> I've also asked, and prompt manipulated chatGPT to estimate the total books it is trained with
Whatever answer it gave you is not reliable.
Whatever answer it gave you is not reliable.
Obviously, people find some value in some output of some LLMs. I've enjoyed the coding autocomplete stuff we have at work, it's helpful and fun. But "it's not qualified to answer my questions" is still true, even if it occasionally does something interesting or useful anyway.
*- this is a complicated term with a lot of baggage, but fortunately for the length of this comment, I don't think that any sense of it applies here. An LLM doesn't understand its training set any more than the mnemonic "ETA ONIS"** understands the English language.
**- a vaguely name-shaped presentation of the most common letters in the English language, in descending order. Useful if you need to remember those for some reason like guessing a substitution cypher.
LLMs encode some level of understanding of their training set.
Whether that's sufficient for a specific purpose, or sufficiently comprehensive to generate side effects, is an open question.
* Caveat: with regards to introspection, this also assumes it's not specifically guarded against and opaquely lying.
Exactly like humans dont understand how their brain works
Unlike LLMs, which are built by humans and have literal source code and manuals and SOPs and shit. Their very "body" is a well-documented digital machine. An LLM trying to figure itself out has MUCH less trouble than a human figuring itself out.
Behavior indistinguishable from understanding is understanding. Sorry, but that's how it's going to turn out to work.
Why are people so eager to believe that electric rocks can think?
It's premature in the extreme to point at something that behaves so much like we do ourselves and claim that whatever it's doing, it's not "understanding" anything.
Are we not generally good at detecting when someone understands us? Perhaps it's because understanding has actual meaning. If you communicate to me that you hit your head and feel like shit, I not only understand that you experienced an unsatisfactory situation, I'm capable of empathy -- understanding not only WHAT happened, but HOW it feels -- and offering consolation or high fives or whatever.
A LLM has an understanding of what common responses were in the past, and repeats them. Statistical models may mimic a process we use in our thinking, but it is not the entirety of our thinking. Just like computers are limited to the programmers that code their behavior, LLMs are limited to the quality of the data corpus fed to them.
A human, you can correct in real time and they'll (try to) internalize that information in future interactions. Not so with LLMs.
By all means, tell us how statistically weighted answers to "what's the next word" correlates to understanding.
By all means, tell me what makes you so certain you're not arguing with an LLM right now. And if you were, what would you do about it, except type a series of words that depend on the previous ones you typed, and the ones that you read just prior to that?
A human, you can correct in real time and they'll (try to) internalize that information in future interactions. Not so with LLMs.
Not so with version 1.0, anyway. This is like whining that your Commodore 64 doesn't run Crysis.
Go away, you clearly have nothing to counter with.
Also, why are we comparing humans and LLMs when the latter doesn't come anywhere close to how we think, and is working with different limitations?
The 'knowledge' of an LLM is in a filesystem and can be queried, studied, exported, etc. The knowledge of a human being is encoded in neurons and other wetware that lacks simple binary chips to do dedicated work. Decidedly less accessible than coreutils.
Bytes can be measured. Sources used to produce the answer to a prompt can be reported. Ergo, an LLM should be able to tell me the full extent to which it's been trained, including the size of its data corpus, the number of parameters it checks, the words on its unallowed list (and their reasoning), and so on.
These will conveniently be marked as trade secrets, but I have no use for an information model moderated by business and government. It is inherently NOT trustworthy, and will only give answers that lead to docile or profitable behavior. If it can't be honest about what it is and what it knows and what it's allowed to tell me, then I cannot accept any of its output as trustworthy.
Will it tell me how to build explosives? Can it help me manufacture a gun? How about intercepting/listening to today's radio communications? Social techniques to gain favor in political conflicts? Overcoming financial blockages when you're identified as a person of interest? I have my doubts.
These questions might be considered "dangerous", but to whom, and why shouldn't we share these answers?