“On February 20, 2024, an optimization to the user experience introduced a bug with how the model processes language.”
ChatGPT spent hours on February 20 spitting out gibberish. OpenAI’s write-up blames inference kernels that “produced incorrect results when used in certain GPU configurations,” so the model picked slightly wrong token numbers. An “optimization to the user experience” broke the product and shipped anyway. Whether ChatGPT made sense depended on which GPU served the request.