🔍 Read the full analysis: AI Helping Build AI: What Should We Expect? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s November 5 assessment says Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with substantial uncertainty. It attributes most of the acceleration to AI assistance for researchers, not fully autonomous research. A separate Google DeepMind example shows a targeted optimization that reduced Gemini training time by 1%, while major limits to broader gains remain.
AI is beginning to assist the research work used to improve AI models, according to a November 5 assessment by independent evaluator METR. METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, while stressing substantial uncertainty and describing the work mainly as AI assistance to human researchers rather than autonomous AI research.
METR said it worked with internal information from Anthropic covering August through October 2026. Its assessment was that the rate of capability improvement on well-specified tasks had accelerated to around twice the previous rate. METR said it would not be surprised by estimates as low as 1.3 times or as high as three times, underscoring that the estimate is not a precise measurement. It judged AI-assisted research likely responsible for most of the acceleration it observed.
The public cannot independently review all the underlying evidence: Anthropic redacted much of the material METR considered most informative. The assessment therefore offers an external evaluator’s judgment, but readers outside the evaluation have limited access to the information behind it. The source article characterizes the reported work as primarily AI helping researchers, not systems independently directing the full research process.
A separate example comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. The company said the change reduced Gemini’s overall training time by 1%. That result illustrates how a large improvement to one component can translate into a smaller gain for an entire training run; it does not by itself establish a general rate of AI-driven research acceleration.
How Research Gains Could Compound
The potential significance lies in a feedback process: AI tools may help researchers build more capable AI, and improved systems may then assist with more of the next development cycle. If those contributions hold up under careful testing, researchers could investigate more ideas or spend less time on routine coding, analysis and troubleshooting. That could raise the amount of research a team completes without requiring every task to be fully automated.
Efficiency gains can also accumulate across stages. The source article gives a simplified illustration: if three independent steps each save 10% of computing, a job would use about 73% of its original compute. Under that illustration’s assumptions, the same budget could run roughly 37% more jobs. This is an example, not a measured result from a laboratory; actual savings may not be independent or apply evenly across a training process.
For businesses, workers and governments, the pace matters because decisions about investment, staffing and infrastructure depend partly on how quickly AI capabilities change. The source argues that the effects will also depend on who can use the systems, who controls the infrastructure and how benefits are shared. Technical progress alone does not settle those questions.
Products Worth Considering
Dell UltraSharp U2724DE 27-Inch WQHD IPS Black Monitor
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
Samsung ViewFinity UJ59 32-Inch 4K UHD Monitor
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successor Models
AI research involves more than producing a promising idea. A team may need to review prior work, inspect code, prepare data, configure experiments, measure outcomes, repeat promising tests and decide whether results are reliable enough to use. AI can potentially help at several points, but an apparent improvement can be misleading: a test may have become easier, runs may use different computing budgets, or a result may vanish under a different random seed.
That makes validation and oversight part of the work being automated. A useful system must not only propose a change or write code; researchers need to determine whether the comparison is fair and the result replicates. METR’s description of current work as mainly assistance indicates that the reported development is not evidence that AI systems have taken over the full research cycle.
The source article connects the METR finding with the AlphaEvolve example as evidence of activity at different levels: a broader assessment of research acceleration and a specific optimization within a model-training process. Neither establishes that progress will continue at the same pace. The article identifies experiment time, hardware availability, differences between small experiments and large systems, research direction, continuity and data quality as possible constraints.
“It would not be surprised by 1.3x or 3x.”
— METR, as summarized in its November 5, 2026 assessment
Limits of the Current Evidence
The size and durability of the acceleration remain uncertain. METR supplied a broad range around its estimate, and Anthropic redacted much of the evidence the evaluator found most informative. The source material does not provide enough public detail to independently assess every underlying result or determine how the estimate was calculated.
It is also unclear how much of the reported assistance will generalize to harder, less well-specified research problems, or how much supervision it requires. A model can produce a plausible result that fails replication or rests on an unfair comparison. The reported examples do not establish that AI can independently select research goals, reliably assess evidence or produce successor models without human oversight.
Further constraints may include the time required to run experiments, access to computing hardware, the gap between small-scale tests and large training runs, and the quality of available data. The source material does not quantify how strongly each factor limits progress or how quickly researchers might address them.
What Researchers Will Need to Verify
The next meaningful evidence will be whether AI-assisted work produces repeatable improvements under independent checks, and whether those gains persist when methods are applied to larger systems. Researchers and evaluators will also need to clarify how much of the work is performed by AI, what human supervision is required and how results are compared with earlier performance.
For now, METR’s assessment is a reason to monitor AI research acceleration, not a firm forecast of when fully automated AI research might arrive. The source article argues that planning should allow for a pace of improvement that can change, while recognizing that the path may be uneven and shaped by technical constraints and choices about deployment and access.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It gave a wide uncertainty range, saying it would not be surprised by estimates of 1.3 times or three times.
Does the assessment show that AI is conducting research autonomously?
No. The source describes the work METR assessed as primarily AI assistance to researchers. It does not establish that AI systems independently run the full research process.
What was the AlphaEvolve result?
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 source article.
Why is the METR finding uncertain?
METR described a broad possible range, and Anthropic redacted much of the evidence METR considered most informative. That limits what outsiders can independently examine about the assessment.
What could slow AI-assisted research?
Possible limits include experiment duration, hardware availability, the difficulty of transferring small-scale results to large systems, research choices, continuity and data quality. The source does not quantify how much each factor will constrain progress.
Source: ThorstenMeyerAI.com
