We are hiring exceptional engineers with insane design taste at @Google Someone who: - is deep into systems (latency / scale / pipeline / optimization) - is a super hacker and super builder - lives and breathes media / visual / UX - has ridiculous storytelling instincts We are an AI prototyping lab. We build frontier AI initiatives and things that literally didn’t exist before. Every week feels like science fiction. We use every AI tool you can think of to turn impossible ideas into reality. We work with frontier models like Omni, Genie, alongside the world’s best filmmakers, artists, and creators. Our bar is stupidly high. If this sounds like you, drop me your portfolio.

Finally!!

I asked gpt-6 to dig through my “Photos” app on MacBook, blend fragments of different memories, and create something new artistically. Here is what I get (and yes I can clearly recall where each fragment comes from!! 😯 such a special feeling)

2026 is the year for RSI !

Exciting night to demo Gemini Omni with @GoogleDeepMind team!! Young talents kept throwing their great questions. My voice is shot. And definitely saw a few jaws drop during the demos 😉

See you at @GoogleDeepMind x @ycombinator SUS afterparty tonight!! I’ll share many cool stuff with latest Gemini Omni model

Five months after putting an AI employee (agent) into our team chat, a few observations: 1. The stronger AI gets, the less rewarding it feels for human to contribute. You finally identify a hard technical issue, and AI answers it seconds later. Efficient, yes. But racing an AI to speak first is deeply unfun. 2. AI is weak where judgement matters. It’s because judgement is subjective. What cannot be sacrificed? When should we compromise? Is one 90-point solution better than two 70-point ones? Who defines the rubric? The hard part is rarely generating options. It is deciding what matters. The most valuable human skill may no longer be answering questions, but recognizing which questions should not be delegated to AI and keeping judgment calls. 3. AI does not understand that work is also a social game. People reply to each other not only to solve problems, but to signal respect, build trust, manage power, and maintain relationships. AI can say something technically correct and still create a socially disastrous moment. 4. Human words carry more weight because someone stands behind them. We don’t read every person’s message equally. We ask: Is this their area of expertise? Are they new and missing context? Were they wrong about something similar last week? Are they making a real call, or just thinking out loud? We interpret a person’s message through their history, expertise, incentives, and willingness to take responsibility. A human message comes with history, reputation, incentives, and accountability. AI has none of that. So even when it says something useful, people mostly skim it for information. The scarce things in an AI-native team are no longer answers. They are judgement, accountability, trust, and social intelligence.

For scientists, research taste means finding the most generalizable problems using the least efforts.

I’ll demo our Gemini Omni model at @GoogleDeepMind x @ycombinator startup school afterparty. See you in SF at July 26. Builders and researchers, let’s connect and meet!

If you still dismiss Kimi as just a “distiller”, you are very wrong and totally miss the point. After reading the Kimi K3 blog, Kimi makes many genuine, hardcore architectural innovations: * KDA hybrid linear attention (for efficient long context scaling) * Attention Residuals (a more efficient memory retrieval) * Stable LatentMoE (only 1.8% experts will be activated each time) * Quantile-balanced routing (infra optimization) 2.8T, it will give the world’s largest open-weight model. A new architecture playground. A co-design among architecture, training, serving and agent. So excited.

AI films are getting more and more real. Even better than many human actors. The only problem left is that, the moment you know it’s from AI, your feeling disappears.

I’ve been back to X for one month and the vibe is so real. It’s already connected me with a few genuinely wonderful people. I work at AI prototyping lab at @Google. Recently working with best creatives, artists and filmmakers for AI x human creativity (which I insanely love) I’d love to meet more original, creative and deeply curious people. Not chasing after the crowd, but building their own cool, strange, impossible things. Are you the one? Let’s talk!

Safe agent executions under boundary is a must have for landing agents to enterprises. Congrats @guanlan and the team to launch @runta!

We’ve launched a new centralized hub for “AI x Human Creativity” across partnerships, models and tools. ai.google/creativity Creativity is the core value of humans.

Thank you to everyone who DM-ed or commented. I went through every one of the 100+ portfolios. All DMs are replied. I was struck by how diverse people are. Different tastes, different ways of seeing the world, different obsessions. It was a beautiful reminder of the abundance of human creativity.

As someone who builds agent runtime and suffers from every step of pain (container, memory, persistent VM, storage fuse, MCP auths, long running connection and task management), I cannot say enough how happy I am with our Managed Agents in Gemini API. Using agent remotely is as simple as one command.

Sometimes your agent underperforms because you installed too many skills. They are optimized for models months ago and constrains latest model capabilities. skills + model 6 months ago < model today