🔍 Read the full analysis: AI Helping Create Its Successor: What Could Happen? on ThorstenMeyerAI.com
Get hardware and tech essentials 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 is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The assessment points mainly to AI assisting human researchers, not fully autonomous AI research; the evidence is partly redacted and the limits to faster progress remain unclear.
METR said on November 5 that Anthropic is likely improving its models’ capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The independent evaluator said the work it assessed was primarily AI assistance to researchers, rather than fully automated AI research—a distinction that matters as laboratories consider how much faster AI systems may help build their successors.
METR’s assessment drew on information from Anthropic covering August through October 2026. It judged that AI-assisted research was likely responsible for most of the acceleration it observed. METR said it would not be surprised by estimates ranging from 1.3 times to three times the earlier rate, underscoring that its central estimate is not a precise measurement.
The public account has an important limitation: Anthropic redacted much of the evidence METR considered most informative. Readers therefore cannot independently inspect all the material behind the evaluator’s judgment. The report concerns improvement on well-specified tasks and research assistance; it does not establish that AI systems can independently conduct an entire research program or reliably produce a more capable successor without human oversight.
A separate example illustrates how AI tools might improve the efficiency of model development. Google DeepMind reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That is a reported optimization in one component and one training process, not evidence by itself that AI research is accelerating across the industry.
How Research Gains Could Compound
If AI tools help researchers investigate more ideas, write or check code, and identify flawed experiments, laboratories may be able to test more approaches with the same staff and computing resources. A model that contributes to improving a later model could then become part of a feedback process: the successor may assist with further research. That possibility is the reason the assessment matters, even though the current evidence describes assistance rather than a fully self-directed loop.
Efficiency gains could add up, but the arithmetic depends on assumptions. In one deliberately simplified illustration, three independent 10% reductions in computing needs would leave a task using about 73% of its original compute. At a fixed budget, that could support roughly 37% more such jobs. This is an illustration, not a measured industry-wide result; real improvements may overlap, fail to generalize, or require additional resources.
The practical effects would reach beyond research labs if more capable systems become cheaper or faster to develop. Businesses and public institutions would have to plan around a pace of change that may vary, while questions about access, infrastructure ownership, and how benefits are shared would depend on deployment and institutional choices as well as technical progress.
Products Worth Considering
LG UltraGear 32GS60QC-B 32-Inch Curved QHD Gaming Monitor
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than generating a promising idea. Researchers review existing work, inspect implementations, prepare data and experiments, compare results, investigate failures, and decide whether an apparent gain holds up. A system that helps with one stage can save time, but a useful contribution must survive checks: the comparison must be fair, the result should not depend on a convenient test setup, and the improvement should hold when experiments are repeated.
For example, an AI agent might change training code and report a 5% improvement. That result could be misleading if the test became easier, the runs used different compute, or the gain disappears under another random seed. Verification is part of research, not an optional step after it. The more reliably systems can propose, implement, test, and evaluate changes, the more of the research process they may assist with.
The current evidence points to a gradual and uneven process. METR described AI assistance to researchers, while the AlphaEvolve example concerns a specific technical optimization. Neither source establishes that every part of AI development is accelerating at the same rate. Small, verified improvements may matter, but their effects depend on where they apply and whether they can be repeated at larger scale.
“The work it described was primarily AI assistance to researchers.”
— METR’s assessment of Anthropic
Limits of the Available Evidence
The size and reliability of the acceleration remain uncertain. METR’s estimate has a wide stated range, and Anthropic redacted much of the evidence evaluators found most informative. The public information does not let readers assess all the underlying records or determine how broadly the findings apply beyond the tasks and period examined.
It is also unclear how much assistance can extend from routine, well-specified tasks to choosing research directions, judging ambiguous results, or developing larger systems. Other potential limits include experiment time, available hardware, the gap between small tests and large-scale training, continuity across research work, and data quality. The source material does not quantify how strongly any one of these constraints may slow progress.
Neither the METR assessment nor the AlphaEvolve example establishes a timetable for fully autonomous AI research, superintelligence, or economy-wide changes in work. Those are possibilities and interpretations, not confirmed outcomes of the reported findings.
What Researchers Must Test Next
The immediate next step is further evaluation of whether AI-assisted gains hold up across more tasks, longer periods, and independent checks. More public detail about evaluation methods and the evidence behind the METR assessment would help readers judge how much weight to place on its estimate. Repeated, comparable measurements will matter more than a single headline rate.
Researchers and organizations will also need to test whether systems can do more than help with isolated tasks: propose useful experiments, carry them out fairly, spot weak evidence, and transfer improvements to larger training runs. Until such capabilities are demonstrated and measured, the current reporting supports a narrower conclusion: AI may already be helping some researchers work faster, but the extent and durability of that effect are not settled.
For companies and policymakers, the development is a reason to monitor both technical evaluations and decisions about access, infrastructure, and oversight—not proof of a fixed pace of change. Progress could be uneven, and plans will need to account for uncertainty as new assessments emerge.
Key Questions
What did METR conclude about Anthropic?
METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It expressed high uncertainty, saying estimates from 1.3 times to three times would not surprise it.
Does the assessment show that AI is building its successor on its own?
No. METR described the work as primarily AI assistance to human researchers. The assessment does not establish that AI systems can independently run an entire research program or build a successor without oversight.
What did Google DeepMind report about AlphaEvolve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. The example concerns a specific component and training process.
Why is the METR estimate uncertain?
METR gave a broad range for its estimate, and Anthropic redacted much of the evidence METR considered most informative. The public report therefore does not allow readers to inspect all the material underlying the judgment.
What could limit AI-driven research acceleration?
Possible constraints include experiment time, hardware availability, scaling results from small tests to large systems, selecting research directions, continuity, and data quality. The available source material does not quantify how much each factor may slow progress.
Source: ThorstenMeyerAI.com
