🔍 Read the full analysis: How AI Could Help Build The Next Generation Of 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 assessment by independent evaluator METR judged that Anthropic’s model capabilities on well-specified tasks were improving at around twice their earlier rate, with high uncertainty, and attributed much of the acceleration to AI assistance for researchers. A separate Google DeepMind result showed an AI-designed change speeding up one component of Gemini’s training process, illustrating a possible route for efficiency gains to compound. The evidence does not establish fully autonomous AI research or a predictable path to superintelligence.
AI assistance may be speeding up some AI research, according to an assessment published November 5 by the independent evaluation organization METR. Reviewing information from Anthropic, METR judged that the company’s model capabilities on well-specified tasks were improving at around twice the earlier rate, while stressing that its estimate carried high uncertainty and that the work described was primarily assistance to human researchers.
METR said it would not be surprised if the acceleration were closer to 1.3 times or three times the earlier rate, reflecting the limits of the available evidence. Its researchers worked with internal information from August through October. Anthropic redacted much of the material METR considered most informative, leaving outside readers unable to inspect key evidence directly. The assessment is therefore an evaluator’s judgment about a particular company and period, not a direct measurement of the whole AI industry.
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%. That result shows how a substantial improvement in one part of a system can translate into a smaller gain across the full process. It does not, by itself, show that AI systems can independently develop successor models.
The proposed feedback loop is straightforward but not automatic: a model helps researchers write or inspect code, design experiments, analyze results, or identify errors; researchers test any suggested change; and validated improvements can be used in later development, as explored in research on AI and scientific progress. An apparent gain must survive fair comparisons, repeated runs, and checks for issues such as different compute budgets or easier test conditions. Verification is part of the research work, not an optional step after it.
How Research Gains Could Compound
If AI tools let a research team test more ideas or spend less time on routine debugging, that team may be able to make progress faster without adding the same amount of human labor. Improvements to training code, data preparation, or experiment design could also reduce the compute or time needed to reach a target capability. Those savings may give teams more room to run further experiments.
A simple illustration in the source article shows the arithmetic, not a measured industry result: if three independent changes each cut resource use by 10%, a task would use about 73% of its original compute. With a fixed budget, that could support roughly 37% more such tasks. Real changes may not be independent, and the illustration does not establish that those savings have occurred together. Its point is that separate efficiency improvements can add up.
The broader consequence is uncertainty about the pace of AI development. Faster research could affect companies, workers, and governments making plans around AI capabilities and infrastructure. But technical progress alone does not determine who can use advanced systems, who controls the computing resources, or how any gains are distributed. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering

Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Bose QuietComfort Wireless Noise-Cancelling Headphones, Petal Pink
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than producing a promising idea. Researchers review prior work, inspect implementations, prepare data and experiments, compare results, investigate failed runs, and decide whether an apparent improvement is real. AI tools can assist with parts of that sequence, but a result that looks better may reflect a flawed comparison, a changed test, or variation between runs.
The METR assessment concerns AI assistance to researchers, not a finding that AI systems are already conducting the full research process autonomously. The distinction matters: helping a person complete well-specified tasks faster is different from independently selecting research goals, evaluating evidence, and building a successor model with little or no supervision.
The AlphaEvolve example offers a concrete case of a tool contributing to a component used in model training, according to Google DeepMind. The reported 23% improvement applied to that component; the company said the effect on Gemini’s overall training time was 1%. Together, the two examples indicate possible mechanisms for research acceleration, but they do not establish that every AI lab is seeing the same effect or that gains will continue at a constant rate.
“METR would not be surprised by an acceleration of 1.3x or 3x.”
— METR, as summarized in its November 5 assessment
Limits of the Current Evidence
The size and cause of the reported acceleration remain uncertain. METR described its estimate as uncertain, and much of the evidence it considered most informative was redacted by Anthropic. The public material does not let readers independently evaluate all the underlying information or determine how well the finding applies beyond the company and period assessed.
It is also unclear how reliably AI contributions transfer from small or well-specified tasks to larger research problems. Experiment time, access to computing hardware, data quality, and the ability to choose productive research directions could all limit progress. The source material does not establish that these constraints have been overcome, or that improvements will continue at the same pace.
Finally, the DeepMind result is a company-reported example, and the available account does not provide enough detail to assess every aspect of its testing. Neither example proves that AI systems can autonomously build increasingly capable successors or that rapid progress toward superintelligence is inevitable. Those remain possibilities and interpretations, not confirmed outcomes.
What Researchers Must Test Next
The next evidence to watch is whether AI-assisted research produces gains that can be independently checked and reproduced across tasks, teams, and time periods. Useful assessments would clarify what work the AI performed, how much human supervision it required, and whether claimed improvements held up under comparable compute budgets and repeated experiments.
METR’s assessment describes work through October 2026, but the source material does not identify a scheduled follow-up or a next publication date. Further evidence could clarify whether the estimated acceleration persists and how much is attributable to AI assistance rather than other changes in research practice or resources.
For now, the practical question is not simply whether AI can help build AI, but which research tasks it can improve reliably and at what cost. Continued evaluation will help distinguish useful assistance from results that fail under scrutiny—and show whether the feedback loop is becoming stronger or is constrained by time, hardware, data, and human judgment.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic’s model capabilities on well-specified tasks were improving at around twice the earlier rate. It gave a wide uncertainty range, saying it would not be surprised by 1.3 times or three times, and said the work described was primarily AI assistance to researchers.
Does the assessment show that AI is building AI on its own?
No. The assessment described AI assisting human researchers. It does not establish that AI systems can independently choose research goals, evaluate results, and build successor models without substantial human involvement.
What did Google DeepMind say AlphaEvolve improved?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. The company said that reduced Gemini’s overall training time by 1%.
Why is the METR finding uncertain?
METR described its estimate as highly uncertain, and Anthropic redacted much of the evidence METR considered most informative. The public account therefore does not provide enough underlying detail for outsiders to independently assess every part of the evaluation.
What could slow AI-assisted research?
Potential limits include the time needed to run experiments, access to hardware, difficulty transferring results from small tests to larger systems, research direction, continuity, and data quality. The source material does not establish how much each factor currently constrains progress.
Source: ThorstenMeyerAI.com
