🔍 Read the full analysis: How AI Could Contribute To Its Successor on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
METR’s November 5 assessment judged that Anthropic’s AI capabilities on well-specified tasks were improving at about twice the earlier rate, with high uncertainty and AI assistance likely responsible for much of the acceleration. A separate Google DeepMind example reported a 23% speedup in one matrix-multiplication component, reducing Gemini training time by 1%. These developments illustrate how AI could help improve future models, but they do not establish a fully autonomous or self-sustaining research cycle.
AI assistance may already be speeding up parts of model research, according to an assessment published November 5 by independent evaluation organization METR. METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, while stressing substantial uncertainty; the finding matters because AI systems that help researchers build and test models could contribute to the development of their successors.
METR said its assessment drew on internal information from August through October 2026. It judged that AI-assisted research was likely responsible for most of the acceleration it observed, but characterized the work as primarily assistance to human researchers, rather than a fully autonomous system conducting AI research. The assessment expressed considerable uncertainty: METR said it would not be surprised by an acceleration of 1.3 times or three times the earlier rate. Those figures are uncertainty bounds described in the source, not a settled measurement of future progress.
The public account has an important evidence limitation. Anthropic redacted much of the material METR considered most informative, leaving readers outside the evaluation with limited ability to examine the underlying evidence directly. The result is an independent assessment, but outsiders must rely substantially on METR’s interpretation of information they cannot see in full.
A separate example comes from Google DeepMind’s report on AlphaEvolve. The company said the system improved a matrix-multiplication component used in Gemini by 23%, contributing to a 1% reduction in Gemini’s overall training time. The component-level improvement was much larger than the end-to-end time saving; the figures describe different parts of the process and should not be treated as equivalent.
How Research Gains Could Compound
The potential significance is a feedback loop: AI helps researchers improve a model, and the improved model may then assist with more research. Even when a system handles only parts of the work, it could let a team examine more ideas or spend less time on routine coding, debugging and analysis. If the contributions hold up under careful testing, that could increase the amount of research a team completes with its existing resources.
Efficiency improvements can also add up, though the outcome depends on where they occur. The source gives a simplified illustration: three independent 10% savings would leave a job requiring about 73% of its original computing resources, allowing roughly 37% more such jobs under the same budget. This is an illustration, not a measured result from a particular AI lab. The AlphaEvolve example likewise shows why a local engineering gain does not translate one-for-one into faster overall training.
For businesses, workers and governments planning around AI, the rate of improvement matters alongside the current capabilities of a model. If research itself becomes more productive, expectations based only on past development speed may not hold. But technical progress alone does not determine who can use the systems, who controls the computing infrastructure or how the resulting benefits are shared; those outcomes also depend on deployment choices and institutions.
Products Worth Considering
DJI Mini 5 Pro Fly More Combo Plus with DJI RC 2
Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
Anker SOLIX C300 DC Portable Power Station with 60W Solar Panel, 288Wh, 300W
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves a chain of tasks, not just generating a promising idea. Researchers may need to review prior work, inspect code, prepare data, configure and run experiments, diagnose failures, compare results and decide whether an apparent gain is real. An AI system can assist at several points, but each point brings opportunities for error as well as potential time savings.
For example, a system might report that a code change improved performance by 5%. That result could be misleading if the test became easier, the runs used different computing budgets, or the improvement disappears with another random seed. Reliable assistance requires fair comparisons and repeatable results, as well as the ability to recognize when evidence is too weak to support a conclusion. METR’s assessment, as described in the source, concerns help to researchers; it does not establish that AI can independently manage every step.
The mechanism behind the broader concern is therefore incremental: some tasks become easier or faster, researchers use the saved time or compute to conduct more work, and later systems may contribute more. That possibility is not proof of runaway progress. It depends on the quality and reliability of assistance, the kinds of research being automated and whether improvements scale beyond small experiments.
“The tool being improved can also contribute to the work of improving it.”
— Thorsten Meyer, source article author
Limits of the Current Evidence
The scale and durability of the reported acceleration remain uncertain. METR’s assessment covers a limited period and, according to the source, relies on information Anthropic partly redacted. The range METR said it would not be surprised by is broad, and the public material does not provide enough detail to resolve that uncertainty independently.
It is also unclear how far AI assistance will extend beyond well-specified tasks and support for human researchers. Choosing promising research directions, interpreting ambiguous results and transferring gains from small experiments to large systems can be harder than executing a bounded task. The source identifies experiment time, hardware, continuity and data quality as possible constraints, alongside the challenge of deciding what work to pursue.
Neither example establishes that AI can autonomously build its successor or that improvements will continue at a fixed rate. The consequences for jobs, access and distribution of benefits also cannot be inferred from technical capability alone.
Evidence to Watch in Future Evaluations
The next useful evidence will show whether AI contributions continue to improve research outcomes when they are independently checked, repeated and compared on consistent terms. Further reporting from evaluators and laboratories could clarify what tasks systems perform, how much human supervision they require and whether gains translate into faster or less costly development at larger scales.
For now, the source presents two different kinds of evidence: METR’s uncertain assessment of research acceleration at Anthropic and Google DeepMind’s reported engineering improvement for Gemini training. Neither settles how quickly AI research will advance. The practical next step is to track the quality, scope and reproducibility of new results, while treating the pace of progress as capable of changing rather than guaranteed to accelerate.
Key Questions
What did METR conclude about AI research at Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate, with high uncertainty. It said AI-assisted research was likely responsible for most of the acceleration it observed, and described the work as primarily assistance to researchers.
Did METR show that AI can build a successor model on its own?
No. The source describes AI assistance to human researchers. It does not establish that a system can independently choose research priorities, run a complete research program and build its successor without human oversight.
What did AlphaEvolve reportedly improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. The component-level and overall figures measure different things.
Why is it difficult to verify the METR assessment?
According to the source, Anthropic redacted much of the evidence METR considered most informative. That limits what outside readers can inspect directly, even though the assessment was made by an independent evaluation organization.
What could slow this feedback loop?
Possible limits include experiment time and hardware, difficulty scaling results from small tests to large systems, selecting useful research directions, maintaining continuity and ensuring data quality. The source does not establish how strongly each constraint will affect future progress.
Source: ThorstenMeyerAI.com
