🔍 Read the full analysis: How AI Might Assist In Building Future AI on ThorstenMeyerAI.com
Get the latest gadgets 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 capabilities on well-specified tasks were improving at about twice the earlier rate, with high uncertainty. It attributed much of the acceleration to AI assistance for researchers, while noting that public evidence is limited by redactions. A separate Google DeepMind example shows how a targeted efficiency gain can reduce training time, but neither finding proves that AI can autonomously build its successors.
An assessment published by METR on November 5, 2026, judged that Anthropic was improving its models’ capabilities on well-specified tasks at roughly twice its earlier rate, while stressing high uncertainty. METR said AI assistance to researchers likely explained much of the acceleration it assessed, bringing fresh evidence to a consequential question: can AI systems help develop the models that follow them?
METR, an independent evaluation organization, said its assessment drew on information from Anthropic covering August through October 2026. Its estimate concerned the rate of improvement on well-specified tasks, not a general measure of all AI progress. METR said it “would not be surprised” by an acceleration as low as 1.3 times or as high as three times, underscoring how imprecise the estimate remains.
The work described was primarily AI assistance to human researchers, not a fully autonomous system directing an AI research program. Assistance could include helping write or inspect code, analyze experiment results, or identify problems in a research process. The source article says Anthropic redacted much of the evidence METR considered most informative, leaving readers unable to inspect a substantial part of the underlying material.
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%, reducing Gemini’s overall training time by 1%. That is a company-reported result about one component and one training process; it does not establish that every research task can be accelerated in the same way.
Why Research Speed Could Compound
If AI tools help researchers test more ideas, troubleshoot routine failures, or improve computing efficiency, a laboratory could make more progress with the same people and resources. The effect could accumulate: a model that contributes useful research assistance may be followed by a stronger model that can contribute more. This is the feedback loop behind concern about rapid capability improvement, though the evidence described so far does not establish that such a loop will become self-sustaining.
Efficiency gains can matter even when they appear small. The source offers a simplified illustration: if three independent changes each save 10% of the compute needed for a job, the work uses about 73% of the original compute. Under that illustration’s assumptions, the same budget could run roughly 37% more jobs. These are illustrative calculations, not a measured result from a particular AI lab; real savings may not be independent or combine so neatly.
The consequences extend beyond research teams. The pace of AI development can affect business investment, workforce planning, and decisions about access to computing infrastructure. How any gains are distributed will depend not only on technical progress, but also on who controls deployment and infrastructure and on the institutions that govern them.
Popular right now
Dyson Ball Animal 2 Upright Corded Vacuum Cleaner, Fuchsia
Vitamix VX1 Professional-Grade Blender, 64 oz, Black
Ninja Air Fryer XL with MaxCrisp, 6.5 Qt, 6-in-1, Dark Grey (AF181)
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
Building a model involves many linked tasks: reviewing research, editing implementations, preparing data, setting up experiments, comparing results, and deciding whether a reported gain is real. AI can be useful at individual steps without being able to conduct an entire research program. The distinction matters because partial automation is not full autonomy.
A proposed improvement must survive careful testing. A change that appears to speed up training might have altered the evaluation, used a different compute budget, or produced a gain that disappears with another random seed. Researchers need to check whether results are reproducible and whether a system is accurately reporting what happened. A useful AI research assistant must do more than generate a plausible change: its contributions need to hold up under fair comparisons and independent checks.
The METR assessment provides an estimate of an emerging trend, while the AlphaEvolve example illustrates a specific efficiency improvement. Together they point to different parts of the issue: assistance with research work and changes that can reduce computational cost. Neither, on its own, establishes that AI systems can select research priorities, evaluate evidence, and build increasingly capable successors without meaningful human direction.
“Would not be surprised”
— METR, as quoted in its assessment of Anthropic’s AI R&D acceleration
Limits in the Public Evidence
The estimate is not a precise measurement of how much AI assistance caused Anthropic’s reported pace of improvement. METR expressed substantial uncertainty, and the public cannot review much of the evidence it considered most informative because Anthropic redacted material. The assessment therefore relies in part on the evaluators’ interpretation of information that is not fully available to outside readers.
It is also unclear how broadly the results apply. The assessment concerns Anthropic and well-specified tasks over a particular period; the source does not establish that other laboratories are seeing the same effect or that AI assistance reliably improves every stage of research. Nor does it show that AI systems can independently choose productive research directions, validate their own findings, or sustain a cycle of increasingly capable models.
Possible constraints include experiment time and available hardware, the difficulty of transferring results from small experiments to large models, the choice of research direction, continuity of work, and data quality. The pace may be uneven, and the size and durability of any acceleration remain unsettled.
Evidence to Watch in Future Evaluations
The next useful evidence will be assessments that describe their methods and uncertainty clearly, and that allow outside readers to examine enough supporting information to judge the conclusions. Follow-up evaluations can test whether the estimated acceleration persists beyond the period covered by METR’s review, and whether it reflects AI assistance rather than other changes in research practice or resources.
Researchers and readers will also need to distinguish routine support from higher levels of autonomy: whether a system can merely help implement a planned experiment, or can reliably propose a question, design a fair test, interpret the outcome, and decide what to do next. For now, METR’s report supports a cautious conclusion: AI may already be speeding parts of model research, but the scale, generality, and longer-term effects are not yet established.
Key Questions
What did METR conclude about Anthropic’s research pace?
METR judged that Anthropic’s capability improvements on well-specified tasks were happening at roughly twice the earlier rate. It emphasized high uncertainty, saying it would not be surprised by an estimate between 1.3 and three times.
Did METR say AI is building new models on its own?
No. The work METR described was primarily AI assistance to researchers. The assessment does not establish that AI independently runs an entire model-development program.
What did Google DeepMind report about AlphaEvolve?
Google DeepMind said AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. This is a company-reported result for a specific component and training process.
Why is the METR estimate uncertain?
METR described a broad plausible range, and Anthropic redacted much of the evidence the evaluators considered most informative. The public therefore cannot independently examine all of the material underlying the assessment.
What evidence would clarify whether AI research acceleration is continuing?
Follow-up evaluations could show whether the effect persists over time, applies across different tasks or labs, and survives outside scrutiny. Clearer access to supporting evidence would help readers judge the scale and reliability of any reported acceleration.
Source: ThorstenMeyerAI.com
