What It’s Like to Brainstorm with a Bot

Contrary to what many of my friends believe, good academics are always working—at least in the sense that when we’re stuck on a problem, which is most of the time, it’s impossible to leave it behind. A worthwhile problem is a brainworm: it stays with you until it’s resolved or replaced by another one. My Dartmouth colleague Luke Chang, a neuroscientist who studies what happens in people’s heads when we communicate, is no stranger to this affliction. One day, on a long drive back to Hanover, he found himself preoccupied with such a worm. The drive up I-89 is usually uneventful—a straight shot north, ideal for letting your mind off the leash. But Luke’s mind snagged on a technical challenge: how to turn a decent model of facial expression into something truly convincing. The aim was to encode the various nuanced ways human faces transmit states of mind, and then to visualize them; smiles and frowns are the barest beginning. The spectrum of human emotions and intentions is embodied in a range of expressions which serve as a basic alphabet for communication. He’d been trying to integrate facial “action unit” measurements into his software. But visualization was proving tricky. Instead of lifelike faces, his code kept spitting out cartoonish sketches. Every recent attempt had ended in disaster, and it was driving him crazy.

Years ago, Luke might have gnawed at the problem alone for the length of the drive. This time, he decided to hash it out with his newest collaborator: ChatGPT. For an hour, they talked. Luke broke down his model and described where things were going wrong. He floated questions, speculated about solutions. ChatGPT, as ever, was upbeat, inexhaustible, and, crucially, unfazed by failure. It made suggestions. It asked its own questions. Some avenues were promising; others were dead ends. We sometimes forget that the machine is less oracle than broad interlocutor. The exchange wasn’t quite spitballing; it was something more organized—human and machine feeling their way through the fog together. Eventually, ChatGPT suggested Luke look into a technique called “disentanglement,” a way of simplifying mathematical models that have grown unwieldy. The term triggered something in Luke. “And then it starts explaining it to me,” he recalled. “I’m, like, ‘Oh, that’s really interesting.’ Then I’m, like, ‘O.K., tell me more—conceptually, and, actually, how would I implement this disentanglement thing? Can you just write some code?’ ”

It could. It did. When Luke got back to his office, the code was waiting in the chat. He copied it into his Python script, hit Run, and went off to a lunch meeting. “It was such a delight to learn a new concept and build it and iterate it,” he told me. “I didn’t want to wait. I just wanted to talk about this then.” And did it work? It did. “That to me has just been such a delightful feeling,” he said. “I feel like I’m accelerating with less time, I’m accelerating my learning and improving my creativity, and I’m enjoying my work in a way I haven’t in a while.” That’s what a good collaborator can do—even, these days, if it happens to be a machine.

Much has been made of the disruptive effects that generative A.I. is having on academic life. As a professor of mathematics and computer science at Dartmouth, I hear the anxiety firsthand. It’s just the latest uneasy chapter in the long history of inventions meant to help us think. These tools have rarely been welcomed with open arms. “Your invention will enable them to hear many things without being properly taught, and they will imagine that they have come to know much while for the most part they will know nothing. And they will be difficult to get along with, since they will merely appear to be wise instead of really being so.” That’s from Plato’s Phaedrus, where Socrates presents, with sympathy, the case against the treacherous technology of writing. It could have been written yesterday, as a warning against gen A.I., by any number of my own colleagues.

The academy evolves slowly—perhaps because the basic equipment of its workers, the brain, hasn’t changed much since we first took up the activity of learning. Our work is to push around those ill-defined things called “ideas,” hoping to reach a clearer understanding of something, anything. Occasionally, those understandings escape into the world and disrupt things. For the most part, though, an “ain’t broke, don’t fix it” attitude prevails. Socrates’ worries reflect an entrenched suspicion of new ways of knowing. He was hardly the last scholar to think his generation’s method was the right one. For him, real thinking happened only through live conversation; memory and dialogue were everything. Writing, he thought, would undermine all that: it would “cause forgetfulness” and, worse, sever words from their speaker, impeding genuine understanding. Later, the Church voiced similar fears about the printing press. In both cases, you have to wonder whether skepticism was fuelled by lurking worries about job security.

We don’t have to look far, in our own age of distraction and misinformation, to see that Socrates’ warnings weren’t entirely off the mark. But he also overlooked some rather large benefits. Writing—helped along by a bit of ancient materials science—launched the first information age. Clay tablets were the original hard drives, and over time writing more than earned its keep: not just as a tool for education and the development of ideas but (to address what Socrates might really have been worried about) as a tremendous engine for employment in the knowledge economy of its day, and for centuries after. For all that, writing never did supplant dialogue; we still bat around ideas out loud. We just have more ideas to talk about. Writing was, and remains, the original accelerator for thought.

Still, for all its creative utility, writing is not much of a conversational partner. However imperfectly, it captures what’s in the writer’s head—Socrates called it a reminder, not a true replication—without adding anything new. Large language models (L.L.M.s), on the other hand, often do just that. They have their own pluses and minuses, and the negatives have received plenty of airtime. But Luke’s story, and those of a growing cohort of “next-gen” professors (Luke was recently tenured), reveal what’s genuinely novel: these new generative-A.I. tools aren’t just turbocharged search engines or glorified writing assistants. They’re collaborators.

A few years ago, Luke would have been driving back from Concord, barely seeing the landscape as he turned his code over in his mind. Some ideas would stick, most would vanish—maybe even a good one or two lost to the ether. That’s just how memory works. Now, with an A.I. assistant riding shotgun, he can talk through the problem in real time. The result isn’t just an idea but an actual, executable script—waiting for him back at the office, ready for immediate testing.

Luke was, it’s natural to say, working with ChatGPT. Some would say that he was merely “using” it, but if you subjected their exchange to a Turing test for collaboration, it would probably pass—even if the “entity” on the other side showed a breadth of knowledge no human colleague could match. Was this co-creation? If Luke had been driving with a friend, we’d likely say yes. Two colleagues, bouncing from prompt to prompt, nudging each other along until someone stumbles onto a key that finally turns a lock. It’s easy to picture Luke restlessly shifting from one idea to the next, until, at last, the “Aha!” arrives. But whose “Aha” is it?

How you answer this may depend on what you think it means to have an idea. Where do ideas come from? There are little ideas and big ones, their size determined by how much they rearrange our understanding—of the world or of ourselves. Some ideas are about forging connections, like Luke’s insight about disentanglement. Others work through analogy: hearing a story in one context and rewriting it for another. We use what we understand to make sense of what we don’t.

In the nineteen-twenties, the challenge of understanding how infectious diseases spread led W. O. Kermack and A. G. McKendrick to develop what are now called the SIR models—short for Susceptibles, Infecteds, and Recovereds. Their key move was analogical: drawing on earlier models for molecules and chemical reactions, the pair mapped those dynamics onto people and disease transmission. It turned out to be a big idea, one still very much alive today not only in public health but in models of misinformation, voting patterns, and the messier corners of human behavior.

Analogical reasoning takes the form of “Hey, that sounds like . . .” We use what we understand as a template for what we don’t. Sometimes it’s enough that one person can pose the problem and another can recast it. Kermack was a biochemist, McKendrick a physician and epidemiologist, and both were trained in mathematics, which provided their common language.

L.L.M.s are well suited to this style of reasoning. They’re quick to spot analogies, and just as quick to translate a story into mathematical form. In my own experiments with ChatGPT, I’ve seen firsthand how adept it is at this kind of model building—quickly turning stories about dynamic, interacting quantities into calculus-based models, and even suggesting improvements or new experiments to try. When I described this to a friend—a respected applied mathematician—his impulse was to dismiss it, or at least to explain it away: this is just pattern-matching, he insisted, exactly the sort of thing these models are engineered to do. He’s not wrong. But this, after all, is the kind of skill we relish in a good collaborator: someone who knows a wealth of patterns and isn’t shy about making the leap from one domain to another.

As machines insinuate themselves further into our thinking—taking up more cognitive slack, performing more of the mental heavy lifting—we keep running into the awkward question of how much of what they do is really ours. Writing, for instance, externalizes memory. Our back-and-forths with a chatbot, in turn, exteriorize our private, internal dialogues, which some consider constitutive of thought itself. And yet the reflex is often to wave away anything a machine produces as dull, mechanical, or unoriginal, even when it’s useful—sometimes especially when it’s useful. You get the sense that this is less about what machines can do than about a certain self-protectiveness. Hence the constant, anxious redrawing of the boundaries between human and machine intelligence. These moving goalposts aren’t always set by careful argument; more often, they’re a kind of existential staking of territory. The prospect of machine sentience hangs over all of this like a cloud. “I think, therefore I am,” Descartes said, trying to solve the mind-body problem. Our trouble now is that if machines can “think,” we’re left to wonder: Who, or what, exactly, gets to say “I”?

Model-building was the first thing I tried. I started with the everyday and moved toward the baroque, trying to link phenomena that seemed, at first, only distantly related. Could the dynamics of chemical bonds, say, help make sense of the ebbs and flows of friendship? The process quickly became addictive, fuelled by the thrill of watching even partial versions of these ideas—some already tossed around with friends, others barely more than a glimmer—take shape, sparking still more ideas in the process.

Sometimes the connections that the machine surfaced were quotidian, or even wrong—as with any collaboration, it’s important to maintain a critical eye. But at other times they bridged to territories I’d visited before but never really explored. That’s when the nature of my interactions with ChatGPT would shift: suddenly, I was drilling down into a bit of differential geometry for use in data analysis, or a concept from quantum mechanics for cognitive science. At this point, it was less like talking to a search engine and more like entering a kind of perpetual office hour—with a professor who never minds interruptions. In academic circles, there’s a choreography of self-sparing politeness: the ritual throat-clearing, “I know this is a dumb question, but . . .” The anxiety about revealing what you don’t know can get to be a little exhausting, and it’s not especially productive. With the L.L.M., I can ask “dumb questions” in private. I encourage my students to do the same—not so they’ll stay out of my office but so that, when they come, their time with me is better spent. I do it when I’m stretching into a new field or collaborating with friends in areas they know much better than I do. The L.L.M. softens my self-consciousness and makes the ensuing conversations richer and more fun.

This style of research—wandering around, then zeroing in—is a version of the ancient fox-hedgehog distinction made famous by Isaiah Berlin. (Archilochus: “The fox knows many things, but the hedgehog knows one big thing.”) In the exploratory phase, I’m the fox, sniffing around in books, conversations, half-baked theories of my own. Then the hedgehog takes over. The L.L.M. amplifies both modes: it makes me a wider-ranging fox and a quicker, more incisive hedgehog.

Sometimes I’m a fox, sometimes a hedgehog, but if I’m being honest I’m mostly a squirrel—increasingly, a forgetful one. I have no systematic method for recording my thoughts or ideas; they’re everywhere and nowhere, buried in books marked by a riot of stickies (colorful, but not color-coded) or memorialized in marginalia, sometimes a single exclamation mark, sometimes a paragraph. The rest are scattered, unmanaged, across desktops both digital and actual. My desks and tables are littered with stray sheets of paper and an explosion of notebooks, some pristine, some half full, most somewhere in between. My favorites are a handful of palm-size flip books I picked up years ago at I.B.M.’s research lab in Yorktown Heights. “THINK” is stencilled on their faux-leather covers. This ragged archive amounts to a record of my thinking, or at least those bits that, for a moment, seemed worth saving. Most I’ll never look at again. Still, I comfort myself with the idea that the very act of marking something—highlighting it, scribbling a note—was itself a small act of creativity, even if its purpose remains mostly dormant. I only wish that I were as good at digging up my acorns as I am at stashing them.

A colleague and collaborator of mine, the neuroscientist Jeremy Manning, is preternaturally good at keeping track of his acorns. His office radiates a rare kind of order, right down to the pristine whiteboard. His digital life is just as organized—a fact that never fails to amaze (and slightly depress) me. In another life, I’d like to be organized by and like Jeremy. But even he has a collection of unrealized ideas. One of them had languished for more than a year on GitHub, the online clearing house where programmers, amateur and professional alike, stow, share, and sometimes abandon their software projects.

I sometimes despair over my own unfinished business. Jeremy, ever optimistic, took a different tack with his. He enlisted Anthropic’s Claude to build what amounted to a “tinkerbot”—a tinkerer let loose in a digital attic packed with its own kind of broken toys, frayed clothes, and battered books, mending and taking inventory as it went. Armed with a technical-design document co-written by Jeremy and Claude, the tinkerbot set about transforming Jeremy’s abandoned code fragments into a working software library—complete with documentation, tutorials, data sets, the works—largely unsupervised, while Jeremy juggled teaching, research, and a newborn at home.

After nearly a month (and several hundred thousand lines of code, most of it written by Claude), Jeremy arrived at Clustrix: a fully functional software library for efficiently running big programming projects across clusters of computers—basically, teams of machines working in concert on problems too large or complex for any single computer to handle. The process wasn’t entirely plug and play. Claude made errors, and now and then got stuck, but as a team it and Jeremy solved the new problems on the way to a finished working product. Clustrix now sits, proudly, on Jeremy’s GitHub page. He would be the first to name Claude as co-creator.

Jeremy’s tinkerbot gives me hope. To what extent are my scattered thoughts like his code fragments—half-finished, abandoned, waiting for rescue? Could a machine revive a box of my broken or discarded ideas, turning them into something that the wider world would find useful and interesting? And if a machine, furnished with a carefully written set of instructions and seeded with the world’s stockpile of realized ideas, could begin generating new ones, would we still insist that true originality belongs only to people? Some cling to the belief that new ideas are conjured from the ineffable depths of the human spirit, but I’m not so sure. Ideas have to come from somewhere, and, for both humans and machines, those somewheres are often the words and images we’ve absorbed.

I’m reminded of the Grimms’ fairy tale “The Elves and the Shoemaker.” A poor but gifted shoemaker is barely keeping his business afloat. He has the talent, but not enough time or resources. Enter a band of cheerful, industrious elves who work through the night, quietly finishing his designs. With the elves in the background, the shoemaker and his wife build a thriving business. They might have simply let the good times roll, but instead, in a gesture of thanks, the shoemaker’s wife—a deft seamstress herself—makes the elves a set of fine clothes, and the elves happily move on. The shoemaker and his wife continue, now on surer footing. No doubt they even learned a thing or two about their craft by observing the elves at work. Maybe they later expanded their shop to produce jerkins and satchels. I like to imagine those elves making the rounds, boosting the fortunes of craftspeople everywhere. “The Elves and the Shoemaker” is one of the few Grimms’ tales where everyone leaves happy.

Is there a future where we simply lay out the thought-leather, rough and unfinished, set the machine going, and return to admire—and take credit for—the handiwork? The shoemaker always had talent; what he and his wife lacked was the means to turn it into a living. The elves didn’t put them out of work; they propelled them to a higher level, allowing them to make custom shoes efficiently, profitably, and cheerfully.

Most of the time, I see our digital assistants as those helpful elves. I’m not naïve about the risks. You can imagine a WALL-E scenario of academia’s future: scholars lounging in comfort, feeding stray ideas to machines and then sitting back to read the output. Though every new tool offers the promise of an easier path, when it comes to creativity, vigilance is required; we can’t let the machine’s product become the unquestioned standard. I bet that even those elves made some shoes that had to be put in the seconds pile. Research, writing, and, above all, thinking have always meant more than simply producing an answer. When I’m working, like Luke, I feel more energized than sidelined by these machine collaborators. As the physicist Richard Feynman once said, “The prize is the pleasure of finding the thing out.” That’s what keeps a lot of us going.

These days, we’re in an uneasy middle ground, caught between shaping a new technology and being reshaped by it. The old guard, often reluctantly, is learning to work with it—or at least to work around it—while the new guard adapts almost effortlessly, folding it into daily practice. Before long, these tools will be part of nearly everyone’s creative tool kit. They’ll make it easier to generate new ideas, and, inevitably, will start producing their own. They will, for better or worse, become part of the landscape in which our ideas take shape.

Will there be ideas that we miss out on because we’re using machines? Almost certainly, but we’ve always missed out on ideas—owing to distraction, fatigue, or the limits of a single mind. The real test isn’t whether we miss fewer ideas but whether we do more with the ones we find. What A.I. offers is another voice in the long, ongoing argument with ourselves—a restless partner in the workshop, pushing us toward what’s next. Maybe that’s what it means to be “always working” now: turning a problem over and over, taking pleasure in the tenacity of the pursuit, and never knowing whether the next good idea will come from us, our colleagues, or some persistent machine that just won’t let the question go. ♦