but LLMs, a very specific and especially inscrutable class of AI which has been designed for “sounding convincing”, without care for correctness or truthfulness.
I think that I’d put it in a slightly less-loaded way, and say that an LLM just produces content that has similar properties to its training content.
The problem is real. Frankly, while I think that there are a lot of things that existing LLM systems are surprisingly good at, I am not at all sure that replacing search engines will be it (though I am confident that in the long run, some form of AI system will be).
What you can’t do with systems like the ones today is to take one data source and another data source that have conflicting information and then have the LLM-using AI create a “deep understanding” of each and then evaluate which is more-likely truthful in the context of other things that have been accepted as true. Humans do something like that (and the human approach isn’t infallible either, though I’d call it a lot more capable).
But that doesn’t mean that you can’t use heuristics for estimating the accuracy of data and that might be enough to solve a lot of problems. Like, I might decide that 4Chan should maybe have less-weight as a solution, or text that ranks highly on a “sarcastic” sentiment analysis program should have less weight. And I can train the AI to learn such weightings based on human scoring of the text that it generates.
Also, I’m pretty sure that an LLM-based system could attach a “confidence rating” to text it outputs, and that might also solve a lot of issues.
I think that I’d put it in a slightly less-loaded way, and say that an LLM just produces content that has similar properties to its training content.
The problem is real. Frankly, while I think that there are a lot of things that existing LLM systems are surprisingly good at, I am not at all sure that replacing search engines will be it (though I am confident that in the long run, some form of AI system will be).
What you can’t do with systems like the ones today is to take one data source and another data source that have conflicting information and then have the LLM-using AI create a “deep understanding” of each and then evaluate which is more-likely truthful in the context of other things that have been accepted as true. Humans do something like that (and the human approach isn’t infallible either, though I’d call it a lot more capable).
But that doesn’t mean that you can’t use heuristics for estimating the accuracy of data and that might be enough to solve a lot of problems. Like, I might decide that 4Chan should maybe have less-weight as a solution, or text that ranks highly on a “sarcastic” sentiment analysis program should have less weight. And I can train the AI to learn such weightings based on human scoring of the text that it generates.
Also, I’m pretty sure that an LLM-based system could attach a “confidence rating” to text it outputs, and that might also solve a lot of issues.