More links:

1. Claude traces the origins of NIST’s elliptic curves. A cryptographic detective story: Claude recovered timestamp-based recipes that reproduce the seeds behind widely used elliptic curves. The finding clarifies their provenance; it does not break ECDSA. Includes reproduction code and remaining mysteries.
https://www.anthropic.com/research/ecdsa-curve-origins

2. Claude designs more efficient quantum circuits. With an automated evaluator checking its proposals, Claude found circuits using fewer Toffoli gates and fewer qubits on seven of eight tasks. On the eighth, it improved the trade-off between the two. A compelling example of AI-assisted algorithm discovery.
https://www.anthropic.com/research/designing-efficient-quantum-circuits

3. AI searches for errors in mathematical papers. Researchers used GPT-6.1 Sol to examine 793 arXiv papers. Among 656 papers published before August 2026, it flagged major theorem errors in 6% and minor errors in another 30%. The authors estimate that slightly under 90% of the major findings are correct.
https://matharena.ai/literature_mistakes/

4. Four Gemini agents complete Minecraft together. Four Gemini 4 Argon agents defeated the Ender Dragon in 27 hours and 14 minutes, writing 2,783 programs and coordinating through chat and shared notes without human intervention during the run. An engaging demonstration of sustained teamwork, using Easy difficulty and a structured game interface.
https://arxiv.org/abs/2610.22319

5. Google demonstrates magic-state cultivation on quantum hardware. On its Willow processor, Google demonstrated a technique for preparing a crucial quantum-computing resource, reducing errors about fortyfold and achieving 99.99% logical fidelity for T-states. A meaningful experimental advance, though larger-scale demonstrations are still needed.
https://blog.google/innovation-and-ai/models-and-research/quantum-computing/magic-state-cultivation-quantum-computing/

6. Timothy Gowers reconsiders AI extinction risk. Gowers explains why he now takes AI extinction arguments more seriously, which objections he finds inadequate, and where he remains uncertain. An accessible personal reassessment rather than a technical risk estimate.
https://gowers.wordpress.com/2026/10/16/how-worried-should-we-be-about-extinction-risk-from-ai/

7. Can models choose rare actions without a written reasoning trace? Experiments find that Astra and Sol can approximate specified low action probabilities without reported reasoning tokens when explicitly given a random seed and extra prompt tokens. Interesting for AI monitoring, with substantial limits on broader claims about covert behavior.
https://www.lesswrong.com/posts/yEdknsLwbooQAnxX6/astra-and-sol-can-take-low-probability-actions-with-no-cot

8. Glimpse: a model built for visual reasoning. Elorian introduces a model that revisits images and video, tracks objects, and uses visual tools while reasoning. Promising demonstrations for video analysis and robotics workflows, with performance claims based on the company’s own evaluations.
https://elorian.ai/blog/introducing-glimpse