The rate of hallucinations depends on the input or the task you're asking it to complete and the length of your conversation. It's also worth mentioning that ho

The rate of hallucinations depends on the input or the task you're asking it to complete and the length of your conversation. It's also worth mentioning that how much hallucinations matter depends on your expectations of the output. What we're seeing right now is people figuring out where ChatGPT can be applied to produce acceptable output an acceptable amount of the time. I think there are a several reasons the projects seem overly ambitious right now. First, a lot of people think the tech will improve exponentially in the next 12 months (I have no idea or opinion on if that's true, but the sentiment is definitely out there). They think that even if the output of the project is crappy today, improvements in ChatGPT will make it worthwhile soon. Second, a lot of people are working with ChatGPT just to get familiar with it and learn about LLMs or ML in general. The point isn't to make a project with a useful application, but to learn from making the project that they did or put a ML project in their portfolio. Third, for a lot of people getting a starting point is worthwhile. If you find it easier to go in and change a paper that someone else drafted or change boilerplate code to fit your needs, then having some ChatGPT output as a starting point has value (even if it's a crappy starting point). Fourth, for a lot of people the expected quality of the output is surprisingly low. I've talked to several people who work in marketing that are using ChatGPT to right bumps for products. That text almost doesn't matter because people aren't really reading the paragraph that promotes a spatula on Amazon or a B movie they've never heard of on Netflix. It doesn't seem like a big difference, but if you sit down to write 10 of those paragraphs, you'll find it's less effort to ask ChatGPT to do it and then just proof read the output. And I think you'll be surprised on where quality matters very little. It's not hard to imagine a company using ChatGPT as the first line of customer support via email or web chat. It can certainly dispense canned troubleshooting steps well. Just have the AI report the tone of the customer's responses and escalate them to a human if they get too upset or argumentative. If the AI said something that was nonsense, the human can apologize and say that they'll "speak with the employee" that the customer was working with before. You already can't trust the output of humans in jobs where you're paying them too little to actually care, so why not use an LLM instead? If the cost of the LLM plus the freebies you give out to appease the customers it makes mad is lower than the cost of hiring humans...