OpenAI · 2022–2025

The road to reasoning.

A first-hand account of how o1 and o3, OpenAI’s first reasoning models, came to be: the models that taught ChatGPT to think before it answers.

  1. GPT-Zero
  2. RLSlow
  3. Berry
  4. o1 & o3

January 2022The beginning

GPT-Zero: the original bet.

In January 2022, I started OpenAI’s reasoning effort with Ilya Sutskever, under the internal name GPT-Zero. The bet was that reinforcement learning could teach language models to reason deliberately: to develop their own strategies over many tokens, and to spend more compute working through a problem before answering.

As the founding researcher and the team’s sole member for the first few months, I shaped the research agenda, built the training system, and produced the early results that gave us conviction to scale. I worked closely with Ilya and, as the effort grew, early collaborators including Łukasz Kaiser.

The aha moment

Building OpenAI o1 · September 2024

Watch the excerptOpenAI · from 03:10

The breakthrough was seeing a model develop its own reasoning through RL and outperform one trained on human-written chains of thought.

That gave us a new path to scale.

2023–2024The research grows

RLSlow to Berry.

I continued the research in RLSlow, working alongside Jakub Pachocki and a small reasoning group. We improved the methods, studied how reasoning scales, and took the experiments into frontier models.

In mid-2023, RLSlow produced further results that strengthened our conviction. I then worked with the Berry team, which scaled this research into o1.

2024–2025The releases

o1 and o3.

The research that began with GPT-Zero and grew through RLSlow and Berry reached the world as o1, OpenAI’s first reasoning model. It showed that capability improves both with more reinforcement learning in training and with more time to reason at inference.

I continued the work through o3, from research and training to release. Together, these models helped establish reasoning as a new direction for frontier AI, and the thinking step now familiar in ChatGPT began here.

These were collective achievements. I’m grateful to Ilya, Jakub, Łukasz Kaiser, Szymon Sidor, and the many research, engineering, and product colleagues who built these models.

Back to selected workNext: TBD Lab & Muse