Top story
Anthropic reportedly pursuing development of its own drugs using AI. The company is moving deeper into scientific R&D beyond its core model business, signaling a major strategic expansion. Source
Research
WARP: Weight-Space Analysis for Recovering Training Data Portfolios. Method that recovers training data portfolios directly from model weight space. Source
AutoMem: Automated Learning of Memory as a Cognitive Skill. Framework that automates learning memory as a cognitive skill for agents. Source
DuoMem: Towards Capable On-Device Memory Agents via Dual-Space Distillation. Distills dual-space memory capabilities into compact on-device memory agents. Source
Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads. Uses logit-contribution scoring to locate non-literal retrieval heads in transformers. Source
Tools
Mistral releases Leanstral-1.5-119B-A6B. Open-source code agent model specialized for Lean 4 proof generation. Source
Particle Scattering Sampler for llama.cpp. Experimental "scatter" sampler that smooths probability mass across nearby ranks within top-k candidates. Source
Jamesob's guide to running SOTA LLMs locally. Detailed guide to multi-GPU local builds in the ~$40–55K range, with cost/quality tradeoffs. Source
LongCat 2 model weights published. Meituan-LongCat releases INT8 and FP8 quantized weights for LongCat-2.0 on Hugging Face. Source
Industry
AI-assisted vulnerability reports spike around Claude Mythos release. Maintainers report Opus dominating new CVE filings, with severity spiking in tandem. Source
Midjourney's medical scanner work raises more questions than it answers. Behind-the-scenes look leaves the ultrasound project's status unclear. Source
Google DeepMind and A24 announce research partnership. First-of-its-kind collaboration between DeepMind and the film studio. Source
Citizen Lab confirms Pegasus spyware on EU Parliament member's iPhone. Espionage incident targeting a member of the committee investigating spyware. Source
Community
Open Source AI Gap Map. Simon Willison publishes a resource mapping gaps in the open-source AI ecosystem. Source
Fable's judgement, tips from Anthropic's Claude Code team. Guidance on letting Fable/Opus exercise judgment rather than dictating behavior, plus token-saving advice. Source
"What does 'Safe AI' look like?". Discussion on whether fine-tuning resistance is a meaningful safety goal for open-weight LLMs. Source
Soatok's Informal Guide to Threat Models. Practical threat-modeling primer with discussion of post-quantum crypto strategies. Source