Why Everyone Is Underestimating Reasoning Models

Some people in the artificial intelligence industry have been skeptical about whether reasoning models such as OpenAI’s o1 represent a paradigm shift. After all, the model hasn’t gotten much usage in ChatGPT. But these skeptics may not fully understand what reasoning models are actually good for.

The queries that ChatGPT typically handles won’t necessarily produce better answers through reasoning models, researchers tell me. But if you talk to hardcore scientists, things get interesting. 

I recently chatted with a senior researcher at Lawrence Livermore National Laboratory, who shared some of the work their team is doing with OpenAI’s reasoning model. This includes using o1-preview and o1-mini to solve questions related to nuclear fusion that would have taken their researchers hours or days to do on their own. (Read more about the lab’s use of o1 and the fast-changing dynamics of the reasoning AI race in our article this morning.)