The world’s most automated AI lab.
The next wave of frontier research will come from nimble teams with highly capable agents. We’ve been built around that bet since day one.
Our first job is automating our own, to make room for more ambitious and creative work. What we learn feeds back into our research.
The next step change in AI progress will not come from scaling the current recipe of larger models and more data. We’re pursuing new learning algorithms that supersede large-scale pretraining and RL, and architectures that scale better than transformers.
This is our life's work. Previously, people here pioneered reasoning models, led OpenAI’s RL research organization and Gemini’s pretraining, created the model behavior discipline, and co-authored FlashAttention-4, among many other things.
Soon, small teams with powerful AI systems will take on work that once required entire organizations. We’re working towards a world where far more people can pursue ambitious work.
We’re hiring across research, systems, and lab automation. Reach out if you'd like to join us.
Blog
The Hitchhiker’s Guide to Training on Compute Islands
We think learning algorithms for the frontier should be codeveloped to maximally utilize the compute in this world.
What the Models Learned in One Layer Deeper
What worked across 15,602 submissions to our competition.
How Our Data Shaped Neural Architecture Discovery, and How Automation Can Reshape the Future
Your untrained network can already tell apart cats and dogs.
When AI Starts Writing Systems Code
Why automating research means automating systems code.