LLM and prime numbers -
>>>>>>>>>>>>>>
Artificial intelligence, hold the intellect.
Is 1 a prime number? For the answer, you might want to ask Google (or a sixth grader) rather than ChatGPT. In March, two researchers from Stanford and one from Berkeley quizzed the AI model’s newest version, GPT-4, and found that it correctly identified prime numbers 97.6 percent of the time. A few months later, when they quizzed it again, it correctly identified prime numbers just 2.4 percent of the time. So what happened? Did ChatGPT get distracted by the cute AI model next door? Did humans literally dumb it down? The researchers say it’s complicated. “It’s very difficult to say, in general, whether GPT-4 or GPT-3.5 is getting better or worse over time,” said James Zou, an assistant professor of data science and a co-author of the preprint study.
It may appear that the AI is getting dumber, but the researchers point out that it also provided fewer potentially offensive responses and was less likely to offer ideas on how to break the law. Plus, they say, it wasn’t actually “smart” to begin with. Since ChatGPT is a large language model trained to generate human-sounding text, it might, in fact, never have assessed primeness at all but rather produced answers based on incidental trends in its training data. ChatGPT’s parent company, OpenAI, doesn’t discuss how it trains ChatGPT, so as it adds new training data—which could change the way the model responds—researchers and users are left to speculate. “While the majority of metrics have improved,” the company said, “there may be some tasks where the performance gets worse.” In the meantime, the Loop can help you out with this minor mathematical mystery: 1 may be lonely, but it’s not prime.
