“AI 洗白”式裁员？+ 为什么 LLM 写不好 + Tokenmaxxing

I just read the most heartwarming news this morning that I wanted to share with you, Kevin.

What's that?

The U.K. government has withdrawn a proposal to let A.I. companies train on copyrighted works after a backlash from artists like Dua Lipa. Did you see this?

No.

Dua Lipa said, “Don’t Start Now with this A.I.” My sugar boo, she’s litigating, Kevin. She’s making some new rules, and she’s saying, “We’re not gonna train on my copyrighted works.”

Wow.

And that’s why she is a queen. And so, Dua Lipa, if you’re listening, we salute you.

Yeah. Dua Lipa, you’re a Dua Keepa.

Yep. Period. Dua Lipa said artist rights.

Wow.

I'm Kevin Roose, a tech columnist at The New York Times.

I'm Casey Newton from Platformer.

And this is Hard Fork.

This week, a big wave of tech layoffs is raising the question: Has A.I. job loss truly begun? Then, writer Jasmine Sun is here to help us answer the question: Why are chatbots bad at writing? And finally, it’s token maxing time. Why are tech companies building leaderboards to measure who is spending the most on A.I.?

1. The AI Layoff Warning

Well, Casey, for years now, we’ve been monitoring for signs of an A.I. job apocalypse.

Yeah, we’ve been monitoring the situation.

It’s true. And over the past few weeks, I think we’ve gotten some early indications that something is happening in the labor market, especially for tech workers.

Yeah, we have certainly heard CEOs of companies announcing layoffs and invoking A.I. as a reason that it is happening, and so that has gotten our attention.

Yeah, so just a couple of examples from the last few weeks. Last week, Atlassian announced a 10 percent reduction in its staff, about 1,600 jobs, that they said were going to help them fund further investment in A.I. and enterprise sales. That came on the heels of a big round of layoffs at Block, the financial tech company formerly known as Square, which said that it was cutting its staff by about 40 percent, or about 4,000 jobs, saying that they were shifting the way that they were working to use smaller and flatter teams.

And then the big one that folks are expecting, maybe as soon as this week, is that Meta is reportedly poised to lay off 20 percent or more of the entire company. This was reported by Reuters last Friday, who said that their sources had told them that Meta was preparing to cut as many as 16,000 jobs, the largest layoffs at that company since late 2022 or early 2023, when they laid off 20,000 people. So as of this recording, that hasn’t happened yet, that we know of, but I know that people at Meta are very on edge and are awaiting further news about their jobs.

Meta, after this story came out, told Reuters that it was, quote, “speculative reporting.”

Which, if you’re not familiar with the language deployed by Meta communications staffers, means this is happening—but we don’t want to tell you it’s happening yet.

Correct.

So, Casey, I want to hear what you make of these layoffs, but first we should do our disclosures. I work for The New York Times, which is suing OpenAI, Microsoft and Perplexity.

And my fiancée works at Anthropic.

So, okay, Casey, what do you make of the fact that all these companies are referencing A.I. in some way as a reason for their layoffs?

Well, I think it’s a little different at each company, Kevin, and I think we can make a decent case for and against the idea that A.I. is really driving the show at each of them, so maybe we should get into that. But at the highest level, I would say companies do continue to tell us now that A.I. is a significant factor in the reduction of these workforces, and sooner or later, I do think we’re going to have to believe them.

Yeah, I think this is the early warning sign for a lot of people, especially in the tech industry, who are, I think it’s fair to say, going to be some of the first people to see their jobs change or disappear because of these new A.I. tools. But let’s get into some of the specifics here.

Yeah.

So, Casey, let’s start with Atlassian, the first company I mentioned. Their CEO, Mike Cannon-Brookes, said in a company blog post that the bar for what great looks like for software companies on growth, on profitability, on speed, on value creation, has gone up. He said, “We are choosing to adapt thoughtfully, decisively and quickly to drive durable, profitable growth.” He claimed that A.I. was not replacing people, but he said it would be disingenuous to pretend that A.I. doesn’t change the mix of skills we need or the number of roles required in certain areas.

Yeah, so I take him at his word. It seems like he himself is trying to walk a middle path there, right? And sort of not denying that A.I. is a factor here, but also not saying, “This is the only reason this is happening.”

I think some other context that is worth having is that Atlassian is one of the companies that could be part of what we’ve been calling the SaaSpocalypse around here, right? This is a company that makes tools for businesses. A lot of its products are essentially structured workflows, and there are those who believe that sooner or later, you’re just going to be able to code your own pretty cheaply.

Now, maybe you will still choose to buy a product from a company like Atlassian, but maybe you’re not going to be willing to pay nearly as much as you would have before. And so the company’s stock price has just been battered over the past year, and I think that has left them, one, hurting for cash a little bit, but two, and probably more importantly, looking for a different story that they can tell the stock market about what they’re doing. And so today that story is, “We’re gonna get rid of some of these workers, and we’re gonna figure out how to make our remaining workers more productive.”

Hmm. So there’s this term that’s been floating around called A.I. washing.

I thought it was when a software engineer finally took a shower.

And basically, the thesis is: These aren’t really layoffs about A.I. This is just sort of a convenient excuse that these companies are using.

Yeah.

Do you think Atlassian qualifies as A.I. washing?

I would like to get a little bit more detail on exactly who they are laying off here, which is a detail that we do have about some of these other companies that helps us answer that question. So I don’t know exactly how it is happening inside of Atlassian, but I think that their CEO was relatively straightforward, as these things go, in saying, “It’s a little bit about A.I., it’s not entirely about A.I.,” but, “Yes, keep your eye on A.I.”

So to me, that just reads as honest, and so I’m gonna give them a pass.

2. Block Shrinks Its Workforce

Okay. Let’s talk about Block. Jack Dorsey, the CEO of Block, gave an explanation about their layoffs. He said, quote, “We’re not making this decision because we’re in trouble. Our business is strong, but something has changed. I had two options: cut gradually over months or years as this shift plays out, or be honest about where we are and act on it now. I chose the latter.” Casey, your take.

So something to know about me and Jack Dorsey is I have a bit of a bias against him as a former Twitter user who misses that website dearly. At this point in 2026, I would not hire Jack Dorsey to run a lemonade stand. But if you want to talk about Block specifically, this is a company that tripled its headcount from about 3,800 people in 2019, in what seems like just classic inattention to what was happening in the business during pandemic-era boom times, right?

And I wonder if you saw this detail, because it truly took me out, Kevin. Five months before they announced the layoffs, Block spent $68 million to fly 8,000 people to an in-person event with Jay-Z.

Come on.

Yeah. So that’s the kind of famous attention to detail that has turned Jack Dorsey into one of the greatest visionaries in tech.

So look, is this about A.I.? Again, what does Block really do? They have those little iPads at the coffee shop—

Yeah.

—and then they have Cash App.

Mm-hmm.

Okay? How many people do you really need to run those products? Probably fewer than 10,000.

Hmm.

Is that about A.I.? I don’t know. Maybe if you squint. But again, this is a company whose stock price was cratering. They needed a different story to tell the market, and I do think you can make a case that A.I. will make the remaining workers more productive. So again, this is another one where it’s like, you could use A.I. to justify what’s happening, but you also could just say, “This company has been mismanaged for a while now.”

Yeah, you could use A.I. washing or Jay-Z washing, which seems to be what they are doing here.

Mm-hmm. Yes.

So this did seem to have an effect on their stock price. In fact, the day after Jack Dorsey announced the layoffs, Block’s stock shot up 17 percent. It’s gone down a little bit since then, but they’re still up from where they were before these layoffs.

And I think we should just say: This is also a part of the equation here, right? These are companies, largely public ones, that have investors’ attention. Right now, there’s this narrative power around AI: If you seem like a company that is investing heavily in AI tools and the AI way of working, your investors say, “Oh, that company is really forward-looking. They must have a plan for how to navigate this transition.” And so I think they’re seeing the power in telling the story that all this is related to AI.

Yeah, which, by the way, reminds me of the peak of crypto mania, when some publicly traded companies would just add a crypto term to their name, and their stock price would shoot up by about 40,000%.

Yes.

It turns out that the public markets actually can just be tricked that easily.

Yes.

That would give me some relief if I were a CEO, just knowing that I could fool people like that.

3. Meta Cuts For AI Infrastructure

So let’s talk about the third large tech company that is reportedly conducting layoffs: Meta. We don’t know exactly who or what teams are being affected by these layoffs, but this is a significant part of their workforce. They seem to be saying in their communications with the public what all of these other companies are saying: “We are going all in on the new way of working, and we are going to have to make some cuts to make that work.”

Yeah. On a recent earnings call, Mark Zuckerberg said, quote, “Projects that used to require big teams now can be accomplished by a single, very talented person.” We should also say that this cut is coming alongside this massive AI infrastructure investment, right? They’re going to spend $135 billion on capital expenditures this year. And even for a company of Meta’s size, that is real money. I know they’re trying to be careful not to spook the stock markets too much. This is obviously the biggest bet in the company’s history, and I think making some substantial cuts is going to signal to the market, “Hey, don’t worry. We’re not completely losing our minds here. We’re going to keep some of these expenses under control.”

Yeah, I think that’s a really important point, because what we’re seeing here at some of these companies is that they are not actually cutting costs in the aggregate by using these tools. They are just shifting the cost from human labor to AI.

Right.

They are plowing this money that they are going to save by laying off these thousands of people into the building of data centers and other AI infrastructure. Basically, the bet they’re making is that these new AI workers are going to be faster, more efficient, maybe cheaper in the long run, maybe not, but they are going to be able to do the work that used to require many thousands of people. And that is a profound shift in the way that companies are talking about their workers.

I recently talked to a venture capitalist who said that a lot of the AI startups that he sees, the most AI-native companies, are spending more on AI tools than they are on payroll. That may be an outlier, but I think that is where these companies believe that we are headed, where the majority of your expenses will not go to paying the salaries of human workers. It will go toward buying the AI tools and the tokens that your company runs on.

Yes, I think that’s absolutely the bet that they’re making. I also think it is worth noting that this is still mostly speculative, right? In the case of Meta specifically, this is a company that has arguably been struggling when it comes to AI. They had to abandon their last model, Behemoth, because it wasn’t very good. The Times reported last week that it’s delaying the release of its latest model, Avocado, because it hasn’t been hitting its performance targets. It’s apparently barely outperformed Gemini 2.5. What is this, last March?

Yeah, that model is really the pits.

That’s an Avocado joke.

That’s very good. Thank you.

So, again, this is not as simple as saying they’re able to cut 20% of their workforce because they’ve just made these massive gains. I’m sure there are individuals there who have made massive gains, but as a company, it still seems like it is somewhat mired in dysfunction. They just did yet another partial reorganization of their AI teams, and that always makes me raise my eyebrows.

Yeah. I will say, one thing that’s been surprising to me about this recent round of layoffs is that the companies that are making them are not the ones on the frontier, right? It is not the OpenAIs, the Anthropics, or the Googles. Those companies are not laying off people en masse because of these AI tools, which they are building and presumably have even better models than the ones they’re releasing to the public. So you have to think that part of this is just companies that are lagging behind their competition saying, “Well, maybe if we just use a bunch of AI, it’ll help us catch up.”

Yes, but also OpenAI and Anthropic are much smaller companies than some of the ones that we’ve been talking about today, at least in number of workers, right? I think it is interesting to think that Atlassian is bigger than OpenAI in terms of the number of people who work there, when you look at the relative value of what they’re generating.

DocuSign has 7,000 employees.

There’s no funnier sentence that is true in all of tech journalism. As somebody who has a paid subscription for DocuSign that I truly resent paying for, get to work over there, people.

Or get not to work.

Get not to work.

4. Workers Consider Union Power

Here’s another question that I would ask, Kevin. We’re seeing a bunch of layoffs. Are these AI-related or not? Does it actually matter if the effect on workers is the same, right? If you’re the worker, whether it’s about AI or not, you’re still out of a job.

Yeah, and it’s not clear to me what workers can or should be doing to protect themselves against these layoffs. One person I talked to said they work at one of these big tech companies, and they’re like, “Well, there’s just a lot of jostling and fear and anxiety right now. People don’t know if they should be using the AI tools a ton because then it shows that they’re getting with the program, or whether that just means that they’re proving that their work can be automated.”

I think there’s a lot of fear, suspicion, and mistrust inside these companies right now, and for good reason. Their executives are planning to lay them off.

Yes, and by the way, I think at least at some of these companies, that may not be an explicit reason for these layoffs, but some of the executives there would see that as a positive byproduct, right? Because if you’re Mark Zuckerberg, you lived through the 2020 era. You had these restive employees who wanted a lot of things from you, and they wanted to have a lot of control over what the company could and could not do and how it did it.

I know that executives over there really resented that sort of thing. And once Meta entered this new era of massive layoffs, employees over there did get really scared for all of the reasons that you would assume. They were like, “Oh, God, maybe I actually am going to lose my job.” All of a sudden, they got a lot quieter, and you started to see a lot fewer protests over there.

So I’m not going to say that these occasional mass layoffs are a way of keeping the workforce in line, but I have noticed that it seems to be having that effect.

Totally. And it makes me wonder whether something that I predicted was going to happen a year or two ago, but did not happen—the sudden and mass unionization of workers at these companies—may actually start to happen in the next year or two.

I think one major difference between what’s happening now at these tech companies and what has been happening for decades at manufacturing companies and car companies, among factory workers, is that those workers were by and large unionized. And so when the employers said, “Hey, we’re going to lay a bunch of you off,” they were able to negotiate. They were able to say, “Hey, maybe instead of laying us all off, maybe you could find other jobs for us. If our jobs are being automated, maybe we should be allowed to retrain to do something else.”

And that was largely successful. There were still layoffs, of course, but not the number that we’re seeing today at these tech companies. So do you think there’s any possibility of that, or is that just a union fever dream?

Here’s what I will say: I cannot think of anything that would make Mark Zuckerberg more mad than a union of software engineers at Meta. And I think the software engineers at Meta should use that information how they will.

You think that would make him more mad than getting booed at a UFC fight?

Absolutely. I think that probably just made him really sad.

Well, there you have it. If you want to make Mark Zuckerberg mad, Meta employees, sign your union card.

When we come back, why aren’t chatbots as good at writing as I am?

We’ll ask Jasmine Sun.

5. LLMs Still Struggle With Writing

Well, Casey, over the last couple of years, we've talked on this show about how AI models are getting better at so many things. They are getting better at coding, at competition math, at solving novel physics problems.

Mass domestic surveillance—autonomous weapons.

Yes. And I think the story of the last few years in AI has been one of sort of rapid, steady progress, but these systems are still sort of jagged, and they have flaws and weaknesses. And one place where they arguably haven't improved that much is in writing.

Now, that's our domain.

Yes. At least that is the argument that Jasmine Sun made in The Atlantic this week. She is a freelance journalist. Her piece was called “The Human Skill That Eludes AI,” and it's her attempt to understand why, despite so much progress in all these different areas, the models of today don't seem to be writing anything particularly good or compelling.

Yeah. And while I think the question of whether LLMs are good at writing is highly subjective and dependent on the use case, I do think Jasmine makes a really interesting technical case for why these models write the way they do.

Yes. And we should say, before we bring her in, Jasmine is a friend of mine. She has also been my researcher on the upcoming book that I'm working on, and I just think she's one of the best people writing about AI today. She writes on her Substack, which is called Jasmine News. It's J-A-S-M-I dot news, and you can read much more of her writing there.

All right. I'll allow it, but I do want to balance it out. By next week, bring me on one of your enemies.

Okay, let's bring her in. Jasmine Sun, welcome to Hard Fork.

Thanks for having me. I'm excited.

Hi, Jasmine.

So you wrote this great piece in The Atlantic this week about the human skill that eludes AI, and I want to start by challenging the subtitle of your piece. Why can't language models write well? Can't language models write well?

So I do say in the piece that most writing, period, is very bad, and so I think that language models are definitely better at writing and language than most humans are. But the question that I was really curious about is, why can't they write at a sort of literary, creative-fiction level?

Because the thing is, if you listen to these AI leaders talk about their aspirations, they say, “We're gonna cure cancer. We're gonna solve physics. We're gonna build a superhuman coder.” They are not shy about saying, “Oh, our AI models are gonna be better than 75 percent of human coders.” They're saying, “No, we will literally build a self-replicating factory tomorrow.”

And then Tyler Cowen asked Sam Altman in an interview from last October, “When do you think GPT will be able to write a Neruda poem?” And Sam Altman says, “Maybe in the future, ChatGPT will be able to write, quote, ‘a real poet's okay poem.’” So that was the thing that fascinated me: Even these guys who are more bullish than anybody else about the capabilities of their technology, they are very reserved about how much literary writing their models can do.

Mm.

And so that was the gap that I was really interested in.

Hmm.

Hmm.

6. GPT Two Had More Voice

And you start your piece with this interesting provocation, which is that, in some ways, GPT-2 was the peak of AI when it comes to creative writing. So explain that.

Part of what got me interested in this piece was I was actually doing research for your book. I was going through all of these previous generations of models and reading the outputs, and the thing that really shocked me was that, in a way, the writing style of GPT-2 and GPT-3 was so much more compelling to me than ChatGPT today.

It doesn't have any of the annoying tics. It doesn't have the em dashes, the tripartite lists, the “it's not this but that.” The tone was much more variable. It would actually surprise you. It would be funny. It would be poetic. And that shocked me, to go back a few generations and realize that maybe they were also lying all the time and all sorts of other things. But from a writing-style perspective, I preferred it, and I wanted to investigate that.

They were weird.

That shocks me. To me, talking to GPT-2 was like talking to somebody who had just fallen down the stairs. You know what I mean? It was like, “Do I need to get you to the hospital? Do you smell toast?”

Yeah, there are these amazing prompts from this early OpenAI prompt library where they would say, “I just won $175,000 in Las Vegas. What do I need to know about taxes?” And GPT-2 would start just writing some short story about an orphanage.

But, yeah, they were surprising.

Yes.

They were nutty. They were weird. They would absolutely be a terrible corporate assistant, a horrible coding intern. It can't do any of the things that modern LLMs can do that I'm very grateful for. But from a pure writing-style perspective, they were very good—GPT-3 in particular.

I found this set of samples that some guy did where it was, “Oh, write in the style of Paul Graham. Write in the style of Richard Dawkins,” whatever, and it could style-match much better than modern LLMs can. And particularly because so much of literary writing comes from voice and style, one of the things I was really interested in was: What did we lose? The LLMs can no longer emulate Paul Graham's style or whoever's style.

Because I would put in the same exact prompt that this guy gave GPT-3 into ChatGPT 5.4 Thinking or whatever, and it would be god-awful.

Hmm.

And I was like, “That's really weird.”

So tell us about what you learned about what happened after the GPT-2 and GPT-3 era that changed the way that these models respond to us.

Yeah, I think the answer is post-training, basically. So they started adding a post-training layer, which is basically saying: We have these crazy, unpredictable, nut-job, concussed models, and they need to learn how to behave because a model that can't behave is a very bad corporate assistant.

And so the AI researchers give them example dialogues and scripts to learn from. They give them words that they can and can't say. They do RLHF, which is a process by which human graders will rate which response is the most helpful-sounding or something like this.

And so now these post-trained models have been trapped, in a way, or trained or guided toward a very particular character or persona that is a very helpful assistant, but might be very bad at writing in creative and surprising ways.

Mm.

I mean, the way that you described it was that there is a phase within the post-training phase where these AI models are evaluated by humans. And that's part of what they call RLHF, or reinforcement learning from human feedback.

And what struck me in your reporting is that you actually talked to some people who have done this kind of feedback, who say that they're just being asked to grade things in ways that don't make sense. Right? Tell us about that.

Yeah, this is super interesting because these job listings you'll see on places like Mercor or xAI, Elon’s company, will list them directly. It'll be like, “Creative writing expert, $45 an hour. Must be a New York Times bestseller,” and have a starred Kirkus review or something like this.

Have you ever gotten a starred Kirkus review, Roos?

I think so.

Okay, good job.

Not sure.

All right.

You might qualify to help Elon—

Yeah.

—to help Annie from Grok write a little bit better.

Yeah, we're gonna get on that job listing. But okay, you were saying.

Yeah, so these companies realize that these AI researchers are really good at knowing what good coding is, but they don't actually know what good writing is, so they're like, “Why don't we hire some humans to find out?”

And so they'll commission MFAs, published authors, and sometimes just random guys with a blog or whatever.

And one of the people I talked to, who was a contractor for Scale AI as a writing evaluator and was doing this for one of the bigger labs, said that the rubric just didn’t make any sense. He would be told things like, “You have to grade them based on the number of exclamation marks there are. If something has 3 exclamation marks, that’s too many, and so you have to ding that one.”

Yeah, and I have to say, generally not bad writing advice.

Yeah.

I guess it depends on the length of the text, but 3 feels like a lot for many scenarios.

This is what they tell women in business communications. It’s like, “Take all those exclamation marks, replace them with periods. We’re just gonna remove all of the exclamation points.”

We teach women to shrink themselves.

Exactly.

Yeah.

He was being asked to grade these things. In another case, he got a bunch of fan fiction, and he was supposed to grade it on its factuality, since that was one of the criteria. I do imagine that one could devise better rubrics than this particular evaluator was given, but I think it does show, at least, that some of these very big companies that are very well-resourced simply do not know how to think about what good writing is.

Briefly, I want to underline that, because to me, that seems like the whole story. We are taking the entire internet, and we are grading it on factuality. So the LLM that you’re gonna get out of that is just probably not gonna be all that creative. And I wonder how much of it is related to this sort of verifiable reward—

Mm-hmm.

—system that a lot of these companies are using, where you have a system generate a bunch of code, and then you have another evaluator model check the code to see whether it’s good or not. That works in domains like programming, where the code either runs or it doesn’t, but creative writing doesn’t work that way. You can’t have an evaluator tell you, with any sort of consistency, whether something is good or not, and so it may just come down to preference. So I guess I’m curious: Do you see this as a technical problem that the labs are frustrated trying to solve, or is this just demand-related? Is this just what people want chatbots to sound like, and in every test where they pit different models against one another, the one that sounds like a bland corporate assistant wins, and so they go with that?

I think both are true. The majority of writing that we are asking the models to do is, “Write this email for me,” right? And they excel at that. They are truly great corporate email writers. They are much better at the whole passive-aggressive thing than I am.

At the same time, I do think, like you said, there is a technical challenge that has to do largely with verifiability. There are people who have spent decades of their lives attempting to articulate what makes Shakespeare Shakespeare, or what makes a Neruda poem a Neruda poem, and they will still not know in any kind of certain way. They will still get into debates with their fellow academics and literary critics about which writer is better than the other, because these things are subjective, because they are ineffable, because they are hard to put in a rubric, and that is the nature of art.

And to that point, you started this segment by talking about Sam Altman saying, “Hey, we just basically can’t write a great poem yet.” Sam Altman, a year ago, said the company had trained a good creative-writing model and posted a short story on X. Many people found it compelling. Is Sam Altman just not being consistently candid with us, Jasmine?

Ooh. Wouldn’t be the first time. But that short story, if you remember, had some great lines, like talking about the seams of mirrors or Thursday, the… What was it?

It was the liminal almost-Friday or something.

Yeah, the liminal day that tastes of almost-Friday.

Wait, I have to actually look this one up—

It’s so good.

—because it was so good.

While you’re looking it up, the thing about AI writing is that it comes up with all of these fun metaphors, and those metaphors are sometimes surprising, but the language is not grounded in life. That was my other thing: Aside from the verifiability, fundamentally, when I think about the writers who I really love—whether it’s journalists or poets or whatever—they are writing from life, right? A journalist goes out and talks to people, and they see stuff and observe the color of the sky in a particular way, or a poet is thinking about personal experiences that they’ve had.

Their writing has stakes. It comes from an emotional place. And the fact that LLMs, while being very talented and grammatically pristine or whatever, don’t have lives means that all of the metaphors they choose, all of the words they choose, and the examples they choose are just ungrounded, right? It’s not coming from a point of view, or a particular experience, or a particular community that makes the writing believable. I think part of what voice and style are is that they are very specific to the life that a person has had, and LLMs cannot get there in the same way a human who hasn’t really lived that life cannot get there.

I don’t know. I feel like it’s case-dependent. I’m a big music fan, and over the past few months, I have enjoyed putting questions about music, and in particular the sounds of certain bands, to an LLM, which sounds like a joke prompt because an LLM has never heard anything—

Mm.

—right? And yet I find that, in general, the models can have good conversations with me about the sound of music. Now, it may be that they are just pattern-matching based on a bunch of public writing on the internet by people who do have ears—

—and have heard, right? I’m very open to that.

Yeah.

But, again, I have just been struck by the way that it is able to write about sensory topics in an evocative way that, at least to me, surpasses what I would predict they would be able to do.

Yeah. I want to pose a couple objections that I think—

Okay.

—someone might make to—

Perfect.

—your article. One of them is: This is cope. This is Jasmine, a writer, a very talented writer, sort of finding the things that AI, in her view, is not good at yet and saying, “This is categorical proof that it will be very hard for AI to do these things.” This is the same reaction that software engineers had when models started getting really good at code. They would say, “Oh, well, it can’t do these other 10 things that I do,” and then, basically, just wait a few years, and the models will be better than all of us at everything, including writing.

I would love for it to be cope, because I try to automate myself away all the time. I have no deep attachment to having to do it. I like writing, but I have tried over and over and over for the past 3 years to automate my own job away and to get Claude to do my job for me. It cannot do it. This is very frustrating.

Mm-hmm.

It’s not out of a lack of trying. Again, I’m going back to the CEOs themselves and the things that they themselves are saying, right? It’s not just me, a writer; it’s Sam Altman saying, “This thing will cure cancer and solve physics, but it will not write better than a real poet’s okay poem.” And so I think that suggests that there is something that is at least perceived as a little bit different.

I think it’s very possible that the models will get much better at writing over the next few years. I don’t think it’s a never thing. I do think that reporting is hard to replicate. I think that having life experiences that are real and verifiable is hard to replicate. I think the style stuff can be improved, especially if you fine-tune the models. But I think what’s also interesting to me about this piece is that it shows how the market incentives, the demand incentives of these companies, do shape what we see as their abilities today.

Mm.

The other objection I’m imagining people might have, who are very AI-pilled, is—

Mm.

—that this is all in the eye of the beholder, right?

Mm-hmm.

There have been several studies now that have shown that if you give people a blind taste test of AI writing versus human writing, they prefer the AI writing until you tell them that it’s AI writing, and then the value in their eyes plummets. I did one of these in a New York Times quiz just recently. So is it possible that the models have already become superhuman at writing, but that the minute we learn that they are AI models generating text and not humans writing words with their fingers, we lose all interest in it just because of the source, not because of the quality of the writing?

I mean, I think it’s definitely interesting and true that people don’t want to like AI writing, and that is part of what bothers them when they see AI text that is obviously AI, even though, as you said, in these quizzes and tests, AI can outperform human writers in those narrow scenarios.

My quibble with a lot of these quizzes and tests is that, as a writer—and you guys are writers too—how much of your job is actually text generation? I think AI is a superhuman text generator, right?

Mm-hmm.

In my job, I am generating text probably 25% of the hours in my day. I spend a lot of time interviewing people. I spend a lot of time coming up with ideas. I spend a lot of time reading, and not just reading indiscriminately, but reading very particular sources that feel like the right ones.

Usually, at the point that you are doing one of these tests, you're saying, “Generate one paragraph very specifically about why Trump won the 2016 election, 500 words or less.” You've already given the prompt, which I think is a critical part of writing: What are you going to write about? You've often supplied some of the evidence and the guidance in the form of saying, “500 words or less,” and at that point, I do think that AI is probably a better text generator than almost all humans are.

But again, when I think about it, AI is still very bad at coming up with ideas for articles. It is still very bad at reporting. The non-text-generation parts of the role feel further away from automation. Again, I'm sort of a “never say never” person. Maybe it'll get there. I would be totally happy if Claude was able to give me good ideas for my next essays, but it's not there yet.

Well, we're already seeing LLMs make huge progress in genre fiction, right? Recently on the show, we talked to the author of a story in The Times about how authors of romance novels are now able to generate dozens of novels a year using LLMs. In fact, much of the discussion that we had was around how you just have to prompt them differently and relentlessly in order to get what you want.

Your piece, Jasmine, made me wonder: How much of getting a model to just write weird can be achieved by repeatedly telling it, in different ways, “Hey, be a little weirder”?

Some of it, but not all of it. I talked to, for example, James Yu, who is the co-founder of Sudowrite, which is one of the earliest creative-fiction AI writing assistants. I talked to some other folks who similarly were in the fiction-writing LLM space.

And like you said, to an extent, a lot of writers are already using these, already leaning on LLMs to generate large amounts of text, and it can be very successful, and it can meet readers' needs and whatever. But even these people who I was talking to were describing to me how freaking hard it is to undo all of the post-training that the labs have done.

They were applying immense amounts of engineering effort, which, in my conversations with them, clearly frustrated them, because it is so hard to get these models to stop being so chirpy, so sycophantic, so PG-13 and everything, in order to get them to this sort of base-model state where they're able to be weird again. So I think it's certainly possible, but I think the labs have made it quite challenging just because of the way that these models are trained.

The other thing that I think is important is that I tend to think that writing and a lot of creative work is actually the perfect use case for these centaur models, right? The idea that the human-plus-AI collaboration is where you can get the furthest. And when I listen to the interviews that you guys did about the fiction authors, I was thinking, “This is a centaur model,” right?

Without the human prompting and bullying the AI into getting weird and getting sensual and whatever, it was not going to do that on its own. I myself do use LLMs as a research assistant. I wrote about that inside The Atlantic piece about the way that Claude has now helped me edit my own work in a way that I found incredibly useful. But I do feel like the collaborative element is important for any domain where the personal perspective, lived experience, whatever, really matters.

Talk about that a little bit. You mentioned your editing process. How are you using AI to help you edit your work, and are you finding it useful?

Yeah, I feel like I really cracked this over the last couple months, which I'm very excited about. Because, again, I've tried to make these things write and edit for me over and over and over, and they've never really been able to do it.

So the thing that I realized was, if I make Claude into an editor that is not just trying to grade and give feedback on my work against some genericized standard of what good writing is, but actually does it against basically what my personal aspirations for writing are, it can give feedback that I find much, much more helpful.

So what I did was basically feed Claude my entire Substack archive of the writing that I've previously done, as well as some of my freelance work.

And just to get real specific, is this inside a Claude project, or how have you set this up? Because I know our listeners are going to want to try this.

Yes. I did it in a project—

Okay.

But on Claude's advice. I was like, “Do I need to Claude Code something?” And Claude was like, “No, that's overkill.”

Okay.

You don't need to code or anything. So, in a Claude project, I gave it my whole archive of writing. I also personally write retro notes to myself after everything I publish. So I have a notes app that's just me writing what was good and bad about everything I've ever written.

Hmm.

Just a few bullet points.

This is why Jasmine's gonna be our boss.

For sure. These are very low-quality bullet points, but I also gave it that because I wanted it to learn my taste. I wanted it to learn: What do I aspire to be? Where do I see myself falling short? And what am I proud of, right?

And so from those two things, plus a little bit more information about, “Here's my audience. This is my beat. These are my goals,” we were able to co-develop a rubric. Instead of asking, “How many exclamation marks does it have?” it would say things like, “Does this take advantage of your, quote-unquote, ‘insider anthropologist position in Silicon Valley?’” That's one of the things that Claude and I think distinguish my voice.

Or it'll also notice, “Oh, Jasmine, you tend to move between registers. You'll switch between startup jargon and internet slang and whatever. And I think the fact that you can do the hi-lo or move from policy to a personal scene is something that is characteristic of your writing.” And so again, we're co-developing these qualitative criteria.

Then I split it into phases: ideation phase, structure rubric, prose rubric, and final fact-checking. What I do now is put this all in a Claude project. I said, “Your job is to evaluate my drafts based on these criteria, but not to do the writing for me, and to make sure to prompt out of me what I can do better.”

I dump the draft into Claude. Claude will run phase-two structure on it. It'll say things like, “Your conclusion is just a summary, and this is really boring. In fact, in your piece about this and that, you actually ended on a scene, and I thought that was much more powerful, so why don't you try ending this one on a scene?”

And Claude will say, rather than inventing a scene, “What were you thinking when the plane took off? What were you feeling inside? Can you think of a scenario where you had a conversation with, say, a kid-safety advocate about AI that really resonated with you? Because right now it sounds like a dry policy explainer.” And that feedback I actually found incredibly useful.

Hmm. It is.

I'm still applying my own judgment to say, “Do I take it or not?” But this is about me becoming the best version of myself as a writer. It's about me self-improving and Claude pushing me to do that, which I found much, much more helpful.

Hmm.

Wow.

I want to ask you both a question as fellow writers. Do you feel the impulse to make your writing weirder because of AI to sort of stand out from the sea of slop? Because I find myself feeling this tug of, “Oh, that's a little weird aside that probably I should cut, but I think I'm gonna leave it in, because Claude would never do that,” right?

Mm.

It's like a marker that I am typing these words, and I feel like that's sort of my imprimatur that I'm leaving. My answer to you is yes, I absolutely feel that way, and I've gone back and tried to edit sentences to make them feel a little bit weirder or, in particular, to make them sound colloquial in a way that I know an LLM generally would not. And yes, it is for that reason.

I think that writing right now, we're all—not all, many of us are—on such high alert for the prospect that we might be reading slop that if you are a writer who does not want to be producing slop, you should be asking yourself that question.

Mm-hmm.

I think it makes me a lot more comfortable writing the way I want to write in the first place. I think maybe, unlike both of you, I didn't sort of come up through newsrooms where I was learning a very specific house style and all of these norms.

I can do news writing now. It’s something I’ve learned now, but I’m actually much more, quote-unquote, Internet- and blogging-native, which is a form that is voicey and irreverent and not as pristine, and will make inappropriate jokes. It’s just a looser form of writing. And so I think what it’s actually done is made me more comfortable doing the bloggy thing instead of always trying to write in a more professionalized journalistic tone.

Hmm. So I think we should leave this with a question for you, Jasmine, which is: Your piece makes the case very convincingly that today’s AIs are not very good at the kind of writing that I think we all value. Do you think they will get there, and what should the companies do to make their models better at writing?

I think that if we separate out text generation from reporting, which I’m not that bullish on the models doing, and we’re just talking about, say, literary fiction, or “Here’s a bunch of interview transcripts; write a magazine feature” or something, I think that if they applied as many resources toward that task as they do toward coding agents and things that actually make the money, I think that they could get there.

Will the companies ever find it financially advisable to spend all their resources on that instead of automating 23-year-old software engineers? Probably not. I would be grateful for that world. I don’t need them to take my job or these folks’ jobs, but I think it’s possible.

Look, they’re going to get around to it eventually. Okay? I hear what you’re saying—

Have you seen what writers make in this economy, Casey?

Eventually, like, they—

Those aren’t going to pay for a lot of data centers.

No, there is economic value in writing. And eventually, the AI companies will want that all to themselves.

You know what would be a very funny outcome of this, taking your point about the sort of guardrails of the models: Maybe the next great American novel will be written by Grok.

Oh, God.

And with that, Jasmine Sang, thank you for joining us. Thank you, Jasmine.

Thank you very much, Kevin and Casey.

Well, Kevin, you’ve recently returned from book leave and are once again writing in The New York Times. How does it feel to see your name in print again?

Feels great. It hasn’t happened yet, but when it does, it’ll be great.

Well, I got to take an early read at a story that you are publishing about the fact that tech companies have now created leaderboards to show which employees are using the most AI tokens in their work.

7. Tokenmaxxing Becomes A Workplace Metric

Yes, it’s a token frenzy out there, and the employees of these companies are competing among their colleagues, informally and for fun, but they’re taking it very seriously. They want to be the people at their company who are using the most AI tokens.

So let me just ask a basic question for listeners who may not be familiar. What is a token, and why is that something you might start keeping track of?

So a token is the basic atomic unit of AI labor. It’s basically a fragment of a word, and it is how AI model providers measure their consumption. So if you type in a prompt, “Help me write this essay,” an old model might have given you a couple hundred tokens in response. That would be a couple hundred words.

What has been happening over the past year or so, as these agentic coding tools have started taking off, is that the models are just much more token-hungry. You can use now hundreds of thousands or even millions of tokens in a single session, and so that is what is propelling these leaderboards: the idea that the more coding you’re doing, the more agentic tools you’re using, the more simultaneous processes you’re running, the higher your token count will be.

One measurement I found useful was that apparently it takes about 10,000 tokens to generate 7,500 words, if that helps to ground you at all. But as you just said, and I want to hear more about this, the more advanced systems are using way more tokens than that. So tell me about some of the numbers that some of the token all-stars are putting up on the boards.

So I don’t know all of the exact numbers, but I did learn that at OpenAI, where they do track this kind of leaderboard, the highest employee token count over a 7-day period recently was a guy who used 210 billion tokens. And this is, for rough scale, about 33 Wikipedias’ worth of text.

Hmm.

Now, all of that is not typing and receiving a response. Some of that is what they call cached tokens. So it’s not all being extruded from the model for the first time. But these are the kinds of numbers that I think even a year ago would have sounded completely insane.

Right. Now, is this guy working on a new mass domestic-surveillance program for the Department of Defense?

I don’t know, and OpenAI did not make him available for interviews.

Oh.

But what I wanted to do in writing this column was to try to call up a bunch of people or talk to a bunch of people who are in this billion-token club, the extreme power users, and just ask them, “Hey, how are you guys using all those tokens, and isn’t that very expensive, and how are you paying for it all?” And I learned a lot.

Yeah. Well, okay. So tell us, first of all, just how expensive it is.

Very expensive.

Yeah.

In fact, I heard that the top user of Claude Code, the top individual user of Claude Code, as measured by Anthropic, spent more than $150,000 on tokens last month. So extrapolate that. That is like an employee making more than $1 million a year.

They are burning that in a month, and I heard similar figures from some of these other extreme coders who are spending something on the order of thousands of dollars a day on tokens from these models. Now, we should also say the employees of these companies get their tokens for free, right?

Right.

So they are not shelling out; their companies are not shelling out. But at other companies, this is starting to become an issue because they are outstripping their budgets for these things.

So there are companies where there are engineers who legitimately are costing their employers maybe $150,000 a week because they’re getting tokens from one of the big providers.

Yeah, I talked to a software engineer in Sweden who said that he probably spends more than his salary on Claude. So this is essentially becoming a very expensive job perk for some of these coders.

8. Leaderboards Create Perverse Incentives

So talk to me about why employers want to create leaderboards to promote this to employees, because I could see other companies saying, “If you spent $150,000 on tokens last month, you actually don’t work at this company anymore, because we’re bankrupt.”

Right. So this was a big question that I had: Why is this going on? And it seems to be some combination of employee motivation and worker tracking, right? There are executives at these companies who think that the more tokens you use, the more productive you probably are.

And as we discussed in a previous segment on this show, these companies are very eager to have their workers start embracing the AI tools. And so at a number of these companies, I talked to people who said, “Yeah, this is just basically them trying to see who is really all in on the new way of programming.”

And you’ve talked to a number of people who are ranking high on these leaderboards. I realize you probably haven’t dug deep into their code, but what is your sense of how productive they actually are? What is the relationship between token usage and taking my company to the next level?

I mean, it’s very unclear, right? Some of these people may be just generating worthless projects.

I think the thing that worries a lot of the people I talk to about these leaderboards is that they just incentivize you to run up your token count, right?

Yes.

Because then you look like the special 10X engineer or 100X engineer who’s outperforming all your colleagues. So I think there are a number of companies that see this leaderboard business as a little strange and maybe counterproductive. But I do think that there is a feeling among the most heavy token users that they are being productive.

Yeah. I have to say, when I read your column, I thought, this just seems like it would create the worst incentives, right?

Yes.

There’s this idea of Goodhart’s law, right? When a measure becomes a target, it ceases to become a good measure. I can’t think of a better way to ensure that token usage becomes a bad measure than creating a leaderboard for it.

Totally.

What are the people inside the company saying about that?

Well, some of them are opposed to this whole leaderboard thing. I also talked with some folks who defended the leaderboards. They said, “Look, it’s never been all that easy to track the productivity of programmers.”

Some people have had their productivity measured by how many lines of code they generate or how many pull requests they made. These are imperfect proxies for how hard you’re working and how much you’re doing.

But the employees of these companies also see this, I think wisely, as a key to their own success. A number of these companies are now using A.I. token use and consumption as part of the performance review cycle. So you go in for your annual review, and your boss says, “Hey, it looks like you only used 70 million tokens last month. What’s going on?”

I think the engineers at these companies are getting wise to the fact that if they want to have a long, successful career, they better start using some tokens.

Yeah, but I imagine that some of them are really nervous about that, though, right? Because it seems clear to me that at least some of these companies want to incentivize token usage because the companies themselves suspect that the more we can get them using this stuff, the less long we will have to employ the humans.

Maybe, although I think it’s less about the A.I. systems replacing the humans and more about it being a radically different way of working, right?

Mm-hmm.

These are people who, most of them, have had long careers in software engineering. They grew up writing code by hand. They maybe grew up using some sort of A.I. assistant, like GitHub Copilot, and what people at these companies are saying is that these agentic engineering systems are just really different.

You have to approach them in a different way, and you have to spend a lot of time with them to understand what they’re good and not good at. To them, this is a way of motivating their employees to say, “Hey, go out and try the new thing.”

Yeah. I don’t know. I’ve been thinking a lot about this question of, if I were an engineer at one of these companies and I had this incentive to get on the leaderboard, how would I approach it? I do think that the instinct to waste a bunch of tokens to rise higher on the leaderboard could ultimately backfire. If you rise too high, people are going to ask you what you did with all the tokens.

Right.

If you’re number 1 at 10 billion tokens and you only managed to vibe-code a calculator or something, people are probably going to get mad at you.

Yeah, and I actually did talk to one person who speculated that the people at the top of the leaderboards are all doing side projects. They’re starting their—

They’re starting a new company.

—their side hustles.

They start a new company with the boss’s money. And if you’re doing that, I just want to say I salute you.

Yeah.

That is the right way to work in 2026.

Yeah. Maybe don’t be number 1 on the leaderboard if you’re doing that. Maybe try to stick around 6 or 7.

Yeah, like middle of the pack—

Yeah.

—is kind of where you want to aim yourself. I mean, let me ask: Is there any kind of token tracking that you think offers a reasonable signal? Do you think that if you’re a tech company, you should create a leaderboard?

No. I think that’s a bad idea for all the reasons that we just talked about, including—

Yeah.

—Goodhart’s law, which is that I think this is just going to lead to people wasting tokens and doing side projects. But if I’m the budget manager at a company and I’m seeing that people are spending multiples of their salary on A.I. tokens, I’m asking them some questions about what they’re doing with all that. If their answer is not, “I built an amazing new product that’s going to generate billions of dollars a year in revenue,” I’m trying to say, “Hey, could you maybe use a little less next month?”

Yeah. I have to say, I have been struck by how this idea of the token leaderboard represents a new incarnation of something that the software industry has been trying to figure out for a long time, which is: How can I figure out if my software engineers are productive?

I was talking recently to this very handsome software engineer who I’m engaged to about your column. He was telling me that he used to be evaluated on how many lines of code he contributed, and he told me about all the games that people used to play back in the day. “Oh, I wrote a quick algorithm to translate a bunch of stuff into some new languages, and it’s completely worthless, but it makes me look like I had a very productive week.”

And so I went back and looked into this, and they were doing this in the ’60s and ’70s. There’s this saying from the early days of computer programming that says, quote, “Measuring programming progress by lines of code is like measuring aircraft-building progress by weight.”

I have to say, I think the same thing kind of applies here, right? If you squint and look at it at the right level of abstraction, it’s probably true that some people who are using a lot of tokens are more productive than some people who aren’t. It just doesn’t quite seem like the right way to measure these things, and I just wonder how quickly the industry is going to figure that out.

Yeah. I think it’s going to be pretty soon, in part because the budgets are just getting very ridiculous. And especially the A.I. model providers are now seeing individual users consuming amounts of their services that entire companies would have consumed just a few months ago.

You know, maybe the last question I have for you about this is: What implications do you think it has for the broader economy, right? Because we know that in so many different sectors of the economy, managers are saying, “I want to incentivize my employees to use A.I., and I want to track how they’re using A.I.”

Do you think that as knowledge of these leaderboards spreads, we’re going to see people in nontechnical fields try to adopt their own version of them?

I hope not. I think it’s really a bad move, not just for tracking actual productivity and output, but just for morale, right?

I remember years ago when Gawker would have a traffic leaderboard at its office, so you could see how many clicks your stories were getting relative to other people. I don’t think anyone who worked there at the time thought that was incentivizing the right things or creating high morale among employees. Basically, everyone was just competing with each other all the time.

And I think in this case it’s even worse because it’s not necessarily even correlated with any success.

Mm-hmm.

It’s just pure, sort of, how many agents can you run in a parallel swarm to work 24/7 doing tasks of uncertain value?

Which is a great question to ask on a first date in San Francisco, too, by the way.

But anyway, I have to say, I worry that this idea of tokenmaxxing is going to spread into the broader economy. I was talking with somebody who works in marketing this week, and she was telling me that her job used to be evaluated solely on creativity. Then recently, the performance review got a new A.I. section, and everyone is being evaluated on how much A.I. they used.

From her perspective, she was like, “This was working fine. I didn’t need to use an A.I. tool to help me, but now my bonus might be based on how much of it I use.”

So I think this thing has already seeped out of the labs and is getting into the water elsewhere. I just hope that managers are really thoughtful about what they are incentivizing, and that maybe A.I. use for the sake of A.I. use is not going to be the boon to your company that you’re hoping it is.

Yeah. I think it’s going to be very case by case. I think there will be people who are tokenmaxxing who are way more productive than their colleagues and doing way more projects way more quickly. I think there will be other people whose managers look at their token budgets and say, “You spent this many tokens on what?” and will have to have some hard conversations.

But I think it’s very hard to draw with a broad brush and say, “All of this tokenmaxxing is pointless productivity theater.” It sounds to me, from my conversations, like some of it really is working for people.

Yeah. Well, on the flip side, I’ve also heard of people in my social circle who have gotten in trouble for spending too much on Claude in their life.

Wait, really?

Yeah. When I heard that, I was like, “Oh, your company’s not going to make it, bro. You’ve got to spend on this stuff.”

Well, what’s so interesting is that now it’s becoming part of job conversations for engineering jobs. People are going into new jobs and saying, “Well, what’s my token budget?” And for the employees of these big AI labs who have unlimited free access to the models, some of them are using so many tokens that they effectively can’t afford to quit their jobs, right? Because anywhere else they would work would have to pay for their tokens, and it would be completely unaffordable to employ them.

Yeah. I mean, those sound like real incentives, and better than the ones at Meta. Do you remember when Meta was spinning up superintelligence labs and they said, “You can sit really close to Mark Zuckerberg”? If I were them, I’d be like, “I’ll take the tokens, thanks.”

All right. Well, just to wrap this up, exactly how many tokens should a person use?

I think that’s something you have to look within yourself for.

Look within yourself?

Yeah.

Okay.

Yeah.

That’s between you and your God.

Yeah.

Yeah.

Do what Marc Andreessen will not: introspect.

Introspect.