AI may already be speeding up the research that improves it
METR’s assessment of AI-assisted research, OpenAI’s security results, elections that could delay data centers, and experiments in editing model weights.

AI-assisted research may be shortening the wait for better models. METR, an external evaluator, estimates that Anthropic’s benchmark scores have been improving at 1.9 times their historical rate, measured against the period before October 2025. METR judges that AI-assisted research likely accounts for much of the acceleration.
That would mean roughly two historical years of progress arriving in one. METR says faster growth in training compute, the computing resources used to train models, did not explain the acceleration. We think a shorter model-development cycle could mean builders need to revisit sooner which tasks still require human help.
The assessment cannot isolate how much acceleration AI caused. It was observational, relied partly on staff reports, and Anthropic redacted much of the evidence METR considered most informative.
Top Stories
OpenAI reports stronger security capabilities and a large Linux-kernel experiment. Its new GPT-6 Cyber model specializes in authorized vulnerability discovery and exploit development; access is restricted through Daybreak. OpenAI reports an ExploitGym score of 48.5%, up from GPT-5.6 Cyber’s 32.1%.
Separately, a new report on an October kernel-security run describes a pipeline that reproduced 1,724 vulnerabilities and produced 948 exploits allowing an ordinary local user to obtain root privileges. Outputs included candidate patches, followed by human review and prioritized reporting. Those are OpenAI’s counts, and candidate patches do not mean every issue has been accepted or fixed upstream.
Data-center opponents’ election wins could delay computing capacity. Reuters reports that gubernatorial results in Texas and Georgia favored campaigns supporting pauses on development. Analysts cited in the article estimate that temporary project halts could affect about 45% of planned US capacity expansion. These are forecasts of possible delays, not enacted restrictions. Our read is that election results now belong alongside technical constraints when assessing how much planned computing capacity will arrive.
Researchers report repairing a small model by editing its weights. In weight-space programming experiments, researchers tested whether language models could interpret and change the numerical weights inside neural networks. In one tool-assisted experiment, Astra edited 0.016% of Qwen3.5-0.8B’s weights, raising its reported score on the MATH-500 math evaluation from 7.9% to 48.1%.
The edit reduced repetitive reasoning loops. Repairing looping does not establish new mathematical ability. Another small model without that failure mode did not meaningfully improve.
Top Links
- FlashProve (anthropic.com): For 4-bit Llama 4 Maverick, Anthropic’s provable-inference prototype reports GPU costs roughly 70–95 times ordinary vLLM inference.
- Benchmark audits (primeintellect.ai): Prime Intellect reports that ambiguous tool interfaces, brittle graders, and publicly accessible answers can distort agent scores.
- Recursa (arxiv.org): The researchers describe agents refining stored skills, instructions, and memories from task experience, rather than demonstrating autonomous improvement of foundation-model weights.
- A patched ChatGPT sandbox flaw (research.checkpoint.com): Check Point’s November disclosure describes a Lockdown Mode bypass already patched in September.
- Reading weights on a laptop (simonwillison.net): In four informal examples, a laptop-running Qwen model identified Pac-Man, Pong, and Breakout from neural-network weights, but failed Freeway.
The week in AI, once a week.