OpenAI has disclosed that it was exploring new reasoning AI models, o3 and o3 mini, that target competing with rivals, including Google, in developing smarter models that can solve more challenging problems.
With reasoning, the o1 model, which began its work in September, takes time to consider responses to prompts from users. The new model is able to provide answers in a more structured, logical fashion of thinking.
ARC-AGI is a test that measures how well an AI model can pass a test after disregarding certain categories of knowledge that the model likely obtained during training. In other words, these o3 models are quite capable of performing some remarkably intricate mathematical and logical computations that they have never encountered before.
During the launch, OpenAI’s chief Sam Altman pointed out that o3 represents the first step into the next phase of artificial intelligence development. He averred that these models can be used to perform other tasks that entail more computations – what he described as “lots of reasoning.”
The performance of the new o3 model is superior to several benchmarks when compared to the o1 model. These are such things as coding-related skills, problem-solving scientific skills, and even mathematical skills often of a quite advanced type. It is said that the model is reported to be three times better at answering the ARC- AGI tests.
The founder of GenAI announced its plan to launch an application for outside researchers to use o3 models and apply before January 10. OpenAI made a global arms race for AI with the release of ChatGPT in November this year. The company enjoys a growing reputation and has made new products which have helped to realize a $6.6bn funding round in October this year.
To date, OpenAI is beginning with public safety testing, which is an indication of the company’s approach to market cautiously. These sound promising if we are to believe the early results and benchmark performances; o3 models could represent a new generation of AI models.
The new o3 and o3 mini models in internal safety testing by OpenAI will be much more potent than the earlier o1 models that were launched, according to the company.