You are currently viewing OpenAI rolls out o1, an AI model with reasoning capabilities

OpenAI rolls out o1, an AI model with reasoning capabilities


OpenAIhas released its much-awaited o1—the first model in the series of AI models—with reasoning capabilities.

The newly-launched AI models are built to tackle complex questions faster than humans. It can excel at reasoning through difficult tasks and solving harder problems in areas such as science, coding, and math, in comparison to earlier models, the company said in a blog post. 

However, the models are currently in preview mode and will expect ongoing updates and enhancements.

“We trained these models to spend more time thinking through problems before they respond, much like a person would. Through training, they learn to refine their thinking process, try different strategies, and recognise their mistakes,” the company said. 

It stated that o1’s reasoning abilities could prove especially useful for researchers and developers across various fields. For instance, healthcare researchers could use it to annotate cell sequencing data, physicists to generate formulas for quantum optics, and developers to build and manage multi-step workflows.

The latest model update also claims to demonstrate performance similar to PhD students on difficult benchmark tasks across physics, chemistry, and biology.

Notably, the new model shines in both math and coding. During the International Mathematics Olympiad (IMO) qualifying exam, GPT-4o solved just 13% of the problems, whereas the reasoning model scored 83%. Its coding skills were assessed in various competitions, where it placed in the 89th percentile in Codeforces rankings.

OpenAI has also introduced a “new safety training method” while developing these models and leveraging their reasoning abilities to ensure they comply with safety and alignment guidelines.

“One of the ways we assess safety is by testing how effectively the model maintains its safety protocols when a user attempts to circumvent them (a tactic called “jailbreaking”). In one of our most challenging jailbreak tests, GPT-4o scored 22 out of 100, whereas our o1-preview model achieved a score of 84,” the company stated. 

To assist developers, the company said it is also launching OpenAI o1-mini—a faster and more affordable reasoning model that’s effective in coding tasks. 

The o1-mini as a smaller model claims to be 80% cheaper than o1-preview, making it a budget-friendly option for applications that require reasoning skills but don’t need extensive general knowledge.

Starting today, ChatGPT Plus and Team users can access the o1 models. Users can choose between o1-preview and o1-mini from the model picker, with initial weekly limits of 30 messages for o1-preview and 50 for o1-mini. Soon, all ChatGPT Free users will also have access to the o1-mini model.





Source link

Leave a Reply