OpenAI Unveils o1-Preview, A New AI Model Trained To Think Like A Human: Here's Who It Will Benefit The Most

OpenAI o1 Preview models employ various strategies, refine their approaches, and learn from mistakes during training, just like a normal human being ideally would.

OpenAI on Thursday introduced the o1 Preview, the first in a new series of AI models designed to tackle complex problems in science, coding, and mathematics. Unlike previous models, o1 Preview spends more time reasoning through tasks, refining its thought process before providing a response.

OpenAI o1 Preview: How To Access

Available now in ChatGPT and through OpenAI’s API, o1 Preview marks the beginning of this new series, with regular updates expected. OpenAI has also released evaluations for the next model, which is still in development.

"Both o1-preview and o1-mini can be selected manually in the model picker, and at launch, weekly rate limits will be 30 messages for o1-preview and 50 for o1-mini," OpenAI said in a blog post. "We are working to increase those rates and enable ChatGPT to automatically choose the right model for a given prompt."

What Makes o1 Preview Stand Out?

These models are trained to think more deeply, similar to human reasoning. They employ various strategies, refine their approaches, and learn from mistakes during training. Tests indicate that the upcoming update performs comparably to PhD students on challenging tasks in physics, chemistry, and biology, and excels in coding and math. For instance, on an International Mathematics Olympiad (IMO) qualifying exam, the reasoning model outperformed GPT-4o, scoring 83 per cent compared to GPT-4o’s 13 per cent. In Codeforces coding competitions, it reached the 89th percentile.

Currently, o1 Preview does not include some of ChatGPT’s existing features, like web browsing and file uploads, but it demonstrates a significant leap in handling complex reasoning tasks, making it a formidable tool for scientific and technical applications.

o1 Safety Measures

With the launch of o1 Preview, OpenAI has also introduced a new safety training approach. This method leverages the model's reasoning capabilities to better adhere to safety guidelines and minimise the risk of rule violations. In safety tests, o1 Preview significantly outperformed GPT-4o, scoring 84 on challenging jailbreak tests compared to GPT-4o’s 22.

OpenAI has strengthened its safety protocols through internal governance, federal collaboration, and rigorous evaluations, including input from the Safety & Security Committee. The company has formalised agreements with AI Safety Institutes in the U.S. and U.K., granting early access to research versions of these models to aid in safety evaluations before public release.

Who Is OpenAI o1 Preview Designed For?

OpenAI’s o1 Preview is particularly useful for tackling complex problems in fields such as science, coding, and mathematics. Potential applications include healthcare research, where it can annotate cell sequencing data; physics, where it can help generate complex formulas for quantum optics; and software development, where it can assist in building and executing multi-step workflows.

This launch represents a new level of AI capability, resetting the series name to OpenAI o1 and setting the stage for future advancements in AI reasoning and safety.