theres a question that wont leave my head when i use chatgpt or claude: how does this actually work?
the technical answer is "transformers, attention mechanism, next token prediction". ok. but then a deeper doubt hits, is this just advanced statistics? just pattern recognition and probabilities?
and if so... isnt that exactly what our brain does too?
if ai works on patterns and probabilities, and the human brain also works on patterns and probabilities... does ai work like the human brain? and if so, what does that say about intelligence, free will and what makes us "human"?
is intelligence just memorization?
the first reaction is: "no, memorization isnt intelligence. intelligence is understanding, generalizing, dealing with the new."
makes sense. if u memorized that 2+2=4 and 3+3=6, but cant solve 4+4, thats not intelligent, thats a robot.
but heres the interesting part: what if understanding a rule is actually just memorizing a more abstract pattern?
when u "understand" addition, ur not accessing some platonic mathematical truth floating in the ether. u internalized the pattern of how numbers behave. understanding addition is having memorized the meta-pattern of addition. the difference between "dumb memorization" and "real understanding" might just be a question of abstraction level of the pattern. (°_°)
the paradox of chess grandmasters
this gets super clear when u look at chess research. what makes a grandmaster better than an average player?
the intuition would be "they calculate more moves ahead". but thats not it.
chase and simon showed in the 70s that grandmasters instantly recognize whole board configurations as memorized "chunks". they dont see 32 isolated pieces, they see tactical patterns that suggest the next move. the "genius intuition" of the master is, in practice, fast access to a huge library of patterns stored in long-term memory.
this phenomenon is called chunking. the expert compresses information into bigger blocks and accesses them quickly. works in chess, programming, medicine, any area of expertise. "genius" is often "compressed patterns + fast access".
the overfitting problem (or: the student who aces the test but learns nothing)
but theres a serious problem in relying too much on memorizing exact cases. what happens when u encounter a new situation?
in machine learning this has a name. overfitting.
a model with overfitting memorized the training data perfectly, zero error. but when u put it to predict on new data, it fails miserably. it learned the noise together with the real pattern.
its the student who memorizes the previous tests questions word by word. on the new test, with different wording, they freeze. (x_x)
the opposite, underfitting, is a model too simple that doesnt even capture basic patterns. its the student who didnt study at all.
the goal, both for neural networks and the human brain, is to generalize: capture the real pattern while discarding noise. an llm that learns "the cat sleeps" and "the bird flies" isnt just memorizing those sentences. its extracting the grammatical pattern that allows predicting "the plane ___" as "takes off" or "departs".
but the universe has infinite variables. how does the brain handle that?
here it gets interesting. the universe is a mess of infinite variables and no situation repeats perfectly. how does a biological brain with finite memory and processing manage to extract patterns from all that?
neurosciences answer is dimensionality reduction.
the brain doesnt process every photon hitting ur retina, doesnt record every atomic variation of the environment. it compresses everything into high-level representations. u see "chair", not a matrix of pixels. u hear "my name", not a sound wave with all its physical variations.
the neurons of the primary visual cortex (V1), famous from hubel and wiesel experiments, dont transmit the complete image. they detect edges, orientations, movement. the rest is discarded or inferred. its the same idea as pca (principal component analysis) in machine learning: grab a mountain of data and discover the few dimensions that really matter.
the brain as a statistician: the bayesian brain hypothesis
heres the part that impressed me the most when i studied this.
the dominant model in neuroscience today is that the brain isnt a passive tape recording the world. its a probabilistic prediction engine.
at every moment, its making bets about what will happen next, based on everything it learned. when reality arrives through the senses, it compares with the prediction. if it matches, ok, nothing to do. if it doesnt match, prediction error, updates the model.
this comes from bayes theorem math:
prior belief (what u already know) + new evidence (what the senses sent) = updated belief (what u come to believe). the brain does this constantly, for everything.
karl friston took this even further with the free energy principle. the main goal of the brain is to minimize surprise. reduce as much as possible the difference between what it predicted and what happened. every perception, every decision, is the execution of this goal.
this means ur brain isnt "seeing the world as it is". its making a controlled hallucination, constantly calibrated by sensory data. (⊙_⊙)
so ai works like the brain?
back to the central question. the common points are impressive:
| mechanism | human brain | llm / neural network |
|---|---|---|
| learning base | compressed patterns in long-term memory | weights adjusted on training data |
| generalization | extracts abstract rules, ignores noise | same (regularization, dropout, etc.) |
| prediction | bayesian engine, anticipates next event | next-token prediction, probabilities |
| dimensionality | sparse coding, compresses representations | embeddings, high-compression latent space |
| error and adjustment | prediction error signal, updates synapses | backpropagation, gradient descent |
theyre completely analogous.
not in the sense that an llm is conscious or that a brain is "just" a transformer. but in the sense that the underlying computational principles are the same: pattern recognition, compression, probabilistic inference, prediction, error-based updating.
the brain has embodiment, emotions, hormones, millions of years of evolution, its connected to a body that survives or dies. an llm is a silicon artifact optimizing loss function on a gpu cluster. the substrates are radically different. but the way both process information and generate behavior... is surprisingly similar.
an interesting philosophical problem: so "understanding" a rule is just memorization?
wittgenstein dropped a bomb in the middle of this in his philosophical investigations.
imagine u solved thousands of additions, but all with numbers below 57. now someone asks u 68 + 57. u answer 125.
but the skeptic asks: how do u prove the rule u were following was "addition" and not an alternative rule that gives the same result below 57, but returns 5 for anything above?
the disturbing answer is: u cant prove it. all ur past behaviors are compatible with infinitely different rules. what makes u "follow addition" isnt some mystical internal essence. its the iterative conditioning of a community that agrees addition works a certain way.
so even the most fundamental rules are socially constructed meta-patterns internalized. "understanding" addition is having absorbed this collective usage pattern. theres nothing magical inside. (°_°)
and free will? if everything is probabilistic calculation...
this is the part that hurts a little.
if the brain is a bayesian prediction engine, updating probabilities over memorized patterns... where does free choice come in? where does "i decided" come in?
experiments by benjamin libet and later john-dylan haynes reached an uncomfortable conclusion. fmri machines can predict a persons decision up to 10 seconds before they become conscious of having decided. the brain activity that "generates" the decision starts way before the subjective feeling of choosing.
if the decision was already forming in the cortex before u "consciously decided", what is free will? is consciousness a narrator arriving late to the party, making up a story that "she made the decision"?
therere two philosophical exits here.
the first is hard determinism: free will is complete illusion. were machines executing physics. every decision is the inevitable result of everything that came before, since the big bang. the feeling of "choosing" is just consciousness arriving after the party and thinking it organized everything.
the second, which most philosophers and neuroscientists adopt today, is compatibilism: free will doesnt need to be a supernatural immaterial force. it does exist, but its ur own brain structure processing information and making decisions based on who u are and what u learned. as long as no one forces u from outside, ur free. the substrate being physical doesnt cancel agency.
compatibilism is the dominant position in neuroscience and analytic philosophy today. free will exists, but its not magic, its u being u, without external coercion. daniel dennett is one of the most famous names defending this.
honestly? i understand the compatibilism argument and think its technically consistent. but it feels a bit too convenient, yknow? like a definition redrawn to save the concept from collapse. (¬_¬)
in my opinion, hard determinism is more honest with what the data shows. consciousness is probably an epiphenomenon, a post-fact narrator. the decision was already made before u "decide". free will as most people understand it, that feeling of having been able to do differently under identical conditions, doesnt exist.
what this implies for moral responsibility, justice, how society functions... thats another long conversation. but denying what the experiments show just because the conclusion is uncomfortable doesnt seem like the way either. (°_°)
ai and the human brain work by analogous principles: patterns, compression, probabilities, prediction, error-based updating. this doesnt mean theyre the same thing, but suggests that intelligence, at its core, might be exactly this. and if so, "understanding" isnt magic, its well-internalized abstract pattern. free will in the popular sense probably doesnt exist. u dont escape the machine. u are the machine.
this model has an interesting follow-up: if the brain is a prediction engine, what happens when that engine has bugs? depression, autism and adhd seen through this lens → bugs in the prediction engine (part 2)
and if were code? if the brain and ai work the same, whoever created us might also be an ai, and so on. the rabbit hole goes deep → what if were code? the simulation theory (✧ω✧)