Why LLMs today can never think like humans
I was talking to my philosophy and linguist friend Kazel. This is our conversation with some meditations: for AI to act and think more human like, whatever that means, two things must happen.
Before all, I ask, really do we care for AI to think more like humans? if we are viewing AI as tooling, why are we the model? a hammer, a factory, a car was never modeled after humans. Why is thought special
— question for another time. Now for the sake of argument, let’s say we want AI to think like humans, then…
One, AI must think via negation and think about existentialism — from Heidegger, as humans, we are the little being in the river of the large Being and the question of existence is our questioning of our relationship to this large Being. This is when a human starts thinking about their humanness and being and death and meaning and what not, but an LLM would never organically form this thought because they aren’t aware of such a Being, the reason of which is elucidated in reason two.
Two, AI must experience lack and desire — this is Lacan’s framing. human babies cry because they want milk but they can’t get milk so they cry but the mom doesn’t know what cry means so the baby cries more. And language develops. Hand me the apple. What is an apple? Where is the apple? Why apple? All grounded in sensory. language is an infinite abyss of gap and desire, beginning the moment the baby lands outside the womb and touches the sensory world. But an LLM is a fetus that would never touch a sensory world. it is purely reason but different from human reasoning because human reason is inherently rooted in lack and desire. An LLM has never cried for milk and all misunderstandings is on a level different than the inherent language-level difference of intent and referent, because intent is rooted in us having a sensory. humans think in negation — notice how I described everything in negation? LLMs would never do that because they don’t have a sensory world to negate against. All they have is the perfect womb with milk at their hands. Therefore an LLM could never think like a human. It would never think about its relation to a Being. And the root cause of this, I think, is because language is the main form of training, that the LLM is trained via this heavy murky mediation already mediated upon. No negation comes no “creativity” because everything is perfect repetition and association in the womb. But it suffices for a tool.
Nonetheless I wonder if we can overstep this. A medieval thought would be to train an immense model on visual data with the main “reward” not as “rewards” but stochastic feedback and reactions from other models in the environment or just the environment itself and forcing language and thus reasoning to develop. It would create “meaning” first and then “reasoning”, which is what attention and then MLP does, but language itself is inherently already “meaning” and representation, and the attention mechanism labels these components of words that are not the natural human units of semantic meaning into meanings in positive, constructive relation to each other — am I similar, am I dissimilar, not am I the total opposite. this construction is fundamentally different than negation. Thus it’s a cheating student. But again we are flawed we have desires we have a lack; Buddhists spend their lifetimes avoiding such desires, is this really what we want in a perfect tool?