Jay McClelland on How Thought Emerges From Neural Networks

Jay McClelland, a cognitive scientist at Stanford and a central figure in the neural-network tradition, opens with Lex Fridman on what has stayed beautiful about these systems, biological and artificial, over decades. The fundamental thing, he says, is how they let us link biology with the mysteries of thought. When he entered the field in the late 1960s and early 1970s, cognitive psychology had just named itself. A 1967 textbook called *Cognitive Psychology* treated the nervous system as of only peripheral interest, something that would not tell us anything about the mind. McClelland did not agree. He is a physical being, dust to dust, and somehow he emerged from that.

He used to talk about awakening from the Cartesian dream. Descartes, walking in the gardens of Versailles, stepped on a stone and a statue moved, then another. The gardeners explained a hydraulic system: contact redirected water, water moved the statue. Descartes built from that a mechanistic theory of animal action, with the body's non-blood-carrying fibers as little hydraulic tubes that sent pressure and produced movement. The human had that animal body, he thought, but some divine something else had to come down and be placed in him to give him the ability to think. The physical world included the body in action. It did not include thought.

Physiology therefore studied sensory and motor systems, things you could measure when you stimulated neurons. Cognition was tied to abstract algorithms. McClelland had learned physiological mechanisms as an undergraduate; sitting in cognitive psychology as a first-year PhD student, he thought: wait a minute, the whole thing is biological. Fridman presses on whether that is magical, that a little biology yields the full beauty of human experience. McClelland says it is obvious and not obvious at the same time. The magic, for him, is how nature could give rise to organisms with the capabilities we have.

He reaches for Darwin. Darwin knew early that none of the ideas on offer explained how evolution could work, spent years collecting hints he did not yet understand, and drew bird after bird that differed slightly. After he figured it out he had nightmares about the complexity of the eye, and about the argument that such an organ could not have emerged from an unguided process. He delayed publishing in part because he was scared of his own ideas. He did not think they could possibly be true. By the twentieth century, Descartes' picture started to look wonky: apes, chimpanzees, and bonobos are pretty smart in some ways.

People still looked for a leftover difference. Someone claimed there was no hippocampus in the monkey brain, only in the human. In the late nineteenth century Thomas Huxley had to perform a surgery in front of a crowd to show that a chimpanzee brain has a hippocampus. Continuity of species is another piece of the idea that we are a total product of nature. Vision science got comfortable with that continuity earlier: a monkey's visual pathway from eye to the first cortical areas is extremely similar to a human's. The leap, Fridman says, seems to happen on the cognition side. Chomsky argued that a genetic fluke about a hundred thousand years ago gave some hominin a tweak that resulted in language, and language then separated us from every other animal.

McClelland thinks there is a lot of truth in the value of language, but that it comes along with sociality, mutual engagement, and rich mechanisms for organizing an understanding of the world, which language then plugs into. Language is a tool for collective intelligence; whatever sits at the core of that collective capacity is the main thing. Evolution, he adds, is not the smooth 500-million-year slope people imagine. The fossil record shows hundreds of millions of years of stasis and then a change that is sudden on geological time, a million years or even ten. Punctuated equilibrium felt right to him about mental stages too: a four-year-old and a seven-year-old think in profoundly different ways, as Piaget's stage theory tried to capture. Insight, transition, a new stage after years of experience: complex-systems research has a language for that, and it joins evolutionary biology, developmental biology, and cognitive development.

The other half of the mystery is assembly, not just learning. Deploying a finished brain into the world is already remarkable; how the learnable mechanism itself is built, from embryo to pathways, is if anything harder from an engineering point of view. Work on brain organoids, growing many miniature brains to watch a mechanism self-assemble from DNA, strikes both of them as almost outrageous. Biological programs generate a mushy system robust enough to learn in an unpredictable world, and to learn a huge number of things that enable survival. From brain to mind, McClelland says, the whole thing seems pretty continuous.

In 1974 McClelland took an assistant professorship at UC San Diego. David Rumelhart and Don Norman had just published *Explorations in Cognition*, chapters on interesting questions about the mind in a completely abstract, non-biological way. He loved the lightness of it, and Norman's spirit of playful exploration, but he was still trying to get from neurons to cognition. At a conference he heard James Anderson, an engineer turned psychology professor, use linear algebra to build neural-network models of perception, categorization, and memory. It blew him out of the water: a model simulating neurons that remembered, recalled, and recognized, not just an abstract stepwise algorithm. Walking across campus in 1977 he felt, he says, like Saint Paul on the road to Damascus. If he thought about the mind as a neural network, it would help him answer the questions about the mind he was trying to answer.

Others were catching the same excitement: Anderson, Steve Grossberg, who had been writing about neural networks since the 1960s, and Geoffrey Hinton. Hinton's PhD dissertation landed in a postdoctoral applicant pool that Rumelhart and Norman were administering; Rumelhart got very excited, and Hinton became one of the first postdocs funded by their grant. Paul Smolensky was another. Hinton and Anderson organized a conference at UCSD in 1979 or 1980 called Parallel Models of Associative Memory, gathering everyone who had been thinking this way. Francis Crick, at the Salk Institute, heard about the group from Hinton, who was known among Brits to be brilliant, and came too.

Fridman asks whether *parallel* was, at the time, almost synonymous with how they thought about neural networks. McClelland says each neuron is an independent computational unit: it gathers data, integrates it, produces a result. A brain has on the order of a hundred billion of these little computers, all capable of working at once, instead of a single central processor going chug, chug, one step after another. That was a revolutionary way to think about computation next to theoretical computer science, which analyzed sequential algorithms on a single machine. A convolutional net makes the picture concrete: layers of thousands of neuron-like units, from a photoreceptor-like pixel array up through stages of abstraction to a classification. Open your eyes and see a cup, a phone, a bottle; these systems now do that. Each unit contributes a tiny bit. That is parallel distributed processing, and that was the excitement: from simple things, with levels of abstraction stacked on them, the beautiful features of cognition might emerge.

Rumelhart came from a tiny town in western South Dakota. His mother was the librarian, his father the newspaper editor; four brothers grew up competing in sports and in mind games, Sudoku and chess. He went to college young at the University of South Dakota, majored in mathematics, and somehow got interested in psychology. Stanford's mathematical psychology program took him. That field used equations, not simulations: stipulate principles, derive a prediction, solve for the probability the subject is correct on the seventh trial. It was a prestigious corner of psychology, and Stanford was its center. In 1967 he became one of the first assistant professors in UCSD's new psychology department. By the time McClelland arrived seven years later, Rumelhart was still doing mathematical modeling but had turned toward cognition, and toward a question that still sits at the center: what does it mean to understand, and how would we know if we really did?

He wanted machines that could hear a couple of sentences and have an insight. A favorite example: Margie is sitting on the front step when she hears the familiar jingle of the Good Humor man; she remembers her birthday money and runs into the house. What is she doing, and why? The natural reading is that the ice-cream truck is coming, she likes ice cream, she needs money, so she runs in to get it. A huge amount of inference has to happen to link those pieces. Rumelhart tried to build good-old-fashioned AI models of language and of content such as *has money*, formal logic and knowledge bases, and by the late 1970s that paradigm was not giving him answers. He started thinking that we needed systems in which multiple simultaneous constraints could influence each other.

He wrote a paper McClelland first shrugged at and later could not get out from under his skin: *An Interactive Model of Reading*. Every level of what comes off the page depends on every other: pixels, letters, words, the ideas the author is trying to convey. At first Rumelhart imagined a set of experts, one for building lines from pixels, one for letters, one for words, one for meanings, all looking at the data and updating hypotheses at other levels. The word expert can tell the letter expert there should be a *T* because the word is probably *the*; the bottom-up feature expert can agree; if they agree, you see a *T*. Top-down and bottom-up, all layers at once, a bidirectional parallel process that is still hierarchical. After reading Hinton's dissertation and hearing about Anderson's models, Rumelhart realized those experts could be replaced by neuron-like units whose connection weights did the job.

He and McClelland built the interactive activation model of letter perception: pixel-level input, line-segment features, letters, and words, all made of neuron-like units linked by weights. The unit for the word *time* connects to *T* in the first position, *I* in the second, and so on, and the connections are bidirectional. Prior knowledge that it might be *time* primes letters and features; without that, the process starts bottom-up. Direction depends on where information arrives first. Context and features together can converge on an emergent perception. That piece of work, McClelland says, got them both completely convinced that the neural-network way of thinking could address the questions they cared about as cognitive psychologists. The algorithmic details were details. The idea that you could get far this way was itself profound.

Fridman asks whether McClelland likes the term *connectionism*. He thinks it is useful because it highlights that the knowledge the system exploits is in the connections between units. There is no separate dictionary. The unit for *time* is not that unit for any reason other than its connections to the letters that make the word. When it is excited it represents support for the hypothesis that *time* is in the input. The word is not written anywhere inside the model; it is only written on the picture they drew, to say that is the *time* unit. If you want to know how the word is spelled, you have to use the connections going out to retrieve the letters. Humans want to think in logic. Connectionism is weird. It is weird that this is how it works.

Go back to the convolutional net. It will say *cat* or *dog* and have no idea why. Between the first layer and the last there are only connections; features get more abstract as you go, in a graded, continuous way, and at the end you still have a pattern of activation over an array of units. McClelland opens his eyes and sees Fridman, or his own dog, a member of a species with many others, known by slightly unique characteristics he cannot describe. He could write a few words about a brand of car and you still would not know which car he meant. Propositional knowledge of what lets us recognize a particular instance never worked. You could not write down the propositions for visual recognition. In that space it always seemed natural that something more implicit is going on. You do not get access to the details of the computation. You get the result. You cannot read the contents of the connections. The connections only cause outputs from inputs.

Does that mean connections can contain the depth of human knowledge, the kind of understanding Rumelhart wanted? McClelland calls it the 64-dollar question, maybe the 64-billion-dollar question with inflation. He used to tell people he was a radical eliminative connectionist, because he did not want anyone to think he wanted to build things into the machine. He dropped *eliminative*. It made it seem wrong to think there is an emergent level of understanding, and he disagrees. He would call himself a radical emergentist connectionist. Higher-level aspects of cognition are real, but they do not exist as such. Douglas Hofstadter's sand-dune example helps: we can treat dunes as entities and even count them, yet a dune is a pile of sand that moves in the wind and reforms. Thoughts are like that. They exist, but we should not treat them as monolithic objects we can understand without the lower-level stuff that lets them change. If the substrate is parallel, distributed, and connectionist, the contents of thought can still be abstract and symbolic. They are just more fluid than a set of logical expressions easily captures.

Rumelhart died about fifteen years before this conversation, and his demise, McClelland says, was one of the most poignant tragedies relevant to the science they had done together. He underwent a progressive neurological condition that is not fully understood in his particular case: certain scans were not done, and the family did not want an autopsy. McClelland had already become interested in a disorder that might have been the one Rumelhart was succumbing to. His colleagues call it semantic dementia: a specific, progressive loss of mind related to meaning. Patients lose the ability to appreciate the meaning of experience from touch, sight, sound, or language. They hear sounds and do not know what they mean.

It starts with similar breeds of dog, or low-frequency, unfamiliar categories. Then it becomes striking. Patients lose pigs, goats, and sheep and call all middle-sized animals dogs; they lose rabbits and rodents and call the little ones cats; they lose hippos and cows and call them horses. One patient reached a point where any four-legged animal was a horse, a dog, or a cat, assigned by size. McClelland's collaborator Karalyn Patterson developed the pyramids and palm trees test: show a pyramid, ask whether a palm tree or a pine tree goes with it. The loss shows up with words or with pictures. Patients cannot relate pictures to each other or match words to pictures. Conceptual grounding from either modality is gone. That is why it is called semantic dementia: the semantics themselves are disintegrating. Distributed representation explains it. A pattern of activation stands for a concept; similar concepts have similar patterns; as you degrade them you lose the differences. The difference between dog and goat drops out of the pattern. Dog is familiar, so dog remains.

Rumelhart was going through this. McClelland likes to honor him by analogy with Stephen Hawking: both suffered degenerative conditions, Hawking in the motor system, Rumelhart in semantics, and perhaps in the semantics of the self. The most focused, goal-directed, competitive, thoughtful person McClelland knew, willing to work for years on a hard problem, started to disappear. For a while it was hard for any of them to appreciate that he was not fully there. Asked whether Rumelhart could introspect the fading of the very mind he had spent a life trying to understand, McClelland says he did not, and that was part of the tragedy from the outside. At some level he sort of did, once it was clear this was a biological condition and not a distracted day.

He retired from Stanford, lived with his brother for a couple of years, then moved into a facility for people with cognitive impairments. In the late 1990s and around 2000 McClelland would spend time with him. They went bowling; Rumelhart could still bowl. Afterward McClelland asked if he wanted Wendy's. No. Where, then? Rumelhart pointed: turn here. He still had enough spatial cognition to get them to the restaurant. At the table he could not come up with any of the words, but he knew where on the menu the thing he wanted was. He could not say what it was. He knew that was what he wanted to eat. Cognition is not monolithic. It is graded, and it is multipartite; partial competencies survive after others go. McClelland has always been fascinated by what used to be called cognitive neuropsychology, especially this gradual disintegration. He is a big believer that losing a human being you value is as powerful as first falling in love with that person. It is all a celebration of the human being. The disintegration itself, too, is a celebration.

In 1982 Hinton came back to San Diego on sabbatical. The interactive-activation papers had just been published. McClelland, Rumelhart, and Hinton started a research group they called the PDP research group. Crick came. Smolensky was still there as a postdoc. Hinton talked to them about learning. The problem with the way they had been approaching it, he said, was that they had been looking to biology for inspiration about how synapses should change the strengths of their connections. That is the wrong way. What you should do is think in terms of how you can adjust connection weights to solve a problem: define the problem, then figure out how the adjustment of the weights will solve it.

Rumelhart heard that and said to himself, all right, I will start thinking that way. Imagine an objective function: I want my machine to classify all these images correctly; I can score how well it does on each one and get a measure of error or loss; then I figure out how to adjust the weights so as to reduce that error. Engineers already knew gradient descent. Bernie Widrow, in Stanford's electrical engineering department, and a collaborator named Hoff had invented the delta rule: for a single layer of weights, define the difference between target and output and change the weights to reduce it. What Rumelhart did was generalize that so he could also change connections from earlier layers into a hidden layer between input and output. He first called the algorithm the generalized delta rule, because it was just an extension of gradient descent.

Interestingly, Hinton thought this was not going to work very well. He had his own alternative, based on the Boltzmann machine; that learning paper came out in 1985. Backprop worked better. The name is opaque until you see the mechanics. To figure out what changes you needed from input to hidden layer, you had to send error signals from the output back through the hidden-to-output connections, so each hidden unit got its own error signal. Rumelhart formulated it that way: we know the error at the output; let us see if we can get a signal at the hidden layer that tells each unit how to change its incoming weights. Stop thinking about biology for a second, Fridman summarizes, and start thinking about optimization. McClelland agrees. The move came from Hinton introducing the idea of an objective function and a derivative, then Rumelhart actually writing the multilayer rule.

What stands out from Hinton, McClelland says, is not only the learning argument. Two 1981 papers were way ahead of their time: one was essentially the idea of transformers, the other an early paper on semantic cognition that inspired Hinton, Rumelhart, and McClelland through the 1980s and still grounds McClelland's thinking about meaning. An unpublished 1977 note, written when Hinton was a PhD student, described how a neural network could do recursive computation. When you call a subroutine you need to save the caller's state. Hinton's idea was to save that state by making fast changes to connection weights, then use those fast changes to reinstate the previous context when the subroutine finished. McClelland always thought that was extremely creative, and Hinton kept exploring it.

Another influence runs through McClelland's later work on mathematical cognition: Hinton does not write too many equations. In Hinton lab meetings, the story goes, you do not get up at the board and write equations the way you do in everybody else's machine-learning lab. You draw a picture. Hinton explains gradient descent by putting his hands together into the shape of a ravine: you are coming down the wall; if you take too big a jump you land on the other side, which is why you turn down the learning rate. It speaks, McClelland says, to the fundamentally intuitive character of deep insight, together with a commitment to understanding that is ultimately explicit and clear, but also intuitive. Fridman compares the effect to being around Richard Feynman, or the chip designer Jim Keller: the process of generating ideas in that person's presence transforms your own work.

Hinton is a descendant of the logician George Boole, from a long line of English academics. Alongside the visual intuition, he has wanted to contribute to the understanding of reasoning itself, not just human reasoning: what can we conclude from what else, and how do we formalize that. The Boltzmann machine work of the early to mid-1980s was his way of connecting that Boolean tradition to a continuous, probabilistic, graded constraint-satisfaction world, ideas linked with theoretical physics as well as with logic. McClelland has always been inspired by it. If neurons are probabilistic rather than deterministic, that may be part of the serendipity of insight: the thought might not have occurred at that instant; it might be partly the result of a stochastic process. That, too, is part of the magic of emergence.

McClelland started using the phrase *computational intelligence* for the field he thinks Hinton and many people at DeepMind are in. He is a human-oriented researcher in that space, interested in the human solution, but a huge amount of the excitement of deep learning is the idea that we may go beyond what our own nervous systems can achieve, scaling the mechanisms of human intelligence. That, he says, is being played out massively at Google Brain, OpenAI, and DeepMind: the computations used to succeed at Go or protein folding. Fridman notes we still have more synapses than the machines. McClelland agrees. Humans are still beating the AIs on that count. They are catching up pretty quickly.

Fridman turns to a paper, *A Parallel Distributed Processing Approach to Mathematical Cognition*, whose introduction cites Tristan Needham: treating mathematics as symbol manipulation is like studying music without ever hearing a note. What, then, is mathematics? McClelland thinks of it as a set of tools for exploring idealized worlds that often turn out to be extremely relevant to the real world, but need not. In those worlds objects have idealized properties, and relations among them can be characterized with precision, so the implications of certain facts let you derive others with certainty. If two triangles share an angle and the adjacent sides match, they are congruent: they have all their properties in common. That is something about triangles, not about formulas. We build bridges out of triangles. We measure heights we cannot climb. We launch a speck of matter from Earth to intersect a tiny body beyond Pluto at a predicted time and date. Those ideas make contact with the physical world.

The same is already true of the natural numbers. He likes to include zero, so he says the non-negative integers. They let you be exact about how many sheep you have: I sent you out with 23, you came back with 22, what happened? It is a fundamental problem of human society that you had better bring back the same number you started with. Numbers allow commerce, contracts, records. They are abstract, idealized, and generalizable, and at the same time potentially very grounded and concrete. One of the things that makes for the incredible achievements of the human mind, he says, is that humans invented these idealized systems that leverage the power of thought. Mathematics is the development of systems for thinking about the properties and relations among sets of idealized objects. The notation we focus on too much is just our way of expressing propositions about those properties, like language: invented for communication, not necessarily the deep representation of the ideas.

How do you model that kind of reasoning? The insights human mathematicians have had, he says, combine intuitive, connectionist-like knowledge that makes something obviously true, so you do not have to think about why, with a next step of pondering something you did not yet have an intuition for. That new thing then becomes part of the intuition the next generation stands on. Walking in the woods with his wife in a Northern California state park, he saw a quotation from Henri Poincaré carved on a bench: it is by logic that we prove, but by intuition that we discover. The essence of his project is to bring intuitive connectionist resources to bear so that discovery can arise from engagement with the formal system. People like Hinton, Newton, Einstein, Rumelhart, Poincaré, Archimedes: a flash of insight is a constellation of simultaneous constraints that settles the mind into a novel state. Then you ask how to prove it, how to write the steps that make it rigorous.

Deep learning has started to give him hope that this can happen in machines. Networks trained on massive text, given a prompt, generate fanciful, creative stories. They seem to have synthesized, from billions of particulars, something like an intuitive sense of what would be a fun little story; it wells up from letting the system play out what someone might say. That is the potential of capturing the intuitive side. Fridman adds AlphaZero: novel chess ideas that feel like brilliant moments, with search on the old-fashioned-AI side and a network that has the intuition of looking at a board and saying how good this is. Grandmasters have that positional sense without deep search, then do a little of what they call calculation. Demis Hassabis has described a move in a Go match against Lee Sedol in a similar way, which made McClelland eager to collaborate with people at DeepMind. What he wants to emphasize about mathematical cognition is older still. Philosophers and logicians going back more than three thousand years constructed the idea of thinking formally. Euclid's *Elements* became the touchstone: a coherent document that laid out a formal system and a method of inference. The ability to think that way is itself the result of immersion, the way language is. Jerry Fodor took systematic thought to be the essential characteristic of the human mind. McClelland takes it as derived and acquired, the result of acculturation in a mode humans invented.

Is it unfair to draw a line between mathematical cognition and linguistic cognition? McClelland thinks it is important to draw the line, then come back and look at the subtleties. Chomsky was born into an academic family; his father taught rabbinical studies at a small college in Philadelphia. He was deeply enculturated in a culture of thought and reason, and he brought a profound engagement with formal systems to the effort to understand natural language. That gave him tremendous power and some amazing insights into linguistic structure. But, McClelland says, the actual intuitive knowledge of these things only goes so far, and not as far as it does in people like Chomsky himself. Lila Gleitman, trained in the same linguistics department, found in her dissertation that linguists' intuitions about even the meaning of a phrase, not just its grammar, were very different from those of an ordinary person who was not a formally trained thinker.

His current hypothesis is that formally trained linguists, philosophers, cognitive scientists, computer scientists, machine-learning researchers, and mathematicians have a mode of engagement with experience that is intuitively organized around systematicity and conformity with the principles of a system, more than is true of the natural human mind without that immersion. The fields through which you study the mind can take you away from its natural operation. That makes it very difficult to be someone who introspects. It is related to beginner's mind, and to a more prosaic plague called the expert blind spot: things so self-evident to the expert that they cannot begin to think how they would explain them, because it is just patently obvious that they must be true.

When Leopold Kronecker said God made the natural numbers and all else is the work of man, he was expressing that intuition: discrete quantities being countable and indefinite in number was not something that had to be discovered. He was wrong. For a long time many cognitive scientists agreed with him. Natural numbers were treated as part of the innate endowment of core knowledge, in the language Elizabeth Spelke and Susan Carey used for innate primitives. They no longer believe that. It is more or less accepted that the natural numbers are a cultural construction, which makes it interesting to study the few people who still exist without those systems. A certain mode of thinking about language or geometry becomes so second nature that you do not know what you need to teach, and in fact we do not teach it very explicitly. In a math class some students catch the way of thinking and can do the problems the professor finds interesting; most who do not engage as deeply never get there. We decide that person must be brilliant, must have a different biological bit. McClelland does not want to dismiss individual biological differences completely. He finds it more interesting that the dinner-table conversation in the Chomsky house may have given him that cast of mind.

Fridman asks how you leap back out of a structure you have spent decades building. One thread of psychological research McClelland has always needed to keep in view is that our explanations of our own behavior are not necessarily part of the causal process, or even valid observations of the constraints that led to the outcome. They are post-hoc rationalizations we can give, from information at our disposal, when asked. The idea comes from a paper by Richard Nisbett and Timothy Wilson on the limits of our awareness of the factors that cause our choices. Insight into why we hold beliefs, make choices, and feel feelings is not something we totally control or observe. It is subject to a culturally transmitted mode of explanation. Even the sense that we can introspect and have access to our own thoughts is a product of culture and practice.

Asked for advice to young people, McClelland says it is about finding the thing you are intrinsically motivated to engage with, then celebrating that discovery. In college he had thought he had to be a psychiatrist, because Freud, Erich Fromm, and popular psychiatry were what he had of the psyche in high school, and psychiatrists went to medical school. First-semester mechanics at Columbia was far from what interested him and too early on a winter morning, so he never made it to physics class. He wandered through freshman year and most of sophomore year until 1968, when the university blew up over Vietnam, a gym in Morningside Heights, Black activism, and SDS. He had happened to take mathematical statistics and Psychology 1. Questions about why some people occupied buildings and others did not ran together in his mind and gave him a path. His psychology professors thought he was great, nominated him for Phi Beta Kappa, and at the ceremony someone asked if he was magna or summa; he was not even getting honors on grades. They had just thought he was interested enough in ideas.

He does not want anyone to be sanctimonious about it. He really enjoyed statistical analysis of data. He really enjoyed running his own experiment, which chemistry and physics had never suggested mere mortals could do except from a textbook. In Psych 1 they had a rat, and after two set experiments you got to do something you thought of. Opportunity to do it himself, plus a set of things that engaged him, is close to why some people become musical geniuses: immersion early enough that it gets into the fabric. His little brother had that for music. At about 13 months he put records on the phonograph and recognized which one he wanted, not by reading labels but by the scratches. Something connected. You are extremely lucky if you have that and can nurture it. Motivational engagement produces the immersion that creates the chance at expertise. McClelland calls the synergy of resonance plus experience the Mozart effect, not a six-week-old handed a violin, but something in the child that the music answered. His siblings all went different ways: a musical brother, an intuitive-engineer brother, a sister passionate about water conservation long before it was the issue it is now. People can be excited about fitting people on bicycles as well as about neural networks.

Fridman asks whether it makes him sad that psychiatry is not the way he explores the mind. McClelland dropped pre-med before he met the biomedical side. Around 2000 he sat on the National Advisory Mental Health Council. A new NIMH director came in saying schizophrenia is a biological illness, a lot like cancer; we have made huge strides in curing cancer, we will find the medications, we are not going to listen to anybody's grandmother anymore, and good old behavioral psychology is not something we will support. It alienated McClelland from the institute. Twenty years later, he says, nothing has changed in our ability to help people with mental illness. That path was a dead end. The flag of deep thinking about the mind, Fridman suggests, moved to people like McClelland, in cognitive psychology with a foot in computation. McClelland found more of the human condition as an undergraduate in sociology and literature than in the pigeon lab. Graduate school and cognitive psychology resolved the dissonance: you could think about the human condition and still do mechanistically oriented work. The richest opportunities now, he thinks, sit in systems and cognitive neuroscience and computational intelligence, not at the very bottom of the neuron.

Is he afraid of death? Less of death than of degeneration, in part because he has watched tragic decline. It is exciting to keep participating, to feel near where the wave is breaking on the shore. Dementia, Alzheimer's, semantic dementia: you gradually lose the thread, and you can live a decade after you have to retire. That is what he fears most, the breaking of the wave, ideas no longer swimming around to be played with, collaboration gone. The edge of the cliff itself, the mystery of the end, feels less discrete to him than it seems to most people, because of his graded, continuous way of thinking about mind. Limited time can motivate. Seven or eight years before this conversation he was doing decision-making experiments that gave satisfying closure, models that fit, predictions you could take to monkeys. He realized he might have only ten or fifteen useful years left, and he did not feel he was getting toward the interesting questions at that level of work. That is when he told himself to pick something hard, and started on mathematical cognition. Imagine it is only ten years, then maybe three or four more, then maybe another five. Give yourself six or eight. A deadline will not go on forever. Keep thinking about the questions that are actually interesting.

What does he hope the paragraph in the encyclopedia says? That he was able to see a path more exciting than the one already laid out for a cognitive psychologist, not for any super-special reason other than having had the right prior context, and that he followed the lead. In a preface he wrote that the joy of science is the moment a partially formed thought in one person's mind gets crystallized a little better in discourse and becomes the foundation of some exciting concrete piece of scientific progress. That happened when he and Rumelhart built the interactive activation model, and when Rumelhart heard Hinton talk about gradient descent and an objective function to guide learning. It happened a lot in that period. He still seeks it with students. He hopes the paragraph includes a person who contributed by finding exciting collaborative opportunities, and by stepping a little off the beaten path.

One episode still energizes him to share. His PhD dissertation was completely empirical. He wrote a paper on the two main experiments, and at the end laid out the beginnings of a theory of the data. He sent it to the *Journal of Experimental Psychology*. The editor wrote back: these are great experiments, we would love to publish them, but please leave the theorizing to the theorists and take that part out. He took it out. He almost found himself labeled a non-theorist. He could have succumbed and decided his job was just to go on doing experiments. That is not what he wanted. In his first year as an assistant professor he still ran experiments, because he needed papers, but he also elaborated that section and sent it to *Psychological Review*, the theoretical journal. They did not accept it either. They said it was interesting, and that this time he should keep thinking about it. That was what got him going: contributing to theory is not a superhuman thing. A mere mortal can do it.

The broader lesson is not to succumb to anybody labeling you. Especially as you become successful, labels get assigned for the thing you were successful at: connectionist, cognitive scientist, not a neuroscientist. Those are stories of the past. Today you are a new person who can revolutionize totally new areas. Do not let the labels hold you back. As for the meaning of it all, the hundred-plus billion people who have lived, why we are busy and directed somewhere: McClelland thinks we make meaning for ourselves, and find inspiration in meaning other people have made. The great religious thinkers of the first millennium BCE, and a few later, laid down important foundations. He believes we are an emergent result of a process that happened naturally, without guidance, and that meaning is what we make of it. Efforts to refine meaning in religious traditions are part of the expression of that goal: not to find out what the meaning is, but to make it ourselves.

It is personal and individual. Meaning comes through the particular combination of synergistic elements that are your fabric, your experience, your context, and it is all made in a local pocket. Here he is at UCSD with this brilliant man Rumelhart, who is having doubts about symbolic AI that resonate with McClelland's desire to see cognition grounded in biology. Make the most of that. From that little pocket came a peculiar emergent process, which is basically each one of us: cells come together, tell fancy stories about themselves, and enjoy the beauty of those stories. The process lives for a time, defined by its local pocket in time and space, tells pretty stories, we write them down, we celebrate how nice they are, and it continues because we build stories on top of each other. Eventually we will colonize other planets, other solar systems, other galaxies, and tell even better stories. It all starts here on Earth.