|
$ ai-news --date 2026-06-25 | mail
AI News — June 25, 2026
// the most viral AI topics of the last 24 hours, curated for your stack
$ cat top-story.md
John Jumper, who shared the 2024 Nobel Prize in Chemistry for co-developing AlphaFold — the AI that predicted over 200 million protein structures — has announced he is leaving Google DeepMind after nearly nine years to join Anthropic. His departure follows Noam Shazeer's recent move from DeepMind to OpenAI, highlighting accelerating talent flight from Google. Anthropic has been building wet-lab partnerships and AI-for-science capabilities, and Jumper is expected to accelerate that research direction.
Why it matters: Jumper brings frontier biology and scientific ML expertise to Anthropic at a pivotal moment; combined with the Shazeer/OpenAI hire, Google DeepMind is losing two of its most visible research leads within weeks — a signal of how fiercely frontier labs are competing for elite talent.
// source: TechCrunch · CNBC
## 📦 Models & Releases
Nearly a month after Google I/O, Gemini 3.5 Pro has moved from announced-but-absent to limited enterprise preview on Vertex AI. It targets a 2-million-token context window (the largest of any announced frontier model) and a “Deep Think” reasoning mode. Prediction markets put general availability at ~50–55% odds before June 30. Why it matters: A 2M-token context enables full-codebase RAG and long-document agent pipelines that are simply infeasible today — if it ships at competitive pricing it directly changes context-window tradeoffs in self-hosted inference setups. // Google Blog / AI Weekly
## 🛠️ Tools & Repos
Google Labs shipped v0.3.0 of DESIGN.md, an open format specification that gives AI coding agents a persistent, structured understanding of a design system. A DESIGN.md file combines YAML front-matter design tokens (colors, typography, spacing) with human-readable rationale; the CLI lints, diffs, and exports to Tailwind / W3C tokens. (17.5K stars, alpha) Why it matters: If you prompt Claude Code or any coding agent on frontend work, this solves the recurring problem of agents silently overriding your design system between sessions. // GitHub
Hermes Agent by Nous Research is trending strongly today (+1,178 stars) with 202K total stars. It is a fully local, self-improving autonomous agent — MIT license, no telemetry, Linux/macOS/WSL2, installs automatically — with a desktop app (v0.15.2) and a Skills Hub of 90,000+ community-contributed skills. NVIDIA selected it as the reference runtime for its 550B Nemotron 3 Ultra model. Why it matters: The most-starred open-source agent runtime in active development; the only one with NVIDIA endorsement and a community ecosystem rivaling commercial platforms. // GitHub / Nous Research
Harness is a Claude Code skill (7.8K stars, trending today) that acts as a meta-skill: given a domain task, it designs a team of specialized sub-agents, writes their individual skill definitions, and wires up orchestration — reducing agent scaffolding from hours to a single prompt. Why it matters: For Claude Code power users, harness is a force-multiplier for building reusable, composable agent pipelines without writing every skill and handoff by hand. // GitHub
## 🤖 Agents & MCP
The June 22–24 releases add claude mcp login <name> / mcp logout for authenticating MCP servers from the CLI without the interactive menu (with --no-browser for SSH), a sandbox.credentials setting that blocks sandboxed commands from reading credential files or secret env vars, org-level model restrictions, and a fix for remote MCP tool calls that previously hung indefinitely (now abort after 5 min, configurable via CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT). Why it matters: The MCP auth CLI makes headless and CI/CD agent workflows fully scriptable; credential blocking closes a real attack surface when agents execute tools inside sandboxes. // Claude Code changelog
Cloudflare now lets you deploy a Workers project with npx wrangler deploy --temporary — no account required. The endpoint lives for 60 minutes; a claim link (valid ~49 hours) lets you optionally promote it to a permanent account. Simon Willison notes the value extends to any rapid prototyping scenario, not just AI agents. Why it matters: For agent workflows that need a public endpoint on the fly — webhooks, MCP tool servers, RAG API bridges — this removes the entire account-setup bottleneck and enables truly ephemeral tool infrastructure. // Simon Willison
## 🏠 Self-Hosting & RAG
Anthropic's Managed Agents now support running tool execution on your own infrastructure (or Cloudflare, Daytona, Modal, Vercel) while orchestration stays on Anthropic's side. MCP tunnels let agents reach private MCP servers inside your network via a single outbound gateway — no inbound firewall rules, no public endpoints, end-to-end encrypted. Self-hosted sandboxes are in public beta; MCP tunnels are in research preview. Why it matters: For operators running vLLM + Qdrant stacks privately, this is the architecture that lets Claude agents call your internal RAG tools and APIs without ever exposing an endpoint to the public internet. // Anthropic Blog / The New Stack
## 📰 Worth Reading
Open-source AI startup Reflection AI will pay SpaceX $150 million per month starting July 2026 through 2029 for access to the Colossus 2 datacenter in Memphis (Nvidia GB300 chips). SpaceX has now sold compute capacity to Anthropic, Google, and Cursor as well — quietly becoming one of the largest AI infrastructure platforms in the US. Why it matters: When a single compute contract exceeds most AI startup valuations, it shows how GPU infrastructure has become the real bottleneck — and how SpaceX's datacenter ambitions may matter as much for AI's future as its rockets do for space. // CNBC / Bloomberg
Ten days after Washington pulled Fable 5 and Mythos 5 from non-US nationals, China blacklisted 10 US firms, Microsoft's CEO warned publicly about model concentration politics, and Anthropic filed suit against the administration. TechPolicy.Press examines whether the Mythos block sets a permanent precedent for model-level export controls — not just chips. Why it matters: If specific model capabilities can be regulated as export-controlled technology, it changes how every frontier lab ships internationally and may accelerate open-weights adoption outside the US. // TechPolicy.Press / CNBC / Al Jazeera
$ echo "automatically researched on 2026-06-25"
> ai-news · news.reichenberg.ruhr
[ unsubscribe ]
[ view in browser ]
|