Understanding AI

Earlier this month, OpenAI released a new product called Deep Research. Based on a variant of the (still unreleased) o3 reasoning model, Deep Research can think for even longer than conventional reasoning models—up to 30 minutes for the hardest questions. And crucially, it can search the web, allowing it to gather information about topics that are too new or obscure to be well covered in its training data.

The coming AI speedup
The success of Deep Research also suggests that there’s a lot of room to improve AI models using “self play.” The big insight of o1 was that allowing a model to “think” for longer leads to better answers. OpenAI’s Deep Research demonstrates that this is true for a wide range of fields beyond math and computer programming.

And this suggests there’s a lot of room for these models to “teach themselves” to get better at a wide range of cognitive tasks. A company like OpenAI or Google can generate training data by having a model “think about” a question for a long time. Once it has the right answer, it can use the answer—and the associated thinking tokens—to train the next generation of reasoning models.