Daniel Watson

I'm a research scientist in Google DeepMind in Montreal. My work spans basic advances in the early days of diffusion models and world models (e.g., one of the very first papers showing 3D generation from a single image with no NeRF-- just diffusion models). These days I work on improving Gemini (though I can't say what exactly 🤫).

I received my undergraduate degree in NYU Abu Dhabi, where I was lucky to be taught the ropes of deep learning by Nizar Habash and Kyunghyun Cho. I then had the great fortune of being admitted to Geoff Hinton's team as part of the last cohort of Google Brain's AI Residency program.

I am very passionate about the engineering and science behind scaling deep neural networks. I ❤️ TPUs, JAX, shard_map, and writing kernels with Pallas to make TPUs burn like they should🔥🔥 I am broadly interested in solving open problems with Transformer-based LLMs such as attention bottlenecking AI chips during autoregressive inference, finding better tokenizers, and designing multimodal architectures embracing the "bitter lesson" as much as possible.

Outside of work, I love reading about pure math, playing the piano (❤️ Chopin, Ravel, Rachmaninoff), perfecting my coffee and cocktails, and enjoying long stays at the beach. I'm originally from Panama 🇵🇦🇵🇦🇵🇦

I am trying to get more into writing and plan to sporadically share fun posts about topics in deep learning (and perhaps completely different ones) in this website.