OpenAI has announced a groundbreaking advancement in artificial intelligence with the introduction of its new AI models, known as o1. These models, set to be integrated into ChatGPT Plus starting Thursday, are designed to showcase significant progress towards humanlike cognition in machines, with claims of enhanced reasoning capabilities that could revolutionize fields like mathematics, coding, and science.
The o1 models mark a critical step in the race among tech giants, including Google DeepMind and Anthropic, to develop increasingly sophisticated AI systems. This fresh wave of innovation focuses on creating AI agents—personalized bots that aim to improve productivity, creativity, and communication while interacting with the digital world.
According to OpenAI, the o1 models demonstrate remarkable performance in solving complex problems. In a qualifying exam for the International Mathematics Olympiad, the o1 models achieved an impressive 83 percent score, a significant leap from the 13 percent scored by the previous generation, GPT-4o.
Mira Murati, OpenAI’s Chief Technology Officer, highlighted the models’ ability to offer insights into their reasoning process. “We get visibility into the model’s thinking . . . we can observe its thought process, step by step,” Murati told the Financial Times.
The o1 models utilize reinforcement learning to tackle problems, which involves a more extensive analysis of queries compared to previous models. This approach, while more costly and time-consuming, yields more consistent and sophisticated responses. Lead researcher Mark Chen explained, “What it’s doing during that time is . . . exploring different strategies for answering your query. If it realizes it’s made mistakes, it can go and correct those things.”
In practical applications, such as online search, OpenAI is exploring how these models could revolutionize research and information retrieval through its SearchGPT tool. Murati suggested that the o1 models could pave the way for “a new search paradigm.”
The advancement of AI systems capable of step-by-step reasoning is seen as a pivotal milestone in the quest for artificial general intelligence (AGI)—machines with cognitive abilities akin to humans. Yoshua Bengio, a renowned computer scientist and Turing Award winner, emphasized that true reasoning capabilities would enhance “consistency of facts, arguments and conclusions made by the AI,” addressing key challenges in AGI development.
However, skepticism remains within the scientific community. Gary Marcus, a cognitive science professor at New York University, cautioned against overestimating AI’s reasoning capabilities, noting that past claims have often been debunked upon closer examination. Similarly, Bengio warned that advanced AI technologies pose increased risks of misuse, prompting OpenAI to reinforce its safety measures. The company has provided early access to a research version of the o1 models to AI safety institutes in the UK and US.
Despite these concerns, technologists are optimistic about the potential of advanced reasoning capabilities to drive AI progress. Aidan Gomez, CEO of AI start-up Cohere, highlighted the dramatic improvements in problem-solving achieved by models that incorporate advanced reasoning. “Models are becoming more expensive in terms of computation, but dramatically better at problem solving,” Gomez said.
