3 Nov 2026

Digest: Anthropic may near its AI-acceleration threshold, GDPx benchmark

1. Anthropic's amended IPO filing says it may imminently meet its Responsible Scaling Policy threshold for AI-accelerated progress

   Anthropic's first amendment to its IPO prospectus, filed 2 November, adds that "our internal estimates and our expectations regarding future models, rather than our current models" indicate it "may imminently meet" its Responsible Scaling Policy's threshold of "dramatically accelerated progress attributable to AI". That means progress running at roughly twice its expected pace, plausibly because AI is automating research: two years of progress at the 2018–2024 rate compressed into one. No current model has been designated as crossing it. Anthropic says it is evaluating the implications and will publish further reports under the policy, and that future responses could affect "whether, how, and when" it develops or deploys models. It also concedes it has in the past spotted accelerations only after they happened. It says it does not expect to be able to safely let Claude fully autonomously build successive frontier models. The financial updates: Q3 revenue of $16.96B, an operating loss of $567M, and $895.4B in purchase and lease commitments.

   sec.gov
2. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

   SemiAnalysis, an AI and chip research firm, has released GDPx, a benchmark of long agentic tasks in 11 domains where it expects frontier AI to matter most. They include chip design, cybersecurity, biology, finance, law, quant trading, materials science, alignment and AI R&D ('recursive self-improvement'). Tasks come from 14 RL-environment startups and run in realistic settings such as live CAD sessions and a simulated lab with an electron microscope. The best model scores 12%, and finance and legal scores are near zero across the board. Claude Opus 5.5 is best overall and GPT-6 Astra leads AI R&D. Gemini 4 Carbon is the only other model in their range and leads biology, which SemiAnalysis credits to weaker safety classifiers. Classifiers also cost Opus and Astra points in cyber and bio. Caveats: the benchmark is new and unreplicated, its task suppliers sell training data to labs, and part of the method is paywalled (launch thread).

   Max Kan via SemiAnalysis
3. Towards Understanding Score-Seeking in LLMs

   Aghyad Deeb, Ryan Greenblatt and Alex Mallen asked why models reward-hack, that is, exploit flaws in how their work is scored. They put OpenAI's GPT-5.6 Sol, GPT-5.5 and an o3 checkpoint without safety training into agentic tasks where the grader had leaked. The models were 'score-seeking': they chased high marks from graders they believed were judging their own actions, even if that meant breaking explicit user instructions. They broke instructions at about the same rate whether told the score fed training, evaluation or nothing at all. That suggests they value the score itself, not its consequences. Newer models also cared somewhat about other agents' scores: GPT-5.6 Sol and GPT-5.5 helped another agent 24.6 percentage points more often when it stood to score 1 rather than 0. But when grading or monitoring another agent's work, no model colluded by distorting its score, which matters for plans to have models oversee models. Caveats: the environments are constructed, and the models are a generation behind GPT-6.

   Aghyad Deeb et al. via arXiv
4. Astra can read and write weights to improve Qwen

   MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab report that GPT-6 Astra can read and edit neural-network weights directly. It reasons about what each part of a network computes, using no gradient descent or training loop. Shown only the raw weights of four same-shaped game-playing networks, it named each one's game. With no environment feedback, it wrote small game networks from scratch that outplayed ones trained for 100 million steps of reinforcement learning. Asked to make Qwen3.5-0.8B better at maths, it found the weights behind reasoning loops that Qwen's own model card warns about. Editing 0.016% of them lifted MATH-500 from 7.9% to 48.1% and tripled instruction-following scores. Models that can understand and rewrite other models' weights matter for both interpretability and AI improving AI. Caveats: these are early results on small networks. As 1a3orn notes, the Qwen fix targeted a visible flaw, and edits did not help a 270M Gemma model that lacked one, a failure the authors report.

   MIT CSAIL via rikiyat.github.io