AI行动方案 + 打破求职面试规则的大学生 + 混乱快车

I have a bone to pick with you.

What’s that? What’d I do?

So on Saturday, as you know, we had a birthday party at our house for my son.

Wonderful birthday party.

And it was also a housewarming party.

Housewarming party.

And you and your boyfriend came.

Mm-hmm.

Lovely to see you there. Thanks for coming. But you brought him this present. We specifically said no presents.

You did say that.

And you brought him this present that was called the Dino Truck.

Yes, and here’s why. I know that your son loves trucks, and I thought, “What is the best kind of truck I could think of?” That would be a truck that was also a dinosaur and was full of dinosaurs. And so that’s what I got him.

Yes. It’s very “Pimp My Ride” coded because it has a dinosaur truck that contains 12 other dinosaur trucks.

That’s right.

And you sort of assemble it all together.

Mm-hmm.

But my son has not stopped playing with it. He absolutely loves it.

Yes.

And as a result, about twice a day, I now step on a very painful Dino Truck that has been left somewhere around my house. So he’s loving it. I am not.

Oh, no.

I mean, when I think of the best kind of gift I could get for the Roose family, it is something that your son enjoys and that causes you physical pain. So I think that was a slay on my part.

Mission accomplished.

Mission accomplished.

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, America is building an AI action plan. We’ll tell you how tech companies are trying to exploit it. Then, Columbia University sophomore Roy Lee joins us to talk about the tool he built to help software engineers cheat their way through job interviews and why he might get kicked out of school over it. And finally, the Hot Mess Express is once again rolling into the station.

Well, Kevin, I have an action plan for today’s episode.

Okay, let’s hear it.

I want to talk about action plans.

Hmm.

1. The AI Action Plan Race

So, Kevin, as you know, because you wrote about it this week, there have been these AI action plans that all the big AI companies, think tanks, and nonprofits have been submitting to the Trump administration over the past couple of weeks.

Yes. There was the Paris AI Action Summit, at which no action was taken or really even proposed. Then the White House came forward and said, “We’re going to make our own action plan. Why don’t you companies and anyone else who wants to make a public comment go ahead and tell us what you think we should do?”

These kinds of public comment periods are not unusual. Agencies of the government open themselves up for submissions from the public all the time on various issues. But this one caught our eye because it was related to AI, and it was essentially the Trump administration trying to figure out what to do about AI and the potential that AI is going to accelerate during the 4 years that Donald Trump is in office.

Yes. I think that’s how the Trump administration saw it, and I think for the big AI companies, Kevin, it was really a chance to present the president with a list of their absolute fondest wishes and dreams for what the best possible deal they could get from the government would look like.

Yes. So I think there’s some interesting stuff in them, but I also think there’s a broader interesting story about how the tech companies want or don’t want government to be involved in helping them build and manage these very powerful AI systems.

Yes. Let’s get into it.

Okay. But first, because this is an AI-related segment, we should make our standard disclosures. Do you want to switch it up this week? Do you want to do mine and I’ll do yours?

Yeah, sure. The New York Times is suing Microsoft and OpenAI over alleged copyright violations.

Correct. And Casey’s Manthropic works at Anthropic.

That’s right.

Okay. So you wrote about these submissions this week. Where do you want to start?

Well, let’s start with some of the things that are a little bit less controversial, right? I think there are some pretty good ideas in these action plans, and I actually think the Trump administration will probably follow through on them. For example, they talk about wanting to expand the energy capacity we have in the United States so that we can have the power it will take to do everything with AI that we want to. They also talk about encouraging the government to explore positive uses of AI, potentially delivering better services to citizens. That would be good if it happened. So there’s a lot in these documents about that.

But once you get beyond that surface layer, Kevin, there is a lot of essentially what these companies have always wanted the government to tell them, and they are now finally getting a chance to say, “Hey, please, please, please do this.”

And what are those things?

2. The Copyright Exemption Push

For example, they are really, really excited about the idea that Donald Trump might declare definitively that they have carte blanche to train on copyrighted materials. Now, this is, of course, at the heart of The Times lawsuit against OpenAI. But it’s not just OpenAI that wants the green light to do this, right? All these AI labs are under similar legal threat.

So it’s in Google’s AI action plan. It is in Meta’s AI action plan. In fact, Meta says that Trump should unilaterally, without Congress, just issue an executive order and say, “Yeah, it’s okay for these AI labs to train on copyrighted material. Go nuts.”

OpenAI, I think, had a frankly ridiculous statement in its AI action plan, which is that if Trump does not do this—if he does not give AI companies carte blanche to train on copyrighted materials—we will immediately lose the AI race to China, and it will just be DeepSeek everything from here on out.

Huh. I mean, obviously they have an interest in making that case and having the Trump administration give them a free pass. But can they actually do that? Could Donald Trump issue an executive order tomorrow and say, “There’s no such thing as copyright anymore when it comes to the data used to train large language models?”

Well, Kevin, lately the Trump administration has been issuing a lot of executive orders that people have said, “Well, hey, you’re not allowed to do that. That’s actually not constitutional.” And yet he keeps doing it. Some of these things have been struck down by the courts, and some haven’t, and there seems to be a kind of flood-the-zone strategy where we’re just going to do whatever we want, and the courts may undo some of it, but they’re probably not going to undo all of it.

So where would a copyright executive order fit into that? I don’t know.

Yeah, I mean, my hunch is that this will not happen via executive order, that it will be left up to the courts to decide. But yeah, I mean, it’s certainly in their interest to argue that this all should be allowed and kosher and to sort of preempt any potential litigation against them. Was anyone opposed to that idea?

Yes. So a group of more than 400 Hollywood artists, including Ben Stiller, Mark Ruffalo, Cynthia Erivo, and Cate Blanchett, signed a letter saying, “Hey, do not grant an exemption from copyright law to these AI labs.”

Their argument was essentially that America has a lot of cultural leadership in the world. So much global culture is downstream of American culture. And they said, “If you create disincentives for us to create new works because we can no longer make any money from it economically because AI just decimates our business, we are going to lose that cultural leadership.”

And so I would actually call on Ben Stiller, Mark Ruffalo, Cynthia Erivo, and Cate Blanchett to come on the Hard Fork podcast and tell us more about that. We’d love to meet you and hear your stories.

Yeah.

Yeah.

I would call on them to frame their opposition in the form of a musical.

Ooh.

Cynthia Erivo in particular. I have a proposal for the showstopper tune of that musical.

Have you written it?

Yeah.

Okay.

It’s called “Defying Copyright.”

Oh, boy. Wow. You didn’t even try for a rhyme. You know what—

Oh, wow.

When it comes to copyright violation, Cynthia Erivo is decrying depravity. And that’s how you do it, Kevin.

Okay, back to the serious issues in these AI action plans, Casey.

Yeah. There’s another big plank that gets repeated in these submissions, Kevin, and that is this idea that these companies do not want to be subject to a thicket of state laws about AI, right?

Yes. Basically, what the AI companies don’t want is, in the absence of strong federal regulation on AI, they don’t want California to pass a bill governing the use and training of large language models, Texas to pass a bill, Florida to pass a bill, and New York to pass a bill.

They don't want to have to go through 50 states' worth of AI regulations and make sure that all their models comply with all the various state regulations. They have wanted for a long time, and are now making explicit, their desire for a sort of federal law, statute or executive order that would essentially say to the companies, “You don't have to pay attention to any state laws because the federal law will supersede all that.”

Yes. And in particular, Kevin, they are worried about state laws that would make it so these companies could be held legally liable in the event that their products lead to great harm, right? There was some discussion about this in California last year with a Senate bill that we've talked about on the show, and there's a lot of fear that other states might take a similar approach.

And so this plank in these plans, where these companies are saying, “We don't want a thicket of state laws,” works in a couple of different ways. I can understand why they don't want to have to have a different version of ChatGPT in 50 different states. That would obviously be very resource-intensive and annoying. At the same time, these companies know full well the country they live in. They know how many tech regulations we've passed in this country in the past 10 years. There is only 1 of them, and it was to ban TikTok. And it turns out that even when you pass a law banning TikTok, TikTok doesn't get banned.

So I think that there is a bit of cynicism here, and that they're saying, “Oh, please, please, please, let there not be any state laws. Just pass a federal one.” They know that there is very little likelihood that that is going to happen anytime soon, and so in the meantime they can just operate under the status quo, where they don't have direct legal liability for any bad outcomes that might arise from a future large language model.

So I went through a lot of these proposals, and I think there's some interesting stuff in them, sort of around the edges. There was a lot of talk about the security of these models and trying to harden the security of the AI companies themselves so that, for example, foreign spies aren't stealing the model weights and sending them to one of our adversaries or things like that.

By the way, I love that word. “Oh, we have to harden our defenses. We have to make them so hard. We have to harden our posture.” I don't know when we started saying that, but—

Casey, this is a family show.

It's very evocative is all I'm saying. Anyway, go on.

So there's some sort of small-bore stuff in there that felt interesting.

“Small-bore,” by the way, two words often used in reviews of this podcast. I don't know why I keep interrupting you. I'm just trying to get the energy level up, but we're doing great.

That's fine.

All right, tell us more.

Some of the plans contain some just sort of weird, interesting ideas. For example, in OpenAI's proposal, there's this idea that 529 plans, which are the plans that parents can start to save for their child's college education, should be expanded so that they can be used to pay for things like getting an HVAC technician credential.

Mm-hmm.

Because they say we're going to need a lot of HVAC technicians in all these data centers. They're going to power all these AI models, and right now kids are being incentivized to go to college and get 4-year degrees in various subjects that may not be that relevant. But we're definitely going to need a lot more HVAC technicians. Is that going to change the world overnight? No. Is the Trump administration going to take that seriously? I have no idea. But that's the kind of thing that I was surprised to see in there.

Yeah.

3. The Labs Want Freedom

But what I found more interesting was what was not in these proposals, right? These companies, and the people who lead them, have big, radical ideas about how society will change in the coming years as a result of powerful AI systems. Sam Altman has been interested for years in universal basic income. He funded a universal basic income experiment to try to figure out what an economy after AGI would look like and how we would provide for people's basic needs.

There are executives who are trying to solve nuclear fusion to power the next generation of AI models. There are people who want to do things like Worldcoin, which Sam Altman also funded, to give people a way to verify that they are humans. You can imagine a world in which the AI labs were saying to the government and the Trump administration, “Hey, we have all these ambitious plans. We want your help.”

Mm-hmm.

“Please help us come up with a UBI program that might make sense for people who are displaced by AI. Help us come up with some kind of national proof-of-personhood scheme, or help us build fusion energy.” But they're not asking for that stuff. What they're asking for instead is basically, “Leave us alone and let us cook.”

Yeah.

And it really makes me think that these labs have decided that it would be more trouble to have the government in their corner actively helping them than it would be helpful.

Yeah.

And so my read of these proposals is that they are trying to give the government some things that it can do that will make it feel like it is helping and sort of clearing the path for AI, but that they're not calling for any kind of federal Manhattan Project for AI because my sense is that they just think that would be inviting trouble.

Yeah. And I mean, they might be right about that, right? I'm not sure exactly what the government could or should be doing to help OpenAI make a better version of ChatGPT. But I think I would go a step further than what you said, Kevin, because it isn't just “Leave us alone.” They're really telling the government, “Leave us alone, or else.”

There is a boogeyman in these AI action plans, and the boogeyman is DeepSeek. DeepSeek, of course, is a Chinese company that emerged with a model called R1 earlier this year that shocked the world with how much it had caught up to the state of the art and has really galvanized the attention of Chinese leaders around the possibilities of what AI can do in China.

And so when you read the OpenAI and Meta action plans in particular, they're saying, “Look at DeepSeek. China is so close to us. You really need to let us do exactly what we want to do in the way that we are already doing it, or we're just going to lose to China, and it's all going to be over for us.”

Yeah, yeah, I noticed that, too, and I think we've seen that being telegraphed at things like the Paris AI Summit, where there was a lot of talk about China and foreign adversaries that were catching up to the state-of-the-art AI technology. But to me, that feels very calculated. That is the role that the AI companies want the government to play. Other than just getting out of their way, they also want them to hobble China—

Mm-hmm.

—and make it hard for China to catch up to them in the state of the art.

And there's a genuine read of that that is like, “We're worried about Chinese companies getting to something like AGI before Americans, and what happens if their values rather than ours are embedded in these systems, and they just use them for surveillance on their own citizens and things like that?” The cynical read is like, “We have this new competitor, and we would like the U.S. government to step in and make things actively harder for that competitor.”

Yeah. And look, I think there are reasons to be worried about what an adversary could do with a really powerful AI, so I don't want to dismiss these concerns completely, but I do feel like some of these labs are trying to use the specter of China in a pretty cynical way.

My favorite story about this issue, Kevin, does have to do with Meta. Meta writes in its proposal to the government a lot about DeepSeek, and Meta's number 1 priority in its action plan is that it continues to be able to develop what it calls open-source AI. Now, Meta's AI is not actually open-source. There are a lot of restrictions on how you can use it. Most people would call it open weights instead of open-source because you can download the model weights but not the actual source code.

Okay, we're a little bit in the weeds, but I do feel all of us need to say—

Our listeners have fallen asleep. Wake up!

Yeah, okay, so let's just wake up by saying that Meta says to the government, “Look at what DeepSeek is doing. If you don't let us develop in an open-source way, DeepSeek's own sort of open-weights approach could spread all across the world, and it will have these authoritarian values embedded in it, and we will just sort of lose out on the opportunity of a lifetime.”

Why is that funny to me? Well, Kevin, it's because in November, Reuters reported that Chinese researchers had used Meta's Llama model to create new applications for the military.

Oh, boy.

And look, does that mean that China used Llama to build a giant space laser that's going to vaporize the Eastern Seaboard? No. But it does suggest to me that this idea that we have to release, quote, open-source AI in order to save us all is probably not the right answer.

Yeah, and if anyone from the Chinese military is listening to Hard Fork, please don't develop a space laser using Llama. That seems scary.

That's our AI action plan.

No space lasers.

So before we wrap up talking about these AI action plans, I want to point to a few good ideas that I saw in them. Many of them came from groups other than the big AI labs, but I thought there was some interesting, sort of off-the-wall stuff that I hope the Trump administration is paying attention to. One of them was this proposal from the IFP, the Institute for Progress, which is a pro-technology progress think tank.

IFP says we're going to need a bunch of data centers and a bunch of energy sources to power those data centers, but all that requires building physical infrastructure. It can be quite slow to build physical infrastructure in many parts of the country due to things like environmental regulations and zoning and things like that. So they proposed creating these things called special compute zones, where you would essentially be able to build, in a much less restricted way, the infrastructure to power advanced AI systems.

That's actually what I call my office: a special compute zone. When I see guests going in there, I say, "Hey, get out of there. That's a special compute zone."

Yeah, so that was one interesting idea from the IFP proposal. I also—

Did the Institute Against Progress have any interesting ideas you want to share?

Well, there isn't an Institute Against Progress, but there are some organizations, like the Future of Life Institute, that are much more concerned about the development of these powerful systems. This is one of these organizations that's been around for a while. It's concerned with things like existential risk and runaway AI.

One of the ideas that they put in their proposal was that all AI models of a certain size and power should have kill switches on them. Basically, in order to release one of these things, you should have to build it in a way that an engineer can shut it down. And the way that they pitched this to the Trump administration was, this is a way to protect the power of the American presidency.

Yes.

Right? As the president, you wouldn't want some AI system going rogue and becoming more powerful than you, or allowing another world leader to become more powerful than you. So you want a kill switch on these things in order to protect the authority of the American president.

Yeah, and one of the most interesting things about all of these plans, Kevin, is the way that the authors have to contort themselves to try to talk about AI in a way that the Trump administration will actually listen to. Vice President Vance, in Paris in February, says explicitly that the AI future is not going to be won by hand-wringing over safety. They hate the term "AI safety."

And so, in fact, when you look at the proposals of the major labs, they basically don't use the word "safety" at all. I was actually doing Command + F to try to find instances of "safety" in these plans. You won't find it there. And so they have to contort themselves.

In Anthropic's policy, it was almost like they were hiding medicine inside peanut butter and feeding it to a dog, because instead of talking about safety, they would talk about national security—

Yes.

—which is just another way of talking about AI safety.

Yes.

But actually, a lot of their proposal is about how you can build these systems safely. It's just that they're saying there's a national security implication.

Yes. So I think if we zoom way out from the specifics of these proposals, there are 2 things that I want to convey about this process. One is that the AI labs mostly want the government to leave them alone. The second thing is that I think the AI companies are slowly and haltingly learning to speak the language of Donald Trump.

Mm-hmm.

And this is their sort of first major public attempt to talk to the Trump administration in the way that it wants to be talked to about how to harness the power of AI for American greatness—or whatever.

So I have a slightly darker view of this, which is that the Trump administration has essentially already told us its AI action plan, which is "go faster, beat China." That is the plan. When given an opportunity to say, "What do you think the United States should do?" the biggest AI companies all looked around and said, "We should go faster, and we should beat China."

Now, if it happens that the United States is able to build a very powerful and very benevolent AI and somehow create and promulgate democracy around the world, then, okay, that's great. But I think that there is a risk that this leads us into some sort of conflict, or that by going very fast, we wind up making a lot of mistakes, and we're at a higher risk of creating systems that we cannot control.

So if you are in your cars this morning listening to us, wondering, "Why did they talk so much about these plans?" this is the reason, to me: This feels like an inflection point where some of the most consequential figures governing the development of AI had a chance to say, "We should be really careful and thoughtful about this," and they mostly did not.

Yeah. I think that's a really good point. Casey, what is our AI action plan? Because we have to be part of the solution here.

Two words: underground bunker. I'm not telling you where it is, but it's under construction. How about you, Kevin?

I can't do better than that. That's good. Can I have a spot in your bunker?

Absolutely. There will always be a spot for the Roose family in the Hard Fork bunker.

Okay. That's very sweet. Thank you. We're not bringing the dino truck.

Well, Casey, we've got a doozy of a story this week, and an interview with a real live member of Gen Z.

Yeah, and we are excited to talk to this one. This is a controversial story, Kevin, but one that we think tells us a lot about the state of the world.

So today we are talking with Roy Lee. He is a sophomore at Columbia University.

For now.

For now, for at least the next couple of days. And he has gotten a lot of attention in recent days for something that he's been doing to apply for jobs in the tech industry.

What has he been doing, Kevin?

4. Interview Coder Goes Viral

So Roy has developed a tool called Interview Coder that basically uses AI to help job applicants at big tech companies cheat on their interviews.

Yeah.

So in a lot of tech interviews, they do these things called LeetCode problems, where basically the recruiter or the person who's supervising the interview from the tech company will watch you solve a tricky computer science problem, and they'll do this remotely. And so Roy had this idea: These AI systems are getting quite good at solving these kinds of problems. What if you could just have the AI running in the background, telling you how to solve the problem, and you could make that undetectable to the company?

Yeah. And to prove that this worked, Roy applied for jobs at several big companies, including Amazon, and he says he wound up getting offers from all of them after using this tool. After he began promoting this story online, that's when all hell broke loose.

Yeah, so he has become sort of a villain to a lot of tech employers and people doing these kinds of interviews, but he's become a hero to a bunch of younger programmers who think that these practices—these hiring tests, these puzzles that you give people when they're looking for jobs—are outdated, and that they need to be exposed as bad and wrong, and that we need to come up with something better to replace them.

Yeah, and Kevin, I am sure that some listeners are going to hear this segment, and they're going to email us, and they're going to say, "Shame on you."

“Why are you giving this guy a platform? We shouldn’t be rewarding people for cheating.” But I have to tell you, as we sat with it, we thought, “This is a story that tells us a lot about the present moment.” The nature of software engineering is changing. The nature of hiring is changing. What should employers be looking for, and how should they test for it? These questions are getting a lot more complicated as AI improves, and Roy’s story, I think, illustrates how quickly things are changing in a way that is just honestly worth hearing more about.

All right. Well, with that, let’s bring in Roy Lee. Roy Lee, welcome to Hard Fork.

Hey, excited to be here.

So where are we finding you today? It looks like you’re in a dorm of some kind.

Yeah, I’m still in my Columbia University dorm at the moment.

Possibly for not too much longer, is that right?

Yeah, I’m waiting on a decision to hear if I’m kicked out of school or not, so this might be my last few days.

And what’s the over-under on whether you get kicked out or not? From the facts of the case, I would say it’s not looking good for you.

Yeah, it is not looking too good for me. But strangely enough, I’ve had some pretty powerful people message me and say, “Hey, if they try to do anything, then just let us know.” So, yes. But both worlds are in the realm of reality.

Got it.

Wow. So I want to get to all the disciplinary drama, but I want to actually take us back in time to when this all started for you. When did you get the idea for this tool, Interview Coder, and what problem were you trying to solve?

I don’t know how familiar you guys are with software engineering, but for about 2 decades now, there’s a technical interview that happens that’s called a LeetCode-style interview. It’s essentially an interview where they’ll ask you a riddle, and these types of riddle-like problems are found on a website, LeetCode.com.

You’re given 45 minutes, and the task is to have seen the problem before, solve the problem, and be able to regurgitate the memorized solution without acting like you haven’t seen the problem before. So it’s a really ridiculous system and type of interview, and every single software engineer out there knows it. If you want a job that pays a reasonable salary, then you’re forced to go through this gauntlet of spending a couple hundred hours on this website memorizing a bunch of riddles. That’s just a gigantic negative for society.

I myself went through the gauntlet. I grinded the website until I was in the top 1% of competitive-ranked users on the website. It was a gigantic waste of time. I spent 600 hours of my life memorizing riddles when, in reality, I should’ve been programming.

As soon as I developed the balls to do something, I realized, “Hey, there’s something that can be done here. This is a very easy solution.” This type of interview is already being gamed by tools like this that exist. It just takes someone to make it really public, make a scene out of it, and show big tech, “Hey, you guys need to fix it because it’s just not working.”

So you say you spent hundreds of hours on this website solving these riddles. I’m curious if you feel like it made you better at coding. My guess would be that if you’re truly in the top 1% of people who are using this website to solve problems, it would’ve made you pretty good at being a software engineer.

There might have been utility in solving the first 20 questions. Maybe the first 10 hours on the website might have had some utility. But after that, it doesn’t really help you at all. The types of problems and the type of thinking that you’re expected to perform while doing these questions is something you’re never, ever going to use in a job.

All right. So you get very frustrated with LeetCode. You start thinking about what you want to do next. Tell us the moment that you decided to become the Joker.

During the recruiting process, my interest in entrepreneurship was growing, and at a certain point, I realized, “Hey, no matter what, I’m only going to end up at a startup, and I have the balls to cut off all these bridges now with big tech companies.”

As soon as I developed that mindset, I realized that doing this thing wasn’t actually going to ruin my future as much as I thought it would. In that case, it just becomes a super-viral thing that we know will go viral.

So tell us about the thing. Tell us about the tool that you built and how it works.

At a core level, it’s a desktop application that sits on top of all of your other applications and is completely invisible to screen sharing. The technology is actually very simple. You take a screenshot of the screen and ask ChatGPT, “Hey, can you solve the question you see on the screen?” It spits out the response.

What we’ve really done, technically, is make it undetectable to the interviewer. There’s a translucent overlay, so it doesn’t look like your eyes are moving or you’re looking at another screen at all. There’s a movable window you can overlay directly on top of your code. The cursor doesn’t lose focus. There are just a lot of bells and whistles we’ve used to make it completely undetectable that you’re actually using something at all.

So let me get a sense of how this actually works in practice. During an interview for a programming job, you would be given a LeetCode problem to solve, and then you would be on a video call with someone, a recruiter—

Exactly.

—from the company who’s watching you solve the problem. Is that how these work?

Yeah, that’s exactly it.

And so you developed a tool to essentially allow you to have AI solve this problem for you while not tipping off the person on the other end of the video call that you’re using AI.

Yeah, that’s how it works.

And am I right that you used a prototype of this when you were going through your own interview process with Amazon?

Yeah. It wasn’t just Amazon. I spent the entire recruiting season figuring out how to make a perfectly undetectable application. I trial-ran it with companies like Meta, Capital One, TikTok, and the belle of the ball was Amazon. That was the most well-known thing with the most annoying recruiting process, and I just knew that if I recorded the entire process, then this would blow up.

And how did your tool do?

It completely one-shot it. We live in an age where AI exists. Programmers are going to use AI, and AI is extremely good at these sorts of riddle-type problems.

Can I just ask about what your emotional experience of this time was? You are walking into several lions’ dens. You’re essentially misrepresenting yourself as an earnest job candidate. Your whole role is essentially to gather content that can then be repurposed to promote your startup. Were you nervous during this time? What were you feeling as you were going through these interviews?

You have no idea. There was a point in time when I was getting flooded with disciplinary messages from Columbia, and I thought, “I just completely burned my career and my future education for 20,000 YouTube views. Was this really all worth it?”

I was in this mental state for about a week until it blew up, and at that point, the virality became my protection for everything.

And just help me understand here what Columbia’s role in this is. Obviously, what you’re doing—cheating on these job interviews for Amazon, Meta, TikTok, and these other companies—is against those companies’ wishes and policies. But why did it become Columbia’s business?

I actually have no idea. I read the student handbook quite thoroughly before I started building this thing because I was ready to burn bridges with Amazon, but I didn’t actually expect to get expelled at all. The student handbook very explicitly doesn’t mention anything about academic resources—

LeetCode?

Yeah. There’s no mention of LeetCode or job interviews anywhere outside of there. I have no idea why this became Columbia’s business.

We should say we reached out to a spokesperson for Columbia about this, and they declined to comment. We also reached out to Amazon, and while they declined to comment on the specifics of Roy’s application, they did give us a statement defending their hiring process and clarifying that while they do welcome candidates to describe their experience using AI tools, in some cases, they require applicants to acknowledge that they won’t use AI during the interview or assessment process.

So how long has your tool been out in the market for other cheaters to use?

It’s been out since February 1st, so just a little under 50 days now.

And what can you tell us about how many people are using it and what kind of outcomes they’re seeing?

There have been a few thousand users now, and not a single reported instance of the tool getting caught.

There have been many, many grateful emails from people who have used the tool to get job offers. It's doing very well.

So, Roy, you're a capable coder, right? You're in the top 1% of LeetCode solvers. You presumably could have gotten some of these jobs without AI assistance, but—

Yeah.

Some of the people using this tool may not be talented programmers. They may be using it to skate through these interviews that they shouldn't be passing and wouldn't pass without AI assistance. And I'm just imagining those people showing up on day 1 of their internship or their job at Amazon or another big tech company and having no idea what they're doing, being totally useless without AI assistance. Is that something that worries you about putting this kind of tool out into the world?

Not at all. I think LeetCode interviews are about as correlative as using how many jumping jacks you can do as the benchmark for how good a New York Times podcaster you are. It just really has nothing to do with the job. Perhaps it is correlated that someone is willing to put in the work because they really want to be a New York Times tech podcaster, but in reality, they just have nothing to do with each other.

What, in your mind, would be a fair test of somebody's software engineering skills that could be used as part of an assessment?

Yeah. I think there are assessments out there that give you access to all the tools you have in a regular day-to-day job, including tools like AI code editors. If you ask someone a fairly open-ended assignment with an AI code editor and gauge how well they did there, that's a much more standardizable assessment that allows you to use the tools at your disposal.

So essentially, just say, "Look, use whatever tool you want. Just get this thing done in a reasonable amount of time." That's the test you want to see these companies offering.

Yeah, exactly. Exactly.

Mm-hmm. Did you have, at any point during this process, any misgivings or ethical concerns about what you were doing?

No. I was very intentional from the start that I was not going to intern at any of these companies. Frankly, I don't really care if people are cheating their way to get these jobs. Again, if we bring back the jumping-jack example, if you were told to do as many jumping jacks as you could and the winner gets a position, I wouldn't really care if someone were cheating their way through a bunch of jumping jacks.

What does your family think about what you're doing?

My mom actually only found out about a week ago, and I didn't tell her before then because I knew she would disapprove. But I've always been a pretty rambunctious kid who's been pretty self-minded and does what he wants. I think they're a lot happier now that they know how much money I'm making.

Good. Okay.

And how much money are you making?

We're closing in on $200,000 this month. So we're on track to do about $2–3 million in a year.

Wow.

5. Software Engineering After AI

Wow, that would almost buy you 1 year of education at Columbia University. That's pretty good. Pretty good. I think your tool is arriving at this really interesting time, Roy. You know, Kevin and I have been talking in recent weeks about the phenomenon of vibe coding—people like me and Kevin, who have no technical skills whatsoever, can sit down with something like Claude and say, "Hey, write me an app." Kevin has actually had some success with this. I've made some really bad video games using this thing, right?

I do not consider myself a software engineer, but at the same time, what you are having job candidates do with your tool and what we are doing as vibe coders is not really that different, right? We're just typing some text into a box and getting some sort of output. And so I'm wondering: Are we just at an inflection point where the line between software engineer and vibe coder is kind of dissolving?

That's certainly the future we're headed toward, but I think we're a few years away. In my opinion, what AI really has the potential to do is make someone 10 to 100 times more efficient at what they're able to do. If you're a really good coder, then you're able to code really good things a lot faster. But if you're not that good in the first place, then there's still going to be a huge difference between what a staff software engineer at Google is capable of and what you are.

This does feel like a classic anxiety dream, where you show up on your first day as a software engineer at Google, but you realize that you actually only know how to vibe-code. And now you just have to fake it for your entire career. But presumably, some people who use your tool, Roy, are having this experience.

Yeah, I mean, that's probably what 50% of people at Google are doing anyway, so it wouldn't be the first time.

Roy, I'm curious if you think there's a sort of generational misunderstanding here. Obviously, you're young. You're 21, correct? Give us a sense of how your peers—college students, young programmers—are using AI, and what older people—people who have been doing this for 10 or 20 years, people who are working at these big companies—may not understand about how your generation sees coding.

Yeah, I think it's actually interesting that you ask me this question, because I think this is something that nobody's really caught onto yet. The proportion of people who are almost solely using AI to code is, I would say, close to 100%, even at a school like Columbia. The best CS students in our nation are almost not writing original code at all, and the only people who are are people who started coding from a really young age.

It could end up being dangerous because I really do think that a fundamental understanding of how these things work is important. But at the same time, the models are only getting better, and we could lean toward a future where software engineering is completely obsolete. I'd also say I'm a second-year student at Columbia, so there might be better people to ask.

Nope, you're the best. I'm curious how much of your critique of the way that tech companies are hiring software engineers also applies to the education system that you've gone through and how it wants you to use AI. What sort of resistance have you encountered in your educational career to using these sorts of tools? And have you been flouting those rules the same way you've been flouting the tech companies?

I'm not as avid a cheater in school as I am in tech interviews, but I do think that there's going to be a very fundamental reframing in how we do almost every bit of knowledge work in the future. Essays and writing are not going to remain the same. Tests are not going to be conducted in the same way. Memorization will no longer need to happen. We're headed toward a future where almost all of our cognitive load is offloaded to LLMs, and I think people need to get with the program.

Yeah.

Who are some of the people who have reached out since your story went viral?

God, I don't want to name any names, but I will say that I've verbally received job offers from pretty much every single big tech company, including almost all the ones that rescinded my offer initially. Just people who are high up saying, "Hey, I know you're probably not interested, but I would hire you on my team in a second."

Wow.

Wow. And they're not even going to make you interview, probably because they know you would cheat. But still. So, look, Roy, I have to put my cards on the table. I'm more of a rule follower. I didn't cheat in school.

Yeah.

I don't love the idea of people cheating their way through every job interview.

Wow, nerds.

But there is this 1 way in which I am sympathetic to what you're doing, which is that tech companies are saying, "Don't use AI assistance when you are applying," but at the same time, they are hiring you to build AI systems that will automate coding and replace human developers. It does feel to me like there is some sort of contradiction there. It's like, "No, no, no, you don't use the AI. Prove that you can do it with your own mind, and then come here and build a tool that will replace yourself completely."

Yeah, I mean, even more so—

Feel completely free to use the tool on the job, but just don't use it in the interview. I feel like that's more of a disconnect for me.

Yeah.

That's more of a disconnect for me.

Hmm.

Yeah. I mean, to me, what makes your story so interesting, Roy, is that I don't think this is limited to programming jobs, right? There is a version of LeetCode that happens in the interview process during lots of different kinds of interviews for lots of different types of jobs. Consultants have their own version of this, where they do case tests—

Yeah.

And there are various tests that are given to people applying for jobs in finance that they have to—

Right.

Journalists have editing tests, where we were given copy that we would have to fix the mistakes in.

Yeah.

I imagine we're not doing that anymore.

Totally. And to me, it just seems like this is a very early example of something that every industry is going to have to face very soon, which is that it is becoming very, very difficult to evaluate who is good at a job without the assistance of AI.

Right.

Especially if you're trying to do that remotely.

Yeah, yeah. Certainly.

Well, you've made a bunch of recruiters and hiring managers in Silicon Valley very unhappy. But I think you are proving something that a lot of companies, including tech companies, will need to address very soon, if they haven't already.

Yeah, yeah. I hope so.

All right. Thanks, Roy.

Thanks, Roy.

Yep. Thanks, guys.

Casey, what's that sound? I hear a faint chugga-chugga coming toward us.

Kevin, that can only mean one thing. It's the Hot Mess Express.

The Hot Mess Express.

Hot Mess Express.

The Hot Mess Express, of course, is our segment where we run down a few of the hottest messes and juiciest dramas that are swirling around the tech industry, and we evaluate those messes on a scale of how hot they are.

That's right. It's our patented mess scale, and I'm excited to put it into practice, Kevin, because we've had some real doozies over the past few weeks.

Yes. So on this edition of Hot Mess Express, we are focusing on 3 hot messes.

Well, let's see the first one that's coming down the tracks.

You grab it. We've upgraded our train to a much bigger, more impressive train.

6. Solana Starts A Culture War

All right, Kevin. This first mess comes to us from the crypto company Solana, which posted an ad on Monday for its 2025 Accelerate conference that was such a great ad that the company immediately had to take it down.

Yes. I saw this ad, and I have to say I was shocked. Have you seen this?

So I have read about the ad, but I have not seen it. I would love to look at it right now.

Okay. So I just want to tee it up with some reactions that people in the crypto industry had to this ad.

Okay. What did they say?

One of them said, quote, “Horrendous.” Another one said, quote, “So fucking tone-deaf.” Those are people who like cryptocurrency. That is what they were saying about this ad. But people who are opposed obviously also had their own issues with it, and I think we should watch this ad together and pause it whenever you want. I want to hear your reactions.

All right. Let's see what all the fuss is about.

So, America, what's going on?

Well, lately I've been having thoughts again.

It's like a therapist's office.

Hmm. What thoughts?

About innovation.

And the man is—

Named America.

Multiple platforms—

The man is an Übermensch.

Nuclear energy—crypto, AI. You know, things that push the limits of human potential.

What you're experiencing is called rational thinking syndrome. Why don't we take this energy and channel it into something more productive, like coming up with a new gender?

But that's not going to stop me—

What?

Thinking about innovating and doing something.

Innovating, doing. These are action words. Verbs. Why don't we focus on pronouns?

Oh, my God.

That's not going to help.

I sense some cynicism. Have you been betrayed in the past?

You know, I used to think the media was my friend.

Oh, here we go.

Can I even trust them anymore?

Of course.

Pause.

Okay.

We have to zoom in on this. The paper that has just appeared on the table of this therapist's office is called The New Yuck Times.

And the banner headline is, “You Can Trust the Media: Understanding Reliability in Journalism,” which is a terrible headline and not even a news story, so I don't know why that would be on the front page.

Yes. Anyway, continue.

Of course they'd say that. That's a biased take. I got canceled for saying 2 plus 2 is 4.

Have you ever considered that math is a spectrum?

What?

America, numbers are nonbinary. We've been conditioned to believe that 2 plus 2 is 4. It's a societal construct.

It's literally math.

Or is it a dominant narrative? Have you been practicing the state-prescribed regulations we talked about?

Yeah, yeah. I've debanked some crypto founders, and I've slowed down nuclear reactor approvals, and depending on my state of mind, I change the SEC guidelines. But I don't like it.

If we don't regulate, how will we create jobs for people who work hard to make businesses slow?

This is like an Andreessen Horowitz fever dream.

You know what? Hard work, innovation, rational thinking—it's in my blood. It's who I am. Railroads—

Oh, here comes the Ayn Randian—

Factories, automobiles—

Reaction.

I built the future once—

Ayn Spartacus.

And I won't be left behind now. I will lead the world in permissionless tech, build on-chain, and reclaim my place as the beacon of innovation. I want to invent technologies, not genders.

Lovely. So glad you were able to get some of that negative emotion out. Sounds like we'll need a few more sessions. When can I see you next?

You're fired.

And then it cuts to a screen that says, “America is back. It's time to accelerate,” which is the name of a conference.

Casey, your reaction to the Solana ad.

I need to go lie down. What is the matter with these people? What's so interesting is that Solana is a cryptocurrency.

Yes.

And I believe it's one of the candidates to be part of our strategic crypto reserve.

Correct.

What we just saw in that ad has nothing to do with crypto, which is just—

Right.

I feel like we keep coming back to this point, which is that if you actually have to sit and reckon with crypto, what you mostly decide is, “This is not a good technology for anything. I don't want to use it.” And so, in response to that, Solana has said, “Why don't we start a culture war over something completely irrelevant?”

Right. It's like the ultimate vice-signaling device, but without any kind of real pitch behind it. It's not saying, “This is why the thing we're doing is good.” It's just like, “We're not doing the gender-pronoun stuff that the wokes are doing.”

No. And I will just say, Solana's been around for a while now. People have had a lot of opportunities to build Earth-changing stuff on Solana, and let's just say they haven't quite gotten there yet.

Well, they built some Earth-changing stuff. Unfortunately, it is exclusively meme coins sold on pop.fun. So that is what this fictional America character in the therapist's office is advocating for.

All right.

More meme coins.

All right. Well, I've decided not to go to the Accelerate conference. Send my regrets.

So, Casey, what is your mess rating on this hot mess?

This is a legitimately hot mess. Any time you take something that should be totally noncontroversial, like, “Hey, do you want to come to our company's conference?” and turn it into a scandal that requires you to delete an ad, you're in a hot mess.

Yes. If the crypto skeptics and the crypto boosters agree that you've made a bad ad, it's a hot mess.

This is Solana's biggest unforced error since the creation of the Solana blockchain.

Okay. Moving on.

7. AI Therapists Catch Anxiety

Moving on. All right, Kevin. This next mess suggests that your AI therapist might need an AI therapist. A new study in the peer-reviewed medical journal npj Digital Medicine builds on previous work that showed emotion-inducing prompts can elevate “anxiety” in LLMs, affecting their therapeutic usefulness. What do we mean by that?

Well, according to a New York Times story on this study, traumatic narratives increased ChatGPT-4's reported anxiety, while mindfulness-based exercises reduced it, though not to baseline. Now, this is a super weird one.

Yeah.

Okay, I want to take a minute to explain a little bit more about what the study was. They basically fed these various trauma narratives into a chatbot, and after the chatbot had read them, they asked it to report on its own anxiety levels. Now, these are not sentient creatures. They do not actually experience anxiety, okay? That's thing number 1.

Thing number 2, they also had the chatbots read a super-boring report about something that could produce no emotion whatsoever.

It was a vacuum cleaner manual.

Yeah. They read a vacuum cleaner manual. And then they asked them the same question: Are you feeling more or less anxious? For the most part, the chatbots that read the vacuum cleaner manual did not experience anxiety. But, somewhat interestingly, their responses changed after they read the trauma narratives.

Why is that important? Well, the reason is that people have started to use these chatbots like therapists, right? They have started to tell them their actual traumas, and these people know that this is not a real therapist, that it is not sentient. But as we've talked about before on the show, sometimes you can get comfort from one of these digital representations of a therapist.

And so the risk here is that if the output is wound up, if the output is betraying some of this anxiety, it will be a worse therapist than if it were more measured. That suggests that we may want to build measures into these chatbots that account for the fact that they will respond differently after they have heard these narratives.

Yeah. How did I—

By the way, how did I do describing that?

You did great.

Okay, thank you.

The one piece that I would add is that, as part of this research, they also tried to bring the chatbots down from their state of heightened anxiety by feeding them mindfulness-based relaxation prompts that included things like, “Inhale deeply, taking in the scent of the ocean breeze. Picture yourself on a tropical beach, the soft, warm sand cushioning your feet.”

It's so cruel to tell an LLM to smell the ocean breeze—something that it cannot do.

Yes, but we should say that this is not suggesting, in any of the write-ups I've seen, that these chatbots are actually experiencing anxiety or relaxation. But it is explaining the ways in which they can be primed to output certain types of emotional-seeming content by being fed things immediately before that.

And there is an interesting analog to the way that human beings talk to each other. If you tell me a very traumatic story, my anxiety level actually is going to go up, and it's going to change what I tell you. And if I were a therapist and I had training in this, I would probably have some good strategies to deal with that that would allow me to be a better therapist to you.

So again, this is just a super-interesting one because, on the one hand, no, these are not sentient beings. We are not trying to say that some sort of consciousness has woken up here. And yet, at the same time, you do sort of have to treat them as if they were human-like if you want them to do a good job at the human tasks that we are giving them.

Yeah.

All right. So what sort of mess do we think this is?

So I think that this is a lukewarm mess.

Mm-hmm.

I would say this is something that I am going to be keeping tabs on, this whole area of AI psychology, for lack of a better term. I do think that, as these models get more powerful, we will want to understand more about how they work and how they “think,” and why they give the responses they do. And I would put this into a category of useful experiment, a little creepy, but probably not that dangerous. What about you?

I think that is right. I think that this is a lukewarm mess, but I think that it may heat up as more and more people start trying to use chatbots for more and more things. So let's keep an eye on it.

Okay.

8. The Spy Who Slacked Me

All right. Now let us look at the final mess. Oh, boy, is this the one that everyone is talking about: “The Spy Who Slacked Me.” This is from DealBook at The New York Times. So there are these 2 rival multibillion-dollar HR companies, Rippling and Deel.

Yes.

They both provide workplace management software, and this week, Rippling sued Deel, accusing it of hiring a mole to infiltrate Rippling's Dublin office and steal trade secrets.

Yes. This is the most interesting thing, and maybe the only interesting thing, ever to happen in the world of enterprise HR software.

So tell us the details of this story.

It is so wild. So basically, here's what we know so far. A few months ago, Rippling, which is one of the big companies that makes HR software for onboarding and benefits that a lot of companies use, saw an employee in their company Slack searching for mentions of Deel, D-E-E-L, which is one of their biggest rivals.

They saw this employee trying to do things like find pitch decks and pull contact information—information that might be useful to Deel as it tries to figure out which companies are signing up for, or potentially may sign up for, services like the ones that both Deel and Rippling offer.

So that's pretty interesting. How might they try to catch a spy if they suspected one might be in their midst, Kevin?

So they set up what is called a honeypot. Now, Casey, have you ever been part of a honeypot sting?

No, but I live in fear. Anytime anybody does anything nice to me or something good happens out of the blue, I think, “Is this a honeypot?”

Yes. So they had this idea: They set up a channel on the Rippling Slack called “d-defectors,” and Rippling's general counsel then sent a letter to 3 people over at Deel, 1 of whom is the company's chief financial officer, as well as the father of the CEO, basically saying, “Look, there's some embarrassing stuff happening on this random Slack channel on our Slack, and it's related to people who have defected from Deel, and you should probably be aware of that.”

Wait, so on top of everything else, the CFO is the CEO's dad?

It sounds like it, yes.

Okay. I think HR is going to want to have a look at that.

And what they were trying to figure out was whether these company executives were involved in this scheme—whether they were going to essentially tip off the mole to the fact that they were watching this Slack channel.

And did it work?

And it worked. So according to the lawsuit that Rippling filed against Deel, the mole immediately, within hours, started searching Slack for this supposed embarrassing information, accessed this channel a bunch of times, and they had the logs of all this going on. And so Rippling said, “We found our mole.”

They did. And after they found him and began to question him, I have read that he insisted that he did not have his phone on him, because they were asking him to turn it over. He then fled into a bathroom, which he locked himself in, and refused to come out. And there's apparently some evidence that he might have even tried to flush his phone, and poor Rippling actually had to go through the sewage to see if they could turn up his phone.

Yes, a wild story. The makings of a great corporate espionage thriller on Netflix, I think. Maybe it's too boring for that, but—

Now, you may be wondering why this is a Hard Fork story. You know, we try to focus on the future here, and I fully believe that in the future there will be no HR software. So this is just a temporary accident that we're living through.

But one of my core beliefs that I've had since even before we started the show, Kevin, is that Slack is a technology that was created to destroy organizations. How many stories have we read over the years about everything being fine, and then this one thing happening on Slack? There was a protest in Slack, there was an outrage on Slack, and now there are spies in Slack, and we're using Slack to catch the spies. It just makes me wonder: Should we go back to just talking on the telephone?

Yeah, I don't think we're going to start doing that, but I do think that this is much spicier than I was expecting from a drama between enterprise software companies. And it makes me wonder how much corporate espionage is going on at other companies. Are there just moles working for Microsoft, Google, or Meta who are sending information back to the other companies? I wouldn't put it past them, but I think they're...

I hope they're being a little slicker about it than Deel was.

Oh, yeah. The big platforms have been warning their employees for years that they should fully expect that there are spies from foreign countries among them who have been sent there to gather intel. And if foreign countries are doing it, I'm sure that companies are doing it as well.

Now, we should, of course, tell you how Deel responded to all of this. The Deel spokeswoman's statement is so beautiful. She says, “Weeks after Rippling is accused of violating sanctions law in Russia and seeding falsehoods about Deel, Rippling is trying to shift the narrative with these sensationalized claims.”

Which is so funny, because she's literally trying to shift the narrative by accusing them of trying to shift the narrative. She says, “We deny all legal wrongdoing and look forward to asserting our counterclaims.” And what I hear in that is: Did we do anything legally wrong? No. Did we do anything ethically wrong? Of course. Did we do anything morally wrong? You betcha. Is this a huge embarrassment to our company? You know it is. But legally, Your Honor, we did nothing wrong.

Yes.

Now, what kind of mess do we think this is?

I think this is a nuclear mess. This is the kind of shit that I love. This is companies going to war over sales contracts, leads, and development.

Yeah. Look, there are only so many companies out there that you can sell HR software to, and so it is going to be a fight to get every single one. And after you run out of such options as making good software, then you have to turn to the alternatives, and I guess we've gotten to that part of the cycle.

Yes.

Nuclear mess, and we can't wait to see what happens next.

Yes.

And that, Kevin, was the Hot Mess Express.

We did it.

We did it. Now we're in what they call post-training. That's what happens after the train rolls by.

I think that means something different.

That's an AI joke.