The American company OpenAI revealed the technical reasons which led to a complete derailment of ChatGPT on the day of February 20, 2024.
An artificial intelligence described as “crazy”, suffering a stroke or showing signs of schizophrenia. If you had visited the Reddit forum dedicated to ChatGPT on February 20, 2024, you could not have missed the numerous threads questioning the mental health of OpenAI’s chatbot.
Indeed, dozens of Internet users have testified to the behavior of the conversational agent, which is, to say the least, strange. For a few hours, the AI talked nonsense. She didn’t just formulate false answers: she constructed nonsense sentences. She also mixed languages, repeated the same words or posted lots of emojis.
Optimization gone wrong
In the hours that followed, OpenAI, the company behind ChatGPT, had not yet spoken to explain the reasons for this departure. It is now done. An explanation appeared on the company’s dashboard, which allows it to indicate to the public the operating status of its various tools.
We learn that the failure of ChatGPT comes from a “ user experience optimization ” went wrong. This improvement disrupted the way OpenAI’s model processes language. This modification took place on February 20, the day when Internet users en masse reported the excesses of the conversational agent.
“ Large language models (LLMs) generate answers by randomly sampling words, partly based on probabilities. Their ‘language’ consists of numbers that correspond to tokens », Explains the company. “ In this case, the problem is at the step where the model chooses these numbers. »
“ As if it were a translation error, the model chose slightly wrong numbers, resulting in sequences of words that made no sense », adds OpenAI. “ More technically, inference kernels produced incorrect results when used in certain GPU configurations. »
Subscribe to Numerama on Google News so you don’t miss any news!