OpenAI has unveiled its latest artificial intelligence model called o1, which the company claims can perform complex reasoning tasks more effectively than its predecessors. The release comes at a time when OpenAI is facing increasing competition in the race to develop more sophisticated AI systems.
OpenAI said on its website that O1 was trained to “spend more time on problems before reacting to them, just like a person would.” “Through training, [the models] learn to refine their thinking process, try different strategies, and recognize their mistakes.” OpenAI believes the new model will be used by healthcare researchers to annotate cell sequencing data, by physicists to create mathematical formulas, and by software developers.
Current AI systems are essentially fancy versions of autocomplete, generating responses through statistics rather than actually “thinking” through a question, meaning they’re less “intelligent” than they appear. For example, when Engadget asked ChatGPT and other AI chatbots to solve the New York Times Spelling Bee, they stumbled and produced nonsensical results.
With o1, the company claims it is “resetting the counter back to 1” with a new type of AI model designed to truly engage in complex problem-solving and logical thinking. In a blog post detailing the new model, OpenAI said it performs similarly to PhD students on challenging benchmark tasks in physics, chemistry and biology, and excels at math and coding. For example, its current flagship model, GPT-4o, only solved 13 percent of problems correctly in the qualifying exam for the International Math Olympiad, while o1 solved 83 percent of problems.
However, the new model doesn’t include capabilities like web browsing or the ability to upload files and images. And, according to The Verge, it’s significantly slower at processing prompts than GPT-4o. Despite having more time to consider its output, o1 hasn’t solved the problem of “hallucinations” — a term for AI models that make up information. “We can’t say we’ve solved hallucinations,” the company’s chief research officer Bob McGrew told The Verge. o1 is still in a nascent stage. OpenAI calls it a “preview” and is making it available starting today only to paying ChatGPT customers, with restrictions on how many questions they can ask per week. In addition, OpenAI is also launching the o1-mini, a smaller version that the company says is particularly effective for coding.
OpenAI has begun previewing a new tool called Operator that can navigate within a web browser. According to a blog post published Thursday, the software is named Computer-Using Agent by the company. “The CUA is trained to interact with graphical user interfaces (GUIs) — the buttons, menus, and text fields that people see on screens — just as humans do,” OpenAI says of the model. “This allows it to perform digital tasks without using OS- or web-specific APIs.”
The current release of Operator is based on OpenAI’s GPT-4o model. It combines that algorithm’s vision capabilities with “advanced reasoning” trained through reinforcement learning. Operator has the ability to “break down tasks into multi-step plans and adaptively improve itself when challenges arise.” According to OpenAI, this capability represents the next step in AI evolution.