Olaf from Frozen joined Jensen Huang on stage at Nvidia GTC. Disney really knocked it out of the park with the mechanics, especially the heel-to-toe walk. I haven’t even seen the movie, but damn, the guy looked so cool (the snowman not the billionaire).

Nvidia also made a bunch of announcements, the centrepiece of which was NemoClaw: an open-source stack which, supposedly, brings security and privacy guardrails to OpenClaw agents. Jensen revealed himself to be 100% OpenClaw-pilled, declaring every company needs an OpenClaw strategy.

OpenAI acquired Astral, the company behind Python tools like Ruff and uv (that reminds me, I need to buy sunscreen). OpenAI is systematically acquiring developer tooling—first Windsurf, now Astral—to own coding workflows end-to-end.

Meta signed a $27B AI infrastructure deal with Nebius, then started planning sweeping layoffs of up to 20% of staff. The layoffs come as AI costs increase. Here we go again.

The Pentagon is developing its own AI models to replace Anthropic after they blacklisted them and made them a supply chain risk. Meanwhile, OpenAI landed a separate AWS deal to sell more classified AI to the government.

Cursor launched Composer 2 with its own frontier coding model based on Kimi K2 — which beats Claude Opus 4.6 in coding benchmarks. They launched it without proper attribution to Kimi, and got called out pretty hard on it by people feeling misled.

OpenAI is merging ChatGPT, Codex, and Atlas into a single desktop superapp with a built-in browser. The company is also cutting back on side projects like Sora and hardware to focus on coding. Interesting move, especially throwing Codex into the mix. Sounds like a lot of different nails to hit with the same hammer?

Google turned AI Studio into a full-stack app builder and began testing a Gemini Mac desktop app. They also launched a Figma competitor, a vibe design tool for UI generation, called Stitch.

Anthropic launched Cowork Dispatch, enabling phone-to-desktop AI task execution. Start a task on your phone and Claude continues it on your desktop. Having used the analogous /remote functionality for Claude Code a few times now, I can recommend testing this if you’re a Cowork user. Don’t use it to code apps at the gym though! I did that once and it’s a thief of joy. Lift weights at the gym. But if you’re on the train or something, dispatch all you want.

Stripe built Minions, an internal one-shot end-to-end coding agent for developer productivity. Have to be honest — I didn’t even read this article but try scrolling down the footer on Stripe’s dev blog; I make no guarantees you will come back the same.

OpenAI released GPT-5.4 Mini and Nano, purpose-built subagent models for fast cheap AI tasks. GPT-5.4 hit 5 trillion tokens per day within one week of launch, generating $1B in net-new annualized revenue. Mini and Nano are designed as cheap workers for the main model. It’s not just Claude Code anymore, we’re all talking to the AI manager now. And AI is doing more, in less time.

OpenAI also launched Codex subagents, allowing Codex to delegate tasks to cheaper specialized models. This mirrors GPT-5.4 Mini/Nano's design where a main model orchestrates smaller workers.

Anthropic made Claude's 1M token context window generally available at the same price with no config needed. The expanded context allows processing entire codebases or long documents in a single conversation. It’s time to throw your 600-page brand guidelines document into Claude, creating a design skill from it, and seeing if it works. There might be massive context decay though – it’s not linear, like RAM, where you can add more and it works just as well.

Xiaomi (yes, the Chinese smartphone maker) released MiMo V2 Pro, an LLM supposedly near GPT-5.2 and Opus 4.6 in performance, but at a fraction of the cost. China is nuts about OpenClaw. This will probably be a popular model to run it on.

Anthropic interviewed 81,000 users on their hopes and fears around AI. It might be the biggest most multi-lingual qualitative study ever conducted!