On February 20, ChatGPT lost control. The very popular generative artificial intelligence (AI) interface that made this technology known was unlocked for several hours this Tuesday. How ? By answering user questions with nonsense sentences.
Other people observed that ChatGPT would start responding in English and then, for no apparent reason, switch to Spanish. Others received answers where each word was highlighted in a different color. In a word, it was weird.
And here the publisher, OpenAI, gives some explanations.
“On February 20, 2024, an optimization of the user experience introduced a bug in the way the model processes language” assures OpenAI in a retex. And to detail: LLMs generate answers by sampling words randomly, partly on the basis of probabilities. Their “language” therefore consists of numbers which correspond to tokens.
“Inference kernels were producing incorrect results”
“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 had no meaning” specifies the editor.
More technically, “the inference kernels produced incorrect results when used in certain GPU configurations,” he adds.
After identifying the cause of this incident, the company claims to have deployed a fix and confirmed that the incident was resolved. In fact, she has said that she had implemented a fix. I think she reverted to an earlier, stable version of LLM.
“These systems have never been stable”
AI specialist Gary Marcus hopes the incident will be seen as a “wake-up call”.
“These systems have never been stable. No one has ever been able to develop security guarantees around these systems,” he wrote Tuesday. According to him, “the need for completely different technologies, less opaque, more interpretable, easier to maintain and debug – and therefore easier to implement – remains paramount”.
This episode, while amusing in hindsight, is a stark reminder of the complexities and potential vulnerabilities inherent in AI technologies. For all that we love about generative AI, it is far from infallible.
I also worry about OpenAI’s deployment model. Almost all SaaS models roll out new versions for a limited number of users. Then, when it becomes clear that the new version works well, the company rolls it out to everyone. That doesn’t seem to be the case here. It appears that many, if not all, users have been affected.
Interestingly, ChatGPT usually limits its deployments. For example, ChatGPT’s new memory feature – where the program remembers conversations you’ve had with it – isn’t available to everyone yet.
Lesson of the day? It’s still too early to rely on ChatGPT – or other AI chatbots – for everyday work.
Source: “ZDNet.com”
Stability AI’s new text-to-image conversion model performs better, offers better image quality and…
Follow all the news from ZDNet on Google News.
Related articles
Partner content
Sponsored content