Russell Mendonca

One of the things we strive to do with each new Gemini release is to make the new Flash model as good or better than the previous model’s Pro model. Gemini 3 Flash exceeds Gemini 2.5 Pro on nearly every metric, often by very large margins, and almost matches Gemini 3 Pro on most benchmarks.

We’ve developed a memory system for our models that provides both short-term visual memory and long-term semantic memory. Our approach allows us to train robots to perform long and complex tasks, like cleaning up a kitchen or preparing a grilled cheese sandwich from scratch 👇

Humans learn by watching. Robots should too.

We got our robots to wash pans, clean windows, make peanut butter sandwiches, and more! Fine-tuning our latest model enables all of these tasks, and this has interesting implications for robotics, Moravec's paradox, and the future of large models in embodied AI. More below!

Amazing capabilities at low latency. Personal news: I’ve joined DeepMind to push on world models! Large-scale video pre-training will unlock generalist robots. Proud of helping drive transfer from human video at Tesla. The team's on a great trajectory for humanoid products.

We discovered an emergent property of VLAs like π0/π0.5/π0.6: as we scale up pre-training, the model learns to align human videos and robot data! This gives us a simple way to leverage human videos. Once π0.5 knows how to control robots, it can naturally learn from human video.

BREAKING: Tesla Model Y Robotaxi with literally no one in the car spotted driving itself on public roads in Austin, Texas! A new age has begun 🤖

Excited to share what we've been brewing at PI! We’re working on making robots more helpful by making them faster and more reliable through real-world practice, even on delicate behaviors like carrying this very full latte cup

Introducing GEN-0, our latest 10B+ foundation model for robots ⏱️ built on Harmonic Reasoning, new architecture that can think & act seamlessly 📈 strong scaling laws: more pretraining & model size = better 🌍 unprecedented corpus of 270,000+ hrs of dexterous data Read more 👇

This is so cool. Tesla can take footage from its massive vehicle fleet and synthetically create new driving scenarios to test edge cases and improve safety for its self-driving software. Tesla can also stitch footage from all 8 cameras into a fully drivable 3D environment—letting engineers steer, brake, and navigate as if they were on real roads, all powered by neural network–generated video streams. • Can simulate 8 Tesla camera feeds simultaneously — fully synthetic. • Used for testing, training, and reinforcement learning. • Allows adversarial event injection (e.g., adding a pedestrian or vehicle cutting in). • Enables replaying past failures to verify new model improvements. • Can run in near real-time, letting testers “drive” inside a simulated world.