Episode 1. For twenty years the advice was learn to code. Then the machines learned English. Nathan argues that language is how you think before it is how you talk, and that people who can control it now hold the most useful skill in technology.
As this episode frames it, for about twenty years the career advice given to students was short enough for a bumper sticker: learn to code. Then, as Nathan puts it, the machines learned ours. In the first episode of This Half-American Life, Julien sits down with Nathan, who studied computer science, runs a small business and is writing a book called What Arrives, to ask whether that advice has flipped.
The machines learned ours
Nathan's starting point is an essay he wrote about what a college student should study now. Its turn is that single line. For decades, he argues, talking to a computer meant learning the computer's language. Now you talk to it in the language you learned as a child.
He doesn't think coding lost. In his view, code and English are two languages doing the same job, turning what a person wants into what a machine does, and the machines switched which one they understand. People who write code will keep writing it. For most people, most of the time, he says, it stopped being the way in.
The question started with a conversation with an AI model. It ranked applied math and statistics first and put English below them, and when Nathan pushed back, it changed its mind. He tells Julien to be suspicious of that, because these systems tend to agree with you if you push hard enough. His narrower point is that a model trained on what people write about careers had picked up their blind spot, ranking the thing it is made of below the things it can do.
Where a thought comes from
The core of Nathan's argument is that language belongs under thinking, not only under communication. A thought, he says, often starts as a shape or a pressure without edges, and finding the word for it is part of thinking it. He calls that move compression: something vast and vague made small enough to carry and hand to someone else.
Julien pushes on this. Professional linguists have almost unanimously rejected the strong form of linguistic relativity, the idea that your language determines what you can think. Research since the 1980s does support a weaker form, in which language influences thought. Nathan concedes the strong version and says the weak one is enough for him: if language shapes what is easy to think, getting better at it makes you a slightly better thinker. He admits his essay states it harder than he would defend it in a seminar, and that the book still has to close that gap.
He is careful about where the compression idea comes from. A paper titled "Language Modeling Is Compression", mostly by researchers at Google DeepMind and published at ICLR 2024, argues that language models are compressors. Nathan says engineers reached the idea from the math side, and that what he adds is the question of what a person should learn.
The mess is doing the work
Julien points out that most people would hear "imprecise" as a flaw. Nathan thinks the gaps in ordinary language are where its power lies. "Make it warmer, but not sappy" is a handful of words that carries a great deal of judgment. The skill he cares about is control: knowing which gaps to leave open and which to close. He calls his way of working with models dialogue engineering, a back-and-forth of writing, reading what comes back and saying it more sharply, and he says the reading half matters as much as the writing half.
He also admits the hole in his essay: it has no concrete story of that skill changing what a machine produced.
The salary problem
Julien raises the objection every listener has. Humanities majors still trail some professions, engineering among them, in lifetime median earnings. Nathan has two answers and says neither fully closes it. Pay may lag a change this recent, and many English programs teach literary analysis rather than language as a precise tool. If the second is true, his advice becomes "whatever you study, master your language", which is where his essay ends up.
He keeps expertise in the picture too. A model can retrieve information, he says, but it doesn't know which information matters for you. Context without language is what he calls trapped expertise; language without context is empty fluency.
What's still open
Nathan ends on what he hasn't settled: how strong the language-and-thought claim can be, and the fact that most of the feedback on his essay came from AI models, which tend to mirror you back to yourself. The human test, he says, has barely started.
Julien: Nathan, you have a computer science degree, and you've spent the last year arguing that the degree that matters most right now is English.
Julien: For about twenty years, the career advice fit on a bumper sticker. Learn to code. Then, fairly suddenly, the machines started taking instructions in plain English. A lot of people who'd spent years apologizing for their degree started wondering whether the story had flipped. Nathan thinks it has, and he thinks the reason has less to do with job listings than with what language does inside your head before anyone hears it. Before we start, one thing you should know. Everything you're about to hear is AI. Our voices are AI copies of our real ones, and we gave permission for that. This conversation was written with AI from Nathan's own book draft and notes. The ideas are his. This is This Half-American Life, and I'm Julien. My guest is Nathan. He has a computer science degree and runs a small business. He's writing a book called What Arrives, and this conversation covers where it starts. Nathan, welcome.
Nathan: Thanks. Glad to be here.
The machines learned ours
Julien: Let's start with the line from your essay that people quote. "Then the machines learned ours." Set it up for me.
Nathan: The setup is the Thanksgiving table. For twenty years, every parent told their kid to become a software engineer because that was success. Programmers built Google and Facebook. They built the infrastructure everybody lives inside now. A technical degree was the only one you didn't have to defend over dinner. And the logic underneath it was clean. If you wanted to talk to a computer, you learned the machine's language. Then the machines learned ours. For fifty years, computers spoke programming languages. Now you talk to them in English, the same language you learned as a kid. The interface flipped. Almost nobody changed the advice.
Julien: So coding lost.
Nathan: I'd frame it differently. Coding and English are two languages doing the same job. Both turn what a person wants into what a machine does. Think of them like Chinese and Arabic. Neither one is the math language or the letters language. The machines just switched which one they understand. Code still matters, and people who write it will keep writing it. For most people, most of the time, it stopped being the door. The bridge between people and machines changed materials.
Julien: There was a long magazine piece in March where a reporter talked to more than seventy software developers about working with AI.
Nathan: Right, and that's the flip showing up in the field. What struck me reading about it was how much of their day is now spent writing English to a machine. These are the people we told everyone to become. Their job turned into describing what they want, clearly, in sentences.
Julien: You studied computer science yourself. Isn't this a strange hill for you to pick?
Nathan: I think it's the only reason anyone should listen. If an English department says English matters, that's a department defending its budget. I started in computer science and I run a small business now. I've stood on both sides of the line this argument is about. And the computer science I studied was mostly theoretical math, in a liberal arts setting. So even my technical degree was really a degree in how to think about structure. Which is kind of my point, if you squint.
Julien: Where did this start for you?
Nathan: With a pretty basic question. What should a college student study in the age of AI? I asked one of these models. It gave me the sensible answer. Applied math, statistics. It ranked English as secondary. And I pushed back, because the thing I was using to have that conversation was language. The whole system runs on it. The model changed its position.
Julien: Hold on. You're citing an AI agreeing with you as evidence?
Nathan: You should be suspicious of that. I am. These systems agree with you if you push hard enough, and that's been a problem through this whole project. My claim is narrower. A model trained on everything people have written about careers picked up the same blind spot everyone has. It ranked the thing it's made of below the things it can do. I'd call that a symptom. It proves nothing on its own.
Julien: Who are you really writing to, then?
Nathan: The person who majored in English and has been apologizing for it for twenty years. The opening of the essay is all admiration on the surface. Programmers built this, programmers built that. If you're that reader, every sentence lands as a loss, and the paragraph never has to say so. Then the turn comes, and it reads like permission. There's a second reader too, the angry engineer. I went back and forth on whether narrowing the ending to English majors would push that person away. Where I landed is that a reader who's furious is as useful as a reader who feels vindicated. Either way they're reacting. I'll take the argument.
Julien: There's a third reader you mention in your notes. The professional who feels like AI belongs to somebody else.
Nathan: That's the one I care about most, I think. A huge number of smart people look at this technology and feel like it's for the engineers. They feel like they missed the train years ago when they didn't learn to code, and now there's a new train, and it's also not theirs. That fear makes them avoid the tools, which makes the fear come true. What I want that person to hear is that they're holding the ticket already. The way in is a sentence. If you can say clearly what you want, and read critically what comes back, you can use this thing as well as anyone in the building. Probably better than a lot of people who can code and can't explain what they're building. I'd like the essay to give that person permission to stop being afraid of it.
Julien: Permission is a strong word for an op-ed.
Nathan: It is. But it's the honest word. The argument is the vehicle. The feeling I'm after is relief.
Where a thought comes from
Julien: In the draft you wrote that thoughts don't begin as words. They begin as, quote, "half-formed images, nameless instincts." Walk me through that.
Nathan: You know the experience of knowing something before you can say it. There's a shape there, a pressure without edges. Then you find the right word, and the thing changes. It gets sharper. It becomes something you can hold and hand to somebody else. I'd go further than "you found a name for it." Finding the word is part of thinking it. Before the word, I had a vague feeling. After the word, I have an idea. That move, from huge and formless to small and held, is compression. You take something vast and make it concise enough to carry around.
Julien: So you're saying language and thinking are the same thing.
Nathan: I'm saying we filed language in the wrong drawer. Everybody puts it under communication, the soft-skill shelf, the write-a-nice-email shelf. I think it belongs under cognition. Somebody who spent four years getting precise with language spent four years upgrading the machinery they think with. Communication comes after that, and it's the smaller payoff.
Julien: Here's the problem. Linguists have mostly thrown out the strong version of that idea, that your language determines what you're able to think. People think about things they have no word for all the time.
Nathan: Yeah. And I can't defend the strong version. If I stood up and said language determines thought, I'd be wrong, and I'd lose the room. The weak version is enough for me. Language shapes what's easy to think and what's natural to reach for. Learn a second language and you get more than a bigger vocabulary. You get a slightly different way of seeing the same thing. If that's true, then getting better at your language makes you a better thinker at the margin. And the margin is where people compete.
Julien: But in the essay you don't hedge. You wrote that the English degree, quote, "was a cognition degree."
Nathan: You caught me. The essay is eight hundred words, and op-eds don't do footnotes. There's a gap between how hard I state it in print and how hard I'd defend it in a seminar, and I know it's there. In my notes it's still an open question, how strong to make the claim. The book has to close that gap, and I haven't written that part yet.
Julien: Where does AI come into the compression story?
Nathan: The same thing happens between two people. I hold an idea. I pick a handful of imprecise words. The words point at the idea without containing it. You fill in the rest from your own head, and the meaning that forms on your side is bigger than what I actually said. The models live inside that same process. There's a paper from Google DeepMind with the title "Language Modeling Is Compression." The engineers got there from the math side. So the compression idea doesn't belong to me, and I'd be embarrassed to pretend otherwise. What I'm doing is carrying it somewhere they didn't go, which is the question of what a person should learn.
The mess is doing the work
Julien: You keep saying imprecise. Most people would hear that as a flaw.
Nathan: That's the instinct, and I think it's backwards. Take "make it warmer, but not sappy." A handful of words. Hand that to a designer or to a model and you've sent an enormous amount of judgment. Taste, emotional register, what you're going for, what you're trying to avoid. If you had to spell that out precisely, you'd need pages. The fuzziness is where the compression happens. A language with no ambiguity is a programming language. You get exactness and you pay for it in length. Human language makes the opposite trade. It trusts the person on the other end to unpack it.
Julien: Then what's the skill? If the fuzz is good, why train anybody?
Nathan: Because the skill is control. Knowing which gaps to leave open and which ones to close. A bad writer leaves the wrong gaps open, and the reader fills them with the wrong thing. A good writer leaves exactly the gaps the reader will fill in correctly. In the essay I wrote that clear communication is the exacting control of imprecision. That sentence took a lot of drafts. The early versions kept saying "be more precise," which contradicted everything in the paragraph above it.
Julien: There's something funny about that, though. You need extremely precise language to explain why imprecision works.
Nathan: Yes, and that bothered me while I was writing. Explaining why the gaps create meaning means closing almost every gap in the explanation. I wrote it down as the precision-imprecision paradox. I don't think it breaks the argument. It shows the skill has two ends. Sometimes you need to be loose and let the listener carry the weight. Sometimes you need to be exact. Mastery is knowing which one the moment is asking for. Most people are stuck at one end. Engineers usually get trained toward the exact end, and a lot of good talkers live at the loose end.
Julien: Give me something concrete. A moment where that skill changed what a machine gave you.
Nathan: That's the hole in the essay. The sharpest note I got on the draft was that it was all argument and no evidence, and the note was right. I never added the story. The closest thing I have is how I actually work. I talk to these models constantly, out loud, and I go back and forth. I read what comes back, I catch where it drifted, I say what I meant more sharply, and I go again. The old idea of prompt engineering was writing one perfect instruction and walking away. What works is closer to a seminar. You argue, you refine, you read closely, you push back.
Julien: You have a name for that.
Nathan: Dialogue engineering. And the reading half matters as much as the writing half. When a model hands you two thousand words, somebody has to tell the confident parts from the hedging, and the grounded claims from the made-up ones. Somebody has to notice when it handed you the median answer instead of the right one. That's close reading. An English degree trains both sides. Writing is the compression. Reading is the decompression. Nobody talks about the second one.
Julien: There's a whole job title built around this now. Prompt engineer.
Nathan: That one makes me smile. Prompt engineering is applied English composition with a technical-sounding name on it. The industry hired people to write clearly and couldn't bring itself to say so. That's a quiet admission that language mastery matters, filed under engineering so nobody had to feel weird about it. I think the simple version of the job is already getting absorbed, though. Wrapping one instruction in better instructions was a phase. The sustained conversation is what's left, and that's harder to teach in a weekend.
Julien: You structure your own paragraphs in a particular way. The claim comes last.
Nathan: Prove before you name. I'd rather the reader feel the conclusion forming before I hand it to them. If I state the thesis first, you spend the paragraph checking whether I earned it. If I build the case and then name it, you've already arrived by the time I say it. Same mechanism we've been talking about. I'm leaving a gap I trust you to close, then closing it a beat after you do.
The salary problem
Julien: Okay, here's the objection everyone listening already has. English majors still earn less than engineers over a lifetime. That's in the data. If language is the most valuable skill in tech, why doesn't the market pay for it?
Nathan: That's the strongest objection, and I've gone back and forth on how to handle it. At one point I cut it from the essay completely, and then I put a sliver back. I have two answers, and neither one fully closes it. The first is that the market lags. The way you talk to machines changed recently, and pay follows skills slowly. The essay frames that gap as opportunity. The second answer is harder for me to say out loud. A lot of English programs don't teach what I'm describing. They teach literary analysis, close reading of novels. That's real work, but it's a narrow use of the instrument. Precision with language as a tool is something else, and plenty of programs never get to it.
Julien: So your answer to "English majors are underpaid" is that maybe they didn't get the education you're praising.
Nathan: Partly, yes. That's the uncomfortable version. And if I'm right about it, the recommendation can't simply be "major in English." It turns into something closer to "whatever you study, master your language." Which is where the essay ends up. It tells parents, no matter what your kids study, tell them to master their language, and if they chose English, tell them they were right. You've put your finger on the part I haven't settled, though. I don't know how many programs actually teach the thing I'm describing. It's sitting in my notes as an open question, and I haven't gone and found out.
Julien: You also wrote that English majors didn't build AI. Engineers did. That sounds like conceding the whole thing.
Nathan: It's a concession, and an honest one. Engineers built the software. But the software was incomplete. It had to learn, and what it learned from was writing. Pretty much everything people have bothered to write down. Writers trained these systems without knowing they were doing it. So the engineers built the machine, and centuries of people putting thought into words filled it. I don't think that makes English majors heroes. It makes the written record the raw material, and the people who can work that material well have an edge with the thing that came out of it.
Julien: Where does expertise come in? A great writer who knows nothing about medicine isn't going to get much out of an AI on a medical question.
Nathan: Agreed, and that's what keeps this from being a "humanities good, engineering bad" piece. I'm not making that argument. Context is what AI doesn't have. Your field, and what actually matters in your situation. The model can retrieve information. It has no idea which information matters for you. Language is the bridge between that context and the machine. Context without language is what I call trapped expertise. Language without context is empty fluency. You need someone holding both ends.
Julien: Trapped expertise. Who's that?
Nathan: The brilliant specialist who can't explain what they know. To a colleague, and now to a model. I'm careful with how I say it. I'm describing a situation anyone can be in, and I don't mean it as a type of person. Plenty of engineers are great writers. But an engineer who can't express an idea can't get a machine to amplify it. And the person who can sit between an expert and a machine and translate in both directions is very valuable right now. Same with sitting between an engineer and a marketer, or a doctor and a patient. The machine just added one more side to translate for.
Julien: In the essay that idea becomes one paragraph with three characters in it, each missing something. A critique of your draft called that paragraph a writing exercise.
Nathan: Yeah, and it stung, because that paragraph went through five drafts before it locked. The critique's point was that by the time you hit it, the essay has already made the argument, so the paragraph is restating instead of moving. I partly agree. I kept the structure anyway. My rule was that I'd change it when a human reader told me the same thing, and the human reader I eventually got didn't flag it. So it stayed. That might be stubbornness. I'm honestly not sure.
Julien: There's a detail in your notes about which profession gets which word.
Nathan: In one draft I gave the doctor the word "precise" and the engineer the word "expressive." It looked like a mistake. You'd expect it the other way around. Then I realized the swap was doing something. The doctor needs exact language because a patient's life hangs on a word. The engineer needs expressive language because the machine will build whatever they describe. Flipping the expected pairing makes the reader notice that every field needs both ends. I flagged it to develop and never did.
Julien: Does better language actually get you a better answer out of the machine? Or does the model meet everyone in the middle?
Nathan: My bet is that the output is bounded by the input. A better language user gets a better result. I don't have a study to point at. I went looking for data on writing skill and AI output quality and couldn't find anything public. If it exists, it's probably sitting inside some company. So I'll say it as a belief from experience. I can't prove it to you today.
Julien: You've described the degree as a gym.
Nathan: Yeah. We ask "what should I major in" as if the degree is the thing you walk out with. What you walk out with is the muscle. So the better question is which muscles a program builds, and whether those muscles are getting more valuable or less valuable. A lot of skills are on a steeper depreciation curve than they used to be. Language moved the other way.
Julien: Which jobs are on the steep part of that curve?
Nathan: I'm cautious here, because it's easy to sound condescending about somebody's profession, and I don't want to rank people. The way I think about it, the most exposed work is skilled and specialized but largely pattern-based. Procedure you learn and then repeat. Real training, real expertise, and still the part AI compresses first. I'd call that partial disruption. The thinking parts of the job get squeezed, and the job itself mostly survives. The safer spot is deep expertise in a field combined with the ability to think across fields and explain yourself. I'll add a warning on my own side, though. "Generalist thinker" is also what mediocre people call themselves.
Julien: If not English, what's the most undervalued major?
Nathan: Philosophy, probably. It trains you to take an argument apart and find the hidden assumption. Its weakness is that it gives you almost no comfort with numbers. Statistics is the most practically powerful, because it trains you to think about uncertainty and about what the data is lying to you about. Its weakness is that it can produce people who run models they can't explain. Notice that the weakness in each case is the other one's strength.
The gap between speaking and speaking well
Julien: If everybody already speaks English, what's actually scarce?
Nathan: The distance between speaking it and speaking it well. Everybody can walk. Not everybody can dance. It's the same body. That's the part I find interesting about the economics. Coding took four years and a lot of tuition. You got English for free, as a child. The thing that replaced code as the way you talk to machines is something you already own. That's the flattering half of the argument. The demanding half is that you probably haven't refined it, and that distance decides how much the machine can do for you.
Julien: That's starting to sound like a pitch for a writing course.
Nathan: It does, doesn't it. I'm not selling one. The narrow version I'll stand behind is this. Writing and basic statistics are the two closest things to cheat codes I know of. They multiply everything else you do, whatever your field is. If I had to tell a kid two things, it would be those.
Julien: Your essay ends on what's almost a joke. "It's called a Large Language Model. The answer is in the name." Is that too cute?
Nathan: Maybe. It's the line people remember, which is why it's there. The point under it is real, though. We spent years having this anxious conversation about what skills matter in the age of AI, and the technology was named after the answer. People read straight past the middle word.
Julien: What about the future where we skip language entirely? Wire straight into the brain. Does the argument survive that?
Nathan: I have an answer, and it's speculative. If thought is tangled up with language the way I think it is, then a device reading your brain is reading language one layer further down. You're catching it before it reaches the mouth. That would make language matter more. I'd flag that as a guess, though. It's in my notes as too speculative for the essay, and I left it out on purpose.
Julien: Let me push on one more thing. You keep saying "language." Do you mean English, specifically?
Nathan: In the essay I mean English because that's the language these models settled on first, and it's the one the reader speaks. The deeper claim is about any human language. Mastery of whatever language you think in. If the models had settled on Mandarin, the essay would be about Mandarin, and an American parent would have a much worse week.
What's still open
Julien: Last question. What are you still unsure about?
Nathan: A couple of things. How strong the language-and-thought claim can be. I backed off it in this conversation, and I'll have to back off it in the book too. The bigger one is that nearly every piece of feedback on this work came from a model. The critique that said the middle sags, the agreement when I pushed, all of it. These systems mirror you back to yourself, and I know that. I did finally give the essay to a friend who majored in English, and I sent it out to a few papers. Nobody's taken it yet. So the human test has barely started. And there's a note in my files I keep going back to. A dentist who loves being a dentist, goes home at five, and coaches Little League, is winning a game the anxious consultant working eighty-hour weeks is losing. None of the skill talk matters if you're miserable. Get good enough at something valuable that you have options. Then spend the options on a life you want. I put that at the bottom of the notes, under everything else.
Julien: Nathan, thank you for doing this.
Nathan: Thanks for pushing on the salary thing. That one's going to stick with me.
Julien: That's our conversation. Everything you heard was AI, and the ideas were Nathan's. Thanks for listening.