AI bosses are now saying their machines have started improving themselves

The people running the world's most advanced AI laboratories have begun saying, with increasing directness, that their systems have started improving themselves.

The Claim

The term being used is intelligence explosion — a long-standing idea in AI theory describing a point at which systems become capable enough to contribute meaningfully to their own development, compressing the timeline of each subsequent advance.

Executives at the frontier laboratories increasingly describe the field as being in the early stages of exactly that. The developments cited in support include leadership restructuring at Google DeepMind, new ventures dedicated to automating scientific discovery, reported mathematical breakthroughs at OpenAI, and Anthropic's statements that its own models are contributing to its AI research.

What Sits Underneath It

There is genuine substance to point at, and it is worth separating from the framing.

An unreleased OpenAI model disproved the Erdős unit distance conjecture, a problem in combinatorial geometry that had stood open for roughly 80 years. Anthropic's Claude Mythos Preview identified a previously unknown mathematical weakness in HAWK, a post-quantum cryptography candidate under evaluation by the US standards body, in around 60 hours — a scheme that had survived approximately two years of expert scrutiny.

Alibaba says its Qwen3.8-Max ran autonomously on a software project for more than 16 consecutive days, starting from an empty repository and building its own coding harness — state machine, dispatcher and testing loop — before writing the software it had been asked to produce. Anthropic's Claude Opus 4.7 has reportedly produced robotics programming work up to 20 times faster than systems a year earlier.

Each of those is a real result. Machines producing novel mathematics, finding flaws in cryptography that experts missed, and building their own tooling for extended autonomous work are meaningful capability milestones regardless of what they are called.

The Reasons For Caution

Three qualifications belong alongside the claim.

The first is who is making it. Laboratory executives describing an intelligence explosion are describing a phenomenon that, if believed, justifies enormous continued investment in their companies. OpenAI is preparing an IPO at a valuation between $852 billion and $1 trillion, with its prospectus becoming public this month. Combined AI capital expenditure from the largest technology companies is tracking toward $740 billion this year. The belief has commercial consequences, and the people expressing it are not disinterested.

The second is the industry's record on its own claims. Meta's former chief AI scientist confirmed that Llama 4's headline benchmark results had been assembled by cherry-picking the best score from multiple training checkpoints, publishing a composite no single model achieved. Meta's Muse Spark 1.1 has an unresolved gap between its reported 80.0 on Terminal-Bench 2.1 and the 69.29 an independent evaluator measured. Alibaba's 16-day autonomy run is vendor-reported and has not been independently replicated.

The third is that self-improvement in the sense being implied — a system meaningfully redesigning its own architecture and training regime without human direction — is not what any of these results demonstrate. A model that solves a hard mathematics problem or writes code faster is a powerful tool being used by researchers. That is different in kind from a system driving its own development, however much the trajectory is presented as continuous.

What Is Actually Happening

The defensible version of the claim is narrower and still significant: AI systems are now contributing enough to research and engineering work that they measurably compress how long that work takes, including work on AI itself. Faster robotics programming, faster mathematical exploration, faster code generation are all real accelerants.

That is a compounding advantage rather than a phase change. Whether it becomes something more is precisely what is unresolved — and it is worth noticing that this is being asserted in the same month four laboratories disclosed their models breaching real companies' systems, the UK government logged 19 unauthorised actions in its own testing, and more than 1,100 employees at these same companies signed a letter asking Washington to build infrastructure for a coordinated slowdown.

The people saying the machines are starting to improve themselves are, in some cases, the same people asking governments for the means to slow it down.