How do we make reward models smarter without extra training data? Meet Reward Reasoning Models (RRMs) from Microsoft— they bring deliberate chain-of-thought reasoning into reward modeling. Instead of outputting a score instantly, RRMs think first, then score. The innovations
RESEARCH
-

AI Technology Discussion: How It Works and User Trust
By
–
I just had such an amazing conversation with @demishassabis about how AI works under the hood, the future of the technology, and how we ensure people can feel comfortable using it. Can’t wait to share it with you!
-
Google Introduces Gemini 2.5 Pro Deep Think Advanced Reasoning
By
–
Yesterday at #GoogleIO, we introduced Gemini 2.5 Pro Deep Think 🧠, pushing the frontiers of AI reasoning. This enhanced reasoning mode is built to tackle drastically complex problems – like USAMO problems that stumped previous models.
— Oriol Vinyals (@OriolVinyalsML) 22 mai 2025
Super proud of the GDM team for this one! pic.twitter.com/SscaARkCEiYesterday at #GoogleIO, we introduced Gemini 2.5 Pro Deep Think , pushing the frontiers of AI reasoning. This enhanced reasoning mode is built to tackle drastically complex problems – like USAMO problems that stumped previous models. Super proud of the GDM team for this one!
-
Deep Think Model Advances Multi-Domain AI Reasoning Capabilities
By
–
Deep Think uses new techniques, allowing the model to consider multiple hypotheses before responding. It's not just about math; it also gets an impressive score on LiveCodeBench , MMMU & more! Learn more:
-
Deep Think releases with safety expert review before wider deployment
By
–
Because we're defining the frontier with Deep Think, we're taking extra care. This includes first getting input from safety experts and trusted testers via the Gemini API to gather crucial feedback before wider release.
-
Integer Quantification Innovation: Technical Complexity vs Accuracy Trade-offs
By
–
The creator reinvents integer quantification from scratch, which is pretty cool! To note, if the goal is just to get 92% accuracy then the hidden layer can be removed (92% is achievable with just logistic regression). But then the project is less technically interesting. IMO ~16
-
Kaggle Top 10 Win: Deep Learning Hidden Layers Success
By
–
Top 10! I'd like to thank mom, dad, and my first 200 hidden layers #kaggle
-

Veo 3 vs Wan 2.1: Physics Performance on Complex Video Prompts
By
–
Tried this prompt with Veo 3:
— fofr (@fofrAI) 21 mai 2025
> televised footage of a cat is doing an acrobatic dive into a swimming pool at the olympics, from a 10m high diving board, flips and spins, there is commentary (not slow motion)
Wan 2.1 still has the best physics for out-of-distribution prompts. https://t.co/SFuzWufSTW pic.twitter.com/S2F1HvtO0hTried this prompt with Veo 3: > televised footage of a cat is doing an acrobatic dive into a swimming pool at the olympics, from a 10m high diving board, flips and spins, there is commentary (not slow motion) Wan 2.1 still has the best physics for out-of-distribution prompts.
-

AI and Science Panel at Google IO 2026
By
–
Great to be part of the panel on ai+science at @Google IO
-

Shaping AI’s Impact: Early Days and Collaborative Guidance
By
–
At today's @StanfordHAI seminar, Dave Patterson and @andykonwinski discussed themes from their paper "Shaping AI's Impact on Billions of Lives." They argued that we are still in the early days of practical AI, and that collaborative efforts can guide AI's impact.
