Let's pay millions to train a model on memes and sarcastic posts, WCGW?
Let's pay millions to train a model on memes and sarcastic posts, WCGW?
Google artificial "intelligence" suggested to add glue to let cheese stick to the pizza, because a decade ago user fucksmith on reddit said so
You're viewing a single thread.
Creating an LLM that can babble plausible nonsense at you, with no particular guarantee of truth although a lot of the time it will be perfectly sensible, is (for as fucking amazing a statement as this is) not overly hard at this point.
Getting it to make sense and not do bad things is the hard part. That was the huge innovation that made ChatGPT different from the nonsense-machines that were its early predecessors.
Rob Miles gave a very unsettling talk where he said, the main danger of AI is that commercial pressures will push it into the implementation phase, once the easy part of making it do stuff is done, before the much harder part is done of making it do reliably what you want it to do. And what we’re seeing now is a pretty solid confirmation that yes, that is absolutely the behavior we can count on in the future, and it will get more and more dangerous as AI models get more and more capable.
Huh? ChatGPT is still a nonsense machine.
Yeah but it's trained with the power of refusing to do anything against ethical rules and procedures which are in place for all of our orderly safety.
It's not trained, it just rejects certain keywords. Which is easy to bypass: https://arxiv.org/abs/2402.11753
It's got keyword rejection.
Part of the problem with Google is it's use of retrieval augmented generation, where it's not just the llm answering, but the llm is searching for information, apparently through its reddit database from that deal, and serving it as the answer. The tip off is the absurd answers are exact copies of the reddit comments, whereas if the model was just trained on reddit data and responding on its own the model wouldn't produce verbatim what was in the comments (or shouldn't, that's called overfitting and is avoided in the training process). The gemini llm on its own would probably give a better answer.
The problem here seems to be Google trying to make the answers more trustworthy through rag, but they didn't bother to scrub the reddit data their relying on well enough, so joke and shit answers are getting mixed in. This is more a datascrubbing problem then an accuracy problem.
But overall I generally agree with your point.
One thing I think people overlook though is that for a lot of things, maybe most things, there isn't a "correct" answer. Expecting llms to reach some arbitrary level of "accuracy" is silly. But what we do need is intelligence and wisdom in these systems. I think the camera jam example is the best illustration of that. Opening the back of the camera and removing the film is technically a correct way to fix the jam, but it ruins the film so it's not an ideal solution most of the time, but it takes intelligence and wisdom to understand that.