Publications & Advancements

We study the relationship between model size, dataset size, and compute budget for language model performance. We find that performance scales as a power-law with each of these factors.

We propose a method for training a harmless AI assistant through self-improvement, without any human labels of harmfulness. The method involves both a supervised learning and a reinforcement learning phase.

This paper provides a comprehensive overview of recent advancements in generative video models, culminating in our work on the Sora model, which is capable of generating high-fidelity, minute-long videos.