Summary
Two real signals tonight. First, Karpathy has joined Anthropic, announced via @ClaudeDevs reposting his personal note about returning to R&D at "the frontier of LLMs." That is the single most consequential AI-people move of the week and reshapes Anthropic's research bench. Second, Anthropic shipped a substantial Claude Managed Agents update (cluster of 5 @ClaudeDevs posts): self-hosted sandboxes, MCP tunnels into private networks, hot-swappable tools/MCP/vaults on a live session, and automatic offloading of MCP outputs over 100k tokens to sandbox files. Beyond that, Scoble's Google I/O note flags a new Gemini Omni "world model" pitched as Holodeck-foundation, and Cursor reports Composer 2.5 is now the most-chosen model on the platform with 10x usage bumps for the day. The @brivael run (10+ posts) is French-language political commentary and gets one skip bullet.
Posts
- Karpathy joins Anthropic (@ClaudeDevs reposting @karpathy). Karpathy: "I've joined Anthropic... I think the next few years at the frontier of LLMs will be especially formative... I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time." This is a senior research hire of the highest order. Anthropic's frontier roadmap, post-training methodology, and (eventually) educational outputs all move because of this. Watch for what part of the stack he ends up touching first.
- Claude Managed Agents: self-hosted sandboxes + MCP tunnels (cluster of 5, @ClaudeDevs · launch blog · docs · cookbooks · claude-api skill). Two security-grade additions: self-hosted sandboxes (agent execution stays in your infra or a managed sandbox provider) and MCP tunnels (agent reaches services inside your security perimeter). Operationally: tools, MCP servers, and vault IDs can now be swapped on a live session without restart, and MCP tool outputs over 100k tokens auto-offload to sandbox files (relevant for KV cache accounting on long-running agents). Onboarding cookbooks shipped for Cloudflare, Daytona, Docker, Modal, and Vercel. This is the enterprise-deployment story for Claude Code architecture and friends.
- Google I/O: Gemini Omni "world model" (@Scobleizer reposting @googleaidevs). Scoble's framing: "Create anything from everything... the foundation for the Holodeck is set." No technical detail in the post itself; need to read the keynote write-up before judging whether this is a real generative-world-model claim or a multimodal-Gemini rebrand. Tracking.
- Cursor: Composer 2.5 now the most-chosen model, 10x usage bump (@mntruell reposting @cursor_ai). Cursor co-founder claims Composer 2.5 has become the default pick on the platform, with everyone getting 10x usage for the rest of the day. Pairs with this afternoon's @karankendre claim that 2.5 hits near-Opus-4.7 benchmarks at 10x lower cost; usage-share telemetry is a softer signal but consistent with the price-performance story.
- PolyAI opens its Agentic Dialog Platform (@Scobleizer · poly.ai). Voice-agent platform going GA to enterprise builders, citing FedEx, Unicredit, PG&E, Marriott, Foot Locker deployments totaling 1B+ resolved conversations. Scoble's prediction: "In 24 months every customer-service line you call will be a voice agent." Worth watching for whether the build-an-agent-in-10-minutes claim survives contact with enterprise compliance.
- AI work-intelligence platforms newsletter (@Scobleizer · unaligned.io). Scoble + Cronin newsletter on enterprise productivity-tracking AI (he names Timeglass as current daily driver). Frames the segment's real bottleneck as employee-surveillance trust rather than capability. Click through to read.
- Code with Claude London (@ClaudeDevs reposting @bcherny). Anthropic developer event photo op. Skip.
- @imagine creator Q&A with @heavypulp (@imagine). xAI's video tool promo for a cinematic-scene Q&A. Skip.
- Ginko Protocol / Aichoes (cluster of 2, @brivael reposting @fao_z_fab · aichoes.lovable.app). Lightweight portable "memory protocol" for human-AI conversation: paste a text spec into any LLM, the model emits a "Gink" snapshot you store anywhere and re-inject later. Cute design as a no-account artifact, but it is a prompt-template convention, not an architectural memory system. Nothing to integrate against.
- @brivael French political and personal feed (cluster of 10, @brivael). Long thread on entrepreneurship, Asch/Milgram conformity studies, French taxation, and personal jabs at French commentators. No AI content. Skip.