To develop Deep Research, we created reinforcement learning tasks to teach the model browsing and data analysis capabilities, and provided it access to tools for training.

AI models are not just about knowledge and synthesis; they can perform tasks like scrolling and clicking in browsers.

We trained a version of GPT-3 specifically for web browsing and data analysis to enhance the capabilities of Deep Research.

The project highlighted the importance of adapting research techniques to ensure practical usability in real-world applications.

The model that powers Deep Research is a fine-tuned version of the O3 model, specifically trained on challenging browsing and reasoning tasks, enhancing its online search and analysis capabilities.

Achieving sustainable scale in AI applications requires aligning models with real-world use cases and optimizing user behavior throughout the product design process.

With large models, you can't predict how users will interact with them, so it's crucial to observe their usage to understand successes and failures.

Clio allows us to analyze real-world usage of our models, helping us design more thoughtful evaluations that match real-world use cases.

The most lucrative application of supervised learning I've worked on is online advertising, where we can predict which ads users are likely to click on.