OpenAI unveiled the o3 reasoner artificial intelligence model

On the last day of the 12-day event introducing new products and capabilities of OpenAI artificial intelligence models, the company announced its biggest news. Last night, the o3 Arguer model was unveiled as a replacement for the o1, but we are still a long way from its general release.

According to TechCrunch, the o3 artificial intelligence model is actually a family of o3 and o3-mini models. The mini model will be smaller and lighter to be more efficient in certain tasks. OpenAI claims that the o3 family approaches the level of artificial general intelligence (AGI), at least under certain circumstances, although there are still many shortcomings and a long way to go before a true AGI model is reached.

AI model o3 is the successor of o1, but why is it not called o2? Probably, legal problems prevented this issue. According to a report by the Information magazine, OpenAI has jumped ship from o2 to avoid a legal battle with the British telecom operator O2. Sam Altman, CEO of OpenAI, sort of confirmed this in the company’s livestream.

OpenAI o3 artificial intelligence model preview release

The o3 and o3-mini models are not yet widely available, but security experts can sign up for preview access to the o3-mini starting today. The o3 preview will be released in the future, but OpenAI has not announced a date yet. Altman says their plan is to have the o3-mini available in late January, then the o3.

The o3 AI model is trained by a method called “reinforcement learning” to think before responding with a process that OpenAI calls a “private chain of thought.” This model can pre-plan reasoning and next steps while working; As a result, it can solve the problem by taking a set of steps.

One of the differences between o3 and o1 is the ability of the new model to adjust the reasoning time. These models can be tuned to have short, medium or long computing (or thinking) times. The more time the o3 model has to think, the better it does.

As for o3’s claim to be approaching AGI levels, the model achieved a score of 87.5% on the ARC-AGI benchmark, which measures how well an AI system can learn new skills off of its training data. Of course, this score has been obtained in the mode of long thinking. o3 performs three times better than o1 in the worst case (with short thinking time).