Research, engineering, and policy.
Threat research, security engineering, and policy analysis from the Cyberou team.
100M Token Context Windows
A research update on ultra-long context models, our partnership with Google Cloud, and the new funding that lets us push context windows two orders of magnitude further.
Latest posts
AGI Readiness Policy
How we evaluate, monitor, and reduce the existential risks of AI capabilities — now and as models scale.
Scaling Domain-Specific Reinforcement Learning
Notes on combining frontier-scale pre-training with domain-specific RL and inference-time compute for real software engineering.
Why We Bet on Inference-Time Compute
Our thesis: automating AI research and code generation improves models and helps solve alignment more reliably than humans alone.
Building HashHop: A New Long-Context Eval
Traditional needle-in-a-haystack tests are too easy. HashHop measures multi-hop retrieval across millions of tokens.
Inside Our GB200 Supercomputing Cluster
Thousands of GB200s, custom networking, and the kernels that keep them fed. A look at the hardware behind the models.
Hiring Researchers Who Ship
We're a small group solving a short list of fundamental problems on a direct path to AGI. Here's how we work.