OpenAI has unveiled a new artificial intelligence (AI) model, dubbed “o1,” also known as “Strawberry,” which the company claims can reason similarly to humans. This development marks a significant milestone in AI technology, as it brings AI closer to solving complex problems in fields such as mathematics, coding, and science.
The new model represents a key step in advancing machine cognition, with OpenAI positioning it as a powerful tool for professionals rather than general users.
New model pushes OpenAI’s capabilities
The release of the o1 model comes right in the middle of an arms race, or rather an AI race of sorts, where companies like Google DeepMind, Anthropic, and OpenAI are working round the clock to create AI systems that can operate independently as so-called “agents.”
These agents are personalised bots designed to assist users in various tasks, such as communication, creative work, and interaction with the digital world.
OpenAI’s o1 model has been integrated into its ChatGPT Plus offering, making them accessible to developers and scientists starting on Thursday. The company claims the new models significantly outperform existing ones, such as GPT-4o.
During a qualifying exam for the International Mathematics Olympiad, the o1 model scored 83 per cent, a vast improvement over GPT-4o’s 13 per cent. OpenAI’s Chief Technology Officer Mira Murati emphasised that these models offer unprecedented visibility into the AI’s thought process, allowing researchers to observe how it tackles problems step by step.
Reasoning capability like never before
The o1 model employs reinforcement learning, a method that allows it to explore different strategies when responding to queries. While the process takes longer, making it more costly compared to previous GPT models, it is also more consistent and capable of handling more sophisticated tasks. Mark Chen, the lead researcher on the project, explained that the model can identify its own mistakes and correct them, improving its responses over time.
One of the most exciting aspects of o1 is its potential to transform online search. OpenAI is testing these capabilities through its SearchGPT tool, which could lead to a “new search paradigm” that enhances research and information retrieval. By applying step-by-step reasoning and planning ahead, the model could drastically improve how users find and interpret information online.
Is this the start of Artificial General Intelligence (AGI)?
Experts in AI believe that the ability to teach models to reason and plan represents a crucial step toward artificial general intelligence (AGI), which is defined as machines exhibiting cognitive abilities similar to humans.
If AI systems could genuinely reason, it would improve the consistency of facts, arguments, and conclusions made by these models, advancing their autonomy and problem-solving skills.
Yoshua Bengio, a prominent computer scientist and Turing Award recipient, noted that models such as GPT, Google’s Gemini, and Anthropic’s Claude have already demonstrated some initial reasoning abilities.
However, the scientific consensus remains that AI systems are still far from achieving true general-purpose reasoning. Bengio highlighted the importance of independent evaluations by scientists to properly assess these advances without conflicts of interest.
Not all experts are convinced by the claims of humanlike reasoning. Gary Marcus, a cognitive science professor at New York University, cautioned that previous claims about AI reasoning have often failed to stand up to scrutiny. He advised a healthy level of scepticism when evaluating new developments in this area.
Safety concerns aplenty
As AI models grow more advanced, the risk of misuse also increases. Bengio pointed out that more capable software could be exploited by bad actors. In response, OpenAI has enhanced its safety protocols, including giving independent safety institutes in the US and UK early access to the research version of the o1 model.
The advancements seen in o1 are expected to drive AI progress over the next several years. Aidan Gomez, CEO of AI start-up Cohere and one of the original Google researchers who helped develop the transformer technology used in chatbots like ChatGPT, noted that teaching models to reason has led to dramatic improvements in their abilities.
However, he also mentioned that these models are becoming more expensive to run due to the computational demands of reasoning and planning before providing an answer.
As AI continues to evolve, models like OpenAI’s o1 could represent the next leap forward in solving complex problems and pushing the boundaries of machine intelligence.