OpenAI launches preview of o1, a new series of AI models designed to think before they respond

OpenAI has just introduced a preview of OpenAI o1, a new series of AI models that take more time to think before responding. According to the company behind ChatGPT, these models are better equipped to handle complex tasks and solve more challenging problems in science, coding, and mathematics than their predecessors.

“We’re rolling out the first of this series in ChatGPT and our API. This is a preview, and we anticipate regular updates and improvements. Along with this release, we’re sharing evaluations for the next update, which is currently in development,” OpenAI shared in a blog post.

This preview follows OpenAI’s release of GPT-4o mini, a smaller, cost-effective version of its chatbot. Described as the “most capable and cost-efficient small model available today,” GPT-4o mini is set to eventually include image, video, and audio capabilities.

We're releasing a preview of OpenAI o1—a new series of AI models designed to spend more time thinking before they respond.

These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math. https://t.co/peKzzKX1bu

— OpenAI (@OpenAI) September 12, 2024

How It Works

According to OpenAI, these new o1 models have been trained to think through problems more thoroughly before responding, similar to human reasoning. They are designed to refine their approach, experiment with different strategies, and identify their own mistakes.

In testing, the upcoming model update performed comparably to PhD students on tough benchmark tasks across physics, chemistry, and biology. It also showed strong results in math and coding. For instance, in a qualifying exam for the International Mathematics Olympiad (IMO), GPT-4o correctly solved 13% of the problems, while the reasoning model achieved 83%. The model’s coding skills were assessed through Codeforces competitions, where it reached the 89th percentile. More details are available in OpenAI’s technical research post.

At this stage, the model lacks some of the features that make ChatGPT widely useful, such as web browsing and the ability to upload files and images. However, for complex reasoning tasks, this release marks a major step forward in AI capabilities. To reflect this new phase, the series has been named OpenAI o1.

Safety Measures

In developing these models, OpenAI introduced a new safety training approach, leveraging their reasoning abilities to better align with safety guidelines. This means the model can more effectively apply safety rules in context.

One method for measuring safety involves testing how well the model adheres to its guidelines, even when users try to bypass them (a practice known as “jailbreaking”). In a challenging test, GPT-4o scored 22 out of 100, while the o1-preview model scored 84. Additional details can be found in the system card and research post.

To align with the capabilities of these new models, OpenAI has strengthened its safety protocols, internal governance, and cooperation with federal authorities. This includes rigorous testing, evaluations through its Preparedness Framework, and oversight from its Safety & Security Committee.

As part of its commitment to AI safety, OpenAI recently formalized agreements with the U.S. and U.K. AI Safety Institutes. These agreements include providing early access to a research version of this model, which serves as a foundation for research, evaluation, and testing before future models are publicly released.

Who Might Benefit

These advanced reasoning capabilities could be particularly valuable for those tackling complex challenges in fields like science, coding, and math. For example, OpenAI o1 could assist healthcare researchers in annotating cell sequencing data, help physicists create intricate mathematical formulas for quantum optics, and enable developers across various fields to build and execute multi-step workflows.