Gregory's Server@bontchev I'm pretty sure I don't know shit about how "AI" works, but...
Isn't it just language statistics? Is there, was there ever any hope of giving it ironclad instructions? It doesn't have any model of the semantics, it only decides that *this* syntax is more likely in the current context than *that* syntax. The *hope* is that the likely syntax will carry some useful semantics along with it, and that the hidden prompt will influence the conversation that follows in a general, average way.
I guess the people putting this service together thought it worked differently. They thought it was actually smart. Like a person. Like someone who *had* to follow orders.
They treated it very like a slave they could order around with impunity. Funny that they were so eager to do that.
@WesternInfidels AI is a HUGE field. Neural network-based LLMs like these chat bots are only a minuscule part of it. The current hype will pass.
LLMs are pretty much just language statistics, yes. But, also, it is possible to make an AI using ironclad rules. It's called an "expert system". It's hard to make, because you can't just throw a humongous amount of data at it and expect it to learn - you have to talk to human experts, extract their knowledge and experience, and codify it in the form of IF/THEN/ELSE rules. Making an LLM is computationally expensive but there is very little human effort involved, so they are much easier to make - that's why they've become so popular.
@WesternInfidels AI is a HUGE field. Neural network-based LLMs like these chat bots are only a minuscule part of it. The current hype will pass.
LLMs are pretty much just language statistics, yes. But, also, it is possible to make an AI using ironclad rules. It's called an "expert system". It's hard to make, because you can't just throw a humongous amount of data at it and expect it to learn - you have to talk to human experts, extract their knowledge and experience, and codify it in the form of...