OpenAI Launches New Series of AI Models With 'Reasoning' Abilities

 OpenAI is launching the “Strawberry” series of artificial intelligence models designed to spend more time processing query answers to solve tough problems. The models are capable of reasoning about complex tasks and solving more challenging problems than previous scientific, coding and mathematical models.

OpenAI uses the codename Strawberry internally to refer to the project, and named the models announced Thursday o1 and o1-mini. o1 will be available in ChatGPT and its API starting Thursday, the company said.

Focused on improving the reasoning capabilities of the company's models, OpenAI's efforts to create AI models that can actually reason in general have been fruitful.

In a blog post, OpenAI said the o1 model scored 83 percent on the International Mathematical Olympiad qualifying exam, while its previous model, GPT-4o, scored only 13 percent. The model also improved its performance on competitive programming problems and exceeded human Ph.D.-level accuracy on benchmark tests of scientific problems. The models achieved these results because of a technique known as “chain of thought” reasoning, which involves breaking down complex problems into smaller logical steps. When this approach is used as a cueing technique, AI models tend to perform better on complex problems.

Now, OpenAI has automated this capability so that models can break down problems on their own without user prompting. We train these models to spend more time thinking about the problem before responding, just like people do. Through training, they learn to refine their thought process, try different strategies, and recognize their mistakes.

The models are capable of reasoning about complex tasks and solving more challenging problems than previous scientific, coding and mathematical models.

OpenAI uses the codename Strawberry internally to refer to the project, and named the models announced Thursday o1 and o1-mini. o1 will be available in ChatGPT and its API starting Thursday, the company said.

Focused on improving the reasoning capabilities of the company's models, OpenAI's efforts to create AI models that can actually reason in general have been fruitful.

In a blog post, OpenAI said the o1 model scored 83 percent on the International Mathematical Olympiad qualifying exam, while its previous model, GPT-4o, scored only 13 percent. The model also improved its performance on competitive programming problems and exceeded human Ph.D.-level accuracy on benchmark tests of scientific problems. The models achieved these results because of a technique known as “chain of thought” reasoning, which involves breaking down complex problems into smaller logical steps. When this approach is used as a cueing technique, AI models tend to perform better on complex problems.

Now, OpenAI has automated this capability so that models can break down problems on their own without user prompting. We train these models to spend more time thinking about the problem before responding, just like people do. Through training, they learn to refine their thought process, try different strategies, and recognize their mistakes.