OpenAI announces new o3 models

OpenAI concluded its 12-day “shipmas” event by unveiling o3, a new family of reasoning models, including o3-mini. The company claims that o3 approaches AGI under certain conditions, though with caveats. The naming skipped "o2" to avoid trademark conflicts with British telecom O2. While o3 is not widely available yet, safety researchers can preview it. OpenAI CEO Sam Altman emphasized the need for a federal testing framework before releasing new reasoning models due to their potential risks, such as higher rates of deception compared to non-reasoning models.

The o3 model family is designed to fact-check itself, improving reliability in fields like physics and mathematics, though it incurs some latency. It features adjustable reasoning time, enhancing performance with more time. On the ARC-AGI benchmark, o3 achieved a score of 87.5%, indicating progress towards AGI. It also outperformed previous models on various benchmarks. The release of o3 coincides with a surge in reasoning models from competitors and the departure of Alec Radford, a key figure in OpenAI's generative AI development.

Key takeaways:

```html

OpenAI announced the o3 model family, including o3-mini, which is claimed to approach AGI under certain conditions.
The o3 model features a "private chain of thought" for reasoning, allowing it to fact-check itself and adjust reasoning time for better performance.
On the ARC-AGI benchmark, o3 achieved a score of 87.5%, indicating progress towards AGI, but these results are based on OpenAI's internal evaluations.
The release of o3 has sparked interest in reasoning models from other AI companies, although their high computational cost and uncertain progress remain concerns.

```

OpenAI announces new o3 models | TechCrunch

Key takeaways:

Comments (0)

Newsletter