LLMs Are Not Autocomplete — They Are Our Cognitive Cousins
LLMs are not fancy autocomplete but ‘cognitive cousins’ that share core principles with human cognition: attention mechanisms recreate production systems, and prediction is the dominant training signal; sample efficiency is three to four orders of magnitude worse, but an architectural change may close the gap.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
LLMs are not autocomplete; they are humans' cognitive cousins
Sripada lays out two competing explanations. LLMs might be nothing more than fancy autocomplete, a statistical approximator. Or prediction as a training signal might be strong enough that it has pressed the core principles of human cognition into the weights as well. He comes down clearly on the second, calling prediction ‘the mother of all training signals’. That choice determines everything that follows: if it is only autocomplete, behavioral resemblance is mere mimicry; if LLMs are cognitive kin, then their psychological effects and architectural features are worth studying as evidence about the human mind.
— Chandra SripadaHuman psychological effects emerge in LLMs on their own
Sripada runs through a list of phenomena that show up robustly in human experiments and also appear in LLMs: the difficulty of center-embedded sentences, garden-path sentences, similarity-based interference, recency, primacy and lost-in-the-middle effects in serial memory, continuity and temporal asymmetry, and disjunctive versus conjunctive visual search. These effects are not forced into existence by prompt engineering; they emerge on their own as the model predicts the next token. He insists the point is not to tally up behavioral similarities but to do mechanistic analysis — which is exactly where the later explanation of Transformers as production systems begins.
— Chandra SripadaSystem 1 is the weights, System 2 is the context
Kahneman's dual-process theory splits thinking into two systems: System 1 is fast, automatic and intuitive; System 2 is slow, serial and controlled. Sripada argues LLMs supply a clean mapping: in-weight processing behaves like System 1, while in-context processing plus chain of thought behaves like System 2. That correspondence turns a vague dichotomy from psychology into a hypothesis that can be checked against Transformer mechanisms.
— Chandra SripadaLLMs supply the mechanism dual-process theory always lacked
Sripada notes that dual-process theory in cognitive science is often criticized as a ‘list of adjectives’: one process is fast, automatic and effortless, the other slow and dependent on working memory, with no mechanistic content. LLMs fill in the mechanism: in-weight and in-context turn out to be a continuum with well-defined poles at each end. For a simple prompt like ‘Michael Jordan plays basketball’, most of the response is already contained in the weights, and no complex structure has to be built in activation space. For a prompt like ‘Einstein is German, Gandhi is Indian, so Marie Curie is?’, the task structure has to be constructed first, which is why it is easy to interfere with.
— Chandra SripadaCondition-action rules resemble the mind more than step-by-step instructions
He contrasts two ways to program a robot to make tea. The traditional fetch-execute approach is a sequence of steps — walk into the kitchen, pick up the kettle, fill it with water — and it breaks the moment something unwritten happens, like someone dropping ice cubes into the kettle. A production system is instead a set of condition-action rules: if the kettle is empty, fill it; if the water is cold, boil it; if the water is hot, pour the tea. Because it judges from a state representation, it is far more robust. Cognitive scientists have long felt production systems fit the human mind better than instruction lists. This setup points directly at the Transformer's attention mechanism.
— Chandra SripadaThe Transformer turns out to be a production system
Sripada translates the Transformer's core mechanisms into the language of production systems: an attention head is a continuous condition-action rule of the form ‘if the key matches the query, add the value’. MLP units work similarly — rows of the first matrix are features, and if one matches the residual stream, the corresponding column feature of the second matrix gets added. Fully 99% of the computation in a Transformer is dot products, that is, a match scalar multiplied by a vector. Computer scientists invented this structure independently, yet in ResNet and Attention Is All You Need they walked back onto the production-system path: the residual stream corresponds to the state representation, and the condition-action rules to production rules.
— Chandra SripadaSample efficiency is four orders off, perhaps one architectural change away
Sripada concedes that LLMs are far behind humans on sample efficiency: a child may need only about 100 million tokens, while an LLM needs 100 billion to 300 billion — three to four orders of magnitude more. But he thinks this may be an innate-structure problem rather than a barrier of principle. He points to ResNet: deep networks without residual connections simply cannot be trained by gradient descent, and adding one simple adjustment — the previous layer plus a function of the previous layer — turned a problem that was unlearnable into a learnable one. There may be another architectural tweak of that kind ahead, one that lifts sample efficiency by several orders of magnitude at a stroke and brings LLMs closer to humans.
— Chandra SripadaThe harder the problem, the more different intelligences resemble each other
Sripada brings in ‘contravariance’ from Dan Yamins and Rosa Cao: the harder a problem is, the smaller the space of possible computational solutions, so independently evolved systems — macaques, humans and LLMs — end up with more similar internal representations. He adds a principle of his own, ‘architectural canalization’: given that Transformers are a production-system architecture, that constraint narrows the solution space further. So the fact that larger models align better with the human brain is not a coincidence; it is forced out by task difficulty and architecture together. He admits this is speculative, but sees it as the key to understanding why LLM and human cognition converge.
— Chandra SripadaIn their own words · checked verbatim
It is the training signal that is the mother of all training signals.
Chandra Sripada18:15
Rick calls this the greatest discovery in the history of cognitive science.
Chandra Sripada35:39
To me, this is screaming, 'Hey, look at me, I'm an LLM. Look at me, look at me. I'm not so alien,' to me.
Chandra Sripada44:01
Prediction may be the mother of all training signals. There is just nothing that can give you the density and the high quality of prediction.
Chandra Sripada1:06:52
Could there be some of these tweaks that give us three or four orders of magnitude? That ain't no thing. You and I know. Three or four orders of magnitude? You're a physicist. That's a joke in your neck of the woods.
Chandra Sripada1:11:21
So I think looking at mechanisms carefully from a cognitive science perspective is a very important project to make sure that we are not creating a dystopian state where we're inflicting massive harms on entities that deserve moral protections.
Chandra Sripada1:31:05
what LLMs right now are the world's greatest teachers. They give each individual high-quality, encyclopedic, aristocratic tutoring.
Chandra Sripada1:37:01
I think that that's... There's legitimately a concern. I don't even wanna call it a worry, because my feeling, which might be wrong in this case, but my feeling is that human beings adapt to technological new capacities. We fill in the other niches that we didn't even know were there.
Sean Carroll1:39:02
Figures
| LLM parameter scale | several hundred billion | 7:00 |
| Year of the IMO gold medal | 2024 | 40:00 |
| Training data for a human child | about 100 million tokens | 1:09:43 |
| Share of Transformer computation that is dot products | 99% | 1:00:00 |
| Three-sigma effect | about three standard deviations | 1:37:01 |
| Erdős problems | someone solved three within two weeks | 1:39:02 |
| Year Sripada was in grad school | 2006 | 1:47:59 |
Glossary
- production system
- A cognitive architecture that triggers actions via condition-action rules; more robust than step-by-step instructions, and used to model the human mind.
- in-weight processing
- The model answers directly from its trained weights, with no new structure built in context — the counterpart of System 1.
- in-context processing
- The model uses the current prompt to construct task structure before reasoning — the counterpart of System 2's controlled processes.
- contravariance
- The harder the problem, the fewer viable computational solutions, so different systems' internal representations converge.
- architectural canalization
- The model architecture itself narrows the solution space, pushing internal representations to converge in a particular direction.
- chain of thought
- The model generates intermediate reasoning steps before answering, resembling System 2's serial thinking.
How to listen
Researchers working on model mechanisms and alignment, investors trying to judge whether LLMs have human-like cognition, and founders who care about how AI rewrites the next generation of education.
The first 16 minutes explaining how LLMs work can be skipped; start at 16:58 with the two possibilities, where the real argument begins.