The AI Friday Brief
Good Wednesday, NOLA. May 20th brings the fallout from Google I/O 2026 — new Gemini models, agentic AI that actually works, and some serious talent moves. The big story: Andrej Karpathy just joined Anthropic, while Google dropped Gemini 3.5 Flash and Gemini Omni with real video understanding. Today's vibe: the frontier labs are getting serious about shipping agent capabilities that work in the real world.
Big Moves & Talent
Andrej Karpathy joins Anthropic
Karpathy announced his move to Anthropic yesterday, bringing his deep expertise in scaling AI systems and mechanistic understanding to the team. This is a major signal: one of the most respected engineers in the field is betting on Anthropic's direction. The timing — right after Anthropic's acquisition of Stainless yesterday — suggests they're building serious infrastructure for production AI.
Karpathy on X
Mistral AI acquires Emmi AI
Mistral picked up Emmi AI, a team focused on embedding models and retrieval. This signals Mistral's move toward full-stack AI — they want to own more of the inference pipeline, not just models. Consolidation is accelerating at the frontier.
Hacker News
New Models & Capabilities
Gemini 3.5 Flash and Omni launched at Google I/O
Gemini 3.5 Flash is Google's answer to speed + quality — optimized for developers who need fast inference without sacrificing reasoning. Gemini Omni handles video natively, meaning you can feed it actual video files and it understands the content directly. Both are available in preview now. This is the kind of capability shift that changes what's possible in production apps.
Google AI Blog
Forge: Guardrails push small models from 53% to 99% on agentic tasks
A Show HN that's legit cool: Forge wraps guardrails around 8B models and gets them performing like much larger ones on structured agent tasks. The kicker: it's open source. If you've been frustrated that only frontier models can reliably follow complex workflows, this changes the game.
Hacker News
Tools & Developer Moves
OpenAI adopts SynthID watermarking for AI images
OpenAI now embeds invisible watermarks in all DALL-E images using Google's SynthID tech. They're also releasing a verification tool so others can check if an image was generated by OpenAI. This is a practical step toward content provenance — useful for filtering generated content at scale.
OpenAI Blog
Remove AI Watermarks: CLI and library for stripping AI-generated markers
The arms race is real: this open-source tool removes AI watermarks from images. It's a sobering reminder that watermarking alone doesn't solve the authenticity problem. The cat-and-mouse game between detection and evasion is just getting started.
Hacker News
llm-gemini plugin updated for Gemini 3.5 Flash
Simon Willison released llm-gemini 0.32 with support for the new Gemini models. If you use his `llm` CLI tool, you can now call Gemini 3.5 Flash from the command line. Good news for anyone building with open tools.
Simon Willison
Worth a Listen
Google I/O 2026 Day 1 recap (30 minutes)
Need to catch up on Google I/O fast? Claire Vo's 30-minute recap walks through every major announcement — Gemini 3.5, Omni, the new agents capabilities, and more. Fast, conversational, and covers the practical takeaways.
How I AI
Also
- Five minutes of LLM progress (six months recap) — Simon's sharp take on what actually moved the needle in the last half-year. Worth reading before diving into all the new stuff.
- Gemini CLI shutting down June 18 — move to Antigravity CLI — If you use Google's command-line tools, heads up: Gemini CLI is being deprecated. Migration guide included.
- Id-agent: Token-efficient UUIDs for AI agents — A small but useful open-source library for agents that need to track state efficiently. Nice tool for building complex agent workflows.
- We blocked AI bot spam using Git's --author flag — A clever trick from Archestra: filtering out AI-generated contributions in GitHub using git history. Practical defense against low-quality automation.
- InsForge: Open-source Heroku for coding agents — A framework for deploying coding agents at scale. Early days, but worth watching if you're building agentic systems.
- Qwen 3.7 Preview released by Alibaba — Alibaba's latest open model. Good to keep tabs on the non-Western frontier in model development.
- Arena AI Model ELO: Real-world model strength tracking — Continuously updated leaderboard tracking how models actually perform in the wild. More useful than benchmarks for picking what to use.
- Bitcoin trader recovers $400K wallet with Claude's help — Real-world use case: Claude helped brute-force through 3.5 trillion password attempts to recover a lost wallet. Practical problem-solving at scale.
- Mozilla patched 423 bugs with AI automation last month — Significant milestone: Mozilla's showing real gains from AI-assisted bug fixing at scale. This is the kind of productivity win that actually sticks.
- Typeless: Dictate at 3x your typing speed — Voice input tool for builders. Still in quick links from Monday, but worth trying if you do a lot of typing.
- Granola: AI note-taking on phone calls — Automatically transcribe and summarize calls. Great for sales, support, or any role with frequent calls.
Pass it on