OpenAI previews more advanced reasoning model

OpenAI introduced o3 on Friday, the latest foundation model focused on multi-step reasoning.

Why it matters: OpenAI and other companies are striving to maintain progress while facing the challenge of having already processed much of the easily accessible content from the internet.

Driving the news: OpenAI provided an update on the new model, which is still in testing.

The company has initiated limited external testing, or "red teaming," where outside security experts evaluate the model for flaws, vulnerabilities, and biases. A smaller version, called o3-mini, is also available.

Yes, but: While o3 isn’t yet ready for widespread use, researchers can apply for access to participate in safety testing.

What they're saying: Both insiders and external experts praised OpenAI's benchmark results, indicating that large generative AI models are far from reaching their limits.

OpenAI’s Product chief, Kevin Weil, shared on LinkedIn, "o3 is incredibly powerful, a massive leap forward from o1 across all our toughest benchmarks." He also noted that o3 is the successor to o1, which debuted earlier this year. (The company avoided calling it "o2" to prevent potential conflicts with British telecom firm O2.)

Box CEO Aaron Levie commented, "OpenAI just announced o3, their new reasoning model that performs exceptionally well across benchmarks. There’s no sign of AI slowing down right now."