At OpenAI, we pursued robotics not for commercial motives, but to demonstrate the power of machine learning, which later influenced large language models.

When we started OpenAI, I was convinced that deep learning was genuinely working and that it improved with scale. This realization was crucial because it seemed like others weren't recognizing its potential.

Our initial goals included figuring out unsupervised learning and solving reinforcement learning, while keeping our team small.

The future of AI research should focus on understanding the scaling laws and capabilities of reinforcement learning models.

Human data might provide a head start for AI, but sustainable progress will depend on reinforcement learning as a continuous source of experience.

OpenAI's approach to research was more startup-like, focusing on opinions about necessary advancements rather than a centralized plan, which allowed for a balance between structured direction and exploratory freedom.

We started building Deep Research over a year ago, driven by progress in reinforcement learning and reasoning models, initially focusing on math, science, and coding tasks.

Current AI systems excel in supervised learning, which is the only paradigm that has shown remarkable success in business applications, while other paradigms like unsupervised and reinforcement learning are still developing.

The distinction between engineers and researchers at OpenAI was minimized to foster collaboration and understanding of the full stack, enhancing the quality of research and implementation.