🔍 Read the full analysis: When AI Helps Build Its Next Generation 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 AI capabilities on well-specified tasks were improving at about twice the earlier rate, with substantial uncertainty. It said AI assistance to researchers was likely responsible for most of the acceleration it observed, while noting that key evidence was redacted. A separate Google DeepMind example shows AI finding a compute efficiency improvement, but how far such gains can scale remains unsettled.
METR, an independent evaluation organization, said in a November 5 assessment that Anthropic’s AI capabilities on well-specified tasks were likely improving at about twice their earlier rate, with high uncertainty. The assessment attributed most of the acceleration it described to AI assistance for researchers, offering a concrete but qualified example of AI systems contributing to the work that develops future models.
METR’s team worked with internal information from August through October. Its assessment said it “would not be surprised” if the acceleration were 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 direct measurement of an overall rate of progress toward more capable AI systems.
The work described was primarily AI assistance to researchers, rather than a fully autonomous system conducting AI research from start to finish. Such assistance can include helping with code, experiments, data preparation, or diagnosing results. Any apparent improvement still needs fair testing: a changed procedure might make a test easier, use different compute, or produce a gain that does not hold across repeated runs.
The public account has a significant evidence limit: Anthropic redacted much of the material METR considered most informative. The assessment therefore gives readers an evaluator’s judgment, but outsiders cannot inspect all of the underlying evidence described. Separately, Google DeepMind reported that its system AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That is a reported efficiency result, not evidence by itself that AI research is accelerating across the field.
How Research Gains Could Compound
If AI systems help researchers complete more useful experiments, the effect could extend beyond a single task: improved tools may help produce models that can assist with later research. That creates a possible feedback loop, but the METR assessment supports a narrower claim: AI assistance likely contributed to acceleration at Anthropic, with substantial uncertainty. It does not establish that AI can independently direct research or that rapid progress is inevitable.
Efficiency improvements can also matter even when they look modest at the scale of a full training run. Google DeepMind’s reported 1% reduction in Gemini training time illustrates how a change in one component may have wider operational value. The source article offers a simplified illustration: three independent 10% savings would leave a task using about 73% of its original compute, allowing roughly 37% more such jobs under the same compute budget. That is a hypothetical calculation, not a reported result from a training system.
For companies, workers, and governments planning around AI, the distinction between assistance and autonomy matters. If researchers can reliably use models to speed parts of development, the pace of capability change may become harder to forecast. Who can access these systems, who controls the computing infrastructure, and who benefits would also depend on deployment and institutions, not only technical progress.
Products Worth Considering

Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray

Bose QuietComfort Wireless Noise-Cancelling Headphones, Petal Pink

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves a chain of tasks: reviewing relevant work, inspecting code, designing an experiment, preparing data, running tests, investigating failures, repeating promising results, and deciding whether a change is real and useful. A model may help with some steps while researchers remain responsible for selecting questions and judging evidence. The source article frames this as a practical way to assess progress: how much of the chain a system can handle, how much supervision it needs, and whether its contributions survive independent checks.
METR’s assessment is one recent evaluation of this process at Anthropic; it should not be read as a finding about every AI laboratory. The separate AlphaEvolve example comes from Google DeepMind and concerns a reported optimization in a component used for Gemini training. Together, the examples show different possible forms of AI contribution, but they do not establish that the same gains are available across organizations or tasks.
The idea of compounding depends on contributions remaining useful as systems and experiments grow. A result that works in a small test may not transfer to a larger model or training run. The source article identifies constraints including experiment time and hardware, the gap between small and large systems, choosing the right research direction, continuity, and data quality. Those limits make the likely path uneven rather than a simple, predictable acceleration.
“would not be surprised”
— METR, in its November 5, 2026 assessment of AI R&D acceleration at Anthropic
What the Assessment Cannot Show
The estimate remains uncertain, and much of METR’s most informative evidence is redacted. The public account does not provide enough detail for outside readers to independently reproduce the assessment or evaluate every underlying measure. METR’s conclusion is a judgment based on information it reviewed, not a definitive public measurement of the pace of AI progress.
It is also unclear how broadly the reported acceleration applies, how much of it came from AI assistance rather than other changes, and whether gains on well-specified tasks will carry over to more complex research. The source material does not establish that systems are choosing research directions, validating results, or building successor models without substantial human involvement. Nor does the AlphaEvolve example establish that its reported efficiency improvement can be repeated or compounded across other training workloads.
Evidence Needed on Scaling Gains
The next useful evidence would clarify how METR measured the rate change, what was redacted, and how AI-assisted work performed against comparable research without that assistance. Further evaluations could show whether the reported effect persists across a longer period, different tasks, and repeated experiments. Publicly inspectable methods and results would help readers distinguish a durable change from a limited or task-specific finding.
For the broader question, the key milestones are whether AI contributions continue to survive independent checks, whether improvements transfer from small experiments to large systems, and whether systems can take on more of the research cycle with less supervision. Until those points are clearer, the source supports a cautious conclusion: AI may already be speeding some research work, but the scale and staying power of that effect remain unsettled.
Key Questions
What did METR conclude about Anthropic’s research pace?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate, while stressing high uncertainty. It said it would not be surprised by an estimate ranging from 1.3 times to three times the earlier rate.
Does the assessment say AI is conducting research autonomously?
No. The work described was primarily AI assistance to researchers. The assessment does not establish that AI systems independently direct the full research process.
What did AlphaEvolve reportedly improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall Gemini training time by 1%. This is a separate reported efficiency result.
Why is the METR finding difficult to verify publicly?
Anthropic redacted much of the evidence METR considered most informative. That limits outsiders’ ability to examine the underlying material, even though METR published its assessment and described its conclusion.
What could limit further acceleration?
Potential constraints include experiment time and hardware, whether results from small tests scale to large systems, selecting useful research directions, continuity, and data quality. The source material does not establish how strongly each factor will limit progress.
Source: ThorstenMeyerAI.com
