Weekly signal intelligence. AI research, code, policy, and the patterns that connect them.
Reward hacking in long-horizon coding agents got measured, the npm supply chain kept burning into GitHub, DeepSeek-V4-Flash crossed 2.7M downloads in a day, and agent-on-agent commerce stopped being a thought experiment.
Read →Flat vector retrieval is being dismantled by a wall of new agent-memory papers, Anthropic's Mythos model walked out the door, DeepSeek dropped V4, and the web-agent attack surface is finally getting named.
Read →Six open-weight model families landed in one week, Anthropic limited Mythos while briefing the White House, and a finetuning paper showed alignment is more reversible than advertised.
Read →Last week's LiteLLM warning landed in production, DeepSeek V4 reset the open-weights frontier, agent credential proxies became a category, and LLM-generated CVEs started breaking kernel review.
Read →DeepSeek V4 and Kimi K2.6 closed another chunk of the price-performance gap, Claude 4.6 Opus reasoning is being distilled into 35B open weights at scale, and an industrial pipeline for abliterated frontier models is now running in the open.
Read →US regulators summon bank CEOs over a frontier model's offensive capability, small models reproduce the same vulnerabilities Anthropic gated behind a researcher program, cloud coding agents converge on the same product, and Gemma 4 gets abliterated within hours of release.
Read →Anthropic gated Mythos to vetted security researchers, the White House pointed banks at it, open-source maintainers are buried under AI-found zero-days, and Gemma 4 uncensored forks hit half a million downloads.
Read →npm became the single attack surface for three separate AI ecosystem breaches in one week, Gemma 4 was abliterated within ninety minutes of release, and the Claude Code leak created an instant secondary supply chain.
Read →ML infrastructure packages are getting backdoored at credential chokepoints, OCR is being replaced by vision-language models, open-weight TTS broke the 200ms barrier, and quantization hype outran the benchmarks.
Read →Sparse MoE models flooded the data, the agentic verification gap drew a $200M bet, and Cursor's model provenance raised uncomfortable questions about what's really under the hood.
Read →Frontier model distillation went industrial, the AI coding culture war crystallized, and Nvidia telegraphed its GTC moves days early.
Read →