Ars TechnicaThe more sophisticated AI models get, the more likely they are to lie
Human feedback training may incentivize providing any answer—even wrong ones.
rob@arstechnica while the logic of what is being said here is reasonable as an account of how AI is not becoming more accurate, the article anthopomorphises AI more than any other I have ever read. "AI models are not really intelligent, not in a human sense of the word." Wrong. There is no "intelligence" at all. That is not how they work.
@arstechnica "... all they are doing is optimizing their performance to maximize reward and minimize red flags". No, they are not doing anything "to maximise reward" - they are not capable of purpose; there is no understanding of what a reward is. Their performance changes as new rules are added. That's all. And that's why they're not getting more accurate, because however much more data is poured in, they have no "knowledge" of "truth". Come on, arstechnica, this is basic.
Wendroid 🇺🇦@robparsons @arstechnica They’re high-calorie sentence construction systems.
MuMind@arstechnica Yes!! I couldn't put my finger on the mechanism but I've been absolutely certain all along that something in the training incentivizes LLMs to BS and always give answers that "sound plausible" vs actually being factually true.
Adlangx@arstechnica I was doing some math with my son the other night and I figured I would put the problem in chatGPT just to see how it did. I was surprised how poorly it performed on Algebra.
Burstaholic@arstechnica that is how models work - you'll ALWAYS get another token, whether it not it makes sense
Somebunny@arstechnica mein Lieblingsbild/ Metapher so far für AI
ALreminds me of is a line from a #startrek movie; the more plumbing you add the easier it is to clog up the drain.