Cong Lu

A recipe for endless self-improvement toward Embodied AGI: 🧬🤖 1️⃣ Genie 3 generates the world. 2️⃣ Gemini (Task Setter) makes a task. 3️⃣ SIMA 2 acts. 4️⃣ Gemini (Reward Model) scores the video. The agent continuously generates its own experience and learns from it! [1/N]

@AcerFur @Liam06972452 @OpenAI @SebastienBubeck @kevinweil Congratulations :))

I guess now is as good a time as any to announce that I shall be joining the AI for Science team at @OpenAI this summer. This has been in the works since January, and I thank @SebastienBubeck and @kevinweil for their personal interest in making this happen.

📍 Can LLMs discover, abstract, and reuse higher-level tool skills across tasks? Existing tool-use benchmarks test solving tasks with fixed tools. But real workflows contain recurring structures where efficiency comes from reusable tool compositions, not isolated calls. We introduce SkillCraft: 126 tasks across 6 domains designed to test whether LLM agents can acquire compositional skills, not just call atomic tools. We also propose Skill Mode, a lightweight protocol with four MCP primitives that let agents compose, verify, cache, and reuse tool chains at test time. Our Key findings across evaluating 8 SOTA models: ⚡Skill Mode enables agents to self-discover and reuse skills, leading to higher success and efficiency than agents without it. The gains are larger for stronger models. 🧠 Stronger models (e.g., Claude) discover more generalizable skills, which transfer across tasks and even across models. 🔍 Deeper composition ≠ better — shallow, well-tested skills generalize best. 🔗 Paper: arxiv.org/abs/2603.00718 💻 Code: github.com/shiqichen17/Sk… 🏠 Page: skillcraft-website.github.io/page (1/7)

Introducing Sovereign AI, the Government’s new £500m venture fund. Sovereign AI will support founders from day one to start here, scale here and win everywhere. sovereignai.gov.uk

The UK is seeding a new frontier AI lab, a home for the world’s very best researchers having the deepest impact. Come join us 🇬🇧🚀 ft.com/content/41f522…

Some thoughts on AI and mathematics, inspired by "First Proof."

Memory is probably the biggest challenge for building practical AI agents. Thrilled to share our work exploring a shift from manually defining memory for each domain → enabling agents to design better memory mechanisms for themselves. Meta-learning memory designs unlocks agents that can learn to continually learn across various tasks. Absolute joy working with @yimingxiong_ and @jeffclune on this!! code: github.com/zksha/alma paper: arxiv.org/pdf/2602.07755

Can AI agents design better memory mechanisms for themselves? Introducing Learning to Continually Learn via Meta-learning Memory Designs. A meta agent automatically designs memory mechanisms, including what info to store, how to retrieve it, and how to update it, enabling agentic systems to continually learn across diverse domains. Led by @yimingxiong_ with @shengranhu 🧵👇 1/

We researchers often say "let us build this cool new technology and there will be many ways the company/society can use it, such as <list>". Here's a great example of that happening! Fun to see Genie help Waymo!!

Super cool use case of Genie 3 simulations!

Great use of Genie 3 from @waymo to create high-fidelity, interactive simulations of rare events that are nearly impossible to capture in the real world.

Super excited to see Genie 3 having impact in autonomous driving. We've talked.... a lot... about this type of application for world models, and its been incredible to see the Waymo team do things we previously dreamed of! Congrats @maxjiang93 and team!!

Genie 3 🤝 @Waymo The Waymo World Model generates photorealistic, interactive environments to train autonomous vehicles. This helps the cars navigate rare, unpredictable events before encountering them in reality. 🧵

Alright, I want to talk more about problem 1051, since, to me, it felt like it was in a different class of LLM results so far. Firstly, I really like the problem! I think it is a very cute statement. 🧵

This is the second paper from our series “Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erdős Problems”. Our agent #aletheia powered by #DeepThink cracked 13 ‘Open’ Erdős problems: 5 novel autonomous solutions, and 8 through existing literature. It was a lot of fun as the world went crazy on Erdős problems while we patiently scanned over 700 problems :) Paper link and more details below.

Project Genie is a prototype web app powered by Genie 3, Nano Banana Pro + Gemini that lets you create your own interactive worlds. I’ve been playing around with it a bit and it’s…out of this world:) Rolling out now for US Ultra subscribers.

Google Genie 3 can one shot an 3d video game in mins.. prompt below

When Genie makes an entire world for you, and then you immediately and accidentally walk off the side of the bridge into the abyss. I did not expect Genie to go ahead and respawn me somewhere new.

Environment: “34th Street–Penn Station” Character: “Discarded pack of cigarettes” Genie 3:

Can Genie 3 run Doom? "Doom, but the walls are made entirely out of screens also running Doom, featuring the marine from Doom but also his head is a screen running Doom"

hey guys, you wanted to try Genie... now's your chance!!!