Thought Leadership
ChatGPT developer OpenAI has presented a new AI model that can solve more complex tasks than previous chatbots. The software, called o1, spends more time “thinking” before giving an answer – “just like a person would”.
The artificial intelligence tries out different approaches and recognizes and corrects its own mistakes, explains OpenAI in a blog post.
Among other things, this is effective in mathematics and software programming. For example, the o1 model solved 83 percent of the tasks in the test for the International Mathematical Olympiad. The current ChatGPT-4o only achieved 13 percent. At the same time, the new model still lacked many of ChatGPT’s useful functions. For example, it cannot search for information on the web and does not support the uploading of files and images – and is also slower so far.
From OpenAI’s point of view, the new model can help researchers with data analysis or physicists with complex mathematical formulas, for example.
Even new software sometimes invents answers
The documents also show that the new model knowingly gave the wrong answer in 0.38 percent of cases in a test selection of 100,000 queries. This mainly happened when OpenAI o1 was asked to refer to articles, websites or books. However, in many cases this was not possible without access to the Internet search. So the software itself invented plausible-looking examples. However, the software only ever wanted to fulfill the user’s wishes. The so-called “hallucinations”, in which AI software simply invents information, are generally an unsolved problem.
ChatGPT is the chatbot that triggered the hype surrounding artificial intelligence over a year ago. Such AI programs are trained with huge amounts of information and can formulate texts at the linguistic level of a human, write software code and summarize information. The principle behind this is that they estimate word for word how a sentence should continue.
dpa
