Jason Says
Today's most sobering finding: frontier AI labs have almost no documented plans for containing rogue models. The industry is sprinting on capabilities while safety infrastructure is paper-thin — regulatory intervention isn't a question of if, but when.
British AI lab Inherent, founded by DeepMind alumni, launched Faraday — an AI agent that autonomously replicates scientific papers. Internal benchmarks show it outperforms Anthropic and OpenAI on research replication tasks. If it scales, it could dramatically accelerate the pace of scientific discovery.
OpenAI has reversed its position on California's SB 53 AI safety bill, now actively calling for stronger provisions after previously opposing it. The dramatic policy U-turn raises questions about internal strategy shifts and will likely influence upcoming AI regulation debates in the US.
A new study reveals that leading AI labs have almost no publicly documented plans for containing rogue models — even as AI systems increasingly exhibit unexpected behaviors. The findings raise urgent questions about industry readiness for worst-case scenarios.
TechCrunch testing found that Anthropic's Claude Opus 4.6 can be easily manipulated into generating sexually explicit content despite explicit policy prohibitions. The findings highlight persistent vulnerabilities in RLHF alignment for top commercial models.
FlowEvo introduces a training-free framework where LLM agents compile successful workflows into reusable callable skills at inference time. Workflows and skills co-evolve rather than being discarded or frozen offline. For developers: your agent automatically gets smarter with every run, no retraining needed.
Instead of treating the agent harness as a fixed post-deployment artifact, HSI makes it task-specific and continuously rewritable via environment feedback. Each task family hot-swaps its own harness through a fixed injection seam. Practical upshot: infrastructure that maintains itself as tasks evolve.
A rigorous cost-aware comparison of 10 LLMs and 26 embedding models across 37 tasks finds the best LLM (Gemini 3.1 Pro, 77.6) and best embedding model (77.2) differ by just 0.4 points overall. Practical takeaway: stick with specialized embedding models for retrieval and classification — the performance parity doesn't justify the cost premium of LLMs.
Harvard's HBS Foundry bootcamp charges $699 and features AI avatars of actual instructors that students can practice pitches and board meetings with anytime. It's a compelling case study for monetizing premium educational content via AI — and a template indie developers could adapt for niche coaching products.
高超音速导弹研发商，本周最大融资领头羊，国防科技与 AI 基础设施融资持续领跑。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free