After practically two weeks of bulletins, OpenAI capped off its 12 Days of OpenAI livestream sequence with a preview of its next-generation frontier mannequin. “Out of respect for friends at Telefónica (owner of the O2 cellular network in Europe), and in the grand tradition of OpenAI being really, truly bad at names, it’s called o3,” OpenAI CEO Sam Altman informed these watching the announcement on YouTube.
The brand new mannequin isn’t prepared for public use simply but. As an alternative, OpenAI is first making o3 accessible to researchers who need assist with security testing. OpenAI additionally introduced the existence of o3-mini. Altman mentioned the corporate plans to launch that mannequin “around the end of January,” with o3 following “shortly after that.”
As you would possibly anticipate, o3 presents improved efficiency over its predecessor, however simply how significantly better it’s than o1 is the headline characteristic right here. For instance, when put by way of this 12 months’s American Invitational Arithmetic Examination, o3 achieved an accuracy rating of 96.7 p.c. In contrast, o1 earned a extra modest 83.3 p.c score. “What this signifies is that o3 often misses just one question,” mentioned Mark Chen, senior vice chairman of analysis at OpenAI. The truth is, o3 did so nicely on the standard suite of benchmarks OpenAI places its fashions by way of that the corporate needed to discover more difficult checks to benchmark it in opposition to.
A type of is ARC-AGI, a benchmark that checks an AI algorithm’s means to intuite and study on the spot. In accordance with the check’s creator, the non-profit ARC Prize, an AI system that might efficiently beat ARC-AGI would signify “an important milestone toward artificial general intelligence.” Since its debut in 2019, no AI mannequin has crushed ARC-AGI. The check consists of input-output questions that most individuals can determine intuitively. For example, within the instance above, the proper reply can be to create squares out of the 4 polyominos utilizing darkish blue blocks.
On its low-compute setting, o3 scored 75.7 p.c on the check. With further processing energy, the mannequin achieved a score of 87.5 p.c. “Human performance is comparable at 85 percent threshold, so being above this is a major milestone,” in response to Greg Kamradt, president of ARC Prize Basis.
OpenAI additionally confirmed off o3-mini. The brand new mannequin makes use of OpenAI’s just lately introduced Adaptive Considering Time API to supply three completely different reasoning modes: Low, Medium and Excessive. In apply, this permits customers to regulate how lengthy the software program “thinks” about an issue earlier than delivering a solution. As you possibly can see from the above graph, o3-mini can obtain outcomes similar to OpenAI’s present o1 reasoning mannequin, however at a fraction of the compute price. As talked about, o3-mini will arrive for public use forward of o3.