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

🔍 Read the full analysis: When AI Helps Design The Technology That Follows on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s November 5 assessment said Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The work described was primarily AI assistance to researchers, not full automation of AI research; Google DeepMind has also reported a targeted efficiency gain from AlphaEvolve.

AI assistance is likely helping Anthropic improve model capabilities on well-specified tasks at about twice its earlier rate, according to an assessment published November 5 by the independent evaluation organization METR. The finding, based on information from August through October, comes with high uncertainty and describes mostly AI support for researchers—not a system independently conducting the full research process.

METR said it would not be surprised if the acceleration were as low as 1.3 times or as high as three times. It judged AI-assisted research likely responsible for most of the change it assessed. Much of the evidence METR considered most informative was redacted by Anthropic, leaving outside readers with limited access to the underlying material and a greater reliance on the evaluator’s judgment.

A separate example comes from Google DeepMind, which reported that its system AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. The company said this reduced Gemini’s overall training time by 1%. The figures describe different levels of the system: a substantial improvement to one component translated into a smaller saving across the full training process.

The article uses a simplified illustration to explain how savings could accumulate: if three independent steps each reduce resource use by 10%, a job would require about 73% of its original computing resources. That is an illustration, not a reported result from Anthropic or Google. Real gains depend on whether improvements are independent, reproducible and applicable to the wider system.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentMETR published an assessment that AI assistance likely contributed to a faster rate of capability improvement at Anthropic, while Google DeepMind reported an AI-designed optimization to Gemini’s training process.

How Research Gains Could Compound

The development matters because AI can be both the product of research and a tool used during research. A model that helps researchers inspect code, plan experiments or analyze results may allow a team to test more ideas with the same time and resources. If some of those ideas improve later models, the improved systems could then assist with subsequent research.

That feedback loop could affect the pace of AI development, but the available evidence does not establish that progress will accelerate indefinitely or that systems can replace research teams. METR’s assessment concerns capabilities on specified tasks and describes predominantly assisted work. It does not by itself show fully autonomous research or predict when more advanced capabilities will arrive.

The consequences would extend beyond laboratories if faster development produces useful systems sooner. But who can access those systems, who controls the computing infrastructure and how gains are shared will depend on deployment choices and institutions, not only on technical progress.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Successors

AI research involves a chain of tasks: reviewing previous work, inspecting implementations, preparing data and code, running experiments, diagnosing failures and deciding whether a result is robust enough to use. AI tools may help with parts of that chain. A reported improvement is not automatically a real one: a test may be flawed, runs may use different amounts of computing, or a result may fail to hold under a different random seed.

That makes independent checks central to the question of whether AI is genuinely accelerating research. A useful system must do more than propose a change or produce a promising number; researchers need to test whether the result survives fair comparison and repetition. Even partial assistance can increase how many ideas a team investigates, while choosing research priorities and judging evidence remain separate challenges.

The METR assessment offers a recent, concrete evaluation of this possibility, while the AlphaEvolve example provides a specific reported optimization in a training system. Neither example establishes that every research task is being automated. Together, they illustrate different parts of the question: whether AI can help research teams work faster, and whether it can contribute improvements to the infrastructure used to train models.

“It would not be surprised by 1.3x or 3x.”

— METR, in its November 5, 2026 assessment

Limits of the Current Evidence

The public account does not provide all the evidence behind METR’s assessment: Anthropic redacted much of the information METR considered most informative. The assessment’s estimate also has a wide uncertainty range, and its described work was primarily AI assistance to researchers. The source does not establish how much of the measured change came from particular tools or tasks.

It is also unclear how well gains from small or well-specified experiments transfer to large-scale training and broader research decisions. Improvements may depend on experiment time, available hardware, data quality, continuity across projects and the ability to choose promising directions. The AlphaEvolve figures are reported by Google DeepMind; the source does not provide independent verification or further details about the method.

These examples do not settle how quickly AI research will advance, whether gains will compound consistently, or whether they will lead to fully autonomous research. The pace could be uneven, and the distribution of any benefits remains dependent on access and deployment.

What Researchers Must Verify

The next step for evaluating the claims is clearer public evidence about how AI tools are used in research and how their contributions are tested. For the METR assessment, the source material does not name a follow-up publication or a date when additional evidence will be released. Further independent evaluations could help establish whether the estimated acceleration persists and how it varies across tasks.

For specific claimed improvements, replication and fair comparisons remain necessary: teams need to determine whether a change holds across repeated runs and under comparable computing conditions. Researchers will also need to judge whether improvements in individual components produce meaningful gains in complete training systems.

For now, the confirmed development is a qualified assessment of faster capability improvement at Anthropic, alongside a company-reported efficiency gain from Google DeepMind. Whether these signals become a sustained feedback loop in which AI materially speeds the creation of its successors remains an open question.

Key Questions

What did METR find about Anthropic?

METR judged that Anthropic was likely improving capabilities on well-specified tasks at about twice the earlier rate. It attached high uncertainty to the estimate, saying it would not be surprised by a rate of 1.3 times or three times.

Did METR say AI is conducting research independently?

No. The work described in the assessment was primarily AI assistance to researchers. The source does not say the systems were independently running the full research process.

What did AlphaEvolve improve?

Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%, according to the company.

Why is the METR estimate uncertain?

METR gave a broad possible range, and Anthropic redacted much of the evidence the evaluator found most informative. The public account therefore does not let readers independently inspect all the underlying information.

Does this show that AI will rapidly improve itself without people?

No. The assessment points to likely acceleration involving AI assistance, not proven full autonomy. How far gains extend depends on testing, research choices, computing resources and whether improvements hold at larger scale.

Source: ThorstenMeyerAI.com

You May Also Like

A Small Business Guide To Putting AI To Work

An OpenAI page is titled “Helping small businesses put AI to work,” but available information does not establish what it describes or when it was published.

Run Full iOS 16.5 Firmware On Android Devices

An independent project boots Apple’s XNU kernel and iOS 16.5 user space on a POCO X5 Pro, without replacing Android or modifying phone storage.

Apple Intelligence Usage Packs: Key Details At A Glance

Apple has begun offering one-time AI usage packs for Creator Studio. The $49.99 price is reported, but Apple does not specify how much usage it buys.

PlayStation Plus Monthly Games For October: F1 25, Hunt: Showdown 1896, Earth Defense Force: World Brothers 2

F1 25, Hunt: Showdown 1896 and Earth Defense Force: World Brothers 2 join PlayStation Plus Monthly Games from October 6 to November 2.