🔍 Read the full analysis: How AI Can Help Build The AI That Follows It on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 5, 2026, assessment by independent evaluator METR judged that Anthropic’s model capabilities on well-specified tasks were improving at about twice their earlier rate, with substantial uncertainty. METR said AI assistance to researchers was likely responsible for most of the acceleration it assessed; a separate Google DeepMind example shows AI helping optimize part of Gemini’s training process.
AI tools are beginning to contribute to the research that improves AI models, according to a November 5 assessment by independent evaluator METR. METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice their earlier rate, while stressing substantial uncertainty and describing the work primarily as AI assistance to human researchers.
METR’s assessment drew on internal information from Anthropic covering August through October 2026. The evaluator said it “would not be surprised” by an acceleration of 1.3 times or three times, conveying a wide range of uncertainty around its estimate. The public account does not provide all the underlying evidence: Anthropic redacted much of the material METR considered most informative, leaving outside readers dependent on the evaluator’s judgment.
The mechanism is not necessarily a model independently inventing its own successor. AI can assist with tasks such as writing or inspecting code, proposing experiments, analyzing results and diagnosing failed runs. Researchers still need to check whether a reported improvement is real—for example, whether comparisons used equivalent computing resources and whether a gain holds across different random seeds. The source article says METR characterized the work it assessed as primarily assistance to researchers, not full automation of AI research.
A separate example comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. These are distinct claims about a specific optimization and an overall training outcome; they do not, by themselves, establish that AI systems can autonomously conduct an entire model-development program.
How Research Gains Can Compound
If AI helps researchers test more ideas or spend less time on routine debugging, a laboratory could make more progress with the same staff and computing budget. If some of those improvements also reduce the resources needed to train models, the resulting systems might then provide more capable research assistance. That creates a potential feedback loop: AI contributes to research, and research produces AI that may contribute more effectively.
The scale and pace of such a loop are not established by the examples in the source. A simple illustration in the article says that three independent savings of 10% would leave a task requiring about 73% of its original compute, allowing roughly 37% more of those tasks on the same budget. That is a mathematical illustration, not a measured result from a particular AI lab. The practical effect depends on whether gains are independent, repeatable and applicable to large training runs.
For companies, workers and policymakers, the issue is that the pace of AI development may itself change. Faster research could affect investment, demand for computing infrastructure and the timing of decisions about adopting AI. But technical capability alone does not determine who can use the systems, who owns the infrastructure or how any gains are shared; those outcomes also depend on deployment and institutions.
Products Worth Considering
Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green
DJI Osmo 360 Motorcycle Handlebar POV Combo
Samsung 57 Odyssey Neo G9 (G95NC) Dual 4K UHD 1000R Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Model Training
AI research involves more than producing a promising idea. A team must review relevant work, modify an implementation, prepare data and experiments, measure results, investigate failures and decide whether an apparent improvement survives careful testing. A system that reports a gain can be wrong if it changes the test conditions, uses a different amount of compute or produces a result that disappears on repetition.
That makes the distinction between assistance and autonomy important. A model can make one part of a researcher’s work faster without choosing the research direction or reliably judging every result. The source article argues that even partial support could matter if it lets a human team investigate more ideas. More extensive automation would require systems to carry out and evaluate more of the research sequence with less supervision.
The METR assessment and the AlphaEvolve example speak to different parts of that picture. METR assessed the pace of capability improvement at Anthropic and attributed much of the acceleration it judged to AI-assisted research, with uncertainty. Google DeepMind’s example concerns a particular efficiency improvement in a component used for Gemini training. Neither source, as presented here, establishes a general rate of progress across the AI industry.
Evidence and Limits Still Matter
The estimate is not a precise, publicly reproducible measure in the material provided. METR expressed substantial uncertainty, and Anthropic redacted much of the evidence the evaluator considered most informative. The source does not provide enough detail to independently assess the underlying data, the exact comparison method or how broadly the estimate applies beyond the tasks and period examined.
It also remains unclear how much of the reported acceleration reflects AI assistance rather than other changes in research practice or resources. The source describes METR’s judgment that AI assistance was likely responsible for most of the acceleration it assessed, but that is an attributed evaluation, not proof that AI was the sole cause. The AlphaEvolve result is a separate company-reported example; the source does not describe an independent assessment of that claim.
Several practical limits could slow a research feedback loop: experiment time, access to hardware, data quality and the difficulty of transferring small-scale findings to large systems. Research teams must also decide what to investigate and maintain continuity across projects. The source gives no confirmed timetable for fully automated AI research or for any transition to superintelligence.
What Future Assessments Must Show
The next evidence to watch is whether independent evaluations can test similar claims with clearer public methods and repeated observations over time. More detail about the tasks measured, the comparison baseline and the role of human supervision would help readers distinguish faster performance on specific work from a broad increase in AI research productivity.
Further reporting would also need to establish whether research gains survive replication and transfer to larger training runs. For the AlphaEvolve example, the reported 1% reduction in Gemini’s overall training time gives a concrete outcome, but the source does not say whether the technique has been applied elsewhere or whether additional improvements followed.
For now, the confirmed development is narrower than the prospect of AI building its successors without people: METR assessed faster capability improvement at Anthropic and said AI assistance likely played a major role, while a separate DeepMind example described a targeted training efficiency gain. How quickly these contributions compound, and who benefits from them, remain open questions.
Key Questions
What did METR report about Anthropic?
In an assessment published November 5, 2026, METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice their earlier rate. It said there was substantial uncertainty and that the work it described was primarily AI assistance to researchers.
Did METR say AI is independently building successor models?
No. The source describes the work as primarily AI assistance to researchers. It does not establish that AI systems are independently choosing research directions, validating all results or building successor models without human involvement.
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%, according to the source. This is a company-reported example of a specific optimization.
How certain is the estimate of faster AI research?
METR expressed substantial uncertainty, saying it would not be surprised by an acceleration of 1.3 times or three times. Anthropic also redacted much of the evidence METR considered most informative, limiting what outsiders can independently review from the material provided.
What could slow AI-assisted research?
Potential constraints include experiment time and hardware availability, the difficulty of applying small-scale results to large systems, data quality and the need to choose productive research directions. The source does not provide a schedule for how quickly these constraints might change.
Source: ThorstenMeyerAI.com
