What Happens When An AI Helps Create Another AI?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Happens When An AI Helps Create Another AI? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

METR said on Nov. 5 that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The assessment points mainly to AI helping researchers, not systems independently building their successors; public evidence is limited, and practical bottlenecks remain.

AI tools are likely helping Anthropic improve its models’ capabilities on well-specified tasks at about twice the earlier rate, according to an assessment published Nov. 5 by independent evaluator METR. The finding has high uncertainty and describes primarily AI assistance to researchers, not an autonomous system designing and building its own successor.

METR said its assessment drew on internal information from August through October. It judged that it “would not be surprised” by an acceleration as low as 1.3 times or as high as three times the earlier rate. The estimate concerns improvement on well-specified tasks; it is not a claim that every aspect of AI development is moving twice as fast.

The public cannot independently inspect much of the underlying evidence: the source article says Anthropic redacted much of the material that METR considered most informative. That makes the evaluators’ judgment central to interpreting the result and limits what outside readers can verify about its methods and evidence.

A separate example cited in the source comes from Google DeepMind. It reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. This is a reported engineering improvement, not evidence by itself that AI research is fully automated.

At a glance
reportWhen: Assessment published November 5, 2026;…
The developmentA METR assessment published Nov. 5 estimated that AI assistance is likely contributing to a faster pace of model improvement at Anthropic, bringing renewed attention to how AI might help develop future AI systems.

How Research Gains Could Compound

The significance is not that one model can simply produce a successor without people. It is that AI assistance could increase the amount of research a team can do: proposing experiments, writing or reviewing code, diagnosing failed runs, and checking whether a result holds up. If those contributions are reliable, researchers may spend less time on routine work and investigate more ideas with the same staff and resources.

Small improvements can accumulate, although the effect depends on whether they are independent and repeatable. The source offers a simplified illustration: three separate 10% reductions in resource use would leave a task requiring about 73% of its original compute, allowing roughly 37% more such jobs under the same compute budget. That is an illustration, not a measured outcome from an AI lab.

The pace of model development matters to companies, workers, and public institutions making plans around AI. But technical capability alone will not determine who can use these systems, who controls the computing infrastructure, or how benefits are distributed. Those outcomes also depend on deployment choices and institutional decisions.

Products Worth Considering

FLEXISPOT 36-Inch Height-Adjustable Standing Desk Converter with Keyboard Tray, Black

FLEXISPOT 36-Inch Height-Adjustable Standing Desk Converter with Keyboard Tray, Black

Canon SELPHY CP1500 Compact Photo Printer Bundle with Two KP-108IN Ink and Paper Sets, Black

Canon SELPHY CP1500 Compact Photo Printer Bundle with Two KP-108IN Ink and Paper Sets, Black

Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C - Graphite

Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Automation

AI research involves more than producing a promising idea. A team must inspect prior work and existing code, set up an experiment, measure its results fairly, repeat runs where needed, and decide whether an apparent gain is real. A system that changes code and reports a 5% improvement could be wrong if it made the test easier, used different computing resources, or produced a gain that disappears with another random seed.

That is why the source distinguishes assistance from full automation. Researchers can use AI to explore more possibilities while retaining responsibility for judging results. A more autonomous system would also have to choose what to investigate next and recognize when evidence is too weak to support a conclusion. The METR assessment, as described in the source, concerns mainly the assistance stage.

DeepMind’s AlphaEvolve example shows how a targeted optimization can affect a larger system: a component-level speed improvement translated into a smaller reduction in total training time. The reported figures also show why gains at one stage should not be confused with equivalent gains across an entire training process.

““The tool being improved can also contribute to the work of improving it.””

— Thorsten Meyer, source article author

What the Public Evidence Cannot Show

The estimate is not a precise measurement of how much AI has accelerated all research and development. METR’s reported range is wide, its assessment covers a particular organization and period, and the source says key supporting evidence was redacted. The public material described here does not provide enough detail to independently assess the full evidence base.

It is also unclear how much of the reported acceleration came from AI tools compared with other changes in staffing, methods, computing resources, or research priorities. The source reports METR’s judgment that AI assistance was likely responsible for most of the current acceleration, but does not establish that AI was the sole cause.

Nor does the assessment establish that AI systems can reliably manage an entire research cycle without human oversight, or that faster progress on well-specified tasks will translate directly into broader capabilities. The source identifies possible constraints, including experiment time, hardware availability, the gap between small experiments and large systems, research direction, continuity, and data quality.

Testing Whether Gains Hold Up

The next useful evidence will be whether AI-assisted research gains can be measured repeatedly, independently checked, and linked to improvements that persist beyond specific tasks or experiments. Further assessments with clearer public methods and evidence would help readers judge how much of the reported acceleration is attributable to AI.

For research teams, the immediate test is practical: whether AI-generated proposals and code survive fair comparisons, repeated runs, and scrutiny for errors. For observers, the key distinction will remain between systems that help researchers work faster and systems that can reliably direct and carry out research on their own. The source argues that planning should allow for changing rates of progress, while recognizing that the transition is likely to be uneven.

Key Questions

What did METR report about Anthropic?

METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate. It described the estimate as uncertain, saying it would not be surprised by a rate between 1.3 and three times the earlier pace.

Does the assessment say AI is building new AI on its own?

No. The source describes the work as primarily AI assistance to researchers. It does not establish that an AI system independently directs and completes the development of a successor model.

What did AlphaEvolve reportedly improve?

Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. Those figures are DeepMind’s reported results, as relayed by the source article.

Why is the METR finding difficult to verify publicly?

The source says Anthropic redacted much of the evidence METR found most informative. As a result, outside readers cannot inspect all of the material behind the evaluator’s judgment.

What could slow AI-assisted research?

The source points to experiment time and hardware, difficulties translating small experiments to larger systems, decisions about research direction, continuity, and data quality. It does not quantify how much each constraint affects progress.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

The Bubble Question, Disentangled: 1999 vs 2026 Category by Category

A detailed analysis compares the 1999 dotcom bubble with the 2026 AI cycle, highlighting differences in valuation, fundamentals, and risks across categories.

Aleph Alpha. The retrospective case.

Analyzing Aleph Alpha’s strategic pivot, founder departure, and merger with Cohere to understand the costs of late structural adaptation in European AI.

The Coding Singularity Is Real — and Steeper Than Clark Presented

Recent data confirms the coding singularity is accelerating faster than previously thought, with AI systems now capable of near-human software engineering tasks.

AI in Business: How a Newcomer Outperformed Western Frontier Models in Company Management Test

A newcomer AI model, Kimi K3, outperformed top Western frontier systems in managing a real company during a live experiment, highlighting the importance of trust and discipline in AI decision-making.