Daily Digest

The Cursor iOS app can check an agent running on a computer, reply to it or start a new task remotely. Remote control makes long-running work more practical, while also increasing the importance of authentication, visible state and reliable cancellation.

Defense Access reduces false-positive blocks for incident response and vulnerability analysis; Red Team Access permits authorized penetration testing; Specialized Access covers safety-critical systems and is reviewed with the US government. Participants accept monitoring and data-retention controls, while ransomware, physical damage and mass-disruption actions remain blocked outside the most restricted arrangements.

Eleven tasks covered agreements, approvals and reusable legal clauses, each scored against 8–50 criteria. OpenAI reports GPT-6 Astra averaged 55.0% versus GPT-5.6 Sol's 41.6%, while simulated time fell from 37.0 to 19.2 minutes; the comparison used each model's best reasoning setting and remains an internal evaluation.

Amin Vahdat explains that at 100,000-accelerator scale something can fail multiple times an hour, so recovery and delivered “goodput” dominate theoretical speed. He also says long-horizon agents sharply raise CPU, storage and networking demand, and describes why Google split TPU 8 into inference and training variants.

The site's maintainer is pausing new problem comments and proof claims, removing open/solved status displays and future ownership language, and prioritising verified formalizations and clear human exposition. Existing material stays archived; the stated goal is to stop automated priority claims from crowding out understanding and collaboration.

It represents tasks as editable concepts and recipes, then uses solver traces and judge feedback to close loopholes and demand more raw-data interpretation. Generated tasks lowered solver accuracy by 22.4 and 25.5 percentage points in biology and materials science while receiving higher quality ratings (7 HF upvotes).

A 2B or 4B multimodal search agent scores candidate reasoning patterns in parallel using a shared KV cache instead of generating a fresh chain each turn. The authors report over 90% lower per-turn reasoning latency and 28–54% lower total inference latency without losing success rate (8 HF upvotes).

Multiple agents generated hypotheses, ran first-principles and many-body calculations, and refined a model against previously unpublished optical measurements of α-RuCl3. The authors call it the first autonomous physical model for new quantum-material observations; peer review and experimental follow-up remain essential (3 HF upvotes).

Compressed learning signals support both updates and policy synchronization, while predicted-KL control limits destabilising steps. The paper reports up to 45.7% lower average memory and stable 27B-model training on one eight-GPU node where dense Adam ran out of memory (11 HF upvotes).

Instead of judging a memory program only by one overall score, it preserves specialists for different skills and integrates complementary gains using paired behavioural differences. The authors report gains of 10.54 and 7.83 percentage points on two long-memory benchmarks (6 HF upvotes).