The creator of ChatGPT has unveiled a new method for preventing hallucinations “Process supervision” is a technique that trains AI models to reward themselves for every right decision they make along the way to an answer. This is distinct from the existing method known as “outcome supervision,” where rewards are distributed following a successful conclusion.
Despite their outstanding capabilities, AI chatbots like ChatGPT are still very unpredictable and challenging to control. They frequently veer off course and produce false information or meandering, incomprehensible statements. In response to this issue, known as AI “hallucinations,” OpenAI has now revealed that it is taking action.
Process supervision, which follows a more human-like path of reasoning, may result in AI that is easier to understand, according to experts. AGI, or intelligence that would be able to comprehend the world as well as any human, would be able to reduce hallucinations, according to OpenAI.
Multiple mathematical examples are provided in OpenAI’s blog post to show the advantages in accuracy that utilising process supervision delivers. The company adds that they will investigate its effects in other fields, but claims that it is “unknown” how well process monitoring will function outside of the realm of mathematics.
From the beginning, OpenAI has made it very clear that users should not blindly trust ChatGPT. The AI bot’s user interface displays a disclaimer that reads that ChatGPT may produce inaccurate information about people, places, or facts.