TL;DR
METR’s November 5 assessment said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, but with high uncertainty. Google DeepMind separately reported that AlphaEvolve made one matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. These examples show ways AI may assist AI research, but they do not establish that a self-reinforcing development loop is already operating at scale.
METR said November 5 that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The assessment describes primarily AI assistance to researchers, not a fully autonomous system building its own successor; alongside a separate Google DeepMind efficiency result, it gives a concrete but incomplete picture of how AI may contribute to further AI development.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
METR, an independent evaluation organization, said it would not be surprised if the acceleration it assessed were as low as 1.3 times or as high as three times the earlier rate. That wide range signals substantial uncertainty. METR also said much of the evidence it considered most informative had been redacted, limiting what outside readers can inspect and making the estimate difficult to evaluate independently from the public account alone.
A separate example comes from Google DeepMind’s AlphaEvolve. The company reported in May 2025 that the system improved the speed of a matrix-multiplication component used in Gemini by 23%. That component-level gain cut Gemini’s overall training time by 1%, according to the company. The figures describe different scopes: a large speed improvement in one part did not translate into a similarly large reduction in total training time.
The proposed feedback loop is straightforward in outline: AI helps researchers suggest changes, write or inspect code, and run experiments; researchers test which changes hold up; successful changes contribute to a later model that can assist with further research. Each step needs evidence. A reported improvement might disappear under repeated trials, reflect an unfair comparison, or fail to transfer to a larger training run. The source essay presents this loop as a schematic argument, not a measured process at any lab.
How Research Gains Could Compound
If AI tools reliably reduce the time or computing needed for research, labs could test more ideas with the same resources. Efficiency gains may also add up. As a simplified illustration, three independent savings of 10% would leave a job requiring about 73% of its original computing; with a fixed budget, that would allow roughly 37% more such jobs. This is illustrative arithmetic, not a measured result from Anthropic or Google DeepMind, and real savings may not be independent or transferable.
The broader consequence depends on whether AI assistance improves the full research cycle, including deciding which questions to pursue and recognizing weak evidence. Faster coding alone would not guarantee faster scientific progress. If capabilities do improve more quickly, the pace could affect how companies plan investment and how institutions prepare for changes in work. Who can access these systems, who controls the infrastructure, and how benefits are shared will also depend on deployment and policy, not only technical performance.
Products Worth Considering
Anker SOLIX C300 DC Portable Power Station with 60W Solar Panel, 288Wh, 300W
LG UltraGear 32GX870A-B 32-inch 4K UHD OLED Gaming Monitor
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
From Lab Assistance to Successors
AI research involves more than producing a promising idea. A team may need to review prior work, understand existing code, prepare data and experiment settings, run trials, investigate failures, repeat promising results, and decide whether a gain is real enough to use. AI can potentially help with parts of that process, but apparent progress can be misleading: a test may have become easier, runs may have used different amounts of computing, or an improvement may disappear with a different random seed.
The two cited examples measure different things. METR’s estimate concerns the rate of improvement in model capabilities on well-specified tasks at Anthropic and carries a broad uncertainty range. AlphaEvolve’s reported result concerns one optimization in a component of Gemini’s training process. Neither alone establishes that models can independently set research priorities, validate their own findings, and produce successive generations with little human direction.
““Would not be surprised” by an acceleration of 1.3x or 3x.”
— METR
Limits of the Acceleration Evidence
METR’s estimate is explicitly uncertain, and substantial parts of the evidence it found most informative were redacted. The public material does not allow readers to fully assess how the estimate was derived or how much of the reported acceleration came from AI assistance compared with other changes in the research process.
It is also unclear how often AI-generated research contributions survive rigorous replication, how well gains on well-specified tasks carry over to larger systems, and whether improvements can continue when experiments require scarce hardware or lengthy run times. The cited material does not establish that AI systems at either lab are autonomously directing a sustained cycle of successor development.
What Future Evaluations Need to Show
Further assessments can clarify whether AI assistance is speeding up research by tracking how much work systems complete, how much human supervision remains necessary, and whether results hold up under fair, repeated tests. More transparent evidence would also help outside researchers judge the scale of any acceleration.
For claims about a self-reinforcing loop, the key question is whether gains persist across successive development cycles and across larger experiments. Until that evidence is available, the METR estimate and AlphaEvolve result support a narrower conclusion: AI has contributed to research and training tasks in reported cases, while the speed, scope, and long-term effect of that contribution remain unsettled.
Key Questions
Did METR find that Anthropic’s AI is building its own successor?
No. METR assessed the rate at which Anthropic was improving model capabilities on well-specified tasks and said the work it described was primarily AI assistance to researchers. Its assessment does not establish a fully autonomous successor-building process.
How certain is METR’s estimate?
METR described the estimate as highly uncertain. It said it would not be surprised by an acceleration of 1.3 times or three times, around its roughly two-times central assessment. It also redacted much of the evidence it found most informative.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. The company said this cut the model’s overall training time by 1%.
Do these examples prove that AI development will accelerate continuously?
No. They show specific reported contributions and an uncertain assessment, not proof of continuous acceleration. Experiment time, hardware availability, research choices, data quality, and whether results transfer to larger systems may all affect progress.
Source: Thorsten Meyer AI
