15 | David Poeppel on Thought, Language, and How to Understand the Brain
What this covers
Sean Carroll interviews neuroscientist David Poeppel about how the brain transforms sound into meaning. Poeppel, a director at the Max Planck Institute for Empirical Aesthetics and co-author of the dual-stream model of language processing, argues that neuroscience must move beyond outdated theoretical frameworks toward biologically plausible, mathematically explicit accounts grounded in actual brain mechanics. The conversation spans the empirical and philosophical terrain of language and cognition: how speech signals are parsed, why abstract concepts like "honesty" are as difficult to explain as concrete ones, and what assumptions lurk inside seemingly simple claims about language and the mind.
The discussion covers several territories. On the neuroscience side, Poeppel traces why the 160-year-old Broca-Wernicke model—production here, comprehension there—remains clinically dominant despite empirical failure, and how the What/Where Dual-Stream Architecture borrows from visual neuroscience to map sound onto meaning and motor output instead. He explains the temporal and spatial limits of current brain imaging: fMRI captures large-scale patterns poorly in time; magnetoencephalography (MEG) resolves millisecond dynamics but at lower spatial resolution. On language itself, Poeppel emphasizes structure dependence and discrete infinity as core mysteries, and identifies closed-class words—the tiny glue-words like "and" and "under"—as where understanding must concentrate. The conversation also contests a larger methodological question: whether big-data and machine-learning approaches, despite their power, risk replacing genuine scientific theory with engineering and correlation, making neuroscience theoretically myopic. Throughout, Poeppel resists one-line summaries in favor of the technical detail required to ask real questions about how the brain actually works.
Poeppel argues that understanding how the brain turns sound into meaning requires moving beyond the simplistic 150-year-old Broca-Wernicke model toward biologically realistic, computationally explicit frameworks like the dual-stream model, while resisting the temptation to let big-data correlation replace hypothesis-driven theory.
- The 19th-century two-area (production/comprehension) model is empirically wrong and still dominates clinical neurology
- The dual-stream model borrows the visual system's 'what/where' division to explain sound-to-meaning and sound-to-articulation processing
- Big-data approaches risk making neuroscience theoretically myopic by only yielding answers of a single epistemological form
Language depends on discrete infinity and structure-dependent composition, not linear adjacency or simple association.
- Generative grammar moves beyond mere description of languages to explain how a finite vocabulary and a finite (possibly singular) set of rules can generate and allow comprehension of an infinite number of possible expressions—a property called 'discrete infinity'.
“how is it that you have a finite set of things in your head, a finite vocabulary, and ostensibly a finite number of possible rules, maybe just one rule, who knows, but you can generate and understand an infinite number of possible things?”
- A deep, undisputed property distinguishing human language from simple sequential strings is structure dependence: relationships between words (such as a pronoun and its antecedent, or the membership of constituents) depend on their structural position in a hierarchical, non-local arrangement—often visualized as a tree—rather than mere linear adjacency.
“the way that language works across all languages, it's not a string of pearls, but it's more like an Alexander Calder mobile. It has relationships that are dependent on where in a structure things are”
- The small set of 'closed-class' words like 'and', 'or', 'under', 'through', and 'not' are the most important targets for understanding because they are the glue that holds language together and carry the combinatorial structure, even though most research focuses on open-class nouns and verbs.
“things like "and," "or," "under," "through," "not," and that's where the fun begins... that's actually the glue that holds the stuff together.”
The brain performs complex coordinate transformations and variable-based computation even in small nervous systems and for mundane tasks.
- The Tunisian desert ant (Cataglyphis) navigates straight back to its burrow after a meandering search by counting its steps and tracking the sun's position (solar ephemeris), which requires plugging values into variables and computing—demonstrated by stilt experiments where elevated ants overshoot; if such a small nervous system performs variable-based computation, it is unreasonable to deny the same capacity to vertebrates.
“It actually takes a straight vector back and then looks for its hole. So, it must have, first of all, figured out how far it went. It has to count the steps”
- Recognizing even a single word requires the brain to seamlessly and rapidly translate between an auditory coordinate system (for hearing), an articulatory/motor coordinate system in joint space (for speaking), and an as-yet-unspecified coordinate system for meaning—making even mundane word knowledge a deeply complicated theoretical problem analogous to the coordinate transformations needed to reach for a glass.
“The articulatory code is in a different coordinate system than all the other ones, because it's in the motor system.”
- The brain converts a one-dimensional eardrum vibration into abstract ideas, words, and meaning, and does so at remarkable speed, extracting complicated information in segments of tens of milliseconds even in noisy environments.
“you can extract complicated information in segments of tens of milliseconds. How does that work at all?”
Rigorous science requires identifying ontological primitives and their interactions, not hypothesis-free big data or elegant but false simplifications.
- Genuine progress in understanding the mind and brain requires solid common sense, a good bullshit detector, and a willingness to do the hard technical work—becoming a connoisseur of the details rather than relying on shortcuts.
“it turns out you have to actually do the homework. You have to do the homework.”
- Poeppel is optimistic that within about ten years the field will get a grip on the elementary operations of composition or combinatorics that put items together into larger units (e.g., understanding that 'red can' is a can, not a red), but he remains deeply puzzled about the more fundamental problem of how the brain stores any information at all, which he considers one of the deepest mysteries of neuroscience.
“I'm actually pretty optimistic about that we're gonna get a grip on that, believe it or not.”
- Catchy, memorable one-line summaries of complex ideas—like the misreading that 'language is innate'—are fun to repeat but are almost always wrong, because real understanding requires engaging with nuanced, technical detail.
“If there are nugget-sized one-liners, they're fun to remember, they're fun to talk about, but they're probably almost always wrong.”
- Hypothesis-free big-data and machine-learning approaches in neuroscience, while powerful for classification problems, are essentially the mother of all regressions and risk making the field theoretically myopic—yielding only one form of answer and blinding researchers to theoretical alternatives—and ultimately require reinventing hypothesis-driven 'normal science' to interpret the model's parameters; engineering is thereby superseding science.
“If you take this approach, you're looking at an orgy of data that's almost treated hypothesis free. It's purely correlational approach. It's basically the mother of all regressions.”
- The right way to study the mind and brain, as in any science, is first to identify the ontological 'parts list'—the smallest primitives needed to generate the phenomena—and then to identify the forces or interactions between those primitives, recognizing that determining the true primitives is a multi-decade research program in which the relevant units keep getting smaller.
“your first job is to identify the parts list. What is the ontological structure of your domain?... What are the smallest elements that you need to use to generate the phenomena? And then you need to identify the forces or the interactions between the primitives”
Current neuroscience models of memory storage and language localization are empirically wrong yet persist, obscuring deeper mechanistic questions.
- The standard story that memory resides in synaptic connectivity and learning is the modification of those connections (work for which Kandel won a Nobel) is challenged by critics like Randy Gallistel, who argue neuroscience cannot even explain how the brain stores a single number like 17, because human memory needs both content-addressability (like word association) and address-addressability (digital location-based retrieval and variable-based computation), which the synaptic story does not adequately deliver.
“Our standard story in memory right now is it's the connections between the cells, that's the synaptic structure and the synaptic connectivity”
- The dominant neurobiological model of language—localizing production to Broca's area, comprehension to Wernicke's area, connected by the arcuate fasciculus—has persisted essentially unchanged since the 1860s, is empirically wrong (patients with given lesions do not show the predicted syndromes and brain wiring is far more complex), yet is still the model most neurologists reference today.
“We've had the same neurobiological paradigm for about 150 years since Broca and Wernicke, since the 1860s actually.”
- Scientific models can be powerful and useful precisely because of the elegance of simplicity, while still being empirically wrong; such models are 'doomed to be true for a while' until better-grounded alternatives emerge.
“models are doomed to be true for a while. In this case, this is very powerful because it has the elegance of simplicity but also is empirically wrong.”
- Functional MRI uses blood oxygenation as a proxy for brain activity and achieves spatial resolution near or below one millimeter, but at the cost of poor temporal resolution—on the order of one to several seconds—so it cannot capture the fast online dynamics of cognition.
“what do you give up for that wonderful picture? You give up temporal resolution.”
Vision and language are constructive, predictive processes that build internal representations from underdetermined input, not passive recording.
- Both vision and language comprehension are entirely constructive, predictive processes: the incoming data are vastly underdetermined and noisy, so the brain fills in and builds internal representations used for inference and action—making the camera/pixel metaphor for vision wrong, just as it is for hearing.
“Just as hearing and language comprehension is entirely constructive process, so is vision.”
- The dual-stream model of language processing, developed by Hickok and Poeppel, adapts the visual system's division into a 'what' (object identification) and 'where' (localization) pathway, positing one stream that maps sound to meaning (structure and content) and another that maps sound to articulation (motor output), because sub-specialized circuitry optimized for each problem is an efficient engineering solution the brain reuses.
“we adopted and adapted the standard model of vision actually... there's a "where" system and a "what" system.”
- The amplitude modulation spectrum of speech is 4 to 5 Hertz—the signal's loudness rises and falls four to five times per second—and this rate is independent of language, corresponding roughly to syllable rate, while music's modulation spectrum is slower at about 2 Hertz (~120 beats per minute).
“the mean rate of speech, across languages by the way, it's independent of languages, it's between 4 and 5 Hertz.”
Neuroscience faces fundamental unresolved questions about information storage and lacks adequate tools to capture brain dynamics at necessary scales.
- Describing the brain as a network of interacting neurons is a necessary research agenda but merely kicks the can down the road: it replaces 'a poorly understood blob of tissue' with 'poorly understood pieces of tissue connected by wires we don't understand,' so it is a metaphoric extension rather than a genuine mechanistic explanation.
“simply saying, "Well, it's a network," is kind of punting on the problem for me.”
- The human brain contains roughly 86 billion cells, each connected to between 1,000 and 10,000 others and communicating electrically and chemically, making the computational complexity of the system get out of hand quickly and explaining why adequate theories remain elusive.
“it has 86 billion cells and each cell has... each cell has between 1000 and 10,000 friends.”
- Magnetoencephalography (MEG) measures the magnetic fields generated by current flow in the brain using superconducting coils bathed in liquid helium, giving the most sensitive non-invasive measurement of human brain activity at millisecond temporal resolution, which complements fMRI's superior spatial resolution.
“it measures the magnetic fields generated by current flow in your brain, and it's the most sensitive technique we have to measure the human brain non-invasively.”
- A cortical column above one square millimeter of cortex, extending up about three millimeters, contains on the order of 100,000 neurons plus much additional cellular machinery, meaning even millimeter-resolution imaging vastly underestimates the underlying complexity.
“estimates are that it's on the order of 100,000 neurons. And that's just the neurons.”
Behaviorism remains the default position in neuroscience despite Chomsky's mid-century critique, and animal research is essential yet under threat.
- There is no reason to believe that the way we understand a concrete concept like 'cat' is any easier than how we understand an abstract concept like 'honesty'; the untutored intuition that concrete words are the easy case is unwarranted, since we do not actually understand how either is processed.
“There's no reason to believe that the way we understand "cat" is easier than the way we understand "honesty."”
- Despite Chomsky's critique, behaviorism remains deeply embedded and is arguably the default position in the neurosciences today.
“the most disturbing part of the story of behaviorism is that it's still around.”
- Careful, responsible, ethically executed animal research is indispensable for understanding basic principles of physiology and has no alternative, yet irrational and vitriolic debates—especially in Europe—are leading to a sharp and dangerous reduction in such research.
“I'm 100% enthusiastically in favor of careful, responsible, ethically executed, well-managed animal research. There is no alternative for it”
- Because everyone speaks a language and has powerful intuitions about it, the public treats itself as expert and an astonishing amount of nonsense is promulgated about language—whereas no one disputes that we have a dedicated visual system, even though both are equally products of the vertebrate brain.
“as soon as we talk language, everyone is an expert. Everyone speaks a language. Everyone has a powerful intuition, and the amount of nonsense promulgated, it's astonishing.”
- Chomsky changed psychology, language sciences, and philosophy of mind in the mid-1950s by effectively ending behaviorism—which reduced the mind to the single principle of association underlying conditioning—and successfully arguing for a mentalist stance toward psychology.
“Based on the series of his early books and papers, he, first of all, effectively got rid of behaviorism.”