Early versions of ChatGPT were not mathematically strong because their initial reward functions were not optimized for mathematical correctness.

Periodic Labs has a team of roughly 30 people.

Periodic Labs uses its physical laboratory to create a "physically grounded reward function," where real-world experimental outcomes serve as the ground truth for training its AI models.

Periodic Labs' near-term commercial strategy is to build AI "co-pilots" for engineers and researchers in advanced industries such as space, defense, and semiconductors.

Liam argues that an AI model's capabilities are specific to its training distribution, meaning a model trained on code will not generalize to solve problems like cancer research without the relevant d...