OpenAI previews more advanced reasoning model
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Sarah Grillo/Axios
OpenAI on Friday detailed o3, the next in its foundation models that focus on the ability to do multi-step reasoning.
Why it matters: OpenAI and others are trying to keep the pace of progress while grappling with the fact that they have already ingested nearly all the readily available content from the Internet.
Driving the news: OpenAI shared details about the new model, which is still in testing.
- The ChatGPT-maker has started limited outside testing, known as "red teaming," in which outside security experts poke for flaws, vulnerabilities and bias.
- There's also a smaller version, known as o3-mini.
Yes, but: It's not ready for broader use, though the company is allowing researchers to apply for access to take part in safety testing.
What they're saying: Insiders and outsiders alike immediately praised the benchmark scores touted by OpenAI as evidence that large generative AI models have not hit a wall.
- OpenAI Product chief Kevin Weil wrote on LinkedIn, "o3 is bonkers good, a massive step up from o1 on every one of our hardest benchmarks."
- Weil also noted that o3 is the successor to the o1 model that debuted this year. (02 is the name of a British telecom firm, so the company may not want to have added another to its list of potential intellectual property disputes.)
- "OpenAI just announced o3, their new reasoning model that appears to perform insanely well across benchmarks," Box CEO Aaron Levie posted. "There are simply no signs of a slow down in AI right now."
