Comment by Jensson
4 hours ago
"Dumb next token predictor" keeps popping up since that is the core way they work. Since they aren't logic engines but prediction engines they will always return a result regardless what you ask it.
Some predictions might be the tokens "I don't know", but that is based on the model mapping your text to those tokens by having seen many similar "I don't know" responses to such contexts, it didn't do any introspective logic to produce that "I don't know", and its possible it actually does know if it followed another branch there so "I don't know" is often not even true.
If they had an introspective part that stops the prediction when its too unreliable it would no longer just be token prediction engine, and I believe we need such a part for them to become what I call smart. I don't think LLM will ever stop being dumb without such an introspective part to them.
And no, that introspective part is not a part of the token predictor. At least not in us humans, the feeling of certainty we have is not a prediction, it is bundled with our thoughts, so we get both "answer is a bear" and "certainty is low", we don't get just one of those as a "prediction".
Will LLM become smart as humans with such an introspective part? I don't know, but I think they will never become as smart as humans without one.
Note: The certainty score has to be per conclusion or response, not per token. You can't evaluate a responses validity by aggregating the weight of each token. Meaning its a logic engine, not token engine, that evaluates the certainty of a statement being correct or not instead of a token being correct or not. That is the level human thinking works at and seems to be dramatically more efficient.
I strongly suspect that if you dug deeply enough into how a human brain does reasoning, the answer would boil down to "it tries a bunch of pathways and picks the one that seems right". Pathways weighted by experience, but ultimately still guess-and-check. What would the alternative be?
Maybe our ability to reason is not as mystical and special as our ego might hope it to be, and discomfort over LLM's bringing that to light is the root of some people's urge to continuously downplay and discredit them.
(As far as I can tell as a non-neuroscientist, the literature on how our default mode network and prefrontal cortex interact agrees with my assessment - free association and logical verification respectively)
You have introspection, you can see a part of your thoughts, you know how that introspective part works since its what we call consciousness, you are conscious about it.
That introspection isn't an illusion, what your consciousness see of your thoughts obviously are tings that has been calculated and aggregated by the brain, so we know the brain calculates and aggregates those thoughts and feelings to produce its results. And we know LLM doesn't do that, it doesn't have a side system that does that sort of introspection.
It actually is an illusion though. The narrative the brain makes up after the fact for why we did something is not perfectly correlated with the actual reason. It's just very good at its job of spinning a convincing narrative, even to itself. This is demonstrated in split brain patients: they're given instructions to do something that are shown only to one eye, the one connected to the non-speaking hemisphere, and when asked why they did it, the speaking hemisphere confidently hallucinates a wrong answer that the person seems to fully believe.
1 reply →
> I don't think LLM will ever stop being dumb without such an introspective part to them.
The very second they add this part, it will "just be a dumb token predictor with introspection", mark my words.
If it keeps doing dumb things, yeah. But if that actually solves it then those opinions will quickly disappear when it replaces all human white collar work since it does it cheaper and better and faster.
AGI is fairly easy to detect for this reason. Does this system make a majority feel you don't have to hire people anymore? If not its not AGI.
> But if that actually solves it then
Solves what? Chess? No thats not AI, its just a chess bot. Turing test? No, thats not AI, its just a dumb token predictor.
Why would the goalposts stop moving at AGI? I am 100% convinced it will somehow still lack the "gusto" or the "taste" or the "timbre" of real intelligence. You can see it in coding right now, AI has bad "taste" in coding, because really, we can't do a better job critiquing that which is obviously (on occasion) just plain good.
1 reply →
Necessary is not the same as sufficient.
If I want my ox-cart to fly, I need to add wings, but that's probably not all that needs changing.