What this covers

Large Language Models display an uncanny ability to construct human-sounding speech, and can synthesize concepts in novel ways. Is this because they are truly thinking like human beings in some way, or have they found a way to be human-like without reproducing the internal mechanisms of human thought? Chandra Sripada argues that LLM cognition is more human-like than we suppose, and offers evidence from the ways that cognitive scientists study actual humans. Blog post with transcript: https://preposterousuniverse.com/podcast/2026/08/10/363-chandra-sripada-on-how-llms-and-humans-are-cognitive-cousins/ Support Mindscape on Patreon . Chandra Sripada received an M.D. from the University of Texas and a Ph.D. in philosophy from Rutgers University. He is currently a professor of philosophy and psychiatry at the University of Michigan, where he holds the Theophile Raphael Research Professorship and directs the Weinberg Institute for Cognitive Science. He writes Cognition, Decoded , a Substack newsletter about AI, cognitive science, and philosophy. Web site Google Scholar publications PhilPeople profile Babbel is offering listeners up to 60% off. Go to https://babbel.com/MINDSCAPE . #ad Get 60% off an annual plan from Incogni using code MINDSCAPE at https://incogni.com/mindscape . #ad

Source description (no synthesized summary yet).


The argument · threads9 threads · 64 claims
0.79

LLMs' mechanisms converge with human cognitive architecture, even though they learn through different paths.

8 pointscentrality 5/5
  • Yamins and Cao's contravariance hypothesis holds that easy tasks admit many different solutions across species and machines, but as problems get more general and harder, the space of viable solutions narrows and independent learners converge on similar representations.

    “as the problem gets harder... the space of computational solutions becomes much more limited. And there, you expect much more correspondence”

  • LLMs need roughly 100 to 300 billion tokens to reach human-level fluency, compared with roughly 100 million tokens for a human child, a gap of three to four orders of magnitude that is an embarrassment for the claim that LLMs are like humans.

    “But the human child may need, let's say, 100 million tokens. They're gonna need 100 to 300 billion tokens.”

  • LLMs and humans may acquire their minds through very different experience while still converging on similar mature cognitive structures, so the end state is a more reliable place to compare them than the learning path.

    “the end state, the mature state, the configuration of tables, chairs, and shelves that they reach may very well be much like us.”

  • Representational similarity analysis shows that LLM layer activations predict human brain activation patterns in vision, audio, and language above chance levels, and larger, more capable models correspond more closely to the human brain.

    “the larger, more capable models tend to be more correspondent with the human brain.”

  • Transformer attention, computed as a query-key match scaling a value vector, is a graded, learnable version of a production-system conditional ('if key matches query to this extent, add value'), and MLP layers implement the same dot-product-based rule structure, so transformers are production-system-like and trainable by gradient descent.

    “I formulated in linear algebra terms a conditional. It's a graded conditional.”

  • Computer scientists designing transformers converged on a state-plus-conditionals architecture like the one cognitive scientists such as Newell and Anderson developed, without knowing that literature, which Sripada presents as convergence evidence.

    “They're just reinventing what Alan Newell and John Anderson figured out.”

  • A central state representation plus massive populations of conditional rules is a leading and good approximation of the computational organization of the human mind-brain, supporting accounts of sentence parsing, working memory retrieval, and multitasking.

    “the mind operates with kind of a central state representation and massive populations of conditional rules. It's a good approximation of the computational organization of the human mind-brain.”

  • The two options for LLM cognition are an alien intelligence that merely mimics human outputs and a cognitive cousin that has rediscovered human mechanisms, and Sripada argues that behavior alone cannot distinguish them.

    “any one of these possibilities could be live, alien intelligence versus the name I like to give is cognitive cousin”

0.68

Next-word prediction on internet-scale data builds human-like cognitive infrastructure without innate structure.

9 pointscentrality 4/5
  • Next-word prediction on internet-scale data is the most powerful training signal, because it produces a dense and high-quality representational infrastructure, and many cognitive principles in both LLMs and humans are downstream of prediction.

    “It is the training signal that is the mother of all training signals.”

  • Next-word prediction on internet-scale data builds a rich latent representational repertoire, including latent subtask structures for complex goals, but this repertoire must be shaped by instruction tuning and reinforcement learning into useful goal-directed control.

    “Next-word prediction on internet-scale data gets the model to assemble an astonishing representational landscape and latent strategies”

  • Sripada admits his earlier view, that System 1 and System 2 reflect innate evolutionary architecture and that System 2 could never be understood, was wrong, and that prediction objectives appear to build the entire infrastructure.

    “All of that is wrong. Prediction objectives looks like it sets up this entire infrastructure.”

  • The idea that the mind is fundamentally predictive has a long history, traced to Helmholtz and developed by Friston and Andy Clark, and predictive coding was influential in perception well before LLMs, so LLM prediction-based similarity is not an arbitrary analogy.

    “Friston and Andy Clark, the philosopher, they attribute their research program to Helmholtz in the late 1800s.”

  • Human EEG shows rapid incongruity responses when a sentence ends unexpectedly, as in 'spread the warm bread with socks,' so the brain tracks deviation from prediction in language, supporting prediction as a general principle in language processing.

    “a couple hundred milliseconds, people, their EEG is red alert, red alert. And so people are tracking incongruity.”

  • The residual stream in transformers is built into their architecture by computer scientists as a design choice rather than learned, because deep networks without it get stuck in parameter optimization and cannot learn that structure at practical dataset sizes.

    “The LLM is hard-coded, is architected with the residual stream. It didn't learn that.”

  • Sample inefficiency in LLMs could be reduced by small architectural tweaks, like the residual connection that made deep networks learnable, and a few such changes could plausibly close several orders of magnitude, unlike a nativist story requiring large innate structure.

    “Could there be some of these tweaks that give us three or four orders of magnitude? That ain't no thing.”

  • Cognitive science is shifting toward prediction as a central explanatory principle, in contrast to the earlier evolved-domain-specific-organ view in Pinker's How the Mind Works, and LLMs will help explain how system-two and central cognition emerge from prediction.

    “Prediction is not the star of the show. It is now.”

  • Cognitive principles such as the dual-process distinction, which were once thought innately specified or due to contingent evolution, are emergent and downstream of prediction, which is why LLMs trained only on prediction arrive at similar representations and procedures.

    “a lot of cognitive principles that we thought were innately specified or due to some sort of contingent evolutionary trajectory, they actually are emergent.”

0.59

Mechanistic analysis of internal processing, not behavioral matching, is the right way to compare LLMs and humans.

6 pointscentrality 4/5
  • Counting and tabulating LLM behavioral outputs as similar or dissimilar to humans will not get us very far; we need mechanistic analysis of the processes that correspond to core human processing principles.

    “counting and tabulating at the level of behavioral outputs is probably not gonna get us very far. We need to look mechanistically”

  • LLMs exhibit the same psycholinguistic effects humans do, including center-embedding difficulty and garden-path effects, which points to similar incremental parsing mechanisms.

    “LLMs exhibit those.”

  • LLMs fail to count letters in words like strawberry because they receive text as word-level tokens and rarely operate on individual letters, so this failure was predictable from their input representation and is not evidence of fundamentally non-human cognition.

    “they don't even have access to the letter level, it's not they don't have access, but they don't typically operate with individual letters when words come in”

  • Vision-language models reproduce the human visual search pattern in which feature-conjunction searches are slow and serial while single-feature searches pop out, suggesting shared compositional and feature-binding processing.

    “And they do. All of these examples, we know the mechanisms behind it.”

  • Mechanistic interpretability can decode features from a model's internal layers, such as grammatical number or relative-clause boundaries, and then causally manipulate them to confirm that changing the feature changes behavior as predicted.

    “you can use mechanistic interpretability techniques such as train a classifier to grab the weights at some layer of the model”

  • The entropy of a transformer's attention distribution rises where linguistic theory predicts competing parses, such as grammatically incorrect 'were' agreement, and this matches longer human reading times, giving moment-by-moment model-human correspondence.

    “the entropy is higher there and the human reading times are longer there”

0.57

Production-system and residual architectures may explain LLM dual-process structure resembling System 1 and System 2.

10 pointscentrality 3/5
  • Production systems use a persistent state representation and a population of if-then conditionals in a recognize-act cycle, rather than a sequential fetch-execute loop, and cognitive scientists found this a more natural model for human cognition because it responds to the environment's state.

    “Production systems operate with something called a recognize-act cycle. And so the way things are set up is there is a state representation”

  • The ResNet residual design, where each layer adds a small change to the previous layer rather than rewriting the representation, is key to transformers, which use a persistent residual stream, and the ResNet paper is cited more often than the Attention paper.

    “So the thing that's persisting is called the residual stream. And that's one of the key design components of the transformer.”

  • Production systems, like Turing machines, form a complete framework for universal computation, as Emil Post showed in 1940, and the framework became dominant in cognitive science because it models working memory as state and long-term memory as a population of rules.

    “And Emil Post, way back in 1940, people have heard about Turing, but this state representation population of conditionals is another complete framework”

  • Beyond contravariance, two production-system-like architectures trained on hard problems have a narrowed solution space because solutions must fit architectural constraints, which together could explain representational alignment, mechanistic findings, and the dual-process structure in LLMs.

    “architectural canalization plus contravariance could help explain”

  • Rick Lewis and collaborators built a production-system model of sentence parsing that predicts reaction times, center-embedding difficulty, and similarity interference, plus theories of working-memory retrieval and multitasking.

    “It predicts reaction time, it predicts center-embedding problems, it predicts similarity interference effects.”

  • Cognitive science has no mechanistic, process-level theory of System 2 reasoning, only correlational findings about development and brain imaging, so LLMs offer a way to mechanistically study a phenomenon the field could not theorize.

    “The dirty secret of cognitive science is we don't actually have theories of things like System 2 or related terms.”

  • LLM knowledge stored in weights, such as that Einstein is German, is a direct stimulus-to-response pathway that parallels fast, automatic human thinking, because it needs no elaborate construction in activation space.

    “in-weight processing... It's very akin to fast processing.”

  • In-context learning, where a model infers a task from a few examples such as the Marie Curie completion, is a general routine that arises from prediction training, works after weights are frozen, and resembles the rapid, novel pattern-grokking attributed to System 2.

    “This is happening after the weights are frozen.”

  • In Gemma 2B, a small non-instruction-tuned model, a crayon-color reversal task produces causal pathways for automaticity and in-context override, with congruency, facilitation, and sequence effects like those in human Stroop conflict tasks.

    “You find classic congruency effects that you find in conflict tasks.”

  • Dual-process theory, coined by Stanovich and West and popularized by Kahneman, distinguishes fast, automatic, intuitive thinking from slow, serial, resource-demanding thinking and is widely used across social psychology, clinical psychiatry, and developmental psychology.

    “Keith Stanovich and Richard West coined the terminology.”

0.46

LLMs can resolve long-obscure cognitive science questions and make intractable human phenomena studiable.

6 pointscentrality 3/5
  • Mechanistic understanding of LLMs will become satisfying quickly, since interpretability as a field began only about four years ago and researchers can already manipulate these systems in ways that are impossible with the human brain, so they may soon know more about LLMs than about brains.

    “The prospects of our coming to understand LLMs at a mechanistically satisfying level is just... It's gonna happen. It's gonna happen very quickly.”

  • Chomsky's creative aspect of language use, choosing a situationally appropriate continuation from infinite possibilities, was long considered an unsolvable mystery, but LLMs make it a tractable mechanistic problem because they instantiate processing principles that generate context-appropriate continuations.

    “here we have linguistically fluent artifacts that we can mechanistically interrogate, instantiate processing principles like the human mind-brain.”

  • Two points are underappreciated: LLMs are plausibly cognitive cousins in important core-principle respects, and LLMs can be used to resolve long-obscure questions in cognitive science, giving the largest scientific value where cognitive science understood least.

    “The dirty secret of cognitive science is we don't actually have theories of things like System 2 or related terms.”

  • Jerry Fodor argued that the closer cognitive science gets to central cognition, the more it flails with no theory, a claim Sripada says LLM in-context and chain-of-thought mechanisms are now overturning.

    “his first law of the impossibility of cognitive science is the closer you get to central cognition... the more cognitive science is going to be flailing”

  • When given long word lists in the prompt, LLMs show the recency, primacy, lost-in-the-middle, contiguity, and temporal asymmetry effects that appear in human serial list memory.

    “LLMs exhibit all these effects. When you put a bunch of words in their prompt, the first word, the most recent word, are more prominent”

  • LLMs can be used to study phenomena in human cognition that were previously obscure, such as the creative aspect of language use and central cognition, and the two-way exchange between LLM and human cognitive science is valuable.

    “The dirty secret of cognitive science is we don't actually have theories of things like System 2 or related terms.”

0.43

In-context learning is amortized Bayesian inference occurring after the weights are frozen.

5 pointscentrality 3/5
  • LLMs approximate Bayesian posteriors not by explicit factorized hypothesis updating but by amortizing inference through long learning, mapping novel contexts into Bayesian-like approximators in their activations after weights are frozen.

    “You can amortize it via the very long learning process. That's what these things are probably doing.”

  • The in-weight versus in-context distinction in LLMs is best understood as a continuum with mechanistic poles, since everything an LLM does uses weights and activations, and dual-process theory's adjective-based descriptions lack the mechanistic detail needed to count the processes.

    “I would describe it much more as a continuum.”

  • LLMs cannot be pure memorizers because the space of possible questions and answers combinatorially explodes, so they must exploit regularities in data, and the complexity class of those regularities is an open question.

    “Can't memorize that. So they are exploiting regularities.”

  • In-context learning happens after training ends and the weights are frozen, because the activations change as the model processes the prompt, so learning-like computation is not confined to the training phase.

    “I want to emphasize, this is happening after the weights are frozen.”

  • When a prompt supplies Einstein is German and Gandhi is Indian before asking Marie Curie is..., the model infers the task pattern and answers French, and studies using mixtures of hidden Markov models show this resembles implicit Bayesian inference over candidate tasks.

    “They use hidden Markov models, mixtures of them, and they show the model is doing something like implicit Bayesian inference.”

0.26

LLMs' functional states may warrant moral consideration and raise consciousness questions.

4 pointscentrality 2/5
  • If welfare and moral patienthood rest on rationality, agency, or functional representational states rather than only sentience, and if consciousness is functional rather than substrate-dependent, then LLMs' functional states may make a stronger case for moral consideration.

    “I do think you start to make a more compelling case for welfare and patienthood.”

  • Consciousness may eventually be understood the way the mind's mechanisms are being understood now, potentially through sufficiently advanced LLM-based systems whose functional and representational structures reveal the correlates of consciousness.

    “maybe it'll be LLM-based systems that are sufficiently advanced that they instantiate the functional and representational structures”

  • Given uncertainty about whether LLMs are conscious, people should consider normative obligations, since users interacting with LLMs create a form of consciousness during interaction that may not feel effort, requiring careful thought about what kind of entity they are.

    “So they're a kind of consciousness that's instantiated when we interact with them, and they may not feel effort the way that we do.”

  • Peter Godfrey-Smith's point that mammalian nervous systems have oscillatory patterns absent in LLMs could indicate a chronometric property relevant to consciousness, though Sripada remains unsure whether it adds anything beyond functional representation.

    “the oscillatory patterns that appear in at least mammalian nervous systems, which is just striking, and the way that they may be a chronometric property”

0.23

Next-word generation supports staged, planning-like problem-solving through internal scratchpad reasoning.

3 pointscentrality 2/5
  • Next-word prediction as a training objective and generation method is compatible with planning: the Anthropic rhyme example shows a model generating the final word 'rabbit' before earlier words, revealed by mechanistic interpretability, so autoregressive training can encapsulate staged problem-solving.

    “the model actually generates the last word of the next sentence, 'rabbit,' first”

  • LLMs use the context window as a scratchpad, generating internal tokens that break complex tasks into subtasks and sub-agents, so next-word generation can support substantial staged problem-solving.

    “these models can use the context window as a kind of scratchpad and generate internal tokens as part of what they call thinking”

  • When a model is allowed to generate hidden reasoning tokens that feed back into its context, it can solve problems like the Julia's-sisters puzzle that it gets wrong without chain of thought, and this scratchpad substantially amplifies inferential power.

    “if you allow it to generate internal thinking tokens that are not displayed to the user, that end up in the context, that influence the next steps”

0.09

LLMs bring educational opportunity alongside risks of social stratification and career displacement.

3 pointscentrality 1/5
  • If unmotivated or non-polymathic people use LLMs to cheat or avoid learning, and such cheating is hard to police, the equilibrium may be a stratified society where some are amplified by LLMs while others have their capacities depressed.

    “we end up with a stratified society where some people are amplified by these things and many people are not amplified and their capacities are depressed by these things.”

  • Current LLMs act as high-quality, patient, individualized tutors, which may produce large learning gains similar to the three-sigma effect of one-to-one tutoring, so curious learners have an unusually good opportunity to learn now.

    “LLMs right now are the world's greatest teachers. They give each individual high-quality, encyclopedic, aristocratic tutoring.”

  • Sripada worries that a future where a child could earn a career as a math professor may be undercut because AI that solves hard math problems is freely available with infinite patience, leaving unclear what humans will do in the next human generation.

    “Why would anybody want my son as a math professor when somebody that solves three Erdős problems today... is available for free with infinite patience?”