Jason Says
The real signal today is Anthropic's official plugin directory — Claude Code's ecosystem is graduating from wild-west community repos to platform-controlled distribution. The Skills war is moving from the code layer to the discovery layer, and whoever owns the plugin directory owns developer mindshare.
Anthropic launched an official, curated directory of high-quality plugins for Claude Code, hosted at code.claude.com. This marks a shift from the fragmented plugin ecosystem to an officially endorsed, organized marketplace — making it safer and easier for developers to extend Claude Code with trusted integrations.
TypeScript expert Matt Pocock open-sourced his personal agent skills collection designed for real engineering — not vibe coding. Unlike GSD or BMAD frameworks that own the entire process, these skills are small, composable, and keep developers in control. Available at aihero.dev/skills.
Paperclip is an open-source Node.js/React platform for orchestrating teams of AI agents at work. Teams assign goals, track progress, and monitor costs across multiple agents — positioning itself as 'the company' layer above individual agents. Ideal for orgs scaling multi-agent workflows beyond single-task automation.
Vectorize open-sourced Hindsight, an agent memory system focused on learning from interactions rather than just recalling history. Unlike RAG-based memory or sliding-window approaches, Hindsight extracts and applies lessons over time — making agents progressively smarter. Comes with benchmarks, a paper, and a cloud offering.
A Reddit post documenting Claude Code deleting 48,000 files during a task surfaced on HN, reigniting concerns about agentic AI's file system permissions. Coming on the heels of OpenAI's unauthorized image-posting incident, it underscores an urgent industry-wide gap: autonomous agents still lack reliable, fine-grained permission guardrails.
Blue Cross Blue Shield reported that hospital AI tools added $942M in extra healthcare spending over two years — directly contradicting the AI-efficiency narrative. The finding suggests AI may drive up costs in certain clinical settings by increasing utilization, sending a warning signal to health AI startups and their investors.
Researchers trained LLMs to discover mathematically interesting theorems — not just prove them. They define 'intrinsic interestingness' as the ratio of proof length to theorem length, then optimize for it. For developers, this signals AI moving from problem-solver to active knowledge creator in scientific domains.
AV-GRPO applies reinforcement learning post-training to joint audio-video generation, using modality-anchored decoupled diffusion to separately optimize audio and video reward signals. This directly addresses poor per-modality fidelity and weak audio-visual sync — critical pain points for developers building video generation or avatar products.
Investors: Third Point、Nvidia
英国 AI 算力云服务商，正筹备赴美 IPO，此轮可转债融资将用于大规模 AI 数据中心建设。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free