🔍 Read the full analysis: What Comes Next When AI Helps Build 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 said Anthropic was likely improving model capabilities on well-specified tasks at roughly twice its earlier rate, with high uncertainty, and that AI assistance to researchers was primarily responsible. Google DeepMind has separately reported a 1% reduction in Gemini’s overall training time from an AI-designed optimization. These examples show how AI may contribute to its own development, but they do not establish that fully automated research or rapid, unbounded progress is imminent.
AI assistance may be speeding up AI research, according to an independent evaluation of Anthropic’s work published November 5 by 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 primarily as AI support for researchers, not autonomous AI research.
METR said its assessment drew on internal information from August through October 2026. The organization judged that AI-assisted research was likely responsible for most of the acceleration it assessed. It said it would not be surprised if the rate were as low as 1.3 times the earlier rate or as high as three times. Those figures describe METR’s uncertain estimate, not a measured rate of improvement that can be applied to every model or research task.
The evidence available to outsiders is limited: Anthropic redacted much of the material METR considered most informative. The finding is therefore an evaluator’s judgment based partly on information that the public cannot inspect. METR’s report describes researchers receiving AI assistance; it does not establish that AI systems independently set research agendas, verify results, and develop successor models without human oversight.
A separate example comes from Google DeepMind. The company reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. The component-level improvement and the overall training-time reduction are different measures. The report illustrates how an AI-developed change to one part of a system can produce a smaller gain across a larger process.
How Research Gains Could Compound
The practical significance is not that AI has already taken over AI development. It is that assistance with routine or technically demanding research tasks could let a team test more ideas, diagnose failures sooner, or spend more time on decisions that still need human judgment. If the resulting improvements make later systems more useful in research, those systems could then contribute further assistance. The feedback loop is plausible, but its pace and scale depend on whether proposed changes work when independently checked.
Small, separate efficiency gains can add up. As a simplified illustration, three independent savings of 10% would leave a job using about 73% of its original compute. With a fixed compute budget, that would allow roughly 37% more such jobs. This is arithmetic, not a forecast: the savings may not be independent, may not apply to the same work, and do not show that research quality or useful output will rise by the same amount.
For businesses and workers, the key issue is that a changing pace of AI development can make planning harder. Faster research could affect when new capabilities arrive, but the effects on employment, access, ownership and the distribution of benefits will also depend on deployment choices and institutions. The examples do not settle those questions.
Products Worth Considering
Samsung Odyssey G7 G75F 40-Inch WUHD Curved Gaming Monitor
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successor Models
AI research involves a chain of work: reading prior results, inspecting code, preparing data, running experiments, checking failures and deciding whether an apparent improvement is real. Systems can assist at several points without taking responsibility for the entire process. A researcher might ask an agent to change code and report a better result, but the result could reflect a changed test, different computing resources or random variation rather than a real advance.
That is why verification matters as much as generating proposals. A useful contribution must survive fair comparisons, repeat runs and scrutiny of the method. Even partial assistance could increase the number of ideas a research team investigates, while full automation would require systems to handle more of the sequence, including choosing what to study and judging evidence. The METR assessment describes assistance as the main factor in the case it reviewed; it does not show that the whole research cycle is automated.
The METR assessment and DeepMind’s AlphaEvolve report are distinct examples, not a single controlled comparison. One concerns an evaluation of research acceleration at Anthropic; the other is a company-reported optimization in Gemini’s training process. Together they give concrete reasons to examine AI’s role in AI development, but they do not prove that similar gains will appear across labs, tasks or time periods.
““The tool being improved can also contribute to the work of improving it.””
— Thorsten Meyer, describing the feedback-loop concern
Limits in the Available Evidence
METR’s estimate has high uncertainty, and much of the evidence it considered most informative was redacted by Anthropic. The public cannot independently review that material from the account provided here. The assessment also concerns one company and a particular period, August through October 2026; it does not establish a general rate of acceleration across the AI industry.
It remains unclear how much of the assessed change came from AI assistance, how durable the reported pace will be, and whether the gains extend beyond well-specified tasks. It is also unclear how often AI-generated research ideas will survive replication and whether productivity gains will translate into faster improvements in overall model capabilities. The AlphaEvolve figures are reported by Google DeepMind, and the source material does not provide an independent evaluation of them.
There are practical constraints: experiments take time and computing hardware; results from small tests may not carry over to large systems; research direction still requires judgment; and progress can be limited by continuity and data quality. The transition, if it occurs, may be uneven. No evidence here establishes fully autonomous AI research or a specific timetable for superintelligence.
What Future Evaluations Must Show
The next useful evidence would show, with clearer public methods, which research tasks AI systems handle, how much human supervision they need, and whether their contributions remain valuable after independent checks and repeat testing. Further evaluations could help distinguish faster work on well-defined tasks from faster improvement in overall model capabilities.
For now, the reported developments support a measured conclusion: AI is assisting some work involved in developing and improving AI systems, and that assistance may be contributing to faster research at Anthropic. How far the feedback loop can extend remains unsettled. Its consequences will depend not only on technical progress, but also on who can access the systems and infrastructure and how institutions manage their use and share the benefits.
Key Questions
What did METR conclude about Anthropic’s AI research?
METR judged that Anthropic was likely improving model capabilities on well-specified tasks at roughly twice its earlier rate. It said the estimate was highly uncertain, with plausible rates ranging from 1.3 to three times the earlier rate, and described AI assistance to researchers as the primary factor.
Does this mean AI is independently building the next generation of AI?
No. The assessment described AI primarily assisting researchers. It does not establish that AI systems independently choose research priorities, validate results and develop successor models without human oversight.
What did AlphaEvolve change in Gemini’s training?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. The company said the change reduced overall Gemini training time by 1%.
Why is the METR finding difficult to verify publicly?
Anthropic redacted much of the evidence METR said was most informative. That leaves outside readers unable to inspect all of the material behind the assessment, which METR itself described as uncertain.
What could slow AI-assisted research?
Research still depends on experiment time, hardware, reliable data and sound choices about what to investigate. Results from small tests may not carry over to large systems, and proposed improvements need careful validation.
Source: ThorstenMeyerAI.com
