Comment by gpm
5 days ago
Most facts about the world can't be deduced from logic. They're just facts, to memorize. The King's lefthanded. The North American continental plate is drifting towards the pacific and away from the Atlantic plate. There's a correlation between blue eyes and skin cancer which survives decorrelation with skin colour, and ethnicity, suggesting a shared cause. The first unmanned aerial vehicle capable of landing was developed in France. A general named Rogers led the British in the war of 1812.
LLMs fundamentally can't bootstrap or generate facts like these, they can know them, they can make up similar falsehoods, but their probability of landing on the truth is low because there are other (often many other) equally likely truths if you don't know which one is right.
(Please note: I made up all the "facts" in this post)
Then a very important first question is how do we (humans) discern facts in such cases?
I was rather explicit about that, you memorize them from trusted sources (or directly observe them). There's no question. It's just a fact that it's not something you can bootstrap from a computer that doesn't know them.
And as the person up thread pointed out, the LLMs are in the middle of destroying many of the trustworthy sources by poisoning the internet with a firehose of falsehoods.
It's all about trust. How do we help machines (and humans) know what to trust?
2 replies →
Are you saying human brain is kind of similarly vulnerable to well-crafted facts? Does it mean any intelligence (human or non-human) needs a large amount of generally factual data to discern facts from fakes, which is an argument toward AIs that can accumulate huge swath of factual data?
I feel like you're trying to twist my words into something they don't resemble at all.
I'm not saying anything is vulnerable to anything. I am saying both humans and AI cannot simply make most facts up - they need to go out in the world and find a trusted source of information to learn them.
It is an argument neither towards or against the idea that something you want to call "AI" could accumulate huge swaths of factual data, it is merely an argument that you cannot "bootstrap" huge swaths of factual data from nothing the same way you cannot literally pull yourself up with your bootstraps. If you want the information, you have to collect it from the environment.