Read the full analysis: AI Improving AI: A Look At What Might Happen on ThorstenMeyerAI.com
TL;DR
METR’s November 5, 2026 assessment judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate, with substantial uncertainty. It attributed most of the acceleration it described to AI assistance for researchers, while noting that much of the supporting evidence was redacted. A separate Google DeepMind example shows a measurable efficiency gain, but the extent to which such improvements will compound remains unclear.
Independent evaluator METR said on November 5 that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate, and judged that AI assistance to researchers likely explained most of the acceleration it described. The assessment carries substantial uncertainty, and Anthropic redacted much of the evidence METR considered most informative, limiting what outsiders can independently verify. The findings add evidence to a consequential question: whether AI tools can help researchers build more capable AI systems faster.
METR said its assessment drew on internal information from August through October 2026. It described the rate estimate as uncertain: the organization said it “would not be surprised” by a rate of 1.3 times or 3 times the earlier pace. The account summarized in the source material says the work was primarily AI assistance to human researchers, rather than a fully autonomous system conducting AI research from start to finish. METR’s estimate concerns progress on well-specified tasks; it should not be read as a direct measure of every part of model development.
There is a separate example of AI contributing to efficiency work. Google DeepMind reported that AlphaEvolve improved a matrix-multiplication component used in Gemini, making that component 23% faster. DeepMind said this reduced Gemini’s overall training time by 1%. The component-level gain was substantially larger than the reported effect on the full training process, a distinction that matters when estimating the practical impact of a technical improvement.
These examples do not establish that AI systems are independently directing their own development. They do show two different kinds of reported contribution: researcher assistance assessed by METR and an optimization attributed by Google DeepMind to AlphaEvolve. The source article uses a simplified illustration to explain compounding: if three independent changes each cut compute needs by 10%, a job would use about 73% of its original compute. That is a hypothetical calculation, not a measured result from either organization.
How Research Gains Could Compound
If AI tools help researchers inspect code, propose experiments, analyze results or diagnose failed runs, a team could investigate more ideas with the same time and resources. If some resulting improvements make later AI systems more capable or efficient, those systems may then assist with further research. This is the feedback loop behind concerns about rapid progress: AI could help improve the tools used to improve AI.
The available examples do not show that this loop is already self-sustaining. The METR assessment describes acceleration with high uncertainty and primarily human-led work. The DeepMind example reports a specific engineering gain, not a general increase in the speed of AI research. Still, even modest savings could matter if they recur across many components or development cycles. Whether gains accumulate depends on their reliability, how widely they apply and the resources needed to test and deploy them.
The stakes extend beyond technical performance. Faster development could affect companies planning investments, workers and institutions adapting to new capabilities, and decisions about access to powerful systems. The distribution of benefits will depend not only on what models can do, but also on who controls the systems and infrastructure, how they are deployed and what rules govern their use.
From Research Assistance to Successors
AI research is a chain of work, not one task. Researchers identify a problem, review previous work, change code or data, run experiments, compare results and decide whether an apparent gain is real. AI assistance at any stage might save time, but a plausible result can be misleading. An experiment may use different computing resources from its comparison, depend on a favorable random seed or inadvertently make its evaluation easier. Reliable research therefore requires fair tests, repeatable results and scrutiny, not just code that runs or an encouraging measurement.
The source article, published by Thorsten Meyer on November 15, frames METR’s assessment as evidence that the process may already be changing, while emphasizing that the evidence is incomplete. METR’s judgment is based partly on internal material that Anthropic redacted from public view. Readers can know what the evaluator concluded, but cannot inspect all the underlying evidence described as most informative.
The source also identifies constraints that could slow progress: the time and hardware required for experiments, difficulty transferring results from small tests to large systems, choosing valuable research directions, maintaining continuity and securing high-quality data. These limits mean development need not accelerate smoothly. Progress may be uneven, with advances in some parts of the process and bottlenecks in others.
“would not be surprised”
— METR, as quoted in its assessment
What the Public Evidence Cannot Show
Important details remain unresolved. Anthropic redacted much of the evidence METR found most informative, so external readers cannot fully evaluate the basis for the organization’s estimate. The source material does not provide the underlying measurements or enough detail to independently assess how the reported pace was calculated. METR’s conclusion is an attributed assessment, not a directly observable measure available for public replication.
It is also unclear how much of the reported acceleration came from AI assistance compared with other changes in research practices or resources. The account says AI assistance likely played the main role in the work METR described, but does not establish that AI was the sole cause. Nor do the examples show whether gains will persist when applied to larger systems, across different research problems or over multiple development cycles.
Finally, the source does not establish a timeline for fully automated AI research or for any particular level of capability. The feedback loop could remain limited by human oversight, experiment costs, hardware, data quality and the challenge of selecting worthwhile directions. Claims about rapid progress toward superintelligence are therefore forecasts and concerns, not outcomes demonstrated by these examples.
Testing the Next Research Gains
The next useful evidence will include further assessments that disclose enough methods and results for outside researchers to judge them, as well as repeated demonstrations that AI-assisted work produces improvements under fair, comparable tests. For specific efficiency claims, the key question is whether component-level gains translate into meaningful system-level savings, as distinct from the reported 1% reduction in Gemini training time in the example provided.
Researchers and organizations will also need to track how much supervision AI-assisted experiments require, whether results hold across repeated runs and whether improvements transfer from small tests to large systems. The source material does not identify a scheduled next report or a specific forthcoming milestone. For now, METR’s assessment and DeepMind’s reported optimization are evidence to monitor, not a settled forecast of how quickly AI development will proceed.
Key Questions
What did METR conclude about Anthropic’s AI research?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate. It said the estimate had substantial uncertainty and that AI assistance to researchers likely accounted for most of the acceleration described.
Did METR say AI is conducting AI research on its own?
No. The source material describes the work METR assessed as primarily AI assistance to researchers. It does not establish that AI systems were autonomously conducting the full research process.
What did AlphaEvolve improve in Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%, according to the company.
Why is METR’s assessment hard to verify independently?
Anthropic redacted much of the evidence METR considered most informative. The assessment’s conclusion is public, but outsiders cannot review all the underlying material described in the source.
Does the evidence prove AI will rapidly improve itself?
No. The examples suggest AI assistance can contribute to research and engineering, but they do not prove that improvements will compound reliably or lead to fully autonomous self-improvement. The pace and scale of future gains remain uncertain.
Source: ThorstenMeyerAI.com