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

🔍 Read the full analysis: How AI Might Help Design Its Next Generation on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 said Anthropic is likely improving AI capabilities on well-specified tasks about twice as fast as before, with high uncertainty, and attributed most of the current acceleration to AI assistance for researchers. Google DeepMind separately reported that AlphaEvolve sped up a matrix-multiplication component used in Gemini, cutting overall training time by 1%. The evidence points to a possible feedback loop, but its pace and limits remain uncertain.

AI systems are beginning to contribute to work that can improve later AI models, according to an independent METR assessment of research at Anthropic and a separate Google DeepMind report on training efficiency. METR judged that Anthropic was likely improving capabilities on well-specified tasks at about twice its earlier rate, but said the estimate carried high uncertainty and described the work mainly as AI assistance to researchers—not autonomous AI research.

METR published its assessment on November 5, 2026, drawing on internal information from August through October. The evaluator said it “would not be surprised” by an acceleration as low as 1.3 times or as high as three times. That range signals substantial uncertainty around the central estimate. METR judged AI-assisted research likely responsible for most of the acceleration it assessed, while the described work was primarily assistance to human researchers.

The public cannot inspect much of the underlying evidence: according to the source article, Anthropic redacted much of the material METR considered most informative. The assessment is therefore a significant outside evaluation, but readers cannot independently review all the evidence behind its conclusion.

A separate example comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. The figures describe different levels of the process: a sizeable improvement to one component, and a smaller reported effect on the full training run.

The source article uses a simplified illustration to explain how improvements could accumulate: three independent 10% reductions in compute would leave a task requiring about 73% of its original compute. Under that assumption, the same budget could run roughly 37% more jobs. This is an illustration, not a measured outcome from either report.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentA November 5 METR assessment and a Google DeepMind optimization report offer concrete examples of AI systems contributing to research and model training.

How Research Gains Could Compound

The potential importance is a feedback loop: a model helps researchers find or test improvements, and those improvements contribute to a later model that may assist with more research. Even partial support could matter if researchers can investigate more ideas, automate routine coding or debugging, and spend less time diagnosing failed experiments. That could increase the amount of work a research team completes without requiring every step to be fully automated.

Efficiency gains may also affect how much progress a fixed computing budget can support. The AlphaEvolve example shows why a local optimization can matter, while its reported 1% reduction in overall training time also cautions against treating a component-level gain as a comparable improvement to the entire system. Whether such gains recur, combine, or translate into faster capability growth is not established by these examples alone.

For businesses and workers, the issue is not just what current models can do, but whether the pace of improvement changes. Faster development could affect planning, investment and the distribution of work. The source article also stresses that who can use these systems, who controls the infrastructure and how benefits are shared will depend on institutions and deployment choices, not only technical progress.

Products Worth Considering

LG UltraGear 32GX870A-B 32-inch 4K UHD OLED Dual-Mode Gaming Monitor

LG UltraGear 32GX870A-B 32-inch 4K UHD OLED Dual-Mode Gaming Monitor

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

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

Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Blush

Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Blush

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Successors

AI research involves a chain of tasks: reading prior work, inspecting code, preparing data, running experiments, investigating failures and deciding whether results are reliable. Systems can help with pieces of that process without independently choosing research goals or validating every result. The distinction matters because assistance is not the same as autonomy.

A reported improvement can be misleading if an experiment is unfairly compared, uses different computing resources, or produces a gain that disappears under another random seed. A useful research contribution must survive testing and independent checks. In the source article’s example, an agent might report a 5% improvement after changing training code, but that result could reflect an easier test rather than a genuine advance.

The METR assessment is presented as evidence that AI assistance may already be affecting research productivity at Anthropic. The AlphaEvolve report offers a separate example of AI-generated optimization in a component of Gemini’s training process. Neither example, as described in the source material, establishes that AI systems can independently conduct the full research cycle or reliably produce their own successors.

““would not be surprised” by an acceleration of 1.3x or 3x”

— METR

Limits in the Public Evidence

The exact scale and cause of the reported acceleration remain uncertain. METR offered a broad range around its estimate, and much of the evidence it found most informative was redacted by Anthropic, according to the source article. The public material described here does not provide enough detail to independently assess every underlying measure or how the estimate was calculated.

It is also unclear how far research assistance can extend. The source identifies possible constraints including experiment time and hardware, differences between small experiments and large systems, selecting promising research directions, maintaining continuity across work, and data quality. A system that helps with routine tasks may not reliably judge whether an apparent result is real or decide what problem to tackle next.

The examples do not establish a timeline for fully automated research, a forecast for superintelligence, or a guaranteed rate of future progress. The source article’s claim that rapid progress deserves serious attention is an interpretation of the evidence, not a confirmed prediction. How improvements affect employment and who receives their benefits also remains dependent on future technical and institutional choices.

Testing the Research Feedback Loop

The next evidence to watch is whether independent evaluations can measure AI research assistance over time, with clearer public methods and enough detail for outside scrutiny. Further reporting on the METR assessment may clarify how researchers define acceleration, which tasks were included and how AI contributions were separated from other changes.

For individual claims of research progress, the key test is replication: whether an improvement survives fair comparisons, repeated runs and independent review. More examples like AlphaEvolve’s reported optimization may show whether gains are isolated or repeat across different parts of model development. The available source material does not identify a specific next evaluation date or a confirmed milestone for autonomous AI research.

For now, the supported conclusion is narrower: AI tools are reported to be assisting some research tasks, and one system was reported to improve a component used in Gemini’s training. Whether those contributions develop into a sustained, self-reinforcing cycle—and how quickly—is still an open question.

Key Questions

What did METR find about AI research at Anthropic?

METR judged that Anthropic was likely improving capabilities on well-specified tasks at about twice its earlier rate. It reported high uncertainty, saying it would not be surprised by an estimate of 1.3x or 3x, and said the work it described was primarily AI assistance to researchers.

Does the assessment show that AI is independently designing successor models?

No. The source material describes AI assistance to human researchers. It does not establish that systems can independently choose research directions, validate results and complete the full process of building a successor.

What did AlphaEvolve reportedly improve?

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

Why could small efficiency gains matter?

If gains are real and repeat across separate parts of a process, they may reduce the computing needed per training run or allow a fixed budget to support more work. The source article’s example of three 10% savings is illustrative, not a reported result from the cited systems.

What is still uncertain about AI-assisted research?

The size and durability of the reported acceleration, how much comes specifically from AI assistance, and how reliably systems can evaluate their own experimental results are unclear. Anthropic redacted much of the evidence METR considered most informative, limiting public review.

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

Search as Code: Perplexity Is Right About the Future — Just Not First to It

Perplexity introduces Search as Code, enabling AI models to assemble custom retrieval pipelines, claiming significant efficiency gains and accuracy improvements.

The Co-Founder’s Black Hole — A Structural Read on Jack Clark’s Automated AI R&D Essay

Jack Clark predicts a >60% chance of fully autonomous AI research by 2028, raising concerns about institutional readiness amid converging technical signals.

The Defender’s Window Is Closing Faster Than Anyone Is Counting

April 2026 saw rapid advancements in AI cybersecurity, with defenders making progress but offensive capabilities advancing even faster, raising urgent concerns.

A Changing Farm: Horses, Tractors, And Millions Of Animals

A historical account traces the decline of farm horses, the rise of tractors and the unresolved limits of comparing automation’s impact on animals and workers.