🔍 Read the full analysis: AI Starts Improving AI: What Might Happen? 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
METR’s November 5 assessment found that Anthropic’s capabilities on well-specified tasks were likely improving at about twice the earlier rate, with substantial uncertainty. It attributed most of the current acceleration to AI assistance for researchers, while noting that much of the underlying evidence is not public. A separate Google DeepMind example shows an AI-designed optimization cutting part of Gemini’s training time, but it does not establish that rapid or self-sustaining progress is inevitable.
AI assistance may be speeding up some AI research, according to an assessment published November 5 by independent evaluation organization METR. It judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate, but emphasized substantial uncertainty and described the work primarily as AI helping researchers—not as an autonomous system building its own successor.
METR said it worked with internal information from August through October 2026 to assess research and development acceleration at Anthropic. Its estimate is not a direct measurement of every part of model development: it concerns progress on well-specified tasks. METR said it would not be surprised by a rate as low as 1.3 times or as high as three times the earlier rate, signaling a wide range around its central judgment.
The public record is incomplete. According to the source article, Anthropic redacted much of the evidence METR considered most informative. That means readers cannot independently inspect all the underlying information described in the assessment and must rely substantially on the evaluator’s account. METR’s finding also does not, by itself, show that AI systems are directing research independently or that acceleration will continue at the same pace.
A separate example from Google DeepMind shows a narrower, measured efficiency gain. The company reported that its AlphaEvolve system improved a matrix-multiplication component used in Gemini by 23%, reducing Gemini’s overall training time by 1%. Those figures describe a specific optimization, not a general increase in model capability or a forecast for future training runs.
How Research Gains Could Compound
If AI tools help researchers investigate more ideas, write or inspect code, analyze experiments, and catch errors, they could increase the amount of useful work a team completes. The potential significance is not simply that a model performs one task faster. It is that some of the work may improve the tools used in later research, which could contribute to further improvements.
That feedback is a possibility, not a guarantee of runaway progress. Experiments must be run and checked; promising results may fail to reproduce, and apparent gains can come from flawed comparisons. The source article illustrates compounding with a deliberately simplified calculation: if three independent steps each reduce a job’s compute needs by 10%, the remaining requirement is about 73% of the original. Under those assumptions, a fixed budget could run roughly 37% more such jobs. This is an illustration, not a measured forecast of AI research productivity.
The practical consequences could extend beyond AI labs if development becomes more efficient. Faster improvements could affect how businesses plan investments and how institutions prepare for changes in work and access to AI. But technical capability alone will not determine who benefits: deployment, infrastructure ownership, and policy choices also matter.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Samsung 57 Odyssey Neo G9 (G95NC) Dual 4K UHD 1000R Curved Gaming Monitor
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)
As an affiliate, we earn on qualifying purchases.
From AI Assistance to AI R&D
AI research includes more than producing a new model. Teams review prior work, change code, prepare data, configure experiments, investigate failed runs, repeat promising tests, and decide whether a result is reliable enough to use. AI systems may assist with parts of that process without taking responsibility for the entire research cycle.
A reported improvement needs careful checking. For example, a change that appears to improve a training result might have made the test easier, used a different amount of computing, or worked only with one random seed. Useful assistance therefore includes testing whether a result holds up, not just proposing a change or producing code. The source article identifies METR’s assessment and Google DeepMind’s AlphaEvolve example as evidence of different kinds: an evaluation of research acceleration and a company-reported engineering optimization.
The evidence supports a measured conclusion: AI is being used in research, and one evaluator believes this assistance is contributing to faster progress on certain tasks at Anthropic. It does not establish that every research activity is accelerating equally, or that current gains will automatically continue.
“The tool being improved can also contribute to the work of improving it.”
— Thorsten Meyer, source article author
Evidence and Bottlenecks Remain
How large and durable the acceleration is remains uncertain. METR’s estimate covers well-specified tasks, carries a broad uncertainty range, and relies on information that Anthropic redacted. The assessment described AI primarily assisting researchers, but the source material does not provide enough public detail to establish how much of the reported acceleration came from particular tools or research activities.
There are also practical limits. Research still takes time, and access to computing hardware can constrain how many experiments teams run. Results from small experiments may not carry over to large systems. Researchers must choose productive directions, preserve continuity across work, and maintain data quality. These constraints could slow progress even if AI assistance improves.
It is also unknown how gains will be distributed. Who can use research systems, who controls the infrastructure, and how any economic benefits are shared will depend on decisions by companies and institutions as well as on technical progress. The available examples do not establish a timeline for superintelligence or show that a self-reinforcing cycle will occur.
What Future Evaluations Should Track
The next useful evidence will come from further evaluations that specify what tasks are being measured, how results compare with earlier rates, and how much human supervision AI-assisted research requires. Independent access to more of the underlying evidence would help readers judge the strength and limits of the findings. Replication across different tasks and organizations would also show whether the pattern extends beyond one assessment.
For now, the source material reports no specific next METR assessment date or confirmed milestone for fully automated AI research. The central question is whether AI contributions remain reliable when independently checked and can be applied to increasingly difficult work. Until more evidence is available, the pace of improvement should be treated as uncertain, not as a settled trajectory.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate. It reported substantial uncertainty and said the work it described was primarily AI assistance to researchers.
Does the assessment show that AI is building the next AI model autonomously?
No. The source says the work METR described was primarily AI assistance to researchers. The assessment does not establish that AI systems are independently running the full research process or building successors without human direction.
What did AlphaEvolve improve in Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. These are company-reported results for a specific component and training process.
Why is the METR estimate uncertain?
METR gave a wide range around its estimate, and Anthropic redacted much of the evidence METR found most informative. The reported assessment also concerns well-specified tasks, not every part of AI development.
What could slow AI-assisted research?
Potential limits include experiment time, hardware availability, difficulty transferring results from small experiments to large systems, research direction, continuity, and data quality. The source does not give a timeline for how these factors will affect future progress.
Source: ThorstenMeyerAI.com
