Jason Says
OpenAI's simultaneous breakthroughs on ten open math problems is today's standout—when AI starts doing original research rather than just solving known problems, the downstream disruption to cryptography and complexity theory will be bigger than most people expect.
OpenAI announces ten advances on long-standing open problems across geometry, cryptography, and complexity theory. This marks a significant step toward AI acting as an original mathematical researcher rather than just a problem-solving tool, with implications for both academia and future model capabilities.
reverse-skill is a trending GitHub project offering an AI-powered cybersecurity skills router for Claude Code, Kiro, Cursor, and Cline. It features automatic routing, on-demand toolchain bootstrapping, and a self-evolving knowledge base—bringing reverse engineering and penetration testing workflows natively into AI coding clients.
last30days-skill lets AI agents research any topic across Reddit, X, YouTube, HN, Polymarket, and the web, ranking results by upvotes, likes, and real-money signals rather than editorial curation. A practical skill for indie developers needing fast competitive or market research.
OpenWork continues trending on GitHub as a free, open-source alternative to Claude Cowork, enabling developers to share skills, MCPs, and connected services across Claude Code, Cursor, and Codex via a single MCP install. Growing community adoption signals cross-tool skill reuse is becoming a core developer need.
DeepSeek quietly released V4-Flash 0731, earning quick praise from developers as a go-to model for lighter tasks. Early adopters are pairing it with stronger models in tiered multi-LLM agent pipelines, making it a cost-efficient default for simple workloads.
xAI is set to launch Grok 4.6 next week. Developers are most focused on whether it will finally support parallel tool use and long-running tasks—currently the main blockers for scaling Grok in complex agentic loops. Delivery on these would mark a meaningful upgrade for automation use cases.
See2Think investigates whether multimodal models genuinely rely on intermediate visual states—sketches, annotations, intermediate images—during reasoning, or merely appear to. The study finds significant gaps in how models generate and actually use visual reasoning steps, a critical warning for developers building multimodal agents and visual reasoning pipelines.
Many deployed LLM agents use directory trees of markdown files as long-term memory, but this default approach has never been rigorously tested. This paper provides the first systematic evaluation of whether agents can keep a growing memory store organized as memories accumulate, conflict, and go stale—essential reading for anyone building file-based agent memory systems.
完成 Wiz 退出后再募 20 亿美元跨三支基金，总可投资本达 35 亿美元，AI 初创是核心押注方向之一。
专为 AI 数据中心激增的能源需求而生的热能储能初创，本轮为年度最大清洁科技融资之一，将加速大规模项目部署。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free