My PhD thesis "Deep Learning with Graph-Structured Representations" is now available for download: hdl.handle.net/11245.1/1b63b… -- It covers a range of emerging topics in Deep Learning: from graph neural nets (and graph convolutions) to structure discovery (objects, relations, events)

I believe that exploring and making mistakes is key to learning and research.

Genuinely amazed by how many generalist visual capabilities one can squeeze out of this model

Gemini Omni can create action replays from different angles. I referenced a video clip with agent in Google Flow. Then asked it to give new angles that follow the original video timing, environment and movement. This test came really close to real-time consistency!

World Models ftw :)

Getting a sloth into the office for this interview was the hardest part I heard.

Super excited to see Gemini Omni finally out in the world! Having been part of this project since its inception, I've seen how its native multimodal capabilities can redefine what's possible. We're truly entering the "Nano Banana era" for video generation. Give it a try!

Gemini Omni allows me to step into an alternative timeline where Graph Convolutional Nets (GCNs) made it to the big stage 🙃 Jokes aside: excited to finally share how far we've come with multimodal reference conditioning.

... and it is *so* fast ⚡️⚡️⚡️

This week on The Information Bottleneck, we are hosting @wellingmax 🥳🥳🥳 Max is one of the most influential ML researchers of the last two decades - Professor at UvA, ex-VP at Qualcomm and MSR, co-founder of ELLIS, two Test of Time awards, and advisor to a long line of people who shaped modern generative modeling (@dpkingma, @TacoCohen , @tkipf). His own work is behind a lot of the machinery the field is still building on - VAEs, graph convolutional networks, and equivariant neural networks, to name a few. He's now CTO and co-founder of CuspAI, designing new materials for carbon capture and PFAS removal using generative models plus physics-based simulation. What would you ask him? Drop questions below.

Introducing Gemini Robotics ER 1.6, our new SOTA robotics model 🤖 which excels at visual and spacial reasoning, now available via the Gemini API!

Veo's Reference-to-Video capability is still #1 👑

We've signed an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, coming online starting in 2027, to train and serve frontier Claude models.

To build multi-player games with video models, we likely need a map. One challenge here is the action binding problem, which we solve with simple RoPE-based attention biasing. While existing multi-actor models specilize in one game, we generalize to 46 games and diverse actions!

Gemma 4 outperforms models over 10x their size! (note the x-axis is log scale!)