On Friday, the 12th day of “12 Days of OpenAI,” OpenAI CEO Sam Altman announced that the o3 is the latest AI “inference” model based on the o1 model launched earlier this year. Announced o3-mini. The company has not released them yet, but plans to make these models accessible for public safety testing and research starting today.
These models use what OpenAI calls “private thought chains.” The model pauses to examine its internal dialogue and plans ahead before responding. This can also be called “simulated reasoning” (SR). This is a form of AI that goes beyond basic big languages. Model (LLM).
According to The Information, the company named this model family “o3” rather than “o2” to avoid potential trademark conflicts with British telecommunications provider O2. During a livestream on Friday, Altman acknowledged his company’s naming weaknesses, saying, “In the grand tradition of OpenAI being really, really bad with names, it’s going to be called o3.”
According to OpenAI, the o3 model achieved record-breaking scores on the ARC-AGI benchmark. This is a visual reasoning benchmark that has not been broken since its creation in 2019. In the low computing scenario, o3 scored 75.7%, while in the high computing scenario, o3 scored 75.7%. In testing, it reached 87.5 percent. This is comparable to human performance at an 85 percent threshold.
OpenAI also reported that o3 scored 96.7 percent on the 2024 U.S. Invitational Mathematics Examination, but only got one question wrong. The model also reached 87.7% on the GPQA Diamond, which includes graduate-level biology, physics, and chemistry questions. In the Frontier Math benchmark by EpochAI, o3 solved 25.2% of the problems, while the other models did not exceed 2%.
