🔍 Read the full analysis: The Tool Helps Build Its Successor: What Could Change? on ThorstenMeyerAI.com
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
TL;DR
A November 5 METR assessment judged that Anthropic was improving model capabilities on well-specified tasks at around twice its earlier rate, with high uncertainty, and said AI assistance to researchers was likely the main factor. Google DeepMind separately reported that AlphaEvolve sped up a matrix-multiplication component used in Gemini by 23%, reducing overall training time by 1%. These examples suggest AI can help improve systems that may then assist with further research, but they do not establish fully automated AI development or a predictable pace of progress.
METR said its assessment drew on internal information from **August through October 2026**. It judged that AI-assisted research was likely responsible for most of the acceleration it observed at Anthropic. The organization described its estimate as uncertain: it said it would not be surprised by a rate of **1.3 times or three times** the earlier rate. Its conclusion concerned progress on well-specified tasks, not a general measurement of every aspect of AI research or model capability.
The public account has a significant evidence limitation. According to the source article, **Anthropic redacted much of the material METR considered most informative**, leaving readers outside the assessment with less access to the underlying evidence. METR’s estimate should therefore be treated as the evaluator’s judgment, not as a directly reproducible public measurement.
In a separate example, **Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster**, cutting the model’s overall training time by **1%**. That is a specific reported engineering improvement; it does not, by itself, show that AlphaEvolve autonomously designed a successor model. The distinction matters: AI can assist with pieces of research and engineering without controlling the full process.
How AI Could Speed Model Development
The potential significance is a feedback loop: a model helps researchers test ideas, improve code or diagnose experiments; some changes survive checking and contribute to a later model; that successor may then be more useful in research. If reliable, such assistance could let a team investigate more approaches or spend less time on routine work. It could change the pace of development even before research is fully automated.
Efficiency improvements can also compound, though illustrative arithmetic should not be mistaken for a measured industry-wide effect. The source article gives a simplified example: **three independent 10% savings** would leave a task requiring about 73% of its original computing resources. With a fixed compute budget, that would allow roughly 37% more of those tasks. Real gains depend on whether savings are independent, apply to the same work and persist when tested at larger scale.
For companies, workers and policymakers, the issue is not only how capable models become but also **who can use them, who controls the computing infrastructure and how benefits are distributed**. Those outcomes depend on deployment choices and institutions as well as technical progress. The METR assessment is a reason to monitor the research process, not proof of a fixed timetable for broad economic change.
Products Worth Considering
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Blush
Apple Mac mini Desktop Computer with M6 Chip (2026)
Dell UltraSharp U2724DE 27-Inch WQHD IPS Black Monitor
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves a sequence of tasks: reviewing prior work, inspecting implementations, preparing data and experiments, running tests, diagnosing failures and deciding whether a result is robust enough to use. Systems may help with parts of this sequence while people continue to set goals, check results and decide what to pursue.
A proposed improvement can be misleading. A code change might make a test easier, use a different amount of computing, or work only under one random seed. A useful research contribution must withstand fair comparisons and independent checks, not merely produce an encouraging result in one run. The source article argues that assessing whether systems can identify such weaknesses is part of measuring their research ability.
The examples discussed here are distinct. **METR assessed AI-assisted research acceleration at Anthropic**, while **Google DeepMind reported a particular optimization by AlphaEvolve**. Neither example establishes that a model can independently choose a research agenda, validate every result and build an entire successor. They provide evidence of assistance and efficiency improvements, with limits on what can be inferred.
““would not be surprised” by 1.3x or 3x”
— METR, as quoted in its assessment of Anthropic’s AI research acceleration
Limits in the Public Evidence
It remains unclear how much of Anthropic’s reported acceleration would persist under independently available data and repeated evaluations. METR’s assessment relied on internal information, and the source says **much of the evidence it considered most informative was redacted**. The public record therefore does not let outside readers fully inspect the basis for the estimate.
The reported rate concerns improvement on well-specified tasks. It does not establish that AI systems can independently manage open-ended research, reliably choose the most valuable problems or build future models without substantial human oversight. The extent of supervision, the reliability of contributions across different research areas and the durability of gains at larger scale are not specified here.
There are also practical constraints: experiments take time and require hardware; results from smaller tests may not transfer to larger systems; research teams must choose directions; work needs continuity; and data quality affects conclusions. The source describes the likely course as uneven. The pace and economic effects remain uncertain.
What Researchers Will Need to Test
The next useful evidence would show, with enough detail for evaluation, how much work AI systems complete, how much human supervision is required and whether proposed gains hold up under fair replication. Further assessments could clarify whether AI assistance is associated with sustained acceleration across tasks, rather than a result limited to a particular team, period or set of well-specified benchmarks.
Readers should also watch for clearer disclosure about the methods and evidence behind future evaluations, especially where confidential information limits independent scrutiny. For reported engineering gains such as AlphaEvolve’s, the relevant follow-up is whether similar improvements recur, how they translate to full training runs and whether their benefits persist as systems scale.
For now, the confirmed development is narrower than a claim of self-improving AI: an evaluator judged that AI assistance likely contributed to faster capability improvement at Anthropic, and Google DeepMind reported a concrete efficiency gain in a Gemini component. **How far those gains can compound—and how quickly—remains an open question.**
Key Questions
What did METR conclude about Anthropic’s AI research progress?
METR judged that Anthropic was likely improving capabilities on well-specified tasks at around twice its earlier rate. It emphasized high uncertainty, giving a range it would not be surprised to see from 1.3 times to three times the earlier rate.
Did METR say AI systems were conducting Anthropic’s research independently?
No. The account says METR described the work primarily as AI assistance to researchers. It does not establish that AI systems independently controlled the research process or built a successor without human involvement.
What did AlphaEvolve change for Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. That is a reported component-level optimization and does not establish full automation of model development.
Why is the METR result difficult for outsiders to verify?
The source article says Anthropic redacted much of the evidence METR considered most informative. That limits public access to the underlying material and means readers must rely more heavily on METR’s assessment.
What could limit AI-driven acceleration in research?
Experiments require time and hardware, smaller tests may not scale, and teams still need to choose research directions. Continuity, data quality, independent validation and the level of human supervision can also affect whether apparent improvements are reliable.
Source: ThorstenMeyerAI.com
