OpenAI has announced its new AI model called o1: the first of a series of models that can ‘reason’. This should provide better answers to more complex questions. According to OpenAI, the new models also suffer less from hallucinations.
OpenAI has its new AI models o1 and o1-mini launched. With this, the company takes another important step towards its goal of a human-like artificial intelligence, although it is still a long way off. The new model, which is still a preview version, should perform better in more complex tasks. For example, the model is better at programming and mathematics, and it is better at solving problems that require multiple steps. In addition, o1 can also explain its own reasoning.
According to OpenAI, o1 solved 83 percent of problems correctly in a qualifying exam for the International Mathematical Olympiad, while its predecessor GPT-4o solved only 13 percent. The new model also scored higher in online programming competitions.
Fewer hallucinations
The name o1 indicates, according to OpenAI, that “the counter has been reset to 1.” Training the model is fundamentally different from its predecessors, but OpenAI CEO Jerry Tworek gives few details to The Verge. He says that the training was done with “a completely new optimization algorithm and a new training dataset that was specifically designed for it.”
According to Tworek, this new model is less prone to hallucinations, where it makes up false information out of thin air. That is a major problem with current AI language models. “We have found that this model hallucinates less, but the problem is not solved yet.”
Also for free users
ChatGPT Plus and Team users will be the first to get access to o1-preview as o1-mini today. Access for Enterprise and Edu users will follow early next week.
OpenAI also wants to give all free ChatGPT users access to o1-mini, but when that is planned, is not yet known. For o1 you have to be a paid user for the time being.
There are also disadvantages
One disadvantage is that the new model is slower to use than GPT-4o. Earlier this week it was leaked that the new model, codenamed Strawberry, often needs between 10 and 20 seconds to ‘think’. While GPT-4o starts working on an answer almost immediately.
Despite o1 hallucinating less often, the model currently still underperforms GPT-4o when it comes to factual information about the world. The model also cannot yet use data from the internet or process images and files.
Read more news about AI and don’t miss anything with our newsletter.
Source: www.bright.nl
