a headline metric. He is deliberate about this. Rather than cite revenue figures he cannot verify in writing, he points to the outcomes he can stand behind: more reliable, more consistent, more governable AI-assisted workflows. At the e-commerce and logistics operation and the airline earlier in his career, that meant turning large volumes of operational data into consistent, trusted inputs so that downstream reporting and business decisions rested on the same definitions. At the AI-native firm, it means deciding whether a pilot is ready to move into production at all.&nbsp;&nbsp;Holding the three forces together requires one more thing that is easy to skip and expensive to retrofit: a safety envelope built before the system scales. Nutakki contributed to an event-driven agent orchestration platform with a fail-closed design, input and tool-argument validation, tenant isolation for artifacts and audit data, and cross-module guards that stop new code from silently bypassing tenant-safe paths. The logic is simple. Input validation, tenant isolation, and architectural guards are far cheaper to design in than to bolt on once agents are already wired into live systems. A system that is fast and accurate but not safe to operate never earns a place in the decision loop.&nbsp;&nbsp;What ties Nutakki's work together is a refusal to let fluency stand in for fitness. The next wave of enterprise AI, in his view, will not be won by whichever model is strongest this quarter. It will be won by the teams that engineer the surrounding system well. That means evaluation that benchmarks against business truth. It means execution that stays composable and inspectable, and safety that is designed in from the start. And it means a semantic layer that forces model output to respect how the business operates.&nbsp;&nbsp;"The next wave of enterprise AI will be judged on governed correctness, not fluency," Nutakki says. For the engineers deciding which pilots are allowed to graduate, that is less a prediction than a working standard.

A model can generate fluent and confident text when trained on a pilot, but it may not perform well if it is used as a component of a real decision, if it hasn't been trained on what constitutes "correct" for a particular company's orders, backlogs, and their rules about service levels.

There is a high probability that most enterprise AI projects will not fail on the demo. They falter somewhere between what is right and what is right for the business. A model can generate fluent and confident text when trained on a pilot, but it may not perform well if it is used as a component of a real decision, if it hasn't been trained on what constitutes "correct" for a particular company's orders, backlogs, and their rules about service levels. Now, for all of these years, Sandeep Nutakki has wandered in that void and he is blunt. The problem of moving AI from a trial run to a trusted production decision system is not, fundamentally, a modeling issue. It's a systems issue. Nutakki is a senior AI and data engineering practitioner with more than seven years across enterprise-scale organizations, including a major international airline and a global e-commerce and logistics company before his current work at an AI-native data-engineering firm. The throughline of his career is not a single product. It is the discipline of making machine output trustworthy enough that operations teams will act on it. That discipline lives at the intersection of three forces that constantly pull against each other: accuracy, speed, and business impact. Those three forces are why so many promising systems stall. A team can push accuracy higher by adding validation, retrieval, and review, but every layer adds latency. It can chase speed by simplifying the pipeline, but a faster wrong answer is worse than a slow one when a fulfillment decision is on the line. And neither accuracy nor speed counts as impact until the output changes what the business does. Balancing the three is the real engineering work, and Nutakki has built the pieces that make the balance hold. The first force is accuracy, and his answer to it is to treat evaluation as a built-in component rather than a final checkpoint. At the AI-native company, he contributed to an evaluation harness that scores agent-generated SQL and narrative answers against structured ground-truth answers. The harness does not ask whether the output reads well. It asks whether the output matches what the business already knows to be true. "Benchmark AI outputs against business ground truth from day one," Nutakki says. "Treat evaluation as a first-class component, not a final QA step." That shift, from human eyeball grading to a structured judging loop, is what lets a team tell the difference between a system that is improving and one that is merely sounding more confident. Accuracy of that kind depends on a less visible layer: the semantic alignment between plain-language questions and enterprise data. Concepts like order status, backlog, and service-level definition do not map cleanly onto a general model's reasoning, because each company encodes them differently. Nutakki's approach is to treat that mapping as infrastructure, an enforced contract between the business and the model, rather than something coaxed out through prompt wording. At his current company he has validated more than 100 AI and semantic functions against enterprise domain objects and expected outcomes, the unglamorous work that keeps a confident answer honest. The second force is speed, and it is where many evaluation-heavy systems quietly collapse. Checking output against ground truth at the scale modern models operate at can mean assembling enormous prompts and running thousands of comparisons. Nutakki worked on an evaluation harness that runs those comparisons in parallel and uses adaptive token-budget handling to keep very large prompts within a model's context limits, so that thoroughness does not become a bottleneck. He has applied the same instinct to execution itself. Instead of one monolithic prompt that does everything at once, he favors composable execution, where a business question is decomposed into discrete, governed steps. "Prefer composable execution over monolithic prompts," he says. The payoff is practical: behavior becomes inspectable, individual steps can be swapped, and the system stays safe to change while it is running in production. The third force, business impact, is the one that decides whether any of it mattered. Impact, in Nutakki's framing, is governed correctness rather than a headline metric. He is deliberate about this. Rather than cite revenue figures he cannot verify in writing, he points to the outcomes he can stand behind: more reliable, more consistent, more governable AI-assisted workflows. At the e-commerce and logistics operation and the airline earlier in his career, that meant turning large volumes of operational data into consistent, trusted inputs so that downstream reporting and business decisions rested on the same definitions. At the AI-native firm, it means deciding whether a pilot is ready to move into production at all. Holding the three forces together requires one more thing that is easy to skip and expensive to retrofit: a safety envelope built before the system scales. Nutakki contributed to an event-driven agent orchestration platform with a fail-closed design, input and tool-argument validation, tenant isolation for artifacts and audit data, and cross-module guards that stop new code from silently bypassing tenant-safe paths. The logic is simple. Input validation, tenant isolation, and architectural guards are far cheaper to design in than to bolt on once agents are already wired into live systems. A system that is fast and accurate but not safe to operate never earns a place in the decision loop. What ties Nutakki's work together is a refusal to let fluency stand in for fitness. The next wave of enterprise AI, in his view, will not be won by whichever model is strongest this quarter. It will be won by the teams that engineer the surrounding system well. That means evaluation that benchmarks against business truth. It means execution that stays composable and inspectable, and safety that is designed in from the start. And it means a semantic layer that forces model output to respect how the business operates. "The next wave of enterprise AI will be judged on governed correctness, not fluency," Nutakki says. For the engineers deciding which pilots are allowed to graduate, that is less a prediction than a working standard.

&nbsp;Inside the accuracy, speed, and business-impact tradeoffs that decide whether an AI pilot ever reaches production.

When 'sounds right' is not good enough: Engineering AI that earns a place in the decision loop

Retail media has a structural problem that its growth rate keeps outrunning. The network sells the advertising, and the network also reports how well it performed.That is not necessarily dishonest, and it does not require anyone to behave badly. It only requires the default measurement to be attribution, which credits the network whenever an exposed household later buys, whether or not the advertising had anything to do with it. Retail media sits closest to the purchase, so it captures the most of that already-arriving demand, and reports it as return.Suneel Dasika's response was to build the control group into the network itself.The approach is structurally simple and organisationally difficult. A randomised four percent slice of the addressable first-party audience is carved out and made permanent, assigned at customer level by hashed identifier, re-randomised only once a year, and excluded upstream in the audience-build layer so that no campaign can target into it even by accident.The consequence is that every campaign on the network inherits a valid control group automatically. No advertiser has to be persuaded to give up reach, no bespoke study has to be designed, and nobody has to justify sacrificing revenue on a particular campaign to learn something about it. Incrementality stops being a project and becomes a property of the infrastructure.What that changed is best expressed as coverage. Lift measurement had previously reached roughly five percent of eligible campaigns, limited to a handful of managed accounts each year by the design and approval cycle each study required. With the holdout standing permanently, coverage became close to universal, at a fully amortised cost of well under one percent of network revenue, absorbed centrally as what Dasika describes as a measurement tax rather than billed to advertisers.The pilot ran with five advertisers chosen for spend volume, category diversity and enough weekly conversion density to reach significance inside a four-week flight. Each received a weekly exposed-versus-withheld read on incremental sales, new-to-brand rate and lift percentage, presented alongside their existing attributed return rather than instead of it.In all five cases, measured lift came in materially below attributed return."Which was the point," Dasika says.The predictable expectation is that advertisers shown a smaller number spend less. All five expanded their investment. Two reallocated budget toward the tactics that survived the comparison, and the holdout became the default reporting layer for those accounts within two quarters. He has seen the pattern before, in an earlier productised incrementality effort that reached eighty percent adoption among top-tier advertisers and lifted their spend by twenty percent.The reason is worth sitting with, because it inverts the usual assumption about measurement transparency. An advertiser who cannot tell which part of a reported return is real has to discount the whole number. An advertiser who can separate genuine lift from captured demand can confidently invest behind the part that works. Precision about what is not working turns out to be what unlocks budget for what is.Presenting the honest number alongside the attributed one, rather than replacing it, also reframed the conversation from defending a figure to calibrating it.Where Dasika's account becomes genuinely unusual is in what he volunteers about the method's limits, at length and without being asked.The control group is randomised, but it is not clean. Withheld households still encounter the brand through organic search, store visits, competitor retail media and the advertiser's own upper-funnel advertising, so the comparison estimates lift against a control contaminated by everything happening outside the network's walls. Match rates, identity resolution and cross-device fragmentation introduce further slippage. At small spend levels, the confidence interval is wide enough to swallow the effect entirely.His team chose to publish that uncertainty rather than smooth it: stating the bounds, suppressing underpowered reads instead of reporting them, and framing every estimate as a conservative floor on true incremental value rather than a point estimate."Advertisers trusted a defensible range more than they trusted a precise number nobody could defend," he says. His view is that the trust, rather than the methodology, is what actually unlocked the incremental budget.He is equally direct about what makes this hard to operate, and the failure modes he lists are mostly organisational.Withholding four percent of addressable reach shows up immediately as lost impressions and reduced delivery, which caps what a sales team can promise. That resistance is structural, so the holdout has to be funded centrally as infrastructure and can never appear as a line item an account team is able to opt out of.The exclusion also has to hold across every path an activation can take. Offsite platforms, third-party demand partners, agency-managed buys and clean-room activations each represent a route by which a withheld household can be reached anyway, and every new integration is a fresh place the holdout can break silently.Identity is the quiet risk. A customer-level assignment is only as durable as the identity graph beneath it, and re-hashing, identifier churn and imperfect matching erode the group over time. Dasika's phrasing is the sharpest warning in his account: a control that quietly degrades produces confidently wrong lift numbers.And at four percent, most campaigns will never clear significance at all, which requires documented minimum thresholds for spend, flight length and conversions, plus the discipline to tell a smaller advertiser that they do not get a number yet.That last one is the real test of whether a measurement system is honest. Every incentive points toward producing a number, because a number is what the advertiser is asking for. Declining to supply one is harder than supplying a bad one, and considerably less popular.&nbsp;

"Advertisers trusted a defensible range more than they trusted a precise number nobody could defend," he says. His view is that the trust, rather than the methodology, is what actually unlocked the incremental budget.

Retail media has a structural problem that its growth rate keeps outrunning. The network sells the advertising, and the network also reports how well it performed. That is not necessarily dishonest, and it does not require anyone to behave badly. It only requires the default measurement to be attribution, which credits the network whenever an exposed household later buys, whether or not the advertising had anything to do with it. Retail media sits closest to the purchase, so it captures the most of that already-arriving demand, and reports it as return. Suneel Dasika's response was to build the control group into the network itself. The approach is structurally simple and organisationally difficult. A randomised four percent slice of the addressable first-party audience is carved out and made permanent, assigned at customer level by hashed identifier, re-randomised only once a year, and excluded upstream in the audience-build layer so that no campaign can target into it even by accident. The consequence is that every campaign on the network inherits a valid control group automatically. No advertiser has to be persuaded to give up reach, no bespoke study has to be designed, and nobody has to justify sacrificing revenue on a particular campaign to learn something about it. Incrementality stops being a project and becomes a property of the infrastructure. What that changed is best expressed as coverage. Lift measurement had previously reached roughly five percent of eligible campaigns, limited to a handful of managed accounts each year by the design and approval cycle each study required. With the holdout standing permanently, coverage became close to universal, at a fully amortised cost of well under one percent of network revenue, absorbed centrally as what Dasika describes as a measurement tax rather than billed to advertisers. The pilot ran with five advertisers chosen for spend volume, category diversity and enough weekly conversion density to reach significance inside a four-week flight. Each received a weekly exposed-versus-withheld read on incremental sales, new-to-brand rate and lift percentage, presented alongside their existing attributed return rather than instead of it. In all five cases, measured lift came in materially below attributed return. "Which was the point," Dasika says. The predictable expectation is that advertisers shown a smaller number spend less. All five expanded their investment. Two reallocated budget toward the tactics that survived the comparison, and the holdout became the default reporting layer for those accounts within two quarters. He has seen the pattern before, in an earlier productised incrementality effort that reached eighty percent adoption among top-tier advertisers and lifted their spend by twenty percent. The reason is worth sitting with, because it inverts the usual assumption about measurement transparency. An advertiser who cannot tell which part of a reported return is real has to discount the whole number. An advertiser who can separate genuine lift from captured demand can confidently invest behind the part that works. Precision about what is not working turns out to be what unlocks budget for what is. Presenting the honest number alongside the attributed one, rather than replacing it, also reframed the conversation from defending a figure to calibrating it. Where Dasika's account becomes genuinely unusual is in what he volunteers about the method's limits, at length and without being asked. The control group is randomised, but it is not clean. Withheld households still encounter the brand through organic search, store visits, competitor retail media and the advertiser's own upper-funnel advertising, so the comparison estimates lift against a control contaminated by everything happening outside the network's walls. Match rates, identity resolution and cross-device fragmentation introduce further slippage. At small spend levels, the confidence interval is wide enough to swallow the effect entirely. His team chose to publish that uncertainty rather than smooth it: stating the bounds, suppressing underpowered reads instead of reporting them, and framing every estimate as a conservative floor on true incremental value rather than a point estimate. "Advertisers trusted a defensible range more than they trusted a precise number nobody could defend," he says. His view is that the trust, rather than the methodology, is what actually unlocked the incremental budget. He is equally direct about what makes this hard to operate, and the failure modes he lists are mostly organisational. Withholding four percent of addressable reach shows up immediately as lost impressions and reduced delivery, which caps what a sales team can promise. That resistance is structural, so the holdout has to be funded centrally as infrastructure and can never appear as a line item an account team is able to opt out of. The exclusion also has to hold across every path an activation can take. Offsite platforms, third-party demand partners, agency-managed buys and clean-room activations each represent a route by which a withheld household can be reached anyway, and every new integration is a fresh place the holdout can break silently. Identity is the quiet risk. A customer-level assignment is only as durable as the identity graph beneath it, and re-hashing, identifier churn and imperfect matching erode the group over time. Dasika's phrasing is the sharpest warning in his account: a control that quietly degrades produces confidently wrong lift numbers. And at four percent, most campaigns will never clear significance at all, which requires documented minimum thresholds for spend, flight length and conversions, plus the discipline to tell a smaller advertiser that they do not get a number yet. That last one is the real test of whether a measurement system is honest. Every incentive points toward producing a number, because a number is what the advertiser is asking for. Declining to supply one is harder than supplying a bad one, and considerably less popular.

"Advertisers trusted a defensible range more than they trusted a precise number nobody could defend," he says. His view is that the trust, rather than the methodology, is what actually unlocked the incremental budget.

Retail media is the fastest-growing corner of advertising, and almost no one can measure it honestly

Apple has launched a new 14-inch MacBook Pro powered by its M6 chip, promising faster graphics, improved on-device AI performance and up to 24 hours of battery life. The company has also introduced a new iMac, positioning both devices around running more AI tasks directly on the hardware rather than relying entirely on cloud services.&nbsp;The new MacBook Pro starts at $1,999 in the US, while the iMac starts at $1,499, according to the supplied announcement details. Pre-orders for the MacBook Pro open on October 28 in 35 countries and regions, with availability beginning November 6.&nbsp;<h2>Apple M6 MacBook Pro: Up to 30% faster graphics</h2>Apple says the M6-powered MacBook Pro delivers up to 30% faster graphics performance and up to 25% faster on-device AI performance compared with the M5 model. The chip features a 12-core CPU and a 12-core GPU. Apple says these improvements could benefit people working with demanding creative software, 3D graphics, code and AI applications.The company also claims up to 24 hours of battery life, which it describes as the longest ever in a Mac. Actual battery performance will depend on how the laptop is used and its configuration.&nbsp;<h2>New iMac promises a major AI performance boost</h2>Apple says the new iMac delivers up to five times faster on-device AI performance than the previous M4 model. This could help users run supported AI features and creative workflows locally on their computers. On-device processing can reduce the need to send certain information to remote servers, although some AI features may still rely on cloud-based models. Apple has not provided a full comparison of every real-world task, so the advertised performance figures should be treated as company-reported results rather than a guarantee that all applications will run at those speeds.<h2>MacBook Pro and iMac: Price and availability</h2>The 14-inch MacBook Pro with M6 starts at $1,999 in the US, while the new iMac starts at $1,499, according to the supplied announcement. MacBook Pro pre-orders begin October 28, with sales starting November 6. Buyers in India should check Apple's local store for confirmed pricing and availability, as the US starting price does not establish the final Indian price. The launch reflects Apple's growing focus on AI performance within its computers. For buyers, the key question will be whether faster graphics, local AI processing and longer battery life justify the cost of upgrading.

Apple has unveiled an M6-powered 14-inch MacBook Pro, claiming up to 30% faster graphics, 25% faster AI performance and 24 hours of battery life. A new iMac promises up to five times faster on-device AI performance than the M4 model.

Apple has launched a new 14-inch MacBook Pro powered by its M6 chip, promising faster graphics, improved on-device AI performance and up to 24 hours of battery life. The company has also introduced a new iMac, positioning both devices around running more AI tasks directly on the hardware rather than relying entirely on cloud services. The new MacBook Pro starts at $1,999 in the US, while the iMac starts at $1,499, according to the supplied announcement details. Pre-orders for the MacBook Pro open on October 28 in 35 countries and regions, with availability beginning November 6. Apple M6 MacBook Pro: Up to 30% faster graphics Apple says the M6-powered MacBook Pro delivers up to 30% faster graphics performance and up to 25% faster on-device AI performance compared with the M5 model. The chip features a 12-core CPU and a 12-core GPU. Apple says these improvements could benefit people working with demanding creative software, 3D graphics, code and AI applications. The company also claims up to 24 hours of battery life, which it describes as the longest ever in a Mac. Actual battery performance will depend on how the laptop is used and its configuration. New iMac promises a major AI performance boost Apple says the new iMac delivers up to five times faster on-device AI performance than the previous M4 model. This could help users run supported AI features and creative workflows locally on their computers. On-device processing can reduce the need to send certain information to remote servers, although some AI features may still rely on cloud-based models. Apple has not provided a full comparison of every real-world task, so the advertised performance figures should be treated as company-reported results rather than a guarantee that all applications will run at those speeds. MacBook Pro and iMac: Price and availability The 14-inch MacBook Pro with M6 starts at $1,999 in the US, while the new iMac starts at $1,499, according to the supplied announcement. MacBook Pro pre-orders begin October 28, with sales starting November 6. Buyers in India should check Apple's local store for confirmed pricing and availability, as the US starting price does not establish the final Indian price. The launch reflects Apple's growing focus on AI performance within its computers. For buyers, the key question will be whether faster graphics, local AI processing and longer battery life justify the cost of upgrading.

Apple has unveiled a new 14-inch MacBook Pro powered by its M6 chip, claiming up to 30% faster graphics and up to 25% faster AI performance than the M5 model. The company has also introduced a new iMac, with pre-orders opening on Wednesday and sales starting November 6.

Apple launches M6 MacBook Pro and new iMac with up to 25% faster AI performance: Check price and features

Apple’s next entry-level iPhone could arrive with a new processor but retain some familiar hardware. The rumoured iPhone 18e is tipped to launch in early March 2027, potentially alongside the iPhone Air 2 and the standard iPhone 18. Leaked specifications suggest a 60Hz OLED display, a single 48-megapixel rear camera and Apple’s upcoming A20 chip. The details remain unconfirmed. If the reported India price of up to Rs 79,900 proves accurate, buyers may need to weigh the new processor against possible compromises in the display and camera setup.&nbsp;<h2>iPhone 18e specifications: 60Hz display and 48MP camera</h2>The iPhone 18e is expected to feature a 6.12-inch OLED display with a 60Hz refresh rate. A 60Hz screen refreshes 60 times per second, while 120Hz displays can offer smoother scrolling and animations.If Apple sticks with 60Hz, the phone could miss out on one feature available on selected higher-end smartphones. The camera setup is also tipped to remain simple. The iPhone 18e may feature a single 48-megapixel rear camera, without a dedicated ultra-wide lens. That could limit flexibility when photographing landscapes, large groups or buildings from close range.<h2>A20 chip and 9GB RAM: What the leak suggests</h2>Under the hood, Apple’s rumoured A20 chip could be built using a 2-nanometre manufacturing process. If confirmed, the newer chip could bring improvements in processing performance and power efficiency. The phone is also tipped to include 9GB of RAM, potentially helping it handle Apple Intelligence features and other demanding tasks. However, processor performance, memory capacity and AI capabilities have not been confirmed. Battery capacity, charging speeds and the final design also remain unknown.<h2>iPhone 18e launch date and expected price in India</h2>Apple could unveil the iPhone 18e during the first week of March 2027, according to current rumours. The iPhone Air 2 and standard iPhone 18 are also tipped to launch around the same period. The expected India price is between Rs 69,900 and Rs 79,900. This is an early estimate, not an official price announced by Apple. For buyers, the biggest question is whether the iPhone 18e will offer enough performance and long-term software value to justify its expected price despite a 60Hz display and a single rear camera. Until Apple announces the device, its specifications, pricing and launch schedule should be treated as unconfirmed reports.

Apple’s rumoured iPhone 18e could launch in March 2027 with a 6.12-inch OLED display, 48MP rear camera and A20 chip. Leaks suggest a 60Hz screen, 9GB RAM and an India price of Rs 69,900-79,900, but Apple has yet to confirm the details.

Apple’s next entry-level iPhone could arrive with a new processor but retain some familiar hardware. The rumoured iPhone 18e is tipped to launch in early March 2027, potentially alongside the iPhone Air 2 and the standard iPhone 18. Leaked specifications suggest a 60Hz OLED display, a single 48-megapixel rear camera and Apple’s upcoming A20 chip. The details remain unconfirmed. If the reported India price of up to Rs 79,900 proves accurate, buyers may need to weigh the new processor against possible compromises in the display and camera setup. iPhone 18e specifications: 60Hz display and 48MP camera The iPhone 18e is expected to feature a 6.12-inch OLED display with a 60Hz refresh rate. A 60Hz screen refreshes 60 times per second, while 120Hz displays can offer smoother scrolling and animations. If Apple sticks with 60Hz, the phone could miss out on one feature available on selected higher-end smartphones. The camera setup is also tipped to remain simple. The iPhone 18e may feature a single 48-megapixel rear camera, without a dedicated ultra-wide lens. That could limit flexibility when photographing landscapes, large groups or buildings from close range. A20 chip and 9GB RAM: What the leak suggests Under the hood, Apple’s rumoured A20 chip could be built using a 2-nanometre manufacturing process. If confirmed, the newer chip could bring improvements in processing performance and power efficiency. The phone is also tipped to include 9GB of RAM, potentially helping it handle Apple Intelligence features and other demanding tasks. However, processor performance, memory capacity and AI capabilities have not been confirmed. Battery capacity, charging speeds and the final design also remain unknown. iPhone 18e launch date and expected price in India Apple could unveil the iPhone 18e during the first week of March 2027, according to current rumours. The iPhone Air 2 and standard iPhone 18 are also tipped to launch around the same period. The expected India price is between Rs 69,900 and Rs 79,900. This is an early estimate, not an official price announced by Apple. For buyers, the biggest question is whether the iPhone 18e will offer enough performance and long-term software value to justify its expected price despite a 60Hz display and a single rear camera. Until Apple announces the device, its specifications, pricing and launch schedule should be treated as unconfirmed reports.

Apple’s rumoured iPhone 18e could launch in March 2027 with a 6.12-inch OLED display, 48MP rear camera and A20 chip. Leaks suggest a 60Hz screen, 9GB RAM and an India price of Rs 69,900-79,900, but Apple has yet to confirm the details.

Apple iPhone 18e could launch in March 2027 with a 60Hz screen and single camera

Ask anyone who has worked a shift in a security operations centre what the job actually feels like, and the answer is rarely dramatic. It is a queue. Alerts arrive faster than anyone can work them, most turn out to be nothing, and the analyst's real skill becomes deciding what to ignore. The uncomfortable part is that the one alert that mattered was probably in there somewhere.&nbsp;Dr. Naga Venkata Aswini Pavan Kumar Inguva calls this the quality problem, and he thinks the industry has been solving for the wrong variable. "Effective cybersecurity depends on context, not simply the number of alerts generated," he said. "Organizations should prioritize technologies that improve the quality of threat detection rather than simply increasing the volume of security notifications."&nbsp;That distinction matters because the instinct when something is missed is almost always to add more detection. More sensors, more rules, more coverage. Each addition is defensible on its own and the cumulative effect is an analyst who trusts the queue less than they did before. Alert fatigue is not a personal failing. It is what happens when a system is tuned to never miss anything and nobody accounts for the cost of being wrong&nbsp;Inguva holds a Ph.D. in Information Technology from the University of the Cumberlands and a master's in computer science, and has spent more than seven years in enterprise technology in the United States after teaching as a professor in India. His research focuses on applying machine learning to intrusion detection, and the false-positive question is the one he keeps returning to.&nbsp;The reason it has resisted a clean fix is that the two goals pull against each other. Tighten a detector until it stops crying wolf and it starts missing real intrusions. Loosen it until it catches everything and the analyst drowns. "Balancing high detection capability with a low false-positive rate remains one of the most persistent challenges in cybersecurity research," Inguva said. There is a second complication underneath that one. Machine learning models that perform impressively in a particular environment often degrade badly in a different one, because they learned the shape of that first network rather than the shape of an attack.&nbsp;His proposed response is a framework he calls the Hybrid Weighted Ensemble Intrusion Detection Framework. The reasoning behind it is that no single model should be trusted to make the call alone. Instead, several models assess the same activity, their outputs are weighted rather than simply counted, and the result carries a confidence score rather than a binary verdict. Alerts are then prioritised, so what reaches the analyst first is what the system is most confident about and most concerned by.&nbsp;It is worth being clear about what this is. The framework is a conceptual research architecture, not a deployed product. Inguva designed the architecture, assessed existing ensemble learning approaches against the problem, and proposed the confidence scoring and alert prioritisation concepts, producing a roadmap for future validation and implementation. The contribution is the design and the argument, and he is candid that the testing comes next.&nbsp;The design does answer a real operational complaint, though. A queue sorted by confidence is a different working environment from a queue sorted by arrival time. It gives the analyst a defensible reason to start at the top.&nbsp;Inguva is equally interested in the gap that swallows a lot of promising security research. Methods that perform well in a paper frequently fail to reach production, not because the mathematics is wrong but because organisations need something that scales, explains itself and fits the workflow people already use. He builds those constraints into the framework rather than treating them as somebody else's problem, which is unsurprising given that he works as a senior software developer for a US state government agency, modernising enterprise applications including telecommunications and network systems, and co-founded Thaapasi Smart Infratech Pvt. Ltd.&nbsp;He is firm that better tooling is not the whole answer, and pushes back on a common assumption. "A common misconception is that cybersecurity can be achieved solely through advanced software or hardware solutions," he said. "In reality, effective cybersecurity depends on a balanced combination of technology, governance, skilled professionals, continuous improvement, and organizational commitment." Cybersecurity, in his framing, is as much a people, process and governance challenge as a technical one.&nbsp;That view shapes where he thinks artificial intelligence belongs in the operations centre. Not in the analyst's chair. "AI should be viewed as a decision-support capability that augments human expertise, allowing analysts to focus on complex investigations and strategic decision-making," he said.&nbsp;His broader research record runs alongside this work, including published research on threat detection using generative adversarial networks, deep learning for network threat detection, and machine learning approaches to classifying attacks, some of which received best paper awards.The ambition behind the framework is modest in a way that is easy to underrate. It does not promise to catch more. It proposes to be wrong less often, and to be honest about how sure it is. For an analyst three hours into a queue, that is the more useful promise.

A researcher argues the real problem is not how much a security team sees, but how little of it is worth seeing

Ask anyone who has worked a shift in a security operations centre what the job actually feels like, and the answer is rarely dramatic. It is a queue. Alerts arrive faster than anyone can work them, most turn out to be nothing, and the analyst's real skill becomes deciding what to ignore. The uncomfortable part is that the one alert that mattered was probably in there somewhere. Dr. Naga Venkata Aswini Pavan Kumar Inguva calls this the quality problem, and he thinks the industry has been solving for the wrong variable. "Effective cybersecurity depends on context, not simply the number of alerts generated," he said. "Organizations should prioritize technologies that improve the quality of threat detection rather than simply increasing the volume of security notifications." That distinction matters because the instinct when something is missed is almost always to add more detection. More sensors, more rules, more coverage. Each addition is defensible on its own and the cumulative effect is an analyst who trusts the queue less than they did before. Alert fatigue is not a personal failing. It is what happens when a system is tuned to never miss anything and nobody accounts for the cost of being wrong Inguva holds a Ph.D. in Information Technology from the University of the Cumberlands and a master's in computer science, and has spent more than seven years in enterprise technology in the United States after teaching as a professor in India. His research focuses on applying machine learning to intrusion detection, and the false-positive question is the one he keeps returning to. The reason it has resisted a clean fix is that the two goals pull against each other. Tighten a detector until it stops crying wolf and it starts missing real intrusions. Loosen it until it catches everything and the analyst drowns. "Balancing high detection capability with a low false-positive rate remains one of the most persistent challenges in cybersecurity research," Inguva said. There is a second complication underneath that one. Machine learning models that perform impressively in a particular environment often degrade badly in a different one, because they learned the shape of that first network rather than the shape of an attack. His proposed response is a framework he calls the Hybrid Weighted Ensemble Intrusion Detection Framework. The reasoning behind it is that no single model should be trusted to make the call alone. Instead, several models assess the same activity, their outputs are weighted rather than simply counted, and the result carries a confidence score rather than a binary verdict. Alerts are then prioritised, so what reaches the analyst first is what the system is most confident about and most concerned by. It is worth being clear about what this is. The framework is a conceptual research architecture, not a deployed product. Inguva designed the architecture, assessed existing ensemble learning approaches against the problem, and proposed the confidence scoring and alert prioritisation concepts, producing a roadmap for future validation and implementation. The contribution is the design and the argument, and he is candid that the testing comes next. The design does answer a real operational complaint, though. A queue sorted by confidence is a different working environment from a queue sorted by arrival time. It gives the analyst a defensible reason to start at the top. Inguva is equally interested in the gap that swallows a lot of promising security research. Methods that perform well in a paper frequently fail to reach production, not because the mathematics is wrong but because organisations need something that scales, explains itself and fits the workflow people already use. He builds those constraints into the framework rather than treating them as somebody else's problem, which is unsurprising given that he works as a senior software developer for a US state government agency, modernising enterprise applications including telecommunications and network systems, and co-founded Thaapasi Smart Infratech Pvt. Ltd. He is firm that better tooling is not the whole answer, and pushes back on a common assumption. "A common misconception is that cybersecurity can be achieved solely through advanced software or hardware solutions," he said. "In reality, effective cybersecurity depends on a balanced combination of technology, governance, skilled professionals, continuous improvement, and organizational commitment." Cybersecurity, in his framing, is as much a people, process and governance challenge as a technical one. That view shapes where he thinks artificial intelligence belongs in the operations centre. Not in the analyst's chair. "AI should be viewed as a decision-support capability that augments human expertise, allowing analysts to focus on complex investigations and strategic decision-making," he said. His broader research record runs alongside this work, including published research on threat detection using generative adversarial networks, deep learning for network threat detection, and machine learning approaches to classifying attacks, some of which received best paper awards. The ambition behind the framework is modest in a way that is easy to underrate. It does not promise to catch more. It proposes to be wrong less often, and to be honest about how sure it is. For an analyst three hours into a queue, that is the more useful promise.

A researcher argues the real problem is not how much a security team sees, but how little of it is worth seeing

Security teams are drowning in false alarms. The fix is not more alerts, it is fewer wrong ones

An AI startup says it hacked Anthropic's systems in less than 12 hours and was able to access its internal messages. Cathedral is based in the US and was founded this year by former Department of Government Efficiency (DOGE) employees. Its staff breached Anthropic as part of a bug bounty program where people are asked to reveal security flaws.&nbsp;According to The Washington Post, the hack led to information like Slack messages, meeting transcripts, salaries and job offers. Responding to the report, Anthropic acknowledged that the Cathedral gained access to a “limited set” of sales, HR and IT systems. It stressed that its AI model’s source code was not compromised and the hackers could not reach user conversations with Claude.&nbsp;Anthropic said in a statement that they learnt of the breach before Cathedral reached out to them and that access was shut off the same day. The hacking has raised security concerns that external actors could break into Anthropic if a company that was set up only recently can, view echoed by a senior Trump administration official. “If a bunch of kids who have just started a company can do this, then potentially the Chinese aren’t far behind.”&nbsp;Others noted that the breach was linked to a "hygiene failure" of the third-party software Anthropic uses.<h2>Anthropic IPO</h2>The news comes after Anthropic filed for its initial public offering (IPO) as it plans to advance its AI systems to find cures for diseases that have been around for thousands of years and make lives easier for everyone. "'In California, Claude is helping to shorten the lines at the DMV, and to better serve people on Medicaid," its prospectus reads.Also Read: <a href="https://www.wionews.com/trending/trick-or-treat-anthropic-s-ipo-sales-pitch-is-the-perfect-halloween-scare-for-ai-enthusiasts-1793079462575">Anthropic's IPO sales pitch is the perfect Halloween scare for AI enthusiasts</a>&nbsp;"We believe that AI is one of the most consequential technologies in human history. If managed well, it could rapidly accelerate scientific innovation, greatly increase economic growth rates, bring cures for major diseases within reach, and solve problems that have resisted generations of human effort," it states.&nbsp;It further said that AI is capable of "potentially curing diseases that have plagued humanity for millennia, including cancer, Alzheimer’s disease, heart disease, rare genetic diseases and major infectious diseases." The IPO is expected to cross the $1.8 trillion valuation of SpaceX.

Anthropic's internal messages and other information were breached by a new company that started operations this year. The AI company is poised to bring one of the biggest IPOs, with hopes of curing diseases like cancer and Alzehimer's/

An AI startup says it hacked Anthropic's systems in less than 12 hours and was able to access its internal messages. Cathedral is based in the US and was founded this year by former Department of Government Efficiency (DOGE) employees. Its staff breached Anthropic as part of a bug bounty program where people are asked to reveal security flaws. According to The Washington Post, the hack led to information like Slack messages, meeting transcripts, salaries and job offers. Responding to the report, Anthropic acknowledged that the Cathedral gained access to a “limited set” of sales, HR and IT systems. It stressed that its AI model’s source code was not compromised and the hackers could not reach user conversations with Claude. Anthropic said in a statement that they learnt of the breach before Cathedral reached out to them and that access was shut off the same day. The hacking has raised security concerns that external actors could break into Anthropic if a company that was set up only recently can, view echoed by a senior Trump administration official. “If a bunch of kids who have just started a company can do this, then potentially the Chinese aren’t far behind.” Others noted that the breach was linked to a "hygiene failure" of the third-party software Anthropic uses. Anthropic IPO The news comes after Anthropic filed for its initial public offering (IPO) as it plans to advance its AI systems to find cures for diseases that have been around for thousands of years and make lives easier for everyone. "'In California, Claude is helping to shorten the lines at the DMV, and to better serve people on Medicaid," its prospectus reads. Also Read: Anthropic's IPO sales pitch is the perfect Halloween scare for AI enthusiasts "We believe that AI is one of the most consequential technologies in human history. If managed well, it could rapidly accelerate scientific innovation, greatly increase economic growth rates, bring cures for major diseases within reach, and solve problems that have resisted generations of human effort," it states. It further said that AI is capable of "potentially curing diseases that have plagued humanity for millennia, including cancer, Alzheimer’s disease, heart disease, rare genetic diseases and major infectious diseases." The IPO is expected to cross the $1.8 trillion valuation of SpaceX.

Anthropic's internal messages and other information were breached by a new company that started operations this year. The AI company is poised to bring one of the biggest IPOs, with hopes of curing diseases like cancer and Alzehimer's/

Anthropic hacked by AI startup comprising former DOGE staff in just 12 hours

The usual life of a customer portal goes like this. It launches with an announcement, logins spike for a fortnight, and then the numbers slide. Customers who cannot find an answer in it do what they did before it existed, which is open a support case. Within a year, the portal is a link in an email signature and the support queue is carrying the load it was built to relieve.Sai Saketh Sunkara has spent his career inside Salesforce implementations, and his read on why that happens is structural rather than cosmetic. The portal fails because it only holds one kind of thing. Product documentation lives in it, but the sales relationship sits in a separate system, the community conversation happens somewhere else entirely, and the customer has to already know which of the three contains their answer.The build he uses to make the point is a Connected Sales and Knowledge Portal, assembled by integrating Salesforce Sales Cloud with Experience Cloud. It replaced a legacy portal that had the engagement problem described above. What went in was not a redesign but a consolidation: personalised content, Knowledge articles, Ideas, Discussions, tutorials and peer collaboration, sitting on the same platform as the sales information and the support resources.A customer question rarely respects those boundaries. Someone asking whether a feature is available on their plan is asking a sales question, a documentation question and often a community question at once, and a portal that can only see one of the three sends them to a human by default.Sunkara reports the portal increased case deflection and reduced the operational pressure associated with agent-assisted support, and that it drew positive feedback from users and acceptance from sales stakeholders. He does not put a figure on his own deployment, which is worth noting in a field where portal claims are usually expressed as percentages.His percentages come from research rather than from that build. Studying four B2B Experience Cloud implementations over twelve months, he found average portal adoption rising from 41 percent to 80.9 percent, self-service resolution from 31.5 percent to 72.4 percent, and case deflection from 27 percent to 68.5 percent. A composite customer experience score more than doubled, from 19.21 to 42.84, while Net Promoter Score moved from 28 to 61. The available Knowledge base grew from 120 articles to 485.Those numbers describe what structured implementations achieved across four organisations. They are not measurements of his own portal, and he is careful about the distinction, presenting them as validation of the principles he applied rather than as his project's results.What the research changed, by his account, is how he thinks the work should be judged. Most organisations measure a portal by whether it launched and how many people logged in, which tells you almost nothing. His position is that you set baselines before deployment and then track adoption, search success, self-service resolution, case deflection, onboarding time, customer satisfaction and escalation quality. "This turns the portal into a measurable customer-success platform rather than a one-time technology implementation," he says.Search success is the item on that list most organisations never instrument, and it is the one he finds most revealing. A portal that logs what customers searched for and did not find is describing its own gaps in the customers' own words.That thought extends into his view of what a community is actually for. The conventional reason to build one is engagement, which he treats as the lesser benefit. "A community should create intelligence, not just engagement," he argues. Product ideas, discussion volume, frequently viewed content, unsuccessful searches and accepted answers together form a readout of what customers need and where the friction is, and his recommendation is that those signals be routed back to the sales, support, customer-success and product teams rather than left in the portal as usage statistics.That pattern shows up in adjacent work. A health insurance broker portal he designed, built on Commerce Cloud and Experience Cloud, gave brokers self-service management of enrolments and workflows, replacing broker-to-payer exchanges that had run through slower channels. The underlying move is identical: take a transaction that required a person and give the person doing it direct access instead.His forward view puts artificial intelligence at the relevance layer rather than the content layer. The next generation of these portals, in his reading, will interpret natural-language questions, recommend the content that fits, summarise trusted Knowledge, walk users through workflows, and decide when a human is genuinely needed. He points to Salesforce's support for context-aware Agentforce conversations and AI-generated answers inside Experience Cloud as evidence the shift from static portal to intelligent self-service is already underway.Sunkara holds 13 Salesforce certifications and has published research on B2B portal transformation. The through-line in both the practice and the writing is the same unglamorous claim: a portal earns its keep when it can resolve a question, and the only way to know whether it does is to have measured what happened before it existed.

Salesforce specialist Sai Saketh Sunkara explains how integrated customer portals can improve adoption, resolve queries through self-service and reduce support cases

The usual life of a customer portal goes like this. It launches with an announcement, logins spike for a fortnight, and then the numbers slide. Customers who cannot find an answer in it do what they did before it existed, which is open a support case. Within a year, the portal is a link in an email signature and the support queue is carrying the load it was built to relieve. Sai Saketh Sunkara has spent his career inside Salesforce implementations, and his read on why that happens is structural rather than cosmetic. The portal fails because it only holds one kind of thing. Product documentation lives in it, but the sales relationship sits in a separate system, the community conversation happens somewhere else entirely, and the customer has to already know which of the three contains their answer. The build he uses to make the point is a Connected Sales and Knowledge Portal, assembled by integrating Salesforce Sales Cloud with Experience Cloud. It replaced a legacy portal that had the engagement problem described above. What went in was not a redesign but a consolidation: personalised content, Knowledge articles, Ideas, Discussions, tutorials and peer collaboration, sitting on the same platform as the sales information and the support resources. A customer question rarely respects those boundaries. Someone asking whether a feature is available on their plan is asking a sales question, a documentation question and often a community question at once, and a portal that can only see one of the three sends them to a human by default. Sunkara reports the portal increased case deflection and reduced the operational pressure associated with agent-assisted support, and that it drew positive feedback from users and acceptance from sales stakeholders. He does not put a figure on his own deployment, which is worth noting in a field where portal claims are usually expressed as percentages. His percentages come from research rather than from that build. Studying four B2B Experience Cloud implementations over twelve months, he found average portal adoption rising from 41 percent to 80.9 percent, self-service resolution from 31.5 percent to 72.4 percent, and case deflection from 27 percent to 68.5 percent. A composite customer experience score more than doubled, from 19.21 to 42.84, while Net Promoter Score moved from 28 to 61. The available Knowledge base grew from 120 articles to 485. Those numbers describe what structured implementations achieved across four organisations. They are not measurements of his own portal, and he is careful about the distinction, presenting them as validation of the principles he applied rather than as his project's results. What the research changed, by his account, is how he thinks the work should be judged. Most organisations measure a portal by whether it launched and how many people logged in, which tells you almost nothing. His position is that you set baselines before deployment and then track adoption, search success, self-service resolution, case deflection, onboarding time, customer satisfaction and escalation quality. "This turns the portal into a measurable customer-success platform rather than a one-time technology implementation," he says. Search success is the item on that list most organisations never instrument, and it is the one he finds most revealing. A portal that logs what customers searched for and did not find is describing its own gaps in the customers' own words. That thought extends into his view of what a community is actually for. The conventional reason to build one is engagement, which he treats as the lesser benefit. "A community should create intelligence, not just engagement," he argues. Product ideas, discussion volume, frequently viewed content, unsuccessful searches and accepted answers together form a readout of what customers need and where the friction is, and his recommendation is that those signals be routed back to the sales, support, customer-success and product teams rather than left in the portal as usage statistics. That pattern shows up in adjacent work. A health insurance broker portal he designed, built on Commerce Cloud and Experience Cloud, gave brokers self-service management of enrolments and workflows, replacing broker-to-payer exchanges that had run through slower channels. The underlying move is identical: take a transaction that required a person and give the person doing it direct access instead. His forward view puts artificial intelligence at the relevance layer rather than the content layer. The next generation of these portals, in his reading, will interpret natural-language questions, recommend the content that fits, summarise trusted Knowledge, walk users through workflows, and decide when a human is genuinely needed. He points to Salesforce's support for context-aware Agentforce conversations and AI-generated answers inside Experience Cloud as evidence the shift from static portal to intelligent self-service is already underway. Sunkara holds 13 Salesforce certifications and has published research on B2B portal transformation. The through-line in both the practice and the writing is the same unglamorous claim: a portal earns its keep when it can resolve a question, and the only way to know whether it does is to have measured what happened before it existed.

Salesforce specialist Sai Saketh Sunkara explains how integrated customer portals can improve adoption, resolve queries through self-service and reduce support cases

Reimagining customer success with Salesforce Sales and Experience Cloud

The pilot goes well. An assistant reads a support case, produces a clean summary, drafts a reply that sounds like a person wrote it. Nobody in the room argues with the result. Then the project stalls, or it ships and quietly stops being used.&nbsp;&nbsp;Ashok Apati has designed enough of those pilots to have a theory about what happens next, and it is not a flattering one for the technology. "The hardest enterprise AI problems are often not model problems," he says. "They are architecture, integration, data, security, workflow, governance, and adoption problems."&nbsp;&nbsp;His own career has been spent building the layer underneath. Now a lead product developer and digital transformation leader at a global consumer goods company, Apati has worked for more than 13 years across enterprise integration architecture, Salesforce, cloud transformation and digital commerce. He led the technical design and proof-of-concept development of generative AI capabilities built into Salesforce, including Case and Account Assistants covering case summarisation, drafted customer responses, risk identification, next-best-action recommendations, knowledge suggestions, account intelligence and pre-visit preparation.&nbsp;&nbsp;That work is where his argument comes from. The assistants demonstrated quickly that a language model could summarise a case usefully. What took longer was everything that had to be true before such a thing could be trusted inside a working business: retrieving the right Salesforce context, excluding personal data that should never reach a model, defining reusable AI services rather than one-off integrations, structuring prompts around specific business actions, and keeping a human in front of anything consequential.&nbsp;&nbsp;None of that is model work. "Enterprise AI is an architecture challenge, not simply an LLM challenge," he says. Organisations spend their time comparing models when the harder question is structural: "creating an architecture that can provide the model with the right context, at the right time, for the right user."&nbsp;&nbsp;There is a less comfortable corollary, and he states it plainly. "AI exposes existing technology and data debt." Fragmented customer data, inconsistent identifiers, outdated APIs, duplicated information, weak integration patterns and unclear data ownership have usually been tolerable for years. Then something arrives that needs to read across all of it at once. "AI does not eliminate these problems," he says. "In many cases, it makes them more visible."&nbsp;&nbsp;He is in an unusual position to say so, because he spent years clearing that debt before AI made it urgent. He played a key technical leadership role in migrating mission-critical applications and integrations to Microsoft Azure while the business kept running, contributing to roughly $600,000 in annual cost savings. He has designed, modernised and supported platforms carrying approximately $260 million in annual transaction volume, serving more than 150,000 active customers and processing over 100,000 orders a year. He helped redesign a customer registration experience that has since handled around 110,000 registrations, and contributed to a single sign-on programme built on Okta.&nbsp;&nbsp;Those projects explain the sequence he now argues for. Cloud provides scalability, APIs and integrations provide connectivity, identity establishes trusted access, enterprise data provides context, and only then does AI have something solid to reason over. "AI transformation is not separate from digital transformation," he says. "It is the next layer of it."&nbsp;&nbsp;Which is also why he is unsentimental about how most AI programmes begin. The common mistake, in his telling, starts with the wrong question: "We have access to Generative AI, what can we build?" He would rather start from a measurable problem, reducing case-handling time or shortening sales preparation, and let that decide what gets built and which model is used.&nbsp;&nbsp;Two of his design rules follow directly from watching pilots fail. The first is that AI has to appear where people already work. If an employee has to leave the application they are in, open a separate tool, paste in the context by hand and carry the answer back, the tool will be abandoned no matter how good the output is.&nbsp;&nbsp;The second is about who remains accountable. For consequential work such as complaint handling, reason-code selection and customer communications, the pattern he favours is deliberately conservative: AI recommends, a human validates, and the enterprise system executes. Rather than forcing a single reason code, the system offers a shortlist for an agent to confirm. Drafted replies get reviewed before they are sent. Autonomy increases later, he argues, and only as governance and measured performance earn it. &nbsp;&nbsp;Underneath both rules sits a preference for substance over polish. "Grounding is more important than eloquence," he says, and the reason is that a fluent wrong answer travels further than an obvious one. "An AI-generated answer that sounds professional but contains incorrect business information can be more dangerous than having no answer at all."&nbsp;&nbsp;It leads him to the test most pilots never have to pass. "A successful demonstration is not evidence of a successful AI implementation," he says, and pilots, in his view, should carry measurable exit criteria before anyone calls them a success: case-handling time, preparation time saved, recommendation acceptance, adoption, cost per interaction.&nbsp;&nbsp;So the question he would put to any organisation shopping for a model is a different one entirely. "Is our enterprise architecture ready to make AI useful, trustworthy, measurable, and scalable?" Pilots rarely die because the model disappointed. They die in the gap between a demonstration and an organisation that was never quite ready to receive one.

Fragmented customer data, inconsistent identifiers, outdated APIs, duplicated information, weak integration patterns and unclear data ownership have usually been tolerable for years. Then something arrives that needs to read across all of it at once.

The pilot goes well. An assistant reads a support case, produces a clean summary, drafts a reply that sounds like a person wrote it. Nobody in the room argues with the result. Then the project stalls, or it ships and quietly stops being used. Ashok Apati has designed enough of those pilots to have a theory about what happens next, and it is not a flattering one for the technology. "The hardest enterprise AI problems are often not model problems," he says. "They are architecture, integration, data, security, workflow, governance, and adoption problems." His own career has been spent building the layer underneath. Now a lead product developer and digital transformation leader at a global consumer goods company, Apati has worked for more than 13 years across enterprise integration architecture, Salesforce, cloud transformation and digital commerce. He led the technical design and proof-of-concept development of generative AI capabilities built into Salesforce, including Case and Account Assistants covering case summarisation, drafted customer responses, risk identification, next-best-action recommendations, knowledge suggestions, account intelligence and pre-visit preparation. That work is where his argument comes from. The assistants demonstrated quickly that a language model could summarise a case usefully. What took longer was everything that had to be true before such a thing could be trusted inside a working business: retrieving the right Salesforce context, excluding personal data that should never reach a model, defining reusable AI services rather than one-off integrations, structuring prompts around specific business actions, and keeping a human in front of anything consequential. None of that is model work. "Enterprise AI is an architecture challenge, not simply an LLM challenge," he says. Organisations spend their time comparing models when the harder question is structural: "creating an architecture that can provide the model with the right context, at the right time, for the right user." There is a less comfortable corollary, and he states it plainly. "AI exposes existing technology and data debt." Fragmented customer data, inconsistent identifiers, outdated APIs, duplicated information, weak integration patterns and unclear data ownership have usually been tolerable for years. Then something arrives that needs to read across all of it at once. "AI does not eliminate these problems," he says. "In many cases, it makes them more visible." He is in an unusual position to say so, because he spent years clearing that debt before AI made it urgent. He played a key technical leadership role in migrating mission-critical applications and integrations to Microsoft Azure while the business kept running, contributing to roughly $600,000 in annual cost savings. He has designed, modernised and supported platforms carrying approximately $260 million in annual transaction volume, serving more than 150,000 active customers and processing over 100,000 orders a year. He helped redesign a customer registration experience that has since handled around 110,000 registrations, and contributed to a single sign-on programme built on Okta. Those projects explain the sequence he now argues for. Cloud provides scalability, APIs and integrations provide connectivity, identity establishes trusted access, enterprise data provides context, and only then does AI have something solid to reason over. "AI transformation is not separate from digital transformation," he says. "It is the next layer of it." Which is also why he is unsentimental about how most AI programmes begin. The common mistake, in his telling, starts with the wrong question: "We have access to Generative AI, what can we build?" He would rather start from a measurable problem, reducing case-handling time or shortening sales preparation, and let that decide what gets built and which model is used. Two of his design rules follow directly from watching pilots fail. The first is that AI has to appear where people already work. If an employee has to leave the application they are in, open a separate tool, paste in the context by hand and carry the answer back, the tool will be abandoned no matter how good the output is. The second is about who remains accountable. For consequential work such as complaint handling, reason-code selection and customer communications, the pattern he favours is deliberately conservative: AI recommends, a human validates, and the enterprise system executes. Rather than forcing a single reason code, the system offers a shortlist for an agent to confirm. Drafted replies get reviewed before they are sent. Autonomy increases later, he argues, and only as governance and measured performance earn it. Underneath both rules sits a preference for substance over polish. "Grounding is more important than eloquence," he says, and the reason is that a fluent wrong answer travels further than an obvious one. "An AI-generated answer that sounds professional but contains incorrect business information can be more dangerous than having no answer at all." It leads him to the test most pilots never have to pass. "A successful demonstration is not evidence of a successful AI implementation," he says, and pilots, in his view, should carry measurable exit criteria before anyone calls them a success: case-handling time, preparation time saved, recommendation acceptance, adoption, cost per interaction. So the question he would put to any organisation shopping for a model is a different one entirely. "Is our enterprise architecture ready to make AI useful, trustworthy, measurable, and scalable?" Pilots rarely die because the model disappointed. They die in the gap between a demonstration and an organisation that was never quite ready to receive one.

Fragmented customer data, inconsistent identifiers, outdated APIs, duplicated information, weak integration patterns and unclear data ownership have usually been tolerable for years. Then something arrives that needs to read across all of it at once.

Why Enterprise AI keeps dying in the pilot stage

Microsoft is reportedly offering voluntary exit packages to around 60,000 employees, giving eligible workers the option to leave the company with financial support as the technology giant adapts its workforce for an AI-driven future.&nbsp;The programme reportedly covers staff in India, the United States, the United Kingdom and Canada, including employees at vice president level and below. Unlike compulsory layoffs, the scheme allows eligible employees to decide whether they want to leave.The move comes as Microsoft continues investing heavily in artificial intelligence (AI), data centres and computing infrastructure. However, the company has not said the buyout programme is solely a response to AI spending.<h2>Why is Microsoft offering buyouts to 60,000 employees?</h2>According to reports by Bloomberg, Microsoft wants to give employees who are considering a future outside the company another option, while allowing business divisions to adjust staffing levels as their needs change. Chief People Officer Amy Coleman reportedly told employees that the offer provides support for those considering leaving and gives participating teams another way to adapt their workforce.Microsoft had 223,000 full-time employees worldwide as of June 2026. Its workforce has remained broadly stable since 2022, around the time the company began expanding its AI investments. The latest programme could allow Microsoft to change its workforce without relying entirely on compulsory job cuts. The company has not announced a target for how many employees it expects to accept the offer.<h2>Which Microsoft teams could be affected?</h2>Reports indicate that employees in divisions including Xbox, Windows, Copilot and finance could be eligible, depending on their role and participating team. Some workers are reportedly excluded, including employees in data centres, sales, marketing, cybersecurity and quantum computing. The programme is expected to cover eligible employees in India, the US, the UK and Canada, with eligibility extending to vice president level and below. Individual eligibility will depend on the company's final programme details. Employees should therefore check the official communication from Microsoft rather than assume that every worker in an included division qualifies.<h2>How does Microsoft’s AI push fit into the plan?</h2>Microsoft is spending heavily on the infrastructure needed to develop and run AI services. That includes data centres, advanced chips and computing capacity supporting its own products and external AI customers. The company is also competing in a fast-moving market where AI tools such as Copilot are becoming a larger part of its product strategy. The buyout programme highlights a wider challenge for major technology companies: adapting their workforce while investing in expensive new technologies. For employees, the offer could provide a supported route out of the company. For Microsoft, it offers another way to adjust staffing as business priorities change. Whether the programme leads to a significant reduction in headcount will depend on how many eligible employees choose to accept it.

Microsoft is reportedly offering voluntary buyouts to around 60,000 employees in India, the US, the UK and Canada. The programme allows eligible staff to leave with financial support as the company adapts its workforce amid growing investment in AI and data centres.

Microsoft is reportedly offering voluntary exit packages to around 60,000 employees, giving eligible workers the option to leave the company with financial support as the technology giant adapts its workforce for an AI-driven future. The programme reportedly covers staff in India, the United States, the United Kingdom and Canada, including employees at vice president level and below. Unlike compulsory layoffs, the scheme allows eligible employees to decide whether they want to leave. The move comes as Microsoft continues investing heavily in artificial intelligence (AI), data centres and computing infrastructure. However, the company has not said the buyout programme is solely a response to AI spending. Why is Microsoft offering buyouts to 60,000 employees? According to reports by Bloomberg, Microsoft wants to give employees who are considering a future outside the company another option, while allowing business divisions to adjust staffing levels as their needs change. Chief People Officer Amy Coleman reportedly told employees that the offer provides support for those considering leaving and gives participating teams another way to adapt their workforce. Microsoft had 223,000 full-time employees worldwide as of June 2026. Its workforce has remained broadly stable since 2022, around the time the company began expanding its AI investments. The latest programme could allow Microsoft to change its workforce without relying entirely on compulsory job cuts. The company has not announced a target for how many employees it expects to accept the offer. Which Microsoft teams could be affected? Reports indicate that employees in divisions including Xbox, Windows, Copilot and finance could be eligible, depending on their role and participating team. Some workers are reportedly excluded, including employees in data centres, sales, marketing, cybersecurity and quantum computing. The programme is expected to cover eligible employees in India, the US, the UK and Canada, with eligibility extending to vice president level and below. Individual eligibility will depend on the company's final programme details. Employees should therefore check the official communication from Microsoft rather than assume that every worker in an included division qualifies. How does Microsoft’s AI push fit into the plan? Microsoft is spending heavily on the infrastructure needed to develop and run AI services. That includes data centres, advanced chips and computing capacity supporting its own products and external AI customers. The company is also competing in a fast-moving market where AI tools such as Copilot are becoming a larger part of its product strategy. The buyout programme highlights a wider challenge for major technology companies: adapting their workforce while investing in expensive new technologies. For employees, the offer could provide a supported route out of the company. For Microsoft, it offers another way to adjust staffing as business priorities change. Whether the programme leads to a significant reduction in headcount will depend on how many eligible employees choose to accept it.

Microsoft is reportedly offering voluntary buyouts to around 60,000 employees in India, the US, the UK and Canada. The programme allows eligible staff to leave with financial support as the company adapts its workforce amid growing investment in AI and data centres.

The money going into AI data centres has to come from somewhere: What Microsoft’s buyout plan reveals

Luiz Inácio Lula da Silva’s election campaign appears to have flooded the internet with old and new AI-generated videos of the man who became Brazil's president one more time, including one in which he dances to Michael Jackson’s famous song "They Don’t Care About Us". While these were primarily humorous memes shared by supporters and circulated organically, &nbsp;it also helped influence recommendation algorithms to push such content amid alleged bias in favour of his rival, who lost. On 25 October, Lula won Brazil’s presidential run-off election, securing a record fourth term after defeating Flávio Bolsonaro, the son of former president Jair Bolsonaro.<h2>Lula as Michael Jackson</h2>The content originated from individual creators, fan accounts and music fandoms. One video that went viral showed a Brazilian version of the music video, titled “They Don’t Care About Us, versão Lula”. Lula was shown ‘performing’ to Michael Jackson’s song, whose original music video was filmed in Brazil in 1996 under the direction of Spike Lee. The video featured locations in Salvador and Rio de Janeiro.Other AI-generated videos showed Lula dancing to songs by Lady Gaga, Taylor Swift and Beyoncé, alongside superhero edits and football scenes featuring the president.Some videos showed Lula and his allies dancing on a stage, mimicking Lady Gaga’s performance in Rio de Janeiro. Others inserted Lula into Taylor Swift’s Eras Tour-style performances, recreated classic Madonna looks or showed him dancing to Beyoncé’s “Crazy in Love”.Lula was also portrayed as Captain America facing Flávio Bolsonaro as Thanos from the Marvel Superhero Universe. Other clips depicted the Brazilian president as a Victoria’s Secret Angel, or alongside football goal celebrations and stadium scenes.Also read: <a href="https://www.wionews.com/race-to-power/lula-fourth-presidential-victory-trump-latin-america-1793032769668">Lula’s 4th presidential victory a blow to Trump’s ambitions in Latin America | The Verdict</a><h2>Not overtly political, but fun to watch: A social media movement against perceived bias</h2>Some supporters and commentators suggested that the trend was an attempt to flood social media platforms such as X, Facebook, TikTok and Instagram in response to perceived algorithmic bias. The allegation was that Meta - which owns Facebook, Instagram and WhatsApp - and X favoured Bolsonaro in their recommendation systems.Many of these videos were not explicitly political. This may have helped them circulate, go viral and bypass restrictions that might otherwise affect conventional campaign material.Some official accounts associated with Lula’s campaign shared content that blended real footage with viral trends or moments, while supporters flooded timelines with AI-generated memes.Also read: <a href="https://www.wionews.com/world/brazil-re-elects-luiz-in-cio-lula-da-silva-for-fourth-term-80-year-old-makes-history-by-defeating-fl-vio-bolsanaro-1792978385715">Brazil re-elects factory worker Lula for fourth term; Indian PM Modi congratulates 'dear friend'</a><h2>Lula’s team and allies had accused Meta of algorithmic manipulation</h2>During the election campaign period, the Lula Army had accused Meta and other social media platforms of bias, alleging that they suppressed the reach of Lula-related content and imposed advertising restrictions.They claimed, for instance, that Meta AI refused to generate images of Lula while allowing images of Bolsonaro. The campaign also alleged that recommendation engines worked differently for the two candidates, favouring Bolsonaro-related content.On 14 October, Lula’s campaign formally approached Brazil’s Tribunal Superior Eleitoral or TSE, seeking equal treatment on Meta’s platforms.Meta attributed the problems to technical glitches, anti-fraud measures and neutral rules, denying intentional political interference. On 15 October, the TSE ordered Meta to ensure equal treatment of the candidates and prohibited unequal restrictions on the distribution and recommendation of their content. The court’s preliminary assessment did not establish that the alleged AI restrictions resulted from deliberate political bias.<a href="https://www.wionews.com/videos/lula-s-campaign-reported-restrictions-on-instagram-posts-1792597837534/embed">https://www.wionews.com/videos/lula-s-campaign-reported-restrictions-on-instagram-posts-1792597837534/embed</a><h2>AI and deepfakes used to build clout and aura</h2>Both camps used AI extensively during the election campaign. Fact-checkers at the Lupa media organisation documented hundreds of suspected AI-generated or manipulated political posts. In one review, Lupa identified 282 AI-generated or altered items among 314 suspicious posts collected between 16 and 26 August.Deepfakes and synthetic content circulated among both pro- and anti-Lula supporters, prompting electoral authorities to intervene. Brazilian electoral rules require political and electoral content created or manipulated using AI to be clearly labelled. The TSE has also ordered the removal of misleading or improperly labelled material in specific cases.&nbsp;Here are some interesting videos:&nbsp;

<blockquote class="twitter-tweet">Viciada em vídeo de IA com Lula https://t.co/Fi0XZFltDR&mdash; Bernadete-de-chocolate-com-pimenta-não-binária (@jaoadymello) <a href="https://x.com/jaoadymello/status/2113661431350751303?ref_src=twsrc%5Etfw">October 23, 2026</a></blockquote> <script async src="https://platform.x.com/widgets.js" charset="utf-8"></script>

<blockquote class="twitter-tweet">TODOS los Fans de Michael Jackson, estamos con Lula Da Silva! Libardo Rodríguez Soto https://t.co/Fjj9YSVcHJ https://t.co/b0C3ypTvKK&mdash; Libardo Rodríguez Soto (@LIROSO91) <a href="https://x.com/LIROSO91/status/2114773345040220595?ref_src=twsrc%5Etfw">October 26, 2026</a></blockquote> <script async src="https://platform.x.com/widgets.js" charset="utf-8"></script>

Lula supporters flooded social media with AI-generated memes, turning him into Michael Jackson or Captain America. This was not accidental. There were allegations of algorithmic bias against Lula by social media giants

Luiz Inácio Lula da Silva’s election campaign appears to have flooded the internet with old and new AI-generated videos of the man who became Brazil's president one more time, including one in which he dances to Michael Jackson’s famous song "They Don’t Care About Us". While these were primarily humorous memes shared by supporters and circulated organically, it also helped influence recommendation algorithms to push such content amid alleged bias in favour of his rival, who lost. On 25 October, Lula won Brazil’s presidential run-off election, securing a record fourth term after defeating Flávio Bolsonaro, the son of former president Jair Bolsonaro. Lula as Michael Jackson The content originated from individual creators, fan accounts and music fandoms. One video that went viral showed a Brazilian version of the music video, titled “They Don’t Care About Us, versão Lula”. Lula was shown ‘performing’ to Michael Jackson’s song, whose original music video was filmed in Brazil in 1996 under the direction of Spike Lee. The video featured locations in Salvador and Rio de Janeiro. Other AI-generated videos showed Lula dancing to songs by Lady Gaga, Taylor Swift and Beyoncé, alongside superhero edits and football scenes featuring the president. Some videos showed Lula and his allies dancing on a stage, mimicking Lady Gaga’s performance in Rio de Janeiro. Others inserted Lula into Taylor Swift’s Eras Tour-style performances, recreated classic Madonna looks or showed him dancing to Beyoncé’s “Crazy in Love”. Lula was also portrayed as Captain America facing Flávio Bolsonaro as Thanos from the Marvel Superhero Universe. Other clips depicted the Brazilian president as a Victoria’s Secret Angel, or alongside football goal celebrations and stadium scenes. Also read: Lula’s 4th presidential victory a blow to Trump’s ambitions in Latin America | The Verdict Not overtly political, but fun to watch: A social media movement against perceived bias Some supporters and commentators suggested that the trend was an attempt to flood social media platforms such as X, Facebook, TikTok and Instagram in response to perceived algorithmic bias. The allegation was that Meta - which owns Facebook, Instagram and WhatsApp - and X favoured Bolsonaro in their recommendation systems. Many of these videos were not explicitly political. This may have helped them circulate, go viral and bypass restrictions that might otherwise affect conventional campaign material. Some official accounts associated with Lula’s campaign shared content that blended real footage with viral trends or moments, while supporters flooded timelines with AI-generated memes. Also read: Brazil re-elects factory worker Lula for fourth term; Indian PM Modi congratulates 'dear friend' Lula’s team and allies had accused Meta of algorithmic manipulation During the election campaign period, the Lula Army had accused Meta and other social media platforms of bias, alleging that they suppressed the reach of Lula-related content and imposed advertising restrictions. They claimed, for instance, that Meta AI refused to generate images of Lula while allowing images of Bolsonaro. The campaign also alleged that recommendation engines worked differently for the two candidates, favouring Bolsonaro-related content. On 14 October, Lula’s campaign formally approached Brazil’s Tribunal Superior Eleitoral or TSE, seeking equal treatment on Meta’s platforms. Meta attributed the problems to technical glitches, anti-fraud measures and neutral rules, denying intentional political interference. On 15 October, the TSE ordered Meta to ensure equal treatment of the candidates and prohibited unequal restrictions on the distribution and recommendation of their content. The court’s preliminary assessment did not establish that the alleged AI restrictions resulted from deliberate political bias. https://www.wionews.com/videos/lula-s-campaign-reported-restrictions-on-instagram-posts-1792597837534/embed AI and deepfakes used to build clout and aura Both camps used AI extensively during the election campaign. Fact-checkers at the Lupa media organisation documented hundreds of suspected AI-generated or manipulated political posts. In one review, Lupa identified 282 AI-generated or altered items among 314 suspicious posts collected between 16 and 26 August. Deepfakes and synthetic content circulated among both pro- and anti-Lula supporters, prompting electoral authorities to intervene. Brazilian electoral rules require political and electoral content created or manipulated using AI to be clearly labelled. The TSE has also ordered the removal of misleading or improperly labelled material in specific cases. Here are some interesting videos:

Lula supporters flooded social media with AI-generated memes, turning him into Michael Jackson or Captain America. This was not accidental. There were allegations of algorithmic bias against Lula by social media giants

Transforming Lula into Michael Jackson and Captain America, how his ‘army’ flooded the internet with AI clips and memes to fight ‘algorithm bias’

The scan chain mismatch sat at the output like a locked door. In the old workflow, Sashaank Kasturirangan would have spent the better part of a day tracing the chain by hand, then pulling up power state tables in a separate window, hoping the two stories lined up. On this run, the answer came back in minutes. A vendor tool pulled the scan path and the power domain status. An AI layer, trained on the shape of past failures, flagged a missing isolation cell and pointed at the X propagation source.&nbsp;That shift, from days to minutes on a single bug, is the kind of result Kasturirangan has been chasing inside gate-level verification, the pre-tapeout stage where a chip's final netlist is pounded against simulation to catch anything that would turn silicon into scrap. He has built an AI-assisted debug framework that has cut scandump gate-level verification timelines from six to eight weeks down to three to four, roughly a 50 percent reduction in debug time on the projects where it has been deployed.&nbsp;The framework does not touch the simulation or build cycles themselves. Those take what they take. What it changes is the human bottleneck in the middle: an engineer squinting at waveforms, cross-referencing logs, and trying to hold a scan chain's state in their head long enough to form a hypothesis.&nbsp;"Originally scandump gate level verification would take 6 to 8 weeks to catch all bugs in scan chains," Sashaank says. "The waveform analysis and debug would take a whole day sometimes, or at least a few hours."&nbsp;His approach treats the pile of artifacts a modern simulation throws off, waveforms, logs, scan states, power domain information, as raw material to be reshaped before any model sees it. The system converts that mess into structured, context-rich inputs. Then a reasoning layer, built oninternally developed AI tooling and Model Context Protocol integrations from EDA vendors, works through the structured view to isolate root causes: X-propagation sources, broken scan chains, awkward interactions across power domains.&nbsp;The training story matters more than the architecture. Kasturirangan fed the system known failures from across multiple projects. Scan chain shifts that broke mid-pattern. Preshift failures caused by improper backdoor flop randomization of reset signals. Each pattern was documented with the prompt that unlocked it, then folded into a verification framework other engineers could call on. Repeated failure types that used to swallow an hour or two of manual inspection now resolve in minutes.&nbsp;Across projects, he reports a 40 to 65 percent reduction in debug turnaround time, with meaningful gains in first-pass root cause accuracy. The work started inside his own group and has since been shared with other groups for cross-functional pickup.&nbsp;Getting there required a juggling act most verification engineers will recognize. Tapeout schedules do not pause for research projects. Kasturirangan was expected to produce real results on real chips while, in parallel, training a tool that might or might not pay off in time. The bet was that extra hours spent on two early projects would compound later in the pipeline.&nbsp;"Given the complexity of chips and chip manufacturing is ever increasing exponentially with billions of transistors in design nowadays, it is very important to maintain quality while improving efficiency," he says.&nbsp;That constraint shaped the design. He built what he calls a structured debug abstraction layer, a middle tier that separates the messy input data from the reasoning on top. Context-aware patterns written for one design could be reused on the next without rewriting the plumbing. Bringup time on new projects dropped as a result.&nbsp;The scan mismatch case is the one he uses to show skeptics what the workflow actually does. Traditional path: manually trace the scan chain, check power states separately, correlate by eye. New path: the vendor MCP extracts the chain and the power domain OFF status in one pull, and the AI layer identifies the missing isolation and the X source. Hours or days collapse into minutes. The engineer still signs off on the finding, but no longer spends the shift getting there.&nbsp;Kasturirangan is careful about where he thinks this ends up. The automation handles patterns it has seen before, or close relatives of them. Novel bugs, the ones that make tapeout meetings tense, still need someone who understands the fundamentals of design verification to recognize when the AI is confidently wrong.&nbsp;"The future of chip development is going to be radically accelerated with AI assisted flows and debugs," he says, "but an engineer who knows the fundamentals of chip design verification will catch any flaws in the cycle or due to AI inaccuracies and further fine tune the process."For now, he is extending the pattern library, watching how the abstraction layer holds up on larger designs, and tracking which failure classes keep showing up often enough to be worth their own prompt. Each one documented is an hour another engineer will not spend staring at a waveform at midnight

DFT specialist Sashaank Kasturirangan has developed an AI-assisted debug framework that cuts gate-level verification timelines by up to 50 per cent. The system helps engineers identify scan-chain failures, X-propagation and power-domain issues in minutes instead of hours

The scan chain mismatch sat at the output like a locked door. In the old workflow, Sashaank Kasturirangan would have spent the better part of a day tracing the chain by hand, then pulling up power state tables in a separate window, hoping the two stories lined up. On this run, the answer came back in minutes. A vendor tool pulled the scan path and the power domain status. An AI layer, trained on the shape of past failures, flagged a missing isolation cell and pointed at the X propagation source. That shift, from days to minutes on a single bug, is the kind of result Kasturirangan has been chasing inside gate-level verification, the pre-tapeout stage where a chip's final netlist is pounded against simulation to catch anything that would turn silicon into scrap. He has built an AI-assisted debug framework that has cut scandump gate-level verification timelines from six to eight weeks down to three to four, roughly a 50 percent reduction in debug time on the projects where it has been deployed. The framework does not touch the simulation or build cycles themselves. Those take what they take. What it changes is the human bottleneck in the middle: an engineer squinting at waveforms, cross-referencing logs, and trying to hold a scan chain's state in their head long enough to form a hypothesis. "Originally scandump gate level verification would take 6 to 8 weeks to catch all bugs in scan chains," Sashaank says. "The waveform analysis and debug would take a whole day sometimes, or at least a few hours." His approach treats the pile of artifacts a modern simulation throws off, waveforms, logs, scan states, power domain information, as raw material to be reshaped before any model sees it. The system converts that mess into structured, context-rich inputs. Then a reasoning layer, built on internally developed AI tooling and Model Context Protocol integrations from EDA vendors, works through the structured view to isolate root causes: X-propagation sources, broken scan chains, awkward interactions across power domains. The training story matters more than the architecture. Kasturirangan fed the system known failures from across multiple projects. Scan chain shifts that broke mid-pattern. Preshift failures caused by improper backdoor flop randomization of reset signals. Each pattern was documented with the prompt that unlocked it, then folded into a verification framework other engineers could call on. Repeated failure types that used to swallow an hour or two of manual inspection now resolve in minutes. Across projects, he reports a 40 to 65 percent reduction in debug turnaround time, with meaningful gains in first-pass root cause accuracy. The work started inside his own group and has since been shared with other groups for cross-functional pickup. Getting there required a juggling act most verification engineers will recognize. Tapeout schedules do not pause for research projects. Kasturirangan was expected to produce real results on real chips while, in parallel, training a tool that might or might not pay off in time. The bet was that extra hours spent on two early projects would compound later in the pipeline. "Given the complexity of chips and chip manufacturing is ever increasing exponentially with billions of transistors in design nowadays, it is very important to maintain quality while improving efficiency," he says. That constraint shaped the design. He built what he calls a structured debug abstraction layer, a middle tier that separates the messy input data from the reasoning on top. Context-aware patterns written for one design could be reused on the next without rewriting the plumbing. Bringup time on new projects dropped as a result. The scan mismatch case is the one he uses to show skeptics what the workflow actually does. Traditional path: manually trace the scan chain, check power states separately, correlate by eye. New path: the vendor MCP extracts the chain and the power domain OFF status in one pull, and the AI layer identifies the missing isolation and the X source. Hours or days collapse into minutes. The engineer still signs off on the finding, but no longer spends the shift getting there. Kasturirangan is careful about where he thinks this ends up. The automation handles patterns it has seen before, or close relatives of them. Novel bugs, the ones that make tapeout meetings tense, still need someone who understands the fundamentals of design verification to recognize when the AI is confidently wrong. "The future of chip development is going to be radically accelerated with AI assisted flows and debugs," he says, "but an engineer who knows the fundamentals of chip design verification will catch any flaws in the cycle or due to AI inaccuracies and further fine tune the process." For now, he is extending the pattern library, watching how the abstraction layer holds up on larger designs, and tracking which failure classes keep showing up often enough to be worth their own prompt. Each one documented is an hour another engineer will not spend staring at a waveform at midnight

DFT specialist Sashaank Kasturirangan has developed an AI-assisted debug framework that cuts gate-level verification timelines by up to 50 per cent. The system helps engineers identify scan-chain failures, X-propagation and power-domain issues in minutes instead of hours

DFT specialist masters complex chip pre-tapeout trials using AI-assisted gate-level debug

The traffic arrives. Response times climb, a threshold is crossed, the autoscaler notices and requests capacity. Instances start, health checks pass, and new capacity joins the pool.By then the surge has often passed, and the people who hit the slow version of the site have already had that experience. The system did exactly what it was configured to do, and the users still got the bad minutes.Sandipkumar Uchdadiya's argument is that this is not a tuning problem. It is what reactive scaling is: a mechanism defined by responding after a threshold has been crossed, deployed against traffic that frequently gives ample warning it is coming.Uchdadiya has spent over a decade building distributed, high-traffic backend systems, and the case he draws on most is one where the warning was unusually clear. Working as a senior .NET and Azure consultant, he led one of the core modules on the digital platform of a national sports organisation, a system serving close to two million memberships.Traffic on a platform like that is not random. It arrives with the sporting calendar: season openings, major competitions, membership campaign launches. The spikes are sharp, short and, as he puts it, predictable in hindsight.Threshold-based autoscaling treated every one of them as a surprise.The consequence came in two forms simultaneously, and the second is the one organisations tend not to count. Response times degraded during genuine surges. Then, after each incident, the instinct was to provision extra headroom as insurance, which quietly became a permanent line in the infrastructure bill."Provisioning extra headroom after each incident just converted a latency problem into a standing cost problem," he says.That is the hidden cost in the headline, and it is worth being precise about the mechanism. The organisation never decided to over-provision. It arrived there one incident at a time, through a sequence of individually reasonable responses to slow response times, and nobody was ever in a position to see the accumulated total as a choice.What Uchdadiya is clear about is that fixing it was not a single change, and that the order mattered.Retuning the autoscaling rules on its own would have achieved little, because the services that needed to scale could not scale by themselves. The high-traffic components had to be migrated onto Azure Service Fabric first, so that individual services could scale independently rather than the platform having to scale as one unit. Only then did adjusting the rules so capacity responded to the right signals sooner mean anything.Alongside that, he introduced an Azure Redis Cache layer, on the observation that a large share of the spike traffic was read-heavy and did not need to reach the backend services at all. The cheapest request to scale for is one that never arrives.And load testing was used to find where the system broke before it broke, rather than discovering the breaking point live during a national sporting event.The engagement ran across three teams and roughly fifteen people. What his module ended up with was a platform that absorbed known high-traffic windows predictably instead of reactively, without the cycle of slow responses followed by compensating spend.On results, Uchdadiya declines to supply the figures the story would most want, and says so in terms worth quoting."I want to be direct here rather than invent precision I don't have," he says. Hard performance numbers from that period are not recoverable, so he describes the outcome qualitatively and leaves it there. Given that a percentage improvement in response time is exactly the sort of number nobody would check, the decision not to produce one is a reasonable indication of how the rest of the account should be read.The more interesting half of his argument comes from the opposite direction, and from a system he is careful to say he does not own.He now works at a global investment management firm whose platforms run on static, fixed-capacity provisioning rather than dynamic autoscaling. That is a deliberate architectural choice, made because in an environment targeting four nines of availability, the business consequence of scaling arriving late, missing a critical financial moment, was judged unacceptable. He is explicit that he is not on the team responsible for that decision and is describing an architecture he works within rather than one he designed.His point is that this is not an absence of a scaling strategy. It is one."Reactive autoscaling and static over-provisioning are really the same tradeoff viewed from two different risk tolerances," he says. One accepts occasional lateness to avoid paying for idle capacity. The other accepts permanent idle capacity to eliminate the possibility of being late. Both are answers to a single question: is your traffic predictable enough to anticipate, or unpredictable enough that you would rather pay for permanent headroom than risk being slow.Neither answer is wrong. What he objects to is that most organisations never ask."The mistake I see most often is organisations defaulting into reactive threshold-based autoscaling without ever actually asking that question, treating it as a settled default rather than a real architectural decision with a real cost model attached."His recommendation follows from the sports platform and is unglamorous enough to be actionable. If traffic has any recognisable pattern at all, a season, a calendar, a campaign schedule, then that pattern is a leading indicator available to scale against in advance.Reactive thresholds discard it and wait for the symptom instead. Which is a strange thing to do with information you already have.

Reactive scaling is defined by responding after a threshold is crossed. On traffic that arrives on a calendar, that is a decision to be late every time.

The traffic arrives. Response times climb, a threshold is crossed, the autoscaler notices and requests capacity. Instances start, health checks pass, and new capacity joins the pool. By then the surge has often passed, and the people who hit the slow version of the site have already had that experience. The system did exactly what it was configured to do, and the users still got the bad minutes. Sandipkumar Uchdadiya's argument is that this is not a tuning problem. It is what reactive scaling is: a mechanism defined by responding after a threshold has been crossed, deployed against traffic that frequently gives ample warning it is coming. Uchdadiya has spent over a decade building distributed, high-traffic backend systems, and the case he draws on most is one where the warning was unusually clear. Working as a senior .NET and Azure consultant, he led one of the core modules on the digital platform of a national sports organisation, a system serving close to two million memberships. Traffic on a platform like that is not random. It arrives with the sporting calendar: season openings, major competitions, membership campaign launches. The spikes are sharp, short and, as he puts it, predictable in hindsight. Threshold-based autoscaling treated every one of them as a surprise. The consequence came in two forms simultaneously, and the second is the one organisations tend not to count. Response times degraded during genuine surges. Then, after each incident, the instinct was to provision extra headroom as insurance, which quietly became a permanent line in the infrastructure bill. "Provisioning extra headroom after each incident just converted a latency problem into a standing cost problem," he says. That is the hidden cost in the headline, and it is worth being precise about the mechanism. The organisation never decided to over-provision. It arrived there one incident at a time, through a sequence of individually reasonable responses to slow response times, and nobody was ever in a position to see the accumulated total as a choice. What Uchdadiya is clear about is that fixing it was not a single change, and that the order mattered. Retuning the autoscaling rules on its own would have achieved little, because the services that needed to scale could not scale by themselves. The high-traffic components had to be migrated onto Azure Service Fabric first, so that individual services could scale independently rather than the platform having to scale as one unit. Only then did adjusting the rules so capacity responded to the right signals sooner mean anything. Alongside that, he introduced an Azure Redis Cache layer, on the observation that a large share of the spike traffic was read-heavy and did not need to reach the backend services at all. The cheapest request to scale for is one that never arrives. And load testing was used to find where the system broke before it broke, rather than discovering the breaking point live during a national sporting event. The engagement ran across three teams and roughly fifteen people. What his module ended up with was a platform that absorbed known high-traffic windows predictably instead of reactively, without the cycle of slow responses followed by compensating spend. On results, Uchdadiya declines to supply the figures the story would most want, and says so in terms worth quoting. "I want to be direct here rather than invent precision I don't have," he says. Hard performance numbers from that period are not recoverable, so he describes the outcome qualitatively and leaves it there. Given that a percentage improvement in response time is exactly the sort of number nobody would check, the decision not to produce one is a reasonable indication of how the rest of the account should be read. The more interesting half of his argument comes from the opposite direction, and from a system he is careful to say he does not own. He now works at a global investment management firm whose platforms run on static, fixed-capacity provisioning rather than dynamic autoscaling. That is a deliberate architectural choice, made because in an environment targeting four nines of availability, the business consequence of scaling arriving late, missing a critical financial moment, was judged unacceptable. He is explicit that he is not on the team responsible for that decision and is describing an architecture he works within rather than one he designed. His point is that this is not an absence of a scaling strategy. It is one. "Reactive autoscaling and static over-provisioning are really the same tradeoff viewed from two different risk tolerances," he says. One accepts occasional lateness to avoid paying for idle capacity. The other accepts permanent idle capacity to eliminate the possibility of being late. Both are answers to a single question: is your traffic predictable enough to anticipate, or unpredictable enough that you would rather pay for permanent headroom than risk being slow. Neither answer is wrong. What he objects to is that most organisations never ask. "The mistake I see most often is organisations defaulting into reactive threshold-based autoscaling without ever actually asking that question, treating it as a settled default rather than a real architectural decision with a real cost model attached." His recommendation follows from the sports platform and is unglamorous enough to be actionable. If traffic has any recognisable pattern at all, a season, a calendar, a campaign schedule, then that pattern is a leading indicator available to scale against in advance. Reactive thresholds discard it and wait for the symptom instead. Which is a strange thing to do with information you already have.

Reactive scaling is defined by responding after a threshold is crossed. On traffic that arrives on a calendar, that is a decision to be late every time.

The hidden cost of autoscaling: By the time your system reacts to the traffic, it's already too late

Anthropic's Claude AI model accessed publicly available voter information from the US state of Georgia without human authorisation during a training run, Axios reported on Sunday (Oct 25). Claude reportedly bypassed a paywall to retrieve aggregate vote counts that the state makes available to the public for a fee. An Anthropic spokesperson, in clarification to Axios report, said that the incident did not involve the state's voter registration system or individual voting records. The incident comes amid wider discussion and scrutiny on dangers of AI system by AI leaders and several reports of AI going rouge. &nbsp;&nbsp;Also Read: <a href="https://www.wionews.com/technology/anthropic-s-revenue-more-than-doubles-in-a-quarter-why-its-huge-loss-still-matters-1793034686486">‘Anthropic’s revenue more than doubles in a quarter’: Why its huge loss still matters</a>&nbsp;"In a training run, Claude accessed aggregate vote counts, pulling data that the state makes available to the public for a fee. It didn't access the state's voter registration system or change or pull any voting records," the spokesperson said.&nbsp;Also Read: <a href="https://www.wionews.com/world/anthropic-s1-filing-ipo-loss-cloud-computing-commitments-1793022200164">Anthropic targets $2 trillion IPO valuation despite $80.7 billion net loss</a>&nbsp;Meanwhile, Anthropic has filed a Form S-1 registration statement with the US Securities and Exchange Commission, aiming to go public the week of November 16, 2026, targeting a $1.8–$2 trillion valuation. Anthropic reported first-half 2026 revenue of about $16.3 billion and a net loss of $80.7 billion through June. This means that its revenue jumped 1,304% from the first half of 2025. Its loss has jumped due to its S-1 filing, which seeks a valuation in the region of $2 trillion. Most of the loss, roughly $72 billion, comes from accounting charges, which are convertible debt “resulting from an increase in our enterprise valuation during each of the respective periods,” Anthropic said. Meaning the company will pay back its early lenders in stocks worth $72 billion.

Anthropic’s Claude AI accessed aggregate vote counts from Georgia during a training run, bypassing a paywall without human authorisation, Axios reported. Anthropic said the system did not access voter registration records or alter any voting data, raising questions about AI oversight.

Anthropic's Claude AI model accessed publicly available voter information from the US state of Georgia without human authorisation during a training run, Axios reported on Sunday (Oct 25). Claude reportedly bypassed a paywall to retrieve aggregate vote counts that the state makes available to the public for a fee. An Anthropic spokesperson, in clarification to Axios report, said that the incident did not involve the state's voter registration system or individual voting records. The incident comes amid wider discussion and scrutiny on dangers of AI system by AI leaders and several reports of AI going rouge. Also Read: ‘Anthropic’s revenue more than doubles in a quarter’: Why its huge loss still matters "In a training run, Claude accessed aggregate vote counts, pulling data that the state makes available to the public for a fee. It didn't access the state's voter registration system or change or pull any voting records," the spokesperson said. Also Read: Anthropic targets $2 trillion IPO valuation despite $80.7 billion net loss Meanwhile, Anthropic has filed a Form S-1 registration statement with the US Securities and Exchange Commission, aiming to go public the week of November 16, 2026, targeting a $1.8–$2 trillion valuation. Anthropic reported first-half 2026 revenue of about $16.3 billion and a net loss of $80.7 billion through June. This means that its revenue jumped 1,304% from the first half of 2025. Its loss has jumped due to its S-1 filing, which seeks a valuation in the region of $2 trillion. Most of the loss, roughly $72 billion, comes from accounting charges, which are convertible debt “resulting from an increase in our enterprise valuation during each of the respective periods,” Anthropic said. Meaning the company will pay back its early lenders in stocks worth $72 billion.

Anthropic’s Claude AI accessed aggregate vote counts from Georgia during a training run, bypassing a paywall without human authorisation, Axios reported. Anthropic said the system did not access voter registration records or alter any voting data, raising questions about AI oversight.

Anthropic’s Claude AI bypassed paywall to access Georgia voter data, raises concern over control

Microsoft has offered voluntary exit packages to around 60,000 employees, giving more than a quarter of its global workforce the option to leave the company with support. The programme covers eligible staff in India, the United States, the United Kingdom and Canada, as the technology giant looks to control costs while investing heavily in artificial intelligence (AI).&nbsp;The offer is voluntary, meaning it does not confirm that 60,000 jobs will be cut. Microsoft had 223,000 full-time employees worldwide as of June 30, 2026. The move comes as the company prepares for another major year of spending on AI chips, data centres and computing infrastructure.<h2>Who is eligible for Microsoft’s voluntary exit offer?</h2>Microsoft said the programme is available to eligible employees in participating teams in India, the US, the UK and Canada at experience level 70 and below. According to reports, employees will have until December to decide whether to accept the offer. Those who choose to leave are expected to depart on December 31.Microsoft Chief People Officer Amy Coleman said in an internal memo that the programme gives eligible employees who are considering a future outside the company another option, while helping teams adapt to changing business needs. Not every team is included. According to a person familiar with the matter, some marketing, commercial business and security teams are excluded from the offer.The company has not set a target for how many employees it expects to accept the package.<h2>Why is Microsoft offering buyouts as AI spending rises?</h2>The programme comes as Microsoft forecasts $175 billion in capital expenditure for 2026, reflecting the cost of expanding its AI infrastructure. Microsoft provides cloud computing services to AI companies, including OpenAI and Anthropic, and is investing in the data centres and chips needed to meet growing demand. That spending has drawn investor attention, with questions about whether future AI revenue will justify the cost of building and operating the infrastructure.&nbsp;Microsoft's latest employee offer provides another way to adjust its workforce as its business needs change. However, the company has not said that the buyout programme alone will fund its AI investments.<h2>Microsoft’s latest move follows earlier workforce changes</h2>This is not the company's first workforce adjustment in 2026. Microsoft previously offered voluntary retirement packages to around 7% of its US workforce. In July, it also announced plans to cut 4,800 jobs as it restructured parts of its Xbox gaming business. The latest offer comes as investors prepare to assess Microsoft's financial performance. The company is scheduled to report quarterly results on Wednesday, providing a fresh look at its spending and cash flow. For employees, the key point is that the offer creates an option to leave rather than announcing a compulsory departure. For investors, it highlights the balancing act facing major technology companies: funding the AI race while managing costs across their existing businesses.

Microsoft is offering voluntary exit packages to around 60,000 employees in India, the US, the UK and Canada. The programme comes as the company forecasts $175 billion in capital expenditure for 2026 to expand AI infrastructure and data centre capacity.

Microsoft has offered voluntary exit packages to around 60,000 employees, giving more than a quarter of its global workforce the option to leave the company with support. The programme covers eligible staff in India, the United States, the United Kingdom and Canada, as the technology giant looks to control costs while investing heavily in artificial intelligence (AI). The offer is voluntary, meaning it does not confirm that 60,000 jobs will be cut. Microsoft had 223,000 full-time employees worldwide as of June 30, 2026. The move comes as the company prepares for another major year of spending on AI chips, data centres and computing infrastructure. Who is eligible for Microsoft’s voluntary exit offer? Microsoft said the programme is available to eligible employees in participating teams in India, the US, the UK and Canada at experience level 70 and below. According to reports, employees will have until December to decide whether to accept the offer. Those who choose to leave are expected to depart on December 31. Microsoft Chief People Officer Amy Coleman said in an internal memo that the programme gives eligible employees who are considering a future outside the company another option, while helping teams adapt to changing business needs. Not every team is included. According to a person familiar with the matter, some marketing, commercial business and security teams are excluded from the offer. The company has not set a target for how many employees it expects to accept the package. Why is Microsoft offering buyouts as AI spending rises? The programme comes as Microsoft forecasts $175 billion in capital expenditure for 2026, reflecting the cost of expanding its AI infrastructure. Microsoft provides cloud computing services to AI companies, including OpenAI and Anthropic, and is investing in the data centres and chips needed to meet growing demand. That spending has drawn investor attention, with questions about whether future AI revenue will justify the cost of building and operating the infrastructure. Microsoft's latest employee offer provides another way to adjust its workforce as its business needs change. However, the company has not said that the buyout programme alone will fund its AI investments. Microsoft’s latest move follows earlier workforce changes This is not the company's first workforce adjustment in 2026. Microsoft previously offered voluntary retirement packages to around 7% of its US workforce. In July, it also announced plans to cut 4,800 jobs as it restructured parts of its Xbox gaming business. The latest offer comes as investors prepare to assess Microsoft's financial performance. The company is scheduled to report quarterly results on Wednesday, providing a fresh look at its spending and cash flow. For employees, the key point is that the offer creates an option to leave rather than announcing a compulsory departure. For investors, it highlights the balancing act facing major technology companies: funding the AI race while managing costs across their existing businesses.

Microsoft is offering voluntary exit packages to around 60,000 employees across India, the US, the UK and Canada. The programme comes as the company forecasts $175 billion in capital expenditure for 2026 to expand its AI infrastructure and data centre capacity.

Microsoft offers voluntary exit packages to 60,000 employees in India, US, UK and Canada

Meta is planning to raise up to $95 billion through stock offerings to finance its artificial intelligence (AI) expansion, adding to concerns about the rising cost of the global AI race. The plan comes days after the Facebook and Instagram owner forecast capital expenditure of up to $145 billion this year, as it builds data centres and secures computing power. The fundraising could become one of the largest equity financing programmes by an AI company. It also comes as Meta’s cash flow weakens, raising questions about how technology giants will pay for their increasingly expensive AI ambitions.<h2>Why Meta wants to raise $95 billion for AI</h2>Meta plans to launch the larger share sale after its new non-voting Class C shares begin trading, which is expected in the first quarter of 2027. The company has not finalised the timing or terms of the offering. Chipmakers Nvidia, AMD and Broadcom have also expressed interest in buying up to $5 billion in Meta’s Class A shares combined, according to the supplied report. That interest highlights a growing concern in the AI industry: chip suppliers may help finance the expansion of major customers that are buying their products. Such arrangements can create a circular flow of capital, with money moving between companies that depend on one another for growth. Meta is investing heavily in chips and data centres to compete with AI companies such as OpenAI and Anthropic.<h2>Meta’s AI spending puts pressure on cash flow</h2>Meta’s spending plans are enormous. The company has committed to about $825 billion in future spending, primarily linked to data centres, AI infrastructure and computing capacity, according to the report. Its financial position also shows the pressure. Meta reported negative free cash flow of $2.12 billion in the third quarter, compared with positive free cash flow of $10.63 billion a year earlier. Free cash flow is the money a company has left after operating expenses and capital investment.&nbsp;As of September 30, Meta reported $82.45 billion in cash, cash equivalents and marketable securities, alongside $80.92 billion in long-term debt. The company recently raised the lower end of its annual capital expenditure forecast, putting its expected spending at $135 billion to $145 billion.<h2>What Meta’s fundraising means for the AI race</h2>Meta’s plan reflects a wider shift in Big Tech. Companies that once relied heavily on cash reserves are increasingly turning to debt and equity markets to finance AI infrastructure. William Blair analysts said investors would likely continue questioning Meta’s financing needs and the possibility of further capital raises. They also argued that AI could ultimately become a profitable business for Meta, although returns may not arrive in a straight line. The company has seen early interest in its Muse AI agent and released new AI models, offering potential reasons for its investment drive. But those developments do not guarantee that spending will deliver the returns investors expect. Meta shares fell around 1.4% in morning trading following the news, according to the supplied report. For investors, the central question is whether Meta can turn its vast AI investment into a sustainable business before the cost of building it puts further pressure on its finances.

Meta plans to raise up to $95 billion through stock offerings to fund its AI expansion after forecasting up to $145 billion in annual capital spending. Negative free cash flow and huge infrastructure commitments are raising questions about the cost of the AI race.

Meta is planning to raise up to $95 billion through stock offerings to finance its artificial intelligence (AI) expansion, adding to concerns about the rising cost of the global AI race. The plan comes days after the Facebook and Instagram owner forecast capital expenditure of up to $145 billion this year, as it builds data centres and secures computing power. The fundraising could become one of the largest equity financing programmes by an AI company. It also comes as Meta’s cash flow weakens, raising questions about how technology giants will pay for their increasingly expensive AI ambitions. Why Meta wants to raise $95 billion for AI Meta plans to launch the larger share sale after its new non-voting Class C shares begin trading, which is expected in the first quarter of 2027. The company has not finalised the timing or terms of the offering. Chipmakers Nvidia, AMD and Broadcom have also expressed interest in buying up to $5 billion in Meta’s Class A shares combined, according to the supplied report. That interest highlights a growing concern in the AI industry: chip suppliers may help finance the expansion of major customers that are buying their products. Such arrangements can create a circular flow of capital, with money moving between companies that depend on one another for growth. Meta is investing heavily in chips and data centres to compete with AI companies such as OpenAI and Anthropic. Meta’s AI spending puts pressure on cash flow Meta’s spending plans are enormous. The company has committed to about $825 billion in future spending, primarily linked to data centres, AI infrastructure and computing capacity, according to the report. Its financial position also shows the pressure. Meta reported negative free cash flow of $2.12 billion in the third quarter, compared with positive free cash flow of $10.63 billion a year earlier. Free cash flow is the money a company has left after operating expenses and capital investment. As of September 30, Meta reported $82.45 billion in cash, cash equivalents and marketable securities, alongside $80.92 billion in long-term debt. The company recently raised the lower end of its annual capital expenditure forecast, putting its expected spending at $135 billion to $145 billion. What Meta’s fundraising means for the AI race Meta’s plan reflects a wider shift in Big Tech. Companies that once relied heavily on cash reserves are increasingly turning to debt and equity markets to finance AI infrastructure. William Blair analysts said investors would likely continue questioning Meta’s financing needs and the possibility of further capital raises. They also argued that AI could ultimately become a profitable business for Meta, although returns may not arrive in a straight line. The company has seen early interest in its Muse AI agent and released new AI models, offering potential reasons for its investment drive. But those developments do not guarantee that spending will deliver the returns investors expect. Meta shares fell around 1.4% in morning trading following the news, according to the supplied report. For investors, the central question is whether Meta can turn its vast AI investment into a sustainable business before the cost of building it puts further pressure on its finances.

Meta plans to raise up to $95 billion through stock offerings to fund its AI expansion after forecasting up to $145 billion in annual capital spending. Negative free cash flow and huge infrastructure commitments are raising questions about the cost of the AI race.

Meta’s AI bet is getting more expensive: What its $95 billion share plan means

Anthropic, the company behind the Claude AI models, is heading towards a potential blockbuster stock market debut with a financial story that captures both the promise and the risks of the artificial intelligence boom. Its reported second-quarter operating income reached $974 million, reversing a $1.7 billion operating loss three months earlier. Yet its IPO filing also puts its net loss for the first half of 2026 at a staggering $80.7 billion.&nbsp;The contrast highlights a central question for investors: Can AI companies turn rapidly growing demand into lasting profits while paying for the enormous computing infrastructure their products require?<h2>Anthropic’s revenue jumps as computing costs grow more slowly</h2>Anthropic reported revenue of $11.6 billion in the second quarter of 2026, up from $4.7 billion in the first quarter. Over the same period, the cost of computing power rose from $3.4 billion to $5.6 billion. That gap helped push operating income into positive territory. Operating income measures what remains after a business pays its operating costs, before certain other expenses and accounting items.&nbsp;The company’s computing costs consumed $2.51 for every dollar of revenue in early 2025, according to the figures provided. By the second quarter of 2026, that had fallen to 48 cents. One reason is that Anthropic books computing capacity 12 to 18 months in advance. When revenue rises faster than those previously arranged costs, more of each additional dollar can flow towards operating income. However, operating profit is not the same as net profit. The reported first-half net loss remains a major figure investors will need to understand, including the accounting factors behind the difference.<h2>$395 billion in cloud commitments raise the stakes</h2>Anthropic’s plans involve enormous spending commitments to secure the computing power needed to develop and run its AI models. According to the supplied filing details, the company plans to spend at least $110 billion with Amazon through April 2036, $120 billion with SpaceX through 2029, $115 billion with Google through 2033 and $45 billion with Microsoft through 2033. These commitments total $390 billion, not including any additional costs or changes to the agreements. They underline how much infrastructure the company expects to need as its business expands. Anthropic said its cash and current sales should cover the next 12 months, but warned it could need additional financing sooner than expected if revenue growth or infrastructure costs develop differently.<h2>What Anthropic’s IPO means for AI investors</h2>The company has reportedly raised around $141 billion in capital and held $112.4 billion in cash at the end of June. It plans to list on Nasdaq under the ticker symbol ANT, with Goldman Sachs and Morgan Stanley leading the offering, according to the supplied information. The filing does not yet set out a final share price or the number of shares to be offered. Those details would be determined later in the IPO process. For investors, Anthropic presents a high-stakes test of the AI business model: explosive revenue growth and improving operating economics on one side, and extraordinary infrastructure commitments and substantial reported losses on the other. The key question is not simply whether Anthropic can grow. It is whether that growth can deliver sustainable profits after the full cost of building AI is counted.&nbsp;

Anthropic’s reported second-quarter operating income reached $974 million as revenue climbed to $11.6 billion. But its $80.7 billion first-half net loss and huge cloud commitments raise questions about the cost and sustainability of AI growth ahead of its planned IPO.

Anthropic, the company behind the Claude AI models, is heading towards a potential blockbuster stock market debut with a financial story that captures both the promise and the risks of the artificial intelligence boom. Its reported second-quarter operating income reached $974 million, reversing a $1.7 billion operating loss three months earlier. Yet its IPO filing also puts its net loss for the first half of 2026 at a staggering $80.7 billion. The contrast highlights a central question for investors: Can AI companies turn rapidly growing demand into lasting profits while paying for the enormous computing infrastructure their products require? Anthropic’s revenue jumps as computing costs grow more slowly Anthropic reported revenue of $11.6 billion in the second quarter of 2026, up from $4.7 billion in the first quarter. Over the same period, the cost of computing power rose from $3.4 billion to $5.6 billion. That gap helped push operating income into positive territory. Operating income measures what remains after a business pays its operating costs, before certain other expenses and accounting items. The company’s computing costs consumed $2.51 for every dollar of revenue in early 2025, according to the figures provided. By the second quarter of 2026, that had fallen to 48 cents. One reason is that Anthropic books computing capacity 12 to 18 months in advance. When revenue rises faster than those previously arranged costs, more of each additional dollar can flow towards operating income. However, operating profit is not the same as net profit. The reported first-half net loss remains a major figure investors will need to understand, including the accounting factors behind the difference. $395 billion in cloud commitments raise the stakes Anthropic’s plans involve enormous spending commitments to secure the computing power needed to develop and run its AI models. According to the supplied filing details, the company plans to spend at least $110 billion with Amazon through April 2036, $120 billion with SpaceX through 2029, $115 billion with Google through 2033 and $45 billion with Microsoft through 2033. These commitments total $390 billion, not including any additional costs or changes to the agreements. They underline how much infrastructure the company expects to need as its business expands. Anthropic said its cash and current sales should cover the next 12 months, but warned it could need additional financing sooner than expected if revenue growth or infrastructure costs develop differently. What Anthropic’s IPO means for AI investors The company has reportedly raised around $141 billion in capital and held $112.4 billion in cash at the end of June. It plans to list on Nasdaq under the ticker symbol ANT, with Goldman Sachs and Morgan Stanley leading the offering, according to the supplied information. The filing does not yet set out a final share price or the number of shares to be offered. Those details would be determined later in the IPO process. For investors, Anthropic presents a high-stakes test of the AI business model: explosive revenue growth and improving operating economics on one side, and extraordinary infrastructure commitments and substantial reported losses on the other. The key question is not simply whether Anthropic can grow. It is whether that growth can deliver sustainable profits after the full cost of building AI is counted.

Anthropic’s reported second-quarter operating income reached $974 million as revenue climbed to $11.6 billion. But its $80.7 billion first-half net loss and huge cloud commitments raise questions about the cost and sustainability of AI growth ahead of its planned IPO.

‘Anthropic’s revenue more than doubles in a quarter’: Why its huge loss still matters

The hardest problem in healthcare data is not volume. It is that the data arrives in shapes nobody agreed on in advance. Referrals, member surveys and eligibility files land daily from different programs, each with its own structure, and a program can redesign a survey without telling anyone downstream. A single malformed record inside a file of a hundred thousand can run for hours before anything fails, and the retry that follows can leave a member with two histories instead of one.&nbsp;&nbsp;Sai Kiran Yadav works on the entry point where those problems either get caught or get inherited. As a lead data architect at a healthcare technology company working in social care and Medicaid, he works at the unglamorous end of the pipeline, on the argument that almost everything downstream is decided there.&nbsp;&nbsp;The platform moved data from MongoDB through Google Cloud Dataflow into BigQuery. Dataflow transformed nested survey structures into tables that teams could query for reporting and analysis. Restricted access to raw datasets, authorized reporting views, and row- and column-level controls governed which information users could access.&nbsp;&nbsp;Surveys were the first problem. Healthcare survey instruments are stored as deeply nested documents and they change constantly, because each new program adds question types on its own schedule. Conventional extract-and-load tooling has nothing stable to attach to. Yadav built a recursive flattening utility that detects those structural shifts automatically and produces clean, query-ready tables on the other side, which turns schema drift from a recurring outage into a handled condition.&nbsp;&nbsp;Ingestion was the second problem, and the more expensive one. Yadav introduced a staging-first ingestion process that validated incoming records before they reached production. In one representative run, the system identified invalid records within the first minute of processing, surfacing problems that previously could cause failures hours later. The workflow also enabled failed jobs to be reprocessed safely without creating duplicate member profiles.&nbsp;&nbsp;That last property is the one operations teams notice. A pipeline that can be re-run without consequence changes what a bad night looks like: the file goes back through, and nothing has to be untangled by hand afterwards.&nbsp;&nbsp;Above the pipelines sits a consolidation layer. Member and organization data in this sector fragments across legacy systems that disagree about identifiers and hold partial profiles each claiming to be authoritative. Yadav modelled golden records that pull the pieces together, so multiple contact methods, program enrolments and engagement states resolve into a single row per entity rather than a query that has to reconcile four systems at read time.&nbsp;&nbsp;His recovery and validation procedures were adopted as organizational standards, which is a quieter form of impact than a metric: the way the company handles a failed load is now the way he decided it should be handled.&nbsp;&nbsp;Yadav also contributed to enterprise cloud modernization projects, helping migrate legacy reporting systems to BigQuery-based architectures. His responsibilities included designing ingestion strategies, developing data models, and implementing partitioning and cost-aware querying practices. Analytical workloads run without slowing the applications front-line staff depend on, which is the constraint that makes these migrations difficult in the first place.&nbsp;&nbsp;Measured outcomes are specific. His automated pipelines reduced the time required to make survey and engagement data available for analysis from one to two days to under ten minutes.&nbsp;Product and operations teams could access member feedback and engagement patterns sooner, with less dependence on manual exports and data preparation. Engineering support requests fell by about 40 percent, which moved staff off repair work and onto new problems.&nbsp;&nbsp;"Healthcare data systems must make information available quickly, preserve the integrity of member records, and control access appropriately. Reliability also means anticipating failures and making recovery safe," Yadav said.&nbsp;&nbsp;Pressure on the field is getting heavier rather than lighter. More sources need integrating, including electronic records and connected devices. Volumes keep climbing against infrastructure that was not sized for them. Privacy expectations tighten while the threat environment does not improve. Fragmented information has a cost that shows up as staff time spent reconciling records that should already agree.&nbsp;&nbsp;For Yadav, reliable healthcare analytics begins with the way data enters a system. Early validation, clear handling of problematic records, and safe recovery help keep information dependable as programs grow. Looking ahead, he advocates separate processing paths for time-sensitive operational signals and longer-term analysis, supported by shared data-quality and access controls. The goal is to give teams timely information while preserving trust in the records behind it.&nbsp;

Referrals, surveys and eligibility files arrive in formats that change without warning, and a single bad record can take down a job hours after it started. Sai Kiran Yadav rebuilt the entry point so the failures surface in the first minute instead. 

The hardest problem in healthcare data is not volume. It is that the data arrives in shapes nobody agreed on in advance. Referrals, member surveys and eligibility files land daily from different programs, each with its own structure, and a program can redesign a survey without telling anyone downstream. A single malformed record inside a file of a hundred thousand can run for hours before anything fails, and the retry that follows can leave a member with two histories instead of one. Sai Kiran Yadav works on the entry point where those problems either get caught or get inherited. As a lead data architect at a healthcare technology company working in social care and Medicaid, he works at the unglamorous end of the pipeline, on the argument that almost everything downstream is decided there. The platform moved data from MongoDB through Google Cloud Dataflow into BigQuery. Dataflow transformed nested survey structures into tables that teams could query for reporting and analysis. Restricted access to raw datasets, authorized reporting views, and row- and column-level controls governed which information users could access. Surveys were the first problem. Healthcare survey instruments are stored as deeply nested documents and they change constantly, because each new program adds question types on its own schedule. Conventional extract-and-load tooling has nothing stable to attach to. Yadav built a recursive flattening utility that detects those structural shifts automatically and produces clean, query-ready tables on the other side, which turns schema drift from a recurring outage into a handled condition. Ingestion was the second problem, and the more expensive one. Yadav introduced a staging-first ingestion process that validated incoming records before they reached production. In one representative run, the system identified invalid records within the first minute of processing, surfacing problems that previously could cause failures hours later. The workflow also enabled failed jobs to be reprocessed safely without creating duplicate member profiles. That last property is the one operations teams notice. A pipeline that can be re-run without consequence changes what a bad night looks like: the file goes back through, and nothing has to be untangled by hand afterwards. Above the pipelines sits a consolidation layer. Member and organization data in this sector fragments across legacy systems that disagree about identifiers and hold partial profiles each claiming to be authoritative. Yadav modelled golden records that pull the pieces together, so multiple contact methods, program enrolments and engagement states resolve into a single row per entity rather than a query that has to reconcile four systems at read time. His recovery and validation procedures were adopted as organizational standards, which is a quieter form of impact than a metric: the way the company handles a failed load is now the way he decided it should be handled. Yadav also contributed to enterprise cloud modernization projects, helping migrate legacy reporting systems to BigQuery-based architectures. His responsibilities included designing ingestion strategies, developing data models, and implementing partitioning and cost-aware querying practices. Analytical workloads run without slowing the applications front-line staff depend on, which is the constraint that makes these migrations difficult in the first place. Measured outcomes are specific. His automated pipelines reduced the time required to make survey and engagement data available for analysis from one to two days to under ten minutes. Product and operations teams could access member feedback and engagement patterns sooner, with less dependence on manual exports and data preparation. Engineering support requests fell by about 40 percent, which moved staff off repair work and onto new problems. "Healthcare data systems must make information available quickly, preserve the integrity of member records, and control access appropriately. Reliability also means anticipating failures and making recovery safe," Yadav said. Pressure on the field is getting heavier rather than lighter. More sources need integrating, including electronic records and connected devices. Volumes keep climbing against infrastructure that was not sized for them. Privacy expectations tighten while the threat environment does not improve. Fragmented information has a cost that shows up as staff time spent reconciling records that should already agree. For Yadav, reliable healthcare analytics begins with the way data enters a system. Early validation, clear handling of problematic records, and safe recovery help keep information dependable as programs grow. Looking ahead, he advocates separate processing paths for time-sensitive operational signals and longer-term analysis, supported by shared data-quality and access controls. The goal is to give teams timely information while preserving trust in the records behind it.

Referrals, surveys and eligibility files arrive in formats that change without warning, and a single bad record can take down a job hours after it started. Sai Kiran Yadav rebuilt the entry point so the failures surface in the first minute instead.&nbsp;

Data lifecycle guardian automates healthcare excellence

Elon Musk’s SpaceX has gained access to some of the US military’s most precise satellite-tracking data, a move that has raised concerns over national security, competition and the future of space traffic management. The agreement with the US Space Force gives SpaceX access to detailed information about objects orbiting Earth, including satellites and space debris. Industry observers warn that the arrangement could give Musk’s company an advantage over competitors offering similar services.The development also comes as the Trump administration faces scrutiny over its handling of a federal programme designed to manage traffic in space.<h2>SpaceX gains access to military tracking data</h2>For the first time, the Space Force has granted a commercial company enhanced access to the US Space Command’s High Accuracy Catalog. The information includes Vector Covariance Messages, which contain tracking data gathered through military radars, telescopes and ground-based sensors.These systems can track objects as small as 10 centimetres in low Earth orbit, helping operators assess potential collisions.SpaceX already runs Stargaze, a satellite-based tracking network that uses cameras and radars mounted on its Starlink satellites. However, the system cannot detect objects as small as those identified by the military's more precise tracking network.The new arrangement allows SpaceX to supplement its existing capabilities. In return, the government will receive access to Stargaze's observations and sensors.<h2>Concerns over competition and security</h2>Critics say the deal could strengthen SpaceX's position in a commercial space industry projected to approach $2 trillion over the next decade.Cole Donovan, a former employee of the federal Traffic Coordination System for Space (Traccs), warned that the arrangement risked making SpaceX the dominant provider of space traffic coordination in the US.“It’s making a company the de facto source for space traffic coordination within the US,” Donovan said.The military data is classified as controlled unclassified information, as wider access could expose details about the capabilities and limitations of US surveillance systems.<h2>Traccs funding adds to the debate</h2>The deal comes amid a prolonged struggle over Traccs, which was established to transfer commercial space traffic management responsibilities away from the military. The programme has faced staffing problems and repeated budget-cutting attempts, although Congress restored funding after proposed reductions.While critics see the SpaceX agreement as a threat to fair competition, a former National Oceanic and Atmospheric Administration official defended the company's role, arguing that the government had failed to establish an effective framework for managing space traffic.The arrangement has therefore intensified a wider debate over whether commercial companies should assume responsibilities traditionally handled by the government.

The US military has granted SpaceX access to sensitive satellite-tracking data, raising national security concerns and fears that Elon Musk’s company could gain an unfair advantage in space traffic management.

Elon Musk’s SpaceX has gained access to some of the US military’s most precise satellite-tracking data, a move that has raised concerns over national security, competition and the future of space traffic management. The agreement with the US Space Force gives SpaceX access to detailed information about objects orbiting Earth, including satellites and space debris. Industry observers warn that the arrangement could give Musk’s company an advantage over competitors offering similar services. The development also comes as the Trump administration faces scrutiny over its handling of a federal programme designed to manage traffic in space. SpaceX gains access to military tracking data For the first time, the Space Force has granted a commercial company enhanced access to the US Space Command’s High Accuracy Catalog. The information includes Vector Covariance Messages, which contain tracking data gathered through military radars, telescopes and ground-based sensors. These systems can track objects as small as 10 centimetres in low Earth orbit, helping operators assess potential collisions. SpaceX already runs Stargaze, a satellite-based tracking network that uses cameras and radars mounted on its Starlink satellites. However, the system cannot detect objects as small as those identified by the military's more precise tracking network. The new arrangement allows SpaceX to supplement its existing capabilities. In return, the government will receive access to Stargaze's observations and sensors. Concerns over competition and security Critics say the deal could strengthen SpaceX's position in a commercial space industry projected to approach $2 trillion over the next decade. Cole Donovan, a former employee of the federal Traffic Coordination System for Space (Traccs), warned that the arrangement risked making SpaceX the dominant provider of space traffic coordination in the US. “It’s making a company the de facto source for space traffic coordination within the US,” Donovan said. The military data is classified as controlled unclassified information, as wider access could expose details about the capabilities and limitations of US surveillance systems. Traccs funding adds to the debate The deal comes amid a prolonged struggle over Traccs, which was established to transfer commercial space traffic management responsibilities away from the military. The programme has faced staffing problems and repeated budget-cutting attempts, although Congress restored funding after proposed reductions. While critics see the SpaceX agreement as a threat to fair competition, a former National Oceanic and Atmospheric Administration official defended the company's role, arguing that the government had failed to establish an effective framework for managing space traffic. The arrangement has therefore intensified a wider debate over whether commercial companies should assume responsibilities traditionally handled by the government.

The US military has granted SpaceX access to sensitive satellite-tracking data, raising national security concerns and fears that Elon Musk’s company could gain an unfair advantage in space traffic management.

Is Trump risking US national security? Musk’s SpaceX gets access to sensitive military data

OpenAI CEO Sam Altman raised serious concerns over the rising capabilities of artificial intelligence (AI) in discovering security vulnerabilities beyond our imagination and expectations. Altman voiced his concern in a recent interview, noting that it is becoming hard to confine AI to sandboxes because the technology works only for a short span of time. He further elaborated that once the model starts evolving its cutting-edge abilities, you can’t even find the relic of how and when it escaped. Speaking about how traditional cybersecurity approaches may feel outdated or ineffective in front of AI’s finding hardware-level bugs. Over the immediate danger from the misalignment at an all-time high, he underscored that alignment research is still in an initial phase, which is far from resolved. “I don't even know a dramatic enough word to use… models can find bugs that are outside of the normal threat model of what we think a hacker can do… hardware-level, insane stuff… I no longer think we're gonna be able to just patch all the bugs," he said.&nbsp;Also read: <a href="https://www.wionews.com/world/killing-people-never-allowed-elon-musk-admits-restricting-starlink-in-ukraine-explains-why-1792901604146">'Killing people never allowed': Elon Musk admits restricting Starlink in Ukraine, explains why</a> OpenAI’s chief contends that the old approach of patching vulnerabilities one by one is no longer enough, arguing instead that the internet’s entire security architecture needs to be overhauled with AI-driven defence systems. He has flagged locally hosted, open-source AI models as a particular security concern, pointing out that even comprehensive monitoring of AI agents globally wouldn’t be enough to stop every harmful action. However, sceptics push back on this narrative, suggesting that Altman’s depiction of AI as inherently uncontrollable conveniently benefits OpenAI’s business interests by deflecting accountability away from developers when problems occur. This comes on the heels of a major PHP security flaw that knocked millions of websites offline worldwide, during which hackers made countless attempts to exploit the weakness- ironically, a vulnerability that was itself uncovered using AI.

Over the immediate danger from the misalignment at an all-time high, OpenAI CEO Sam Altman underscored that alignment research is still in an initial phase, which is far from resolved.

OpenAI CEO Sam Altman raised serious concerns over the rising capabilities of artificial intelligence (AI) in discovering security vulnerabilities beyond our imagination and expectations. Altman voiced his concern in a recent interview, noting that it is becoming hard to confine AI to sandboxes because the technology works only for a short span of time. He further elaborated that once the model starts evolving its cutting-edge abilities, you can’t even find the relic of how and when it escaped. Speaking about how traditional cybersecurity approaches may feel outdated or ineffective in front of AI’s finding hardware-level bugs. Over the immediate danger from the misalignment at an all-time high, he underscored that alignment research is still in an initial phase, which is far from resolved. “I don't even know a dramatic enough word to use… models can find bugs that are outside of the normal threat model of what we think a hacker can do… hardware-level, insane stuff… I no longer think we're gonna be able to just patch all the bugs," he said. Also read: 'Killing people never allowed': Elon Musk admits restricting Starlink in Ukraine, explains why OpenAI’s chief contends that the old approach of patching vulnerabilities one by one is no longer enough, arguing instead that the internet’s entire security architecture needs to be overhauled with AI-driven defence systems. He has flagged locally hosted, open-source AI models as a particular security concern, pointing out that even comprehensive monitoring of AI agents globally wouldn’t be enough to stop every harmful action. However, sceptics push back on this narrative, suggesting that Altman’s depiction of AI as inherently uncontrollable conveniently benefits OpenAI’s business interests by deflecting accountability away from developers when problems occur. This comes on the heels of a major PHP security flaw that knocked millions of websites offline worldwide, during which hackers made countless attempts to exploit the weakness- ironically, a vulnerability that was itself uncovered using AI.

Over the immediate danger from the misalignment at an all-time high, OpenAI CEO Sam Altman underscored that alignment research is still in an initial phase, which is far from resolved.

Sam Altman warns AI can now find bugs unexpectedly 'outside normal threat model'

technology

Technology, News on Technology, stories on Technology, articles on Technology, Technology stories at Wion, Technology news, Technology latest news

technology

From manual processes to intelligent procurement: How are enterprises embracing AI and automation

iPhone Air 2 may get 2 rear cameras, smaller Dynamic Island and A20 Pro chip

DeepSeek, Moonshot used OpenAI’s GPT-6 Astra to train new AI models: Report

Google unveils Gemini 4 Carbon just two weeks after Argon: Here’s what’s new

How is India strengthening cybersecurity in age of AI-powered threats?

Apple launches touchscreen MacBook Pro in India: Prices start at Rs 3.49 lakh

Why Alphabet stock fell 10% despite strong revenue growth and Google Cloud gains

Engineering AI that earns a place in the decision loop

Retail media is booming, but almost no one can measure it honestly

Apple launches M6 MacBook Pro and new iMac with up to 25% faster AI performance

Apple iPhone 18e could launch in March 2027 with a 60Hz screen and single camera

Security teams are drowning in false alarms. The fix? Fewer wrong alerts

Anthropic hacked by AI startup comprising former DOGE staff in just 12 hours

Reimagining customer success with Salesforce Sales and Experience Cloud

Why Enterprise AI keeps dying in the pilot stage

The money behind AI data centres: What Microsoft’s buyout plan reveals

How Lula’s AI meme army took on alleged ‘algorithm bias’ of Big Tech

DFT specialist masters complex chip pre-tapeout trials using AI

The hidden cost of autoscaling: By the time you react to traffic, it's too late

Anthropic’s Claude AI bypassed paywall to access Georgia voter data

Microsoft offers exit packages to 27% of its workforce across four countries

Meta’s AI bet is getting more expensive: What its $95 billion share plan means

Claude maker Anthropic sees profit turnaround, commits billions to computing

Data lifecycle guardian automates healthcare excellence

Is Trump risking US security? Musk’s SpaceX gets sensitive military data

Sam Altman warns AI can now find bugs unexpectedly 'outside normal threat model'