OpenAI has unveiled its latest artificial intelligence model, o1, which the company says can perform complex inference tasks more efficiently than its predecessors. The announcement comes as OpenAI faces an intensifying race to develop more sophisticated AI systems.
OpenAI explains on its website that O1 was trained to “increase the amount of thought time it takes before answering a question, just like humans do. Through training,[the model]refines its thought processes, tries different strategies, and learns to recognize mistakes.” OpenAI envisions the new model being used by medical researchers to annotate cell sequence data, by physicists to generate mathematical formulas, and by software developers.
Current AI systems are essentially more sophisticated versions of autocomplete, generating responses from statistics rather than actually “thinking” about the question, so they’re not as “intelligent” as they may seem. For example, when Engadget tried to get ChatGPT and other AI chatbots to solve the New York Times Spelling Bee, they didn’t do well and produced gibberish.
With o1, the company claims it is “resetting the counter to one” and developing a new kind of AI model designed to actually tackle complex problem-solving and logical thinking. In a blog post detailing the new model, OpenAI said it performed on par with PhD students on difficult benchmark tasks in physics, chemistry and biology, and outperformed in math and coding. For example, the company’s current flagship model, GPT-4o, correctly solved 83% of the problems in the International Mathematical Olympiad qualifying exams, compared to just 13% for o1.
However, the new model does not include web browsing or the ability to upload files or images. It also processes prompts much slower than GPT-4o, according to The Verge. Despite being able to spend more time considering the output, o1 has not solved the problem of “hallucinations,” a term that refers to an AI model making up information. “We haven’t solved hallucinations,” Bob McGrew, the company’s chief research officer, told The Verge.
O1 is still in its early stages: OpenAI is calling it a “preview,” and it’s only available to paid ChatGPT customers starting today, with a limit on the number of questions they can ask per week. OpenAI is also releasing the o1-mini, a slimmed-down version that the company says is especially effective for coding.
