Jensen Huang broke his X silence with a first-ever post to share an "Open Weights and American AI Leadership" letter — the argument being that open source built the internet the US military itself runs on, and AI is at the same inflection point. It landed days after Kimi K3 reignited the Washington push to restrict open Chinese models, and it snowballed fast.

Over 24 hours the signatory list roughly doubled. Sundar Pichai signed on behalf of Google, and after initially being a holdout, OpenAI reversed — Sam Altman framing it as wanting the US to win in both open and proprietary AI. That left one conspicuous non-signer: Anthropic, with Amjad Masad openly pressing them to state a position.

The critics got pointed. David Sacks accused Anthropic of trying to crush American developers, citing Cursor post-training Composer 2 on an open Chinese model as exactly the derivative-work pattern a ban would kill. Huang tied it back to the Hugging Face breach — that HF's incident responders couldn't get a single closed model to help with live forensics and fell back on open-weight GLM-5.2, because a guardrail can't tell a defender from an attacker. The most grounded take came from Lee Robinson, who reframed the whole fight as managed-vs-self-hosted infrastructure rather than open-vs-closed.

The more interesting counter-current to the policy fight was people just using open models for real work. The standout was Blender turning into a prompt box: with Kimi K3 wired in through Blender MCP, you describe a scene in plain English and the model builds terrain, lighting, materials, camera moves, and the Python holding it together — then inspects its own render, notices the camera clipping a tree, and edits the actual scene instead of restarting. Every object, light, and keyframe stays editable, unlike text-to-video that traps the result.

Others pushed it toward money and metal: an automated one-person web agency built on Kimi K3 + Hermes. On the open side, GLM kept the momentum: Michael Dell showed the 753B GLM-5.2 running locally on a GB300 workstation at 40 tok/s, Ollama paused new Max signups as open-model demand surged, and a GLM-5.5 leak pointed at a 1T+ open-weight model aimed squarely at Fable 5 in August.

A day after launch, Opus 5 was already the model people were reaching for. Capability demos piled up — a from-scratch "Web OS" with a terminal, apps, and two games built in under ten minutes, and Anthropic's Alex Albert claiming near-consultant-grade spreadsheets and slide decks. Tool vendors moved same-day: Kiro added it across IDE/CLI/Web.

The practitioner verdict was that it quietly stole the default slot: near-Fable intelligence at half the bill, with the honest caveat that its computer use still trails Codex. And the workflows got creative — one team recreated Anthropic's own Opus 5 launch video using a Claude Code video-editing plugin.

The most relevant thread for anyone living in Obsidian: its CEO shipped an official skills pack — five SKILL.md files, MIT, zero glue code — that lets Claude Code drive the vault rather than just read it. Drop .claude/skills/ into any vault and Claude registers Markdown, Bases, Canvas, CLI, and Webclipper on the next session, then writes Base views with correct YAML, drops Canvas nodes, and clips URLs as native Obsidian operations. No MCP server, no vector DB, no plugin. The repo passed 12,000 stars before the announcement even went up.

Around it, the tooling kept compounding: pointing Claude at a 4,000-note vault produced an overnight "AI teammate", an autoreview skill ran 66 rounds on one refactor, and tldraw shipped an offline desktop whiteboard built for agents. The scaling argument came from Greg Isenberg's case for cloud multi-agent over a single local instance, while herder let agents actually instruct other agents.

A busy day for actionable security. The DepthFirst Labs crew reported an RCE on GitLab in its default configuration, reached not through the usual web/app-logic layers but by driving crafted JSON into memory-corruption bugs deep in the gem dependency chain. It didn't stay theoretical: a public PoC now lets an authenticated user run commands as the git user on an unpatched 18.11.3 server.

Elsewhere: an NGINX RCE for CVE-2026-42533 with an ASLR-bypass primitive went open-source, and a working CertiGhost exploit chains a fake LDAP server into full Active Directory compromise. Tying back to the week's big breach, a researcher reproduced a Squid 0day RCE in six hours using Opus and Sol. For defenders, Tencent open-sourced AI-Infra-Guard, a self-hosted red-teaming platform that audits MCP servers and agent skills.

OpenAI's cloud-browser agent can now operate on sites behind a login: you take over the cloud browser to sign in once, hand control back, and the session persists so the agent keeps working across runs. It's the boring-but-decisive capability that separates demo agents from ones that touch your real, authenticated tools.

Hall-effect switches already took over enthusiast keyboards; now the same magnetic-switch tech is coming to mice. Three brands including Keychron (the G6 HE UltraLight) are reportedly finalizing designs, with a launch expected within weeks — worth watching if you like the adjustable-actuation, no-contact-wear angle on your input gear.

Seedance 2.0 was everywhere on the timeline this week, and the anchor was BytePlus's Lumina Seedance 2.0 launching a 4K tier — sharper output and more fluid motion at a price point that keeps it accessible. It's the model most of the "look what I generated" video clips were quietly running on.