Augustus Odena

I'm an AI researcher and entrepreneur.

I co-founded Adept, where I was responsible for the research program. We built the first Computer Use Agents and the Fuyu family of multi-modal models. I am now at Amazon, where I am now working on (IMO) the most important unsolved problem in AI research.

Along with my Adept co-founder Max Nye, I discovered/invented the notion of test-time-compute for Transformers. Our Scratchpad Technique (popularly known as chain-of-thought) is the basis for all modern reasoning systems. I also hold the patent for Chain-of-Thought prompting.

Prior to Adept, I was at Google Brain for ~5 years. In addition to the Scratchpad work, I did a lot of research on Program Synthesis. I led Brain's work on Program Synthesis with Large Language Models and co-created Google Sheet's SmartFill Program Synthesizer,

Before I worked on program synthesis, I worked on generative modeling. Along with collaborators, I did a lot of the early work on scaling up Generative Adversarial Networks. That includes the first GAN to generate real ImageNet samples, the paper that made Self-Attention work in GANs, and work that diagnosed and fixed the frequency artifacts that plagued early synthesis models.

I have also worked on Semi-Supervised Learning and other, miscellaneous topics, such as Coverage-Guided Fuzzing of Neural Networks.

I came to Brain by being in the first batch of Google Brain Residents, before which I worked at a startup called Nervana Systems. Before all of that, I was a trader: I worked at Five Rings Capital and Millennium. Before I was a trader, I was a Math major at Columbia University.