LLMs Are Not Autocomplete, They're Our Cognitive Cousins
LLMs are not fancy autocomplete but ‘cognitive cousins’ that share core cognitive principles with humans: attention mechanisms recapitulate production systems, prediction is the dominant training signal; sample efficiency is off by three to four orders of magnitude, but an architectural tweak could 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're humans' cognitive cousins
Sripada offers two competing explanations: LLMs may just be fancy autocomplete, a statistical approximator; or the prediction training signal may be strong enough to compress the core principles of human cognition into the weights. He clearly leans toward the latter, calling prediction ‘the mother of all training signals.’ This choice determines everything that follows: if it's just autocomplete, behavioral similarity is mere imitation; if it's a cognitive relative, then LLMs' psychological effects and architectural features are worth studying as evidence about the human mind.
— Chandra SripadaHuman psychological effects emerge spontaneously in LLMs
Sripada lists a string of phenomena that reliably show up in human experiments and all appear in LLMs: the difficulty of center-embedded sentences, garden-path sentences, similarity interference, the recency/primacy/middle-loss effects in serial memory, continuity and temporal asymmetry, and disjunctive vs. conjunctive search in visual search. These effects are not hard-coded through prompt engineering; they emerge spontaneously as the model predicts the next word. He stresses that one cannot stop at counting behavioral similarities—the key is mechanistic analysis, which is exactly the starting point for explaining the Transformer with production systems later.
— 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, intuitive; System 2 is slow, serial, controlled. Sripada argues LLMs give a clear correspondence: in-weight processing is like System 1, in-context processing plus chain of thought is like System 2. This correspondence turns psychology's fuzzy dichotomy into a hypothesis that can be tested against Transformer mechanisms.
— Chandra SripadaLLMs supply the mechanism dual-process theory was missing
Sripada points out that dual-process theory in cognitive science is often criticized as ‘a list of adjectives’: one is fast, automatic, effortless, the other slow and dependent on working memory, but with no mechanistic content. LLMs help fill in the mechanism: in-weight and in-context are actually a continuum with clear poles at either end. For a simple prompt like ‘Michael Jordan plays basketball,’ the weights already contain most of the response and no complex structure needs to be built in activation space; for a prompt like ‘Einstein is German, Gandhi is Indian, so Madame Curie is?’ the task structure must first be constructed, making it vulnerable to interference.
— Chandra SripadaConditional rules model the human mind better than step-by-step instructions
Compare two ways of programming a robot to make tea: traditional fetch-execute is step-by-step instructions—‘walk into the kitchen, pick up the kettle, fill it with water’—and it collapses when it hits an unwritten situation like ‘someone put ice cubes in the kettle’; a production system is a set of conditional rules—‘if the kettle is empty, fill it; if the water is cold, boil it; if the water is hot, pour the tea’—making judgments based on state representations, and is therefore far more robust. Cognitive scientists have long felt production systems model the human mind better than itemized instructions. This setup points directly at the Transformer's attention mechanism.
— Chandra SripadaThe Transformer is actually a production system
Sripada translates the Transformer's core mechanisms into production-system language: attention heads are continuous conditional rules—‘if the key matches the query, add the value’; MLP units are similar, where the rows of the first matrix are features and, if they match the residual stream, the features of the second matrix's columns are added. 99% of the computation in a Transformer is dot products, i.e. ‘a match scalar times a vector.’ This structure was invented independently by computer scientists, yet in ResNet and Attention Is All You Need it walked back onto the production-system path: the residual stream corresponds to a state representation, and the conditional rules correspond to production rules.
— Chandra SripadaSample efficiency is off by four orders of magnitude, maybe one architectural change
Sripada concedes that LLMs' sample efficiency is far behind humans: a child may need only about 100 million tokens, while an LLM needs 100 billion to 300 billion—three to four orders of magnitude apart. But he thinks this may be an innate structural problem rather than a fundamental barrier. He cites ResNet: deep networks without residual connections simply cannot learn via gradient descent, and adding a simple adjustment—‘the previous layer plus a function of the previous layer’—turns an unlearnable problem into a learnable one. Perhaps a similar architectural tweak lies ahead that could raise sample efficiency by several orders of magnitude at once and bring LLMs closer to humans.
— Chandra SripadaThe harder the problem, the more different intelligences look alike
Sripada introduces Dan Yamins and Rosa Cao's ‘contravariance’: the harder the problem, the smaller the space of possible computational solutions, so independently evolved systems like macaques, humans and LLMs should end up with more similar internal representations. He adds his own ‘architectural canalization’: since the Transformer is a production-system architecture, this constraint further narrows the solution space. So larger models aligning better with the human brain is not a coincidence but something forced by task difficulty and architecture together. He admits these are speculations, but considers them key to understanding the convergence between LLMs and human cognition.
— 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 | hundreds of billions | 7:00 |
| IMO gold medal year | 2024 | 40:00 |
| Human child training data volume | 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 in 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 conditional rules, more robust than step-by-step instructions, used to model the human mind.
- in-weight processing
- The model answers directly using its trained weights without constructing new structure in context, corresponding to System 1.
- in-context processing
- The model builds task structure from the current prompt before reasoning, corresponding to 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, making internal representations converge in a particular direction.
- chain of thought
- The model generates intermediate reasoning steps before answering, akin to System 2's serialized 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 will rewrite the next generation's education.
The first 16 minutes of LLM working-principles primer can be skipped; the real discussion starts at 16:58 with the two possibilities.