Efficient Computation Inc.
We believe understanding how capabilities form during pretraining and building principled reinforcement learning environments are the path to solving science with AI.
Our beliefs
- We need to understand when capabilities form during pretraining and how data affects them. Understanding this will lead to sample-efficient capability learning.
- Reinforcement learning is the clear pathway to superintelligence. It is essential to find scalable ways to make high-quality RL environments that models believe are real and that have non-gameable rewards.
Products / active research
- Causal biology RL environments where models predict genetic and drug effects graded on the viable grounding of their conclusions
- Formally verifying security exploits: attempting to create formally verifiable grounded exploit systems to help models prove vulnerabilities in software
- Computed-state use: do models know things they fail to use?
- Timing windows: does when a model learns something decide whether it can retrain it?