🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

1/6 We introduce RPG, a principled framework for deriving and analyzing KL-regularized policy gradient methods, unifying GRPO/k3-estimator and REINFORCE++ under this framework and discovering better RL objectives than GRPO: Paper: arxiv.org/abs/2505.17508 Code:

Mind the GAP! we've had a few works proposing techniques for enabling scaling in deep rl, such as MoEs, tokenization, & sparse training. Ghada Sokar and i looked further & found a bit more clarity into *what* enables scaling, leading us to simpler solutions (see GAP in figure)! 1/

The most exciting finding I learned recently! RL intrinsically leads to sparse updates, while SFT updates densely. We are still investigating if the updated gradients relate to “critical” params. Hope our findings help better understand RL and motivate thoughts on efficiency

High capital costs (e.g., in filmmaking) push products towards popular culture to recoup investments, while low capital costs (e.g., in painting) allow for more avant-garde and niche pursuits. If AI massively decreases production costs and other barriers to entry, I would expect

We're thrilled to announce SignGemma, our most capable model for translating sign language into spoken text. 🧏 This open model is coming to the Gemma model family later this year, opening up new possibilities for inclusive tech. Share your feedback and interest in early

Introducing MedGemma, our most capable open model for multimodal medical text and image comprehension. 🩻 MedGemma is available now as part of Health AI Developer Foundations → goo.gle/medgemma

“We believe that one of the ways we can take our civilization forward is to really understand about nature, and the best way we can do that is through science." WATCH Pushmeet Kohli discuss Google DeepMind's mission to use AI to benefit humanity: ai.princeton.edu/news/2025/watc…

Happy to share that we’re giving UK uni students access to our best models - including Gemini 2.5 Pro and NotebookLM. They’re amazing tools for research, writing, exam prep… I wish I’d had them while I was at uni :) Enjoy, and good luck with this exam season!

It was a pleasure to visit the Institute for Advanced Study (Institute for Advanced Study (IAS)) and the CS dept (Princeton Computer Science) at Princeton University (Princeton University). I greatly appreciated the interactions with passionate students and esteemed professors. Thanks Sanjeev Arora for hosting my talk.

Excited to introduce the Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. We harness the power of open-ended algorithms to search for agentic systems that get better at coding, including improving their own code. It’s the Automated Design of Agentic Systems

Self-referential self-improvement - wow!

One of the first steps in a very exciting direction!

The IAS always has great material on its Youtube channel, including this very interesting playlist titled 'Workshop on Quantum Information and Physics' I recently found out about (which came out 8 months ago). Mathematicians, physicists and computer scientists, you have to

Amidst the massive demand for Gemini 2.5 and Veo 3 models, wanted to also give a big shout out to our world-class infrastructure, chip and SRE teams, who work tirelessly to keep our wonderful TPUs from melting, and without whose incredible work none of this would be possible.

90% of doing science is being open to new ideas.

A great post as always from Chaitanya, pls go and read it 👇 It also made me reflect on a point... we've (as a field) done the equivariant approach a disservice by calling a model 'equivariant' only if it has 3D spatial symmetries -- and this is a (relatively) simple action

What an enormous privilege to give the opening lecture at the OxML summer school this morning. Never have I had such a thought-provoking set of audience questions! Here's to the automation of innovation towards human flourishing alongside the next generation of researchers.

Applications are open for Google DeepMind’s newest Academic Fellowship, focusing on realising AI’s amazing potential to tackle antimicrobial resistance. Looking forward to supporting this important research in partnership with Fleming Initiative and @ImperialCollege!

boom! 🚀

In nature, researchers from Google Quantum AI, TU München, & UoN Physics & Astronomy simulated a (2+1)D lattice gauge theory and visualized string dynamics, revealing the deconfined-to-confined excitation transition as the effective electric field increases. → goo.gle/4kwssqk