🔍 Read the full analysis: What Changes When AI Helps Build Future AI? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 5 METR assessment judged that Anthropic’s model capabilities on well-specified tasks were improving at roughly twice an earlier rate, with high uncertainty, and attributed much of the acceleration to AI assistance for researchers. Google DeepMind has separately reported that AlphaEvolve made one Gemini matrix-multiplication component 23% faster, reducing overall training time by 1%. These examples point to possible feedback between AI tools and AI development, but they do not establish fully autonomous research or a predictable path to superintelligence.
AI-assisted research may be speeding up AI development, according to an independent METR assessment published November 5, 2026. METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice an earlier rate, but stressed substantial uncertainty; the assessment described mainly AI assistance to researchers, not an autonomous system building its successor.
METR said its assessment drew on internal information from August through October 2026. It judged that AI-assisted research was likely responsible for most of the acceleration it observed at Anthropic. The evaluators said they would not be surprised by estimates ranging from 1.3 times to three times the earlier rate, underscoring that the roughly twofold figure is an uncertain assessment rather than a precise measurement of a settled trend.
The public record is incomplete: Anthropic redacted much of the evidence METR considered most informative. That means readers cannot independently inspect all the material behind the assessment and must rely substantially on the evaluator’s account. METR’s finding concerns the rate at which model capabilities improve on well-specified tasks; it should not be read as a direct estimate of the pace toward general intelligence or superintelligence.
A separate example comes from Google DeepMind’s report on AlphaEvolve. The company said the system improved a matrix-multiplication component used in Gemini by 23%, cutting Gemini’s overall training time by 1%. The component-level gain was much larger than the end-to-end change. It illustrates how a local optimization can matter to a large training run, but does not by itself show that AI systems can conduct an entire research program independently.
How Research Gains Could Compound
The development matters because research itself is a major input to building better AI. If models reliably help researchers write and test code, inspect results, find errors, or explore more approaches, a team may be able to complete more work with the same staff and computing resources. A stronger model could then assist with later research. That creates a possible feedback loop: AI contributes to improving AI, which may increase the capacity available for subsequent development.
Efficiency gains can also accumulate across stages. In a simplified illustration, three independent 10% reductions in resource use leave a task requiring about 73% of its original compute. With a fixed budget, that could support roughly 37% more of those jobs. This is an illustration, not a reported result from Anthropic or Google; real gains may not be independent, repeatable, or transferable to larger systems.
For businesses, workers and governments, the distinction between gradual and accelerating improvement affects planning. Faster progress could change the timing of new capabilities and the resources needed to develop them. But technical capability alone does not determine who can use the systems, who owns the infrastructure, or how any gains are distributed. Those outcomes also depend on deployment choices and institutions.
Popular right now
eufy C30 Robot Vacuum and Mop
Ninja Air Fryer XL with MaxCrisp, 6.5 Qt, 6-in-1, Dark Grey (AF181)
De'Longhi Classic Signature Espresso Machine
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Automation
AI research involves more than proposing an idea. A team may need to review prior work, inspect an implementation, prepare data and experiment settings, run tests, investigate failures, repeat promising results, and decide whether a measured improvement is real. AI tools can potentially assist with several of these steps without taking responsibility for the whole process.
That distinction is important when evaluating claims of research acceleration. A system might suggest a code change that appears to improve performance, while the result actually reflects a different test, unequal compute, or random variation. Useful assistance depends on fair testing and independent verification, as well as on generating promising ideas. The METR assessment described AI primarily assisting researchers; it does not establish that the systems choose research priorities or validate their own findings without human oversight.
Google’s AlphaEvolve example offers a separate measure of what optimization can look like in practice: a 23% improvement in one component corresponded, according to Google DeepMind, to a 1% reduction in total Gemini training time. Component-level results and overall training outcomes are different measures, so the larger figure should not be presented as the gain to the full training process.
“It “would not be surprised” by an acceleration of 1.3x or 3x.”
— METR, as summarized in its November 5 assessment
Evidence Gaps and Practical Limits
The METR estimate remains uncertain, and much of its most informative evidence is redacted. The public summary does not let outsiders independently assess all the underlying data or determine precisely how much of the observed change came from AI assistance rather than other factors. The estimate also applies to capabilities on well-specified tasks, not to every kind of AI research.
It is not clear how readily gains from small experiments carry over to large training runs, or how often apparent improvements survive replication. The process may be constrained by experiment time, hardware, data quality, and the difficulty of selecting fruitful research directions. Continuity also matters: an AI tool that helps with one task may not retain useful context or perform consistently across a longer research program.
Neither cited example establishes fully autonomous AI research, a guaranteed compounding rate, or a timeline for superintelligence. The source article argues that rapid progress deserves serious attention, but that is an interpretation of the evidence. The pace and consequences of further development remain uncertain.
What to Watch in Future Assessments
The next useful evidence will include additional assessments that explain their methods, specify comparison baselines, and allow readers to judge how much AI assistance contributed. Replications across teams and tasks could help establish whether METR’s estimate reflects a broader change or a result limited to particular work at Anthropic. More detail about supervision and verification would clarify how much of the research process is being automated.
For reported efficiency gains, the key measure is not only whether a component improves, but whether the improvement holds under comparable conditions and changes the cost or performance of a full system. Until evidence answers those questions, the most supportable conclusion is limited: AI is already assisting some AI research, and a recent independent assessment suggests that assistance may be affecting the pace of improvement, but the scale and durability of that effect are not settled.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice an earlier rate, with substantial uncertainty. It said AI-assisted research likely accounted for most of the acceleration it assessed.
Does this show that AI can build the next AI model on its own?
No. The METR assessment described work that was primarily AI assistance to researchers. It does not establish autonomous research or a system independently designing, testing, and validating a successor model.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. The company said this reduced Gemini’s overall training time by 1%.
Why is the METR estimate uncertain?
METR gave a wide range of outcomes it considered plausible, and Anthropic redacted much of the evidence the evaluators found most informative. The public cannot independently review all the material behind the assessment.
What could limit AI-driven research acceleration?
Potential constraints include experiment time, hardware, data quality, the challenge of choosing useful research directions, and whether results from small tests carry over to large systems. Apparent improvements also need fair testing and replication.
Source: ThorstenMeyerAI.com
