AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: When AI Takes A Hand In Improving AI on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s November 5 assessment said Anthropic was likely improving capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The work described was mainly AI assistance to researchers, not full automation; a separate Google DeepMind example shows how a targeted efficiency gain can reduce training time.

METR said on November 5 that Anthropic was likely improving its AI models’ capabilities on well-specified tasks at about twice the earlier rate, with substantial uncertainty. The independent evaluator attributed much of the apparent acceleration to AI assistance for researchers, offering a concrete but qualified example of AI systems helping people develop more capable AI.

METR based its assessment on internal information from August through October 2026. It said it “would not be surprised” by an acceleration as low as 1.3 times or as high as three times. The estimate concerns the rate at which capabilities improve on well-specified tasks; it is not a claim that every aspect of AI development has doubled in speed.

The account of METR’s findings says the research work was primarily AI assistance to human researchers. That can include helping with experiments and other parts of development, but the source does not establish that systems are independently running AI research from start to finish. Anthropic redacted much of the evidence METR considered most informative, leaving outside readers with limited access to the underlying material.

A separate example from Google DeepMind illustrates a narrower kind of gain. The company reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. The figures are company-reported and apply to that component and training process; they do not establish a general rate of improvement across AI systems.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentMETR published an assessment estimating that AI-assisted research is likely contributing to faster capability improvements at Anthropic.

How Research Gains Could Compound

The significance is not that any single reported improvement proves runaway progress. It is that AI assistance may affect the process that produces future models. If a system helps researchers test ideas, write or inspect code, or identify experimental problems, a team could investigate more work with the same resources. A stronger later model might then provide more useful assistance. That possible feedback loop is the central concern raised by the assessment.

Efficiency improvements can also accumulate, though the arithmetic depends on the gains being independent and applicable to the same work. The source gives a simplified illustration: three separate 10% savings would leave a job requiring about 73% of its original computing, allowing roughly 37% more jobs on the same compute budget. This is an illustration, not a measured outcome from Anthropic or Google DeepMind.

For companies and public institutions, the pace of AI development affects decisions about infrastructure, staffing, investment and oversight. But technical progress alone does not determine who can use these systems, who controls the computing resources, or how gains are shared. Those outcomes also depend on deployment choices and institutions.

Products Worth Considering

Canon SELPHY CP1500 Wireless Photo Printer Bundle with 108 Sheets of KP-108 Paper, 3 Ink Cartridges, and Cleaning Cloth

Canon SELPHY CP1500 Wireless Photo Printer Bundle with 108 Sheets of KP-108 Paper, 3 Ink Cartridges, and Cleaning Cloth

FlexiSpot E6 Electric Standing Desk, 55 × 28 in, Bamboo Top, White Frame

FlexiSpot E6 Electric Standing Desk, 55 × 28 in, Bamboo Top, White Frame

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Model Development

AI research includes a chain of tasks: reviewing relevant work, changing code, preparing experiments, measuring results, investigating failures and deciding whether an apparent gain holds up. AI tools can assist with parts of that chain, but a reported improvement still needs fair testing. A change that appears beneficial could instead reflect a different test, unequal computing time or an effect that disappears when the experiment is repeated with a different random seed.

That distinction helps explain the limits of the METR assessment. Its estimate concerns capability improvements on well-specified tasks and describes work that was mainly researcher assistance. It does not show that AI systems can reliably choose research priorities, validate every result or operate a laboratory without human direction.

The AlphaEvolve example provides a specific efficiency result, rather than evidence about the entire research process. A 23% faster matrix-multiplication component translated, according to Google DeepMind, into a 1% reduction in overall Gemini training time. The difference between those figures shows why a large local improvement need not produce an equally large change in a full system.

Evidence Behind the Acceleration Estimate

The available account does not disclose the underlying evidence in full. Anthropic redacted much of what METR considered most informative, so readers cannot independently inspect all of the material behind the evaluator’s judgment. METR itself described substantial uncertainty, including a range from 1.3 times to three times the earlier rate.

It also remains unclear how much of the reported acceleration resulted specifically from AI assistance, how the estimate would vary across other research tasks, and whether the gains will persist as projects become more demanding. The source identifies possible constraints: experiment time and hardware, differences between small experiments and large systems, the choice of research direction, continuity of work and data quality. It does not quantify how strongly any one factor limits progress.

The reported examples do not establish that AI systems can independently direct research or that progress toward superintelligence is inevitable. The source author says the possibility of rapid progress should be taken seriously; that is an interpretation of the evidence, not a finding that METR’s estimate proves such an outcome.

What Researchers Must Test Next

The next test is whether AI-assisted contributions continue to produce gains when experiments are repeated, results are checked independently and methods are applied to larger systems. Researchers and evaluators will also need clearer evidence about how much supervision these tools require and which parts of the research process they can handle reliably.

Further reporting or evaluation could clarify the information withheld from public view and whether the estimated acceleration holds across a longer period. Until then, the assessment is best read as a qualified estimate of a recent trend, not a precise forecast. The pace may change as technical limits, access to computing and human decisions shape how these systems are used.

Key Questions

What did METR report about Anthropic?

METR estimated that Anthropic was likely improving capabilities on well-specified tasks at about twice its earlier rate. It stressed substantial uncertainty and said the observed work was mainly AI assistance to researchers.

Does the assessment show that AI is conducting research autonomously?

No. The source says the work METR described was primarily assistance to human researchers. It does not establish that AI systems can independently conduct the full research process.

How uncertain was METR’s estimate?

METR said it would not be surprised by an acceleration of 1.3 times or three times the earlier rate. It also had access to information that Anthropic later redacted from public view.

What did AlphaEvolve improve in Gemini?

Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall training time by 1%. These are company-reported figures for a specific component and training process.

What could slow AI-assisted research?

Possible constraints cited in the source include hardware and experiment time, difficulty transferring small-scale results to larger systems, research direction, continuity and data quality. Their individual effects are not quantified.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

Gold Dip Draws Asian Buyers As North American Funds Lose Cash

BMO reports higher Indian bullion premiums and stronger Shanghai gold-contract trading as global ETF outflows, led by North America, weigh on gold.

What AI Looks Like Through The Eyes Of Religious Scholars At Anthropic

A New York Times headline reports a meeting between religious scholars and Anthropic, but the available details do not explain what was discussed.

AnySpeechRecognizer And AnyDecisionModel Bring Speech And Decision AI To Apple

Hugging Face announced two Swift packages for selecting speech-recognition and decision models, including options that can run on Apple devices.

What The Basic Income Pilots Say About A Future Powered By AI

Larry Ellison proposed taxing robot-added value to fund universal high income. South Korea’s rural pilot offers a current test, with results still developing.