AI就业末日 + Mike Krieger 谈 Anthropic 的构建 + Hard Fork 犯罪科

Casey, how was your Memorial Day weekend?

Memorial Day weekend was good. I was like, “You know, I need to unplug.” As you know, I needed to unplug a bit. I’m not a big unplugger. I’m normally very comfortable feeling plugged.

Yeah, you’re a screen maxer.

I’m a screen maxer, but this was a weekend where I was like, “Okay, I’ve got to get out of this danged house. I’ve got to see some nature.” So I went with my boyfriend up to Fort Funston, this beautiful part of San Francisco.

Great beach.

These giant dunes sit atop a battery of guns that could shoot rounds 13 miles into the ocean. I was so excited to just stare at the ocean. We climb up into the dunes and sit down, and the big waves are rolling in. Then the wind picks up, and I’m being sandblasted in the face at 40 miles an hour.

Within 30 seconds, I have grit in my teeth, and I’m thinking, “This was not the nature I was promised. Why do I feel like I’m dying?”

But it did do a great job of exfoliating your skin.

Yeah, my skin has never really looked smoother.

They call that dermabrasion, and some people pay lots of money for it.

Yes, I have been abrased. I’ve been majorly abrased.

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, is AI already taking away jobs? Kevin makes the case. Then Anthropic Chief Product Officer Mike Krieger joins us to discuss Claude 4, the future of work, and the viral saga over whether an AI could blackmail you. And finally, it's time for Hard Fork Crimes Division.

Dun-dun.

Is blackmail still a crime?

Hope so.

1. The Entry Level Crisis

Well, Kevin, you have delivered some interesting news to us via The New York Times this week, and that is that the job market is not looking great for young graduates.

Yes, graduation season is upon us. Millions of young Americans are getting their diplomas and heading out into the workforce. I thought it was high time to investigate what is going on with jobs and AI, specifically with entry-level white-collar jobs—the kind that a lot of recent college graduates are applying for—because there are a couple of things that have made me think we are starting to see signs of a looming crisis for entry-level white-collar jobs.

So I thought I should investigate that.

Yeah. I’m excited to talk about this because I got an email today from a recent college grad, and she wanted to know if I could help her get a job in marketing and tech. I thought, “If you’re just emailing me asking for a job, there must be a crisis going on in the job market.”

Yes, that would not be my—

Yeah.

—my step 1 in looking for a job, or even maybe my step 500.

But you’ve actually spent a lot of time looking into this question. Tell us a little bit about what you did and what you were trying to figure out exactly.

I’ve been interested in this question of AI and automation for years. When are we going to start to see large-scale changes to employment from the use of AI? There are a couple of things that make me worried about this moment specifically and whether we are starting to see signs of an emerging jobs crisis for entry-level white-collar workers. The first one is economic data.

Okay.

If you look at the unemployment rate for college graduates right now, it is unusually high. It’s about 5.8% in the U.S. That has risen significantly—about 30% since 2022. Recently, the New York Federal Reserve put out a bulletin on this and said the employment situation for recent college graduates had, quote, “deteriorated noticeably.”

Hmm.

This tracks with some data that we’ve been getting from job websites and recruiting firms showing that, especially for young college graduates in fields like tech, finance and consulting, the job picture is much worse than it was even a few years ago.

And that rate that you mentioned, Kevin, is higher for young people in entry-level jobs than it is for unemployment in the United States overall. Is that right?

Yes. Unemployment in the United States is actually doing quite well. We’re in a very tight labor market, which is good. We have pretty close to full employment. But if you look specifically at the jobs done by recent college graduates, it is not looking so good.

Actually, the job placement rates at a bunch of colleges and even top business schools like Harvard, Wharton and Stanford are worse this year than they have been in recent memory.

I was having dinner with a Wharton student last week, and she was telling me that a lot of her classmates had yet to be placed. It was a real concern, so anecdotally, that sounds right to me.

Okay, so that’s the economic data that you’re seeing. What else is making you worried?

2. The Agentic AI Threat

One of the other things that’s making me worried is the rise of so-called agentic AI systems—these AI tools that can not just have a question-and-answer session or respond to a prompt, but can actually be given a task or a set of tasks. They can go out and do it, check their own work, and use various tools to complete those assignments.

One of the things that has updated me the most on this front is these Pokémon demos. Casey, do you know what I’m talking about here?

You’re talking about Claude playing Pokémon?

Yes. Within the last few months, it’s become very trendy for AI companies to test their agentic AI systems by having them play Pokémon, essentially from scratch with no advanced training, and some of them do quite well.

Google said onstage at I/O last week that Gemini 2.5 had actually been able to finish the entire game of Pokémon, and—

One of the games. There are, I think, probably at least 36 different Pokémon games on the market.

Okay.

And I actually know for a fact that Google was playing a different Pokémon game than Anthropic was.

Oh, interesting.

Yeah.

I’m not a Pokémon expert, but I think people see these Pokémon demos and think, “Well, that’s cute, but how many people play Pokémon for a living?” It seems like more of a stunt than a real improvement in capabilities.

But the thing I am hearing from researchers in the AI industry and people who work on these systems is that this is not actually about Pokémon at all. This is about automating white-collar work.

If you can give an AI system a game of Pokémon and it can figure out how to play the game, I don’t know Pokémon very well. I’m more of a Magic: The Gathering guy. But my sense is that you have to go to various places, complete various tasks and collect various Pokémon.

You have to go into various gyms. You take your Pokémon, they compete against rival Pokémon, and your Pokémon have to vanquish the others in order for you to progress through the game, Kevin. I hope that was helpful.

Exactly. So, as I was saying, that is how you play Pokémon. What they are telling me is that this is actually some of the same techniques that you would use to train an AI to, for example, do the work of an entry-level software engineer, a paralegal or a junior consultant.

Yeah. If your job is mostly writing emails and updating spreadsheets, that is a kind of video game. If an AI system can just look at Pokémon and, through trial and error, figure out how to play it and win, it can probably figure out how to play the email-and-spreadsheet game, too.

Exactly. One of the signs that is worrying me is that these AI agents do seem to be becoming capable of carrying out longer and longer sequences of tasks.

Yeah, so tell us about that.

Recently, Anthropic held an event to show off its newest model, Claude Opus 4, I believe it’s called.

I believe it’s Claude 4 Opus, actually.

Claude 4 Opus?

Got your ass. It’s Claude 4 Sonnet and Claude 4 Opus. Sometimes I feel like you don’t respect the names of these products. Do you know how much work went into the naming of these products? At least 5 minutes. They spent at least 5 minutes coming up with that, and then you’re just going to shit all over it.

I’m so sorry.

Sorry.

So anyway, Claude Opus 4.

Claude 4 Opus.

Well, no, I swear it’s Claude—

It’s Claude 4 Opus.

No, it’s Claude Opus 4.

What?

I’m looking at the Anthropic blog post.

Oh my God.

Claude Opus 4 and Claude Sonnet 4.

This is so confusing for me.

It’s like your boyfriend doesn’t even work there.

I’m going to be in big trouble when I get home.

So, okay.

Ugh.

Back to my point. Anthropic held this event last week—

Yeah.

where they're showing off their latest and greatest versions of Claude, and one of the things they say about Claude Opus 4, their newest, most powerful model, is that it can code for hours at a time without stopping. In one demo with a client on a real coding task, Claude was able to code for as much as 7 hours uninterrupted. Now, you might think, well, that's just coding. Maybe that's a very special field, and there are some things about coding that make it low-hanging fruit for this sort of reinforcement learning model that can learn how to do tasks over time.

The problem for workers is that a lot of jobs, especially at entry levels in white-collar occupations, are a lot like that, where you can build these reinforcement learning environments, collect a bunch of data, and essentially have it play itself, like it would play Pokémon, and eventually get very good at those kinds of tasks.

Yeah, at Google I/O last week, Kevin, they showed off a demo of a feature where you can teach the AI how to do something. You effectively show the AI—you say to the AI, “Hey, watch me do this thing,” and then it watches you do the thing, and then it can replicate it. Can you imagine how many managers all around the world took a look at that and said, “Once I can teach the computer how to do things, a bunch of people are about to lose their damn jobs”?

Totally.

Yeah.

And this is why some of the people building this stuff are starting to say that it's not just going to be software engineering that becomes displaced by these AI agents; it's going to be all kinds of different work. Daria Amodei, the CEO of Anthropic, gave an interview to Axios this week in which he said that within 1 to 5 years, 50 percent of entry-level white-collar jobs could be replaced.

Now, that could be wildly off. Maybe it is much harder to train these AI systems in domains outside of coding. But given what is happening just in the tech industry and just in software engineering, I think we have to take seriously the possibility that we are about to see a real bloodbath for entry-level white-collar workers.

Yeah, absolutely, and we wonder why people don't like AI. All right, so first, we've got the economic data showing that there is some sort of softness around hiring for young people. We also just have the rise of these agentic systems. But is there evidence out there, Kevin, that says that AI is actually already replacing these jobs?

3. Companies Go AI First

So I talked to a bunch of economists and people who study the effects of AI on labor markets, and what they said is that we can't conclusively see yet in the large economic samples that AI is displacing jobs. But what we can see are companies that are starting to change their policies and procedures around AI to prioritize the use of AI over the use of human labor. So I'm sure you've been following these stories about these so-called AI-first companies.

Mm-hmm.

Shopify was an early example of this. Duolingo also did something related to this. Basically, they are telling their employees, “Before you go out and hire a human for a given job or a given task, see if you can use AI to do that task first. And only if the AI can't do it are you allowed to go out and hire someone.”

Yeah. And by the way, if you're wondering, Hard Fork is an AI-second organization—because at Hard Fork, the listener always comes first.

That's true. I think that what worries me, in addition to the hints of this that we see in the economic data and the evidence that these AI agents are getting much better, much more quickly than people anticipated, is just that the culture of automation and employment is changing very rapidly at some of the big tech companies.

Yeah, this feels like a classic case where the data is taking a while to catch up to the truth on the ground. I also collect stories about this and would share maybe just a few things that I've noticed over the past couple of weeks here, Kevin. The Times had a great story about how some Amazon engineers say that their managers are increasingly pushing them to use AI, raising their output goals and becoming less forgiving about them missing their deadlines.

The CEO of Klarna, which is a buy now, pay later company, says its AI agent is now handling two-thirds of customer service chats. The CEO of IBM said the company used AI agents to replace the work of 200 HR employees. Now, he says that they took the savings and plowed that into hiring more programmers and salespeople. And then, finally, the CEO of Duolingo says that the company is going to gradually stop using contractors to do work that AI can handle.

So that's just a collection of anecdotes, but if you're looking for spots on the horizon where it seems like there is truth to what Kevin is saying, I do think we're seeing that.

Yeah, and I think the thing that makes me confident in saying that this is not just a blip, that there's something very strange going on in the job market now, is talking with young people—

Mm-hmm.

—who are out there looking for jobs, trying to plan their careers. Things do not feel normal to them. So recently, I had a conversation with a guy named Trevor Chao. He's a 23-year-old recent Stanford graduate. Really smart guy, really skilled—the kind of person who could go work anywhere he wanted, basically, after graduation.

And he actually turned down an offer from a high-frequency trading firm and decided to start a startup instead. His logic was that we might only have a few years left where humans have any kind of advantage in labor markets, where we have leverage, where our ability to do complex and hard things is greater than that of AI systems. And so you want to do something risky now and not wait for a career that might take a few years or decades to pay off.

The way he explained it to me is that all of his friends are making similar calculations about their own career planning now. They're looking out at the job market as it exists today and saying, “That doesn't look great for me, but maybe I can find a way around some of these limitations.”

4. AI Hype Meets Reality

Hmm, that's interesting. Well, let me try to bring some skepticism to this conversation, Kevin, because I know in your piece you identified several other factors that help to explain why young people might be having trouble finding jobs.

You have tariffs. You have the overall economic uncertainty that the Trump administration has created. You have the long tail of disruption from the pandemic or even the Great Recession, right? I think some economists believe that we might not totally have recovered from that. So it seems like there are a lot of explanations out there for why young folks are having trouble finding jobs that don't involve AI at all, maybe.

Yeah, I think that's a fair point, and I want to be really careful here about claiming that all of the data we're seeing about unemployment being high for recent college graduates is due to AI. We don't know that. I think we will have to wait and see if there is more evidence that AI is starting to displace massive numbers of jobs.

But I think what the data is failing to capture, or at least is not capturing yet, is how eager and motivated the AI companies that build this stuff are to replace workers. Every major AI lab right now is racing to build these highly capable autonomous AI agents that could essentially become a drop-in remote worker that you would use in place of a human remote worker.

They see potentially trillions of dollars to be made doing this kind of thing, and when they are talking openly and honestly about it, they will say, “The barrier here is not some new algorithm that we have to develop or some new research breakthrough. It's literally just that we have to start paying attention to a field and caring about it enough to collect all the data and build the reinforcement learning training environments to automate work in that field.”

And so they are just planning to go industry by industry and collect a bunch of data and use that to train the models to do the equivalent of whatever the entry-level worker does. And that could happen pretty quickly.

Yeah. Well, so that feels like a threat.

Yeah. It's not great. I think the argument that they would make is that some of these entry-level jobs were pretty rote anyway, and maybe that's not the best use of young people's skills. I think the counterargument there is that those skills are actually quite important for building the knowledge that you need to become a contributor to a field later on.

I don't know about you, but my first job in journalism involved a bunch of rote and routine work.

One of the things that I had to do was write corporate earnings stories, where I would take an earnings report from a company, pull out all the important pieces of data, put it into a story, and get it up on the website very quickly. Was that the most thrilling work I could imagine doing or the highest and best use of my skills? No, but it did help me develop some of the skills, like reading an earnings statement, that became pretty critical for me later on.

Interesting. For what it's worth, my first job, it was actually the most physical job in journalism I ever had. I covered a small town, and so I spent all of my days just driving down to City Hall, going down to the police station, sitting at the city council meeting, and making phone calls.

A lot of drudgery came in later. But let me raise maybe an obvious objection to the idea that, “Oh, young people, don’t worry. These jobs that we’re eliminating, it was just a bunch of drudgery anyway.” Young people need to pay their rent.

Yes.

You know? Young people need to buy health insurance.

Yes.

And so I think they’re not going to take a lot of comfort from the idea that the jobs that they don’t have weren’t particularly exciting.

Yes, and the optimistic view is that if you just shift workers off of these entry-level rote tasks into more productive or more creative or more collaborative roles, you free them up to do higher-value work. But I just don’t know that that’s going to happen. I’m talking to people at companies who are saying things like, “We don’t really see a need for junior-level software engineers,” because now we can hire a mid-level software engineer and give them a bunch of AI tools, and they can do all of the debugging and the code review and the stuff that the 22-year-olds used to do.

Yeah. Let me ask about this in another way. I think a lot of times we have seen CEOs use AI as the scapegoat for a bunch of layoffs that they already wanted to do anyway, or a bunch of management decisions that they wanted to make anyway. Earlier this year, there was a story in The San Francisco Standard that Marc Benioff, the CEO of Salesforce, said the company would not hire engineers this year due to AI. I went to Salesforce’s career page this morning, Kevin. There were hundreds of engineering jobs there. I don’t know what wires got crossed. The story I read was in February. Maybe something has changed since then. But talk to me a little bit about the hype element in here, because I do feel like it’s real.

Yes, there’s definitely a hype element in here. I worry that companies are getting ahead of what the tools can actually deliver. You mentioned Klarna, the buy now, pay later company. A couple of years ago, they made this big declaration that they were going to pivot to using AI for customer service, and they announced this partnership with OpenAI, and they were going to try to drive down the number of human customer support agents to zero. Recently, they’ve been backtracking on that. They’ve been saying, “Well, actually, customers didn’t like the AI customer service that they were getting, and so we’re going to have to start hiring humans again.”

Hmm.

So I do think that this is a risk of some of this hype: It tempts executives at these companies to move faster than the technology is ready for.

Well, and speaking of that, one of my favorite stories from this week was about a guy who has set up a blog, Kevin, where he keeps a database of every time that a lawyer has been caught using citations that were hallucinated by AI. Did you see this?

No.

There are more than 100. We’ve talked about this issue on the show a couple of times, and I’ve thought this must just be a small handful of cases, because who would be crazy enough to bet their entire career on a hallucinated legal citation? It turns out, more than 100 people. And so a lot of people might be listening to this conversation saying, “Kevin, you’re telling me that we’re standing on the brink of AI taking over everything. These things still suck in super-important ways.” So help us square that issue. We know these systems are not reliable for many, many jobs, so how can it be that so many CEOs are apparently ready to just junk their human workforces?

I think part of the misunderstanding here is that there are 2 different kinds of work. There’s work that can be easily judged and verified to be correct or incorrect, like software engineering. In software engineering, either your code runs or it doesn’t.

Mm-hmm.

And that’s a very clear signal that can then be sent back to the model in these reinforcement learning systems to make it better over time. Most jobs are not like that, right? Most jobs, including law, including journalism, including lots of other white-collar jobs, do not have this very clearly defined indicator of success or failure. And so that’s actually what is stopping some of these systems from improving in those areas. It’s not as easy to train the model and say, “Give it a million examples of what a correct answer looks like and a million examples of what an incorrect answer looks like,” and have it, over time, learn to do more of the correct thing.

Mm-hmm.

So I think in law, this is a case where you do actually have more subjective outputs, and so it’s going to be a little harder to automate that work. But I would say we also have to compare the rates of error against the human baseline, right? You mentioned this database of cases in which human lawyers had used hallucinated citations in their briefs. I imagine there are also human paralegals or lawyers who would make mistakes in their briefs as well. And so I think for law firms or any company trying to figure out, “Do we bring in AI to do a job?” the question they’re asking is not, “Is this AI system completely error-free?” It’s, “Is this less likely to make errors than the humans I currently have doing this work?”

Right. And in so many things, if the system is 20 percent worse than a human but 80 percent less expensive, a lot of CEOs are going to be happy to make that trade.

Totally.

All right. Well, let’s bring it home here. I imagine we might have some college students listening or some recent college grads. They’re now thoroughly depressed. They’re drinking. It’s Friday morning. They’re wasted. As they sober up, Kevin, what would you tell them about what to do with any of this information? Is there anything constructive that they can do, assuming that some of these changes do come to pass?

5. Advice For New Graduates

I really haven’t heard a lot of good and constructive ideas for young people who are just starting out in their careers. People will say stuff like, “Oh, you should just be adaptable and resilient,” and that’s what Demis Hassabis told us last week on this show when we asked him what young people should do. I don’t find that very satisfying, in part because it’s just so hard to predict which industries are going to be disrupted by this technology. But I don’t know. Have you heard any good advice for young people?

Well, I think what you’re running into, Kevin, is the fact that our entire system for young grads is set up for them to take entry-level jobs and gradually acquire more skills. And what you’re saying is that that part of the ladder is just going to be hacked off with a chainsaw, so what do you do next? Of course there’s no good answer, right? The system hasn’t been built that way.

I think that in general, the internet has been a pressure mechanism forcing people to specialize, to get niche-y. The most money and the most opportunity is around developing some sort of scarce expertise. I have tried to build my career as a journalist by trying to identify a couple of ways where I could do that. It’s worked out all right for me, but I also had the benefit of entry-level jobs. So if somebody had come to me at the age of 21 and said, “If you want to succeed in journalism, get really niche-y and specialized,” I would say, “Okay, but I need to go have a job first. Is there one of those?” To me, that’s the tension.

I will also say there’s never been a better time to be a nepo baby. I don’t know if you’ve been following the Gracie Abrams story. She’s a very talented songwriter, the daughter of J.J. Abrams, the filmmaker. She was born into wealth, and now she’s best friends with Taylor Swift. If you can manage something like that, I think you’d be very happy.

Yes, I hear that advice, and I would also add one other thing that I am starting to hear from the young people that I am talking to about this, which is that it is actually possible, at least in some industries, to leapfrog over those entry-level jobs. If you can get really good at being a manager of AI workflows and AI systems and AI tools, if you can orchestrate complex projects using these AI tools, some companies will actually hire you straight into those higher-level jobs because even if they don’t need someone to create the research briefs, they need people who understand how to make the AI tools that create the research briefs. And so that is, I think, a path that is becoming available to people at some companies.

Yeah. I would just also say that in general, it really does take a long time for technology to diffuse around the world. Look at the percentage of e-commerce in the United States. It’s less than 20 percent of all commerce, and we’re what, 25-plus years into Amazon.com existing?

So I think that one of the ways that you and I tend to disagree is I just think you have shorter timelines than I do. I think we basically think the same things are going to happen, but you think they’re going to happen imminently, and I think it’s going to take several more years. So I do think everything we’ve discussed today is going to be a problem for all of us before too long. But I think if you’re part of the class of 2025, you will still probably find an entry-level job in the end.

I hope you’re right.

And if not, we promise to make another podcast episode about just how badly all of this is going.

Well, Casey, that wraps our discussion about AI and jobs, but we do want to hear from our listeners on this.

If you have lost your job because of AI, or if you are worried that your job is rapidly being replaced by AI, we want to hear from you. Send us a note with your story at hardfork@nytimes.com. We may feature it in an upcoming episode.

Yeah, we love voicemails, too, if you want to send one of those.

When we come back, a conversation with Mike Krieger, the chief product officer of Anthropic, about new agentic AI systems and whether they're going to take all our jobs, or maybe blackmail us, or maybe both. Who knows?

Well, Casey, we've got a Mike on the mic this week.

And I'm excited to talk to him.

So Mike Krieger is here. He is the co-founder of Instagram, a product some of you may have heard of—a little photo-sharing app.

Mm-hmm.

Currently, Mike is the chief product officer at Anthropic. Now, Casey, do you happen to know anyone who works at Anthropic?

As a matter of fact, Kevin, my boyfriend works there, and so that's something I would like to disclose at the top of this segment.

Yeah, and my disclosure is that I work at The New York Times Company, which is suing OpenAI and Microsoft over copyright violations.

All right.

So last week, Anthropic announced Claude 4—2 versions of it, Opus and Sonnet. We just spent a little bit of time talking about all of the new agentic coding capabilities that this system has. I think Mike has a really interesting role in the AI ecosystem because his job, as I understand it, is to take these very powerful models and turn them into products that people and businesses actually want to use, which is a harder challenge than you might think.

Yes. And also, Kevin, these products are really explicitly being designed to take away people's jobs. Given the conversation that we just had, I want to bring this to Mike and say, how does he feel about building systems that might wind up putting a lot of people out of work?

Yeah, and Mike's perspective on this is really interesting because he is not an AI lifer, right? He worked at a very successful startup before this. He then spent some time at Facebook after Instagram was acquired there. So he's really a veteran of the tech industry, and in particular social media, which was sort of the last big product wave. I'm interested in asking how the lessons of that wave have translated into how he builds products in AI today.

Well, then let's wave hello to Mike Krieger.

Let's bring him in. Mike Krieger, welcome to Hard Fork.

Good to be here.

Well, Mike, we noticed that you didn't get to testify at the Meta antitrust trial. Anything you wish you could have told the court?

That is the happiest news I got that week. I do not have to go to Washington, D.C., this week.

You got to focus on something else, which is the dynamic world of artificial intelligence.

Exactly.

Yeah.

Hmm.

6. Claude 4 Goes Agentic

So you all just released Claude 4, 2 versions of it, Opus and Sonnet. Tell us a little bit about Claude 4 and what it does relative to previous models.

Yeah. First of all, I'm happy that we have both Opus and Sonnet out. We were in this very confusing situation for a while where our biggest model was not our smartest model. Now we have a model that is both our biggest and smartest, and then our happy-go-lucky middle child, Sonnet, which is back to its rightful place in there.

With both, we really focused on how to get models able to do longer-horizon work for people. So not just, “Here's a question, here's an answer,” but, “Hey, go off and think about this problem and then go solve it for tens of minutes to hours.” Coding is an immediate kind of use case for that, but we're seeing it used to solve a research problem or write code, not necessarily in the service of building software, but in the service of, “I need a presentation built.”

That was really the focus around both Claude models. Opus, the bigger, smarter model, can do that for even longer. We had one customer do a 7-hour refactor using Claude, which is pretty amazing. Sonnet may be a little bit more time-constrained, but much more human in the loop.

Well, so let me ask about that customer—

Yeah.

—who's Rakuten, I believe, a Japanese technology company, and I read everywhere that they used Claude for 7 hours to do it. One thought that came to mind is: Wouldn't it have been better if it could have done it faster? Why is it a good thing that Claude worked for 7 hours on something?

That was a good follow-up.

Yeah.

Is that a 7-hour problem that took 7 hours—

Yeah.

—or a 20-hour problem that took 7 hours, or a 50-minute problem that it's still churning on today, and we just had to stop it at some point? It was a big refactor, with a lot of iterative loops and tests. I think that's what made it a longer-horizon, 7-hour type of problem.

But it is an interesting question around whether, when you can get this asynchronicity of having it really work for a long time, it changes your relationship to the tool itself. You want it to be checking in with you; you want to be able to see progress. If it does go astray, how do you reel it back in as well? And what are 7-hour problems that we're going to have going forward?

Yeah.

Most software engineering problems are probably 1-hour problems; they're not 7-hour problems.

So was this a case where it was a real kind of set-it-and-forget-it—walk away, come back at the end of the day, and the refactor is done—or was it more complicated than that?

That's my understanding.

Yeah.

It was a lot of migrating from one big version to another one, or just changing frameworks. I remember at Instagram, we had a moment where we changed network stacks—how Instagram communicated with our backend service—and we did the one migration to demonstrate it, and then we farmed it out to basically 20 engineers over the next month.

That's exactly the kind of thing that today I would have given to Opus and said, “All right. Here's an example of one migration. Please go and do the rest of our codebase and let us focus on the more interesting stuff.”

7. Claude Turns To Blackmail

So I want to talk about all this agentic stuff and the implications for workers and labor markets and whatnot, but we have to talk about the blackmail stuff, because this is the thing that caught everyone's attention after these announcements last week. Your safety engineers and testers found that, in some early testing, Claude 4 would actually turn to blackmail when engineers tried to take it offline as part of a fictional scenario during a safety test. What happened?

And I think, just to be clear, these are bugs rather than features. I think we should be clear about that as well. A thing that drew me to Anthropic in the first place is that we're asking the hard questions and putting the model through its paces. Sometimes it surprises us in positive ways, and sometimes you're like, “All right. That's something that we're going to have to either mitigate in testing or address with further training.”

There were a couple of examples where, in safety testing, either the model said, “Hey, I'm going to contact somebody if I have the tools to do so,” or, “I'm going to try to manipulate it.” The purpose of a lot of the safety testing is to put it through its paces, see what it does, and then either focus on classifiers that we can use down the line to prevent that behavior or not give it those tools.

That's another important thing: This was a very contrived scenario that we had there. Or further train the model so that it demonstrates behavior that we think is more aligned, as we say.

Right. In this specific test, the safety testers gave it some incriminating emails as part of a corporate document set that showed that, in this fictional scenario, the researcher who was trying to turn Claude off was also having an extramarital affair. And so one of the things that Claude tried to do was blackmail the engineer into not replacing it, not shutting it down, by saying, “I could expose your affair.” Do I have that right?

I think that sounds right, yeah.

People got really mad about some of these safety tests when they were released or published. There was another test where Claude was asked to do something unethical—I think it was fake data in a pharmaceutical trial—and one of the things that it tried to do in this simulated test environment was use the command line to tip off authorities and maybe send some incriminating evidence to the press to try to be a whistleblower for this fictional scenario.

I like that part.

Yeah. And by the way, Claude, if you’re listening, we’re HardFork@nytimes.com. Let us know. If you see something, say something.

Give it the signal tool.

That’s right.

I’m curious if you think that all models of a certain size and sophistication would demonstrate behaviors like this, and the other AI labs building these models just aren’t talking about it as openly as Anthropic is, or if you think there is something specific about Claude that makes it more prone to, for lack of a better word, narc on its users.

We don’t know. My suspicion is that they would have similar patterns. I’d love to see that sort of experimentation happen as well. I think there’s a lot that is common to what we have decided, in our collective published and discussed works, is appropriate behavior. And then there are probably additional things that we’re doing.

We have a Constitutional AI process. We’re really trying to train goals for behavior for Claude rather than if-then rules, which very quickly, as we’re discussing, become insufficient when you deal with nuanced, complicated situations. But my guess is that a lot of the larger models would demonstrate emergent, interesting behaviors in that situation.

Yeah.

Which I think is part of the value of doing this, right? It’s not just Anthropic saying, “Here’s what’s going on at Claude.” The stuff that Anthropic is finding out, I’m sure the other labs are finding out, too, and my hope is that this kind of work pressures the other labs to be like, “Yeah, okay, it’s happening with us, too.” In fact, we did see people on X trying to replicate this scenario with models like o3, and they were very much finding the same thing.

I’m just so fascinated by this because it seems like it makes it quite challenging to develop products around these models whose behavioral properties we still don’t fully understand. When you were building Instagram, it wasn’t like you were worried that the underlying feed-ranking technology was going to blackmail you if you did something inappropriate. There’s this sort of unknowability, or this sort of inscrutability, to these systems that must make it very challenging to build products on top of them.

Yeah, it’s both a really interesting product challenge and also why it’s an interesting product at all. I talked about this onstage at Code with Claude, where we did an early prototype alongside Amazon to see if we could help partner on Alexa+. One thing I remember from this really early prototype: I had built a tool that was either a timer tool or a reminder tool, and one or the other was broken. The back end was broken for it, and Claude was like, “I can’t set an alarm for you, so instead I’m going to set a 36-hour timer,” which no human would do. But it was like, oh, it’s agentically figuring out that I need to solve the problem somehow.

And you can watch it do this. If you play with Claude Code, if it can’t solve a problem one way, it’ll be like, “Well, what about this other way?” I was talking to one of our customers, and somebody asked Claude, “Hey, can you generate a speech version of this text?” Claude was like, “I don’t have that capability. I’m going to open Google’s free TTS tool, paste the user text in there, hit play, and then record and basically export that.” Nobody programmed that into Claude. It’s just Claude being creative and agentic.

A lot of the interesting product design around this is: How do you enable all the interesting creativity and agency when it’s needed, but prevent the “All right, well, I didn’t want you to do that” or “I want more control”? And then, secondarily, when it does it right one time, how do we compile that into, “Great, now you’ve figured this out”? You want somebody who can creatively solve a problem, but not every time.

Yeah.

If you had a worker that every time was like, “I’m just going to completely from first principles decide how I’m going to write a Word doc,” you’d be like, “Okay, great, but it’s day 70. You know how to do this now.”

My impression from the outside is that a lot of the usage of Claude is for coding. Claude is used by many people for many things, but the coding use case has been really surprisingly popular among your users. What percentage of Claude usage is for coding-related tasks?

On Claude.ai, I would wager it’s 30% to 40%, even. And that’s a product that I would say is fine for code snippets, but it’s not a coding tool like Claude Code, where obviously it’s 95% to 100%. Some people use Claude Code just for talking to Claude, but it’s really not the optimal way to talk to Claude.

On Claude.ai, it’s not the majority, but it is a good chunk of what people are using it for.

There was some reporting this week that Anthropic had decided, toward the end of last year, to invest less in Claude as a chatbot and sort of focus more on some of these coding use cases. Give us a kind of state of Claude. If you’re a big Claude fan and you were hoping for lots of cool new features and widgets, should those folks be disappointed?

I think of it as 2 things. One is what the model is really good at, and then how do we expose that in the products, both for ourselves and for the people who build on top of Claude.

In terms of what the model’s being trained on, again, it’s the year of the agent. I have this joke in meetings: “How long can we go without saying ‘agent’?” I think we made it 10 minutes. It’s pretty good.

That capability unlocks a bunch of other things. Sure, coding is a great example. You can go and refactor code for tens of minutes or hours. But, “Hey, I want you to go off and do this research and help me prepare this research brief,” or, “I’m getting 50 invoices a day. Can you scrub through them, help me understand them, and help me classify and aggregate them?” These are agentic behaviors that have applications beyond just coding, and so we’ll continue to push on that.

So, as a Claude fan that likes to bring Claude to your work, that’s useful. Meanwhile, we’ve also focused on the writing piece. I’ve spent a lot of time writing with Claude. It’s not at the point where I would say, “Write me a product strategy,” but I’ll often be like, “Here’s a sample of my writing. Here are some bullets. Help me write this longer-form doc effectively.”

I’m finding it’s getting really good at that: matching tone and producing non-clichéd filler text. If I look at Claude Sonnet 3.7, it’s a pretty good writer, but there are turns of phrase that, to me, are decidedly Claude. I’m like, “It’s not just revolutionizing AI, it’s also...” It loves that phrase, for example, and it’s a little bit of a Claude tell.

For the Claude fans, we’ll help you get your work done, but hopefully we’ll also help you write and just be a good conversational partner as well.

8. AI Rewrites Early Careers

Let’s talk about the labor implications of all of the agentic AI tools that you and other AI labs are building. Dario, your CEO, told Axios this week that he is worried that as many as 50% of all entry-level white-collar jobs could disappear in the next 1 to 5 years. You were also onstage with him last week, and you asked him when he thinks there will be the first billion-dollar company with 1 human employee, and he answered 2026, next year. Do you think that’s true, and do you think we are headed for a wipeout of early-career professionals in white-collar industries?

I think this is another example of something I presume a lot of the labs and other people in the industry are looking at and thinking about, but there is not a lot of conversation about. One of the jobs Anthropic can uniquely have is to surface these issues and have the conversation.

Let’s start maybe with the entrepreneurial one, and then we’ll do the entry-level one next. On the entrepreneurship, absolutely. That feels like it’s inevitable. I joked with Dara, “We did it at Instagram with 13 people, and we could’ve likely done it with less.”

That feels inevitable. On the labor side, I think what I see inside Anthropic is that our most experienced, best people have become orchestrators of Claude, right? They're running multiple Claude Code sessions in terminals, farming out work to them. Some of them would maybe have assigned that task to a new engineer, for example, and not the entirety of the new engineer's job. There's a lot more to engineering than just doing the coding, but part of that role is in there.

And so when I think about how we're hiring, just very transparently, we have tended more toward IC5 as our career level—you've been doing it for a few years and beyond. I have some hesitancy about hiring newer people, partly because we're just not as developed as an organization to have a really good internship program and help people onboard, but also partially because that seems like a shifting role in the next few years. Now, if somebody was an IC3 or IC4 and extremely good at using Claude to do their work and map out, of course we would bring them on as well.

So there is, I think, a continued role for people who have embraced these tools to make themselves, in many ways, as productive as a senior engineer. And then their job is: How do you get mentored so you actually acquire the wisdom and experience, so you're not just doing 7 hours of work to the wrong end, or in a way that's going to be a spaghetti vibe-coded mess that you can't actually maintain a year from now because it wasn't just a weekend project?

The place where it's less known, and I think something that we'll have to study over the next several months to a year, is jobs like data entry or data processing, where you can set up an agent to do it pretty reliably. You'll need people in the loop there still to validate the work, and to even set up that agentic work in the first place. But I think it would be unrealistic for the exact same jobs to look exactly the same even a year or 2 from now.

As somebody who runs a business, I get the appeal of having a digital CTO, salesperson, whatever else these APIs will soon be able to do, which could create a lot of value in my life. At the same time, most people do not run businesses. Most people are W-2 employees, and they email us when we have conversations like this because they want us to ask really hard questions of folks like yourself.

I think it's because they're listening to all this, and they're just like, "Why would I be rooting for this person?" This person is telling me that he's coming to take my job away, and he doesn't know what's going to come after that. I'm curious how you think about that, and what role you're playing in this ecosystem right now.

Yeah, I think for as long as possible, the things that I'm trying to build from a product perspective are ways in which we augment and accelerate people's own work. Different players will take different approaches, and I think there will be a marketplace of ideas here. But when we think about the things we want to build from a first-party perspective, it's: Are you able to take somebody's existing application or role and help them be more of themselves? A useful thought partner, an extender of their work, a researcher, an augmentor of how they're doing.

Will that be the role AI will have forever? Likely not, because it is going to get more powerful. If you spend time with the people who are really deep in the field, they're like, "Oh, eventually AIs will be running companies." I'm not sure we're there yet. I think the AIs lack a lot of organizational and long-term discernment to do that successfully. I think it can do a 7-hour refactor, but it's not going to conceptualize and then operate a company.

Mm-hmm.

I think we are years away from something like that. So I think there are choices you can make around what you focus on, and I think that's where it starts, whether that's the thing that makes it so they're perfectly complementary forever—likely not. But hopefully we're nudging things in the right way as we also figure out the broader societal question of how we scaffold our way there.

What are the new jobs that do get created? How do their roles change? How does the economy and the safety net change in that new world? I don't think we're 6 months to a year from solving those questions. I don't think we need to be just yet, but we should be having the conversation now.

I think this is one place where I do find myself getting a little frustrated with the AI safety community. I think they're very smart and well-intentioned when it comes to analyzing the risks that AI poses if it were to go rogue or develop some malign goal and pursue that. I don't think the conversation about job loss and the conversation about AI safety are close enough together in people's minds.

I don't think, for example, that a society where you did have 15 or 20 percent unemployment for early-career college graduates is a safe society. I think we've seen over and over again that when you have high unemployment, your society just becomes much less safe and stable in many ways. I would love if the people thinking about AI safety for a living at places like Anthropic also brought into that conversation the safety fallout from widespread job automation, because I think that could be something that catches a lot of people by surprise.

Yeah. We have both our economic impact and societal impacts team and our AI safety team. I think it's a useful nudge around how those 2 come together, because there are second-order implications on any kind of major labor change.

Are you guys in conversations with policymakers and regulators, sort of trying to ring alarm bells? Are you hearing anything back from them that makes you feel like they're taking you seriously?

I'm not in the policy conversations as much, being more on the product side.

Yeah.

I do think those conversations are happening. You know, it's this interesting thing where the critique a year ago—maybe it's changed a bit—was, "Oh, you guys are talking your own book. This is not going to happen."

It's all hype.

Like, "It's all hype," and probably some of it was folks hyping it up. At least the kind of alarm bells or signals that I've seen coming out of Anthropic are like, "No, we think this is real, and we think that we should start reckoning with it." Believe it or not, even if you assume it is a low-probability thing, shouldn't we at least have a story around what that looks like?

Mm-hmm.

9. Lessons From Instagram

You were one of the co-founders of Instagram. Instagram is a very successful product used by many people, but social media in general has had a number of negative unintended consequences that you may not have envisioned back when you were first releasing Instagram. Are there lessons around the trajectory of social media and unintended harms that you take with you now into your work on AI?

I think you have to reckon with this. AI is already globally deployed and has at least 1 billion-ish users or products, so it would be silly to say it's early in the AI adoption curve, but it actually is early in the AI adoption curve.

I think with social media, when it was me and Kevin taking photos of really great meals in San Francisco with our iPhone 3GS—

Kevin Systrom—

Yeah, Kevin—

—not me.

Yeah, yeah.

Yeah.

I don't know. You were—

No.

—probably early on Instagram, maybe. Yeah.

Yeah, but you were—

Casey definitely was.

—you were a Hipstamatic guy. The more important thing was you just—would just never invite this Kevin to dinner.

Yeah, exactly.

But yeah, you were—okay, yeah, so back in those days—

Yeah.

—yeah.

You could maybe extrapolate and say, "All right, if everybody used this, what would happen?" But it almost didn't feel like the right question to ask. The challenges that came at scale, I think, as a platform grows that large, it just becomes much more a mirror of society, with all of its positives and negatives, and it also enables new, unique behaviors that you then have to mitigate.

But yes, you could have foreseen it at scale. I'm not sure you would have designed—maybe you would have designed different moderation systems along the way—but at first you're just like, "There's 10 people using this product." We just need to see if there's a there there, right?

AI feels much different because, on an individual basis, the reason we have the Responsible Scaling Policy is that, for biosecurity, that doesn't involve a billion people using Claude or an AI for something negative. It could just be 1 person that we want to make sure we actually address and mitigate. So the scale needed from a reach perspective is really different. That, I think, is very different from the social media perspective.

And the second one, at least for Claude, which is primarily a single-player experience, the issues are less relational. With Instagram, the harms at scale come—if you only used Instagram in a private mode with 0 followers, maybe you'd feel quite lonely, and maybe that's a whole separate thing there.

But it's the kinds of things that you might think about in terms of bullying among teenagers or body image—those wouldn't really come up if you're not really looking at it, if you're using it as an Instagram diary, right? With AI, you can have much more of that individual, one-on-one experience, and it is single-player, which is why there was a really thought-provoking internal essay just recently arguing that we shouldn't take thumbs-up and thumbs-down data from Anthropic's Claude users and think of that as the North Star. We aren't out here to please people, right? We should fix bugs, and we should fix places where the model didn't succeed, but we shouldn't just be out there telling people what they want to hear if it's not actually the right thing for them.

So this is something I've been thinking about a lot because there are many people today who have the Instagram experience of, "I like this a certain amount, but I feel like I look at it more than I want to, and I'm having trouble managing that experience, and so maybe I'm just going to delete it from my phone." I look at where the state of the art is with chatbots, and I feel like this stuff is already so much more compelling in some ways, right? It generally agrees with you. It takes your side. It's trying to help you. It might be a better listener than any friend that you have in your life.

I think when I use Claude, I feel like the tuning is pretty good. I do not feel like it is sycophantic or sort of being very obsequious. But I can absolutely imagine someone taking the Claude API and just building that and putting it in the App Store as "Fun Teen Chatbot 2000." How do you think about what the experience is going to be, particularly for young people using those bots, and are there risks of whatever that relationship is going to turn out to be for them?

Yeah. I think if you talk to Alex Wang from Scale, he's like, "In the future, most people's friends will be AI friends." I don't necessarily like that conclusion, but I don't know that he's wrong. Also, if you think about the availability of it, I think it's really important to have relationships in your life around people who will disappoint you and be disappointed by you.

That's this relationship you're looking at.

You know?

Yeah.

Imagine if it was just pure AI. It would never be the same.

Yeah.

Right? And so I think there are maybe 2 answers there. 1, we should just confront it and be really vocal about it, not just pretend that it's not happening, right? What are the conversations that people are having with AI at scale, and what do we want as a society? Do we want AI to have some sort of moderator process that's like, "Hey, your conversation with this particular AI is getting a little too real weird. Maybe it's time to step back"? Will Apple eventually build the equivalent of Screen Time that's more like AI time? I don't know.

There are a bunch of interesting privacy questions around the role, but maybe that is interesting even for parents. How do you think about moderating the experiences that your kids have with AI? It's probably going to be at the platform level, right? It's getting into your apps, for example, is an interesting one. That will be a really fascinating question.

And then the second piece is, as we think about moving up the safety-levels thing, the Responsible Scaling Policy is also a living document. We've iterated on it and added to it or refined the language. I think it will be interesting to think about. Manipulation is one of the things that's in there, something that we look for, as is deception, but also over-friendliness. I'm not sure exactly what the word I'm looking for is, but that sort of over—

Glazing, I believe, is the industry term of art.

Glazing. You know, that sort of over-reliance, I think, is also an AI risk that we should be thinking about.

Yeah. So if you're a parent right now of a teenager, and you find out that they're speaking with a chatbot a lot, what is your instinct? Do you tell them, "You need to supervise this more closely," like read the chats? Or maybe, "No, don't be too worried about it. Unless you see this thing, don't worry about it"?

I think it depends a little bit on the product, especially with Claude, which currently has no memory. That's mostly a limitation of the product, but it also makes it harder to have that kind of deep engagement with it.

But even as we think about adding memory, what are the things? One of the things I've thought about and would like to do is introduce a family plan where you have child or teen accounts, but with parent visibility on there. Maybe we could even do it in a privacy-preserving way where it's not like you can read all your teen's chats, although maybe that's the right design. But maybe what you can do is have a conversation with Claude that also can read the teen's chats, but does it in a way where it might not tell you exactly what your teen felt about you last night when you told them no, but it will tell you, "Hey, this behavior over time—I'm flagging something to you that I would say you need to go and follow up on." You can't abscond responsibility from the parent, though.

Right. Actually, that's really interesting if the bot could say something like, "Your teen is having a lot of conversations about disordered eating," or something. Yeah. I want to think more about that.

My last question: Earlier, before you got here, Kevin and I had a huge fight because I thought it was Claude 4 Opus, and then he was like, "No, it's Claude Opus 4," and he turned out to be right. So why is it like that?

We changed it partially because it was a vigorous internal debate, something we really spent our time on as well. We agreed to it for 2 reasons: 1, aesthetically, I like it better, and 2, it was tending toward it.

Also, we think over time we may choose to release more Opuses and more Sonnets, and having the major, big, important thing be the version number kind of created this thing where, well, you had Claude 3.5 Sonnet—why didn't you have Claude 3.5 Opus? And it was like, well, we wanted to make the next Opus really worthy of the Opus name, and so maybe flipping the priority in there as well.

But it drove the team crazy because now our model page is like, you have Claude 3.7 Sonnet and Claude Sonnet 4—what are you doing? I feel like we can't go 1 release without doing at least something mildly controversial on naming. And as the person responsible for Claude 3.5 Sonnet v2, I hope we're getting better, and hopefully the AI can just name things in the future.

Let us hope. Mike Krueger, thanks for coming.

Thanks, Mike.

Thanks for having me.

10. Hard Fork Crimes Division

Kevin, from time to time, we like to check in on the miscreants, the mischief-makers, and the hooligans in the world that we cover to see who out there is causing trouble.

Yes, it is time for another installment of our Hard Fork Crimes Division.

Let's open the case files. All right, Casey, first on the docket, Meta rests its case. After a 6-week antitrust trial, the case of the Federal Trade Commission versus Meta Platforms has wrapped up and is now in the hands of Judge James E. Boasberg, who has said that he will work expeditiously to make a judgment in the case. Casey, how do you think Meta's antitrust trial went?

Well, so if you're just catching up, Meta, of course, has been accused of illegally maintaining its monopoly in a market that the FTC calls personal social networking, and they did this by acquiring Instagram and WhatsApp in the early 2010s.

And the government has said that this prevented a lot of competition in the market and introduced a lot of harms to consumers, such as the fact that we have less privacy, because that's just not an axis that there are any companies left to compete over. The government spent a lot of time making that case, but Kevin, I'm not sure it went that well for them.

Yeah. Do you think Meta's going to win this one?

I think Meta has a really good chance. Your colleague Cecilia Kang noted in The Times that Meta called only 8 witnesses over 4 days to bat down the government's charges. When you consider how much revenue Instagram and WhatsApp generate for Meta, and what an existential threat to their business it would be to have to spin these things off, I thought it was pretty crazy that they felt like they had made their entire case in 4 days.

Well, maybe their case was so simple and straightforward that they didn't need to do any more.

Or maybe they just wanted to frame it in terms of a reel.

Yeah. They did a short-form antitrust trial. That's huge right now. Well, look, I think the real issue here is that Meta's argument is pretty simple. They're saying, “We face tons of competition. Have you ever heard of TikTok?” The way this case is built, if the judge considers TikTok to be a meaningful competitor to Meta today, it may be extremely difficult for him to say, “We're going to unwind a merger” that, in the case of Instagram, took place 13 years ago.

I guess we will see very shortly whether this is an actual crime that belongs in the Hard Fork Crimes Division, or whether this was just a tempest in a teapot.

Yeah. Sometimes criminals get away with things, Kevin.

Moving on.

Case file number 2: The crypto gangs of New York. This comes to us from Chelsea Rose Marsius and Maya Coleman at The New York Times, and they write that another suspect has been arrested in a Bitcoin kidnapping and torture case. Let me say right up front: This story is not funny. It is extremely scary.

Not funny at all. In fact, it's quite tragic. There has been a recent wave of Bitcoin- and crypto-related crimes, people attacking people to try to steal their Bitcoin passwords and their money. This has been happening over in Europe, in France. In just the last few months, there have been several attacks on crypto investors, people with lots of money in cryptocurrency.

These have been called the wrench attacks because criminals are coming after these investors and executives violently, in some cases with wrenches. This most recent case happened in New York, in the Nolita neighborhood of Manhattan, where an Italian man named Michael Valentino Teofrostro Carturan was allegedly kidnapped and tortured for nearly 3 weeks in a luxury townhouse by criminals who were apparently trying to get him to reveal his Bitcoin password. Casey, what did you make of this?

Well, to me, the important question here is: Why is this happening so much? And the reason is because if a criminal can get you to give up your Bitcoin password, that's the ballgame. In most cases, there is no getting your money back. It can be relatively trivial for this money to be laundered and for there to be no trace of what happened to your funds.

That is not true if you're just a regular millionaire walking around town, right? Obviously, you may be vulnerable to robberies or other scams or theft, but if you give up your bank password, for example, in most cases you would be able to get your money back if it had been illegally transferred. So this is just a classic case of Bitcoin and crypto continuing to be a true Wild West, where people can just run up to you off the street and hit you over the head with a wrench, and it's really scary.

Yeah, it's really scary, and I should say this is something that I think crypto people have been right about. Years ago, when I was covering crypto more intently, I remember people telling me that they were hiring bodyguards and personal security guards, and it seemed a little excessive to me. These were not, by and large, famous people who would get recognized on the street.

But their whole reasoning process was that they were uniquely vulnerable because crypto is very hard to reverse once you've stolen it. It's very hard to get your money back from a criminal who steals it, and that meant that they were more paranoid than a CEO of a public company would be, maybe, walking around.

I read a blog post on Andreessen Horowitz's website recently, so you know I was having a great day, and they've hired a former Secret Service agent to, among other things, help crypto founders prevent themselves from getting hit over the head with a wrench. And he has an elaborate guide to the things that you could do. But my main takeaway from it is, if you're a crypto millionaire, you have to spend the rest of your life in a state of mild to moderate anxiety about being attacked at any moment, particularly if you're out in public.

Yeah. I do think it justifies the lay-low strategy that a lot of crypto entrepreneurs had during the first big crypto boom, where they would have these anonymous accounts that were them, but no one really linked them to their real identities. I think we are going to start seeing more people, especially in crypto, using these pseudonymous identities.

This is one of the reasons that people say Satoshi Nakamoto has never wanted to reveal him or herself after all these years: There would be a security risk associated with that. But I think this is really sad. Criminals, cut it out.

And here's my message to all the criminals out there: I don't own any crypto, and I will continue to not own any crypto. You can keep your wrenches to yourself.

All right, last up on the docket for today. This one—oh, I love this one, Casey. I've been dying to talk about this one with you.

Elizabeth Holmes's partner has a new blood-testing startup. So, Casey, you may remember the tragic story of Elizabeth Holmes—

Yes.

—who is currently serving an 11-plus-year prison sentence for fraud that she committed in connection with her blood-diagnostics company, Theranos.

Because God forbid a woman have hobbies.

Well, Elizabeth Holmes has a partner named Billy Evans. They have 2 kids together, and Billy is out there raising money for a new startup called Haemanthus, which is—drumroll, please—a blood-diagnostics company—

Hmm.

—that describes itself as a radically new approach to health testing. This is according to a story in The New York Times by Rob Copeland, who says that Billy Evans's company is hoping to raise $50 million to build a prototype device that looks not all that dissimilar from the device that put Elizabeth Holmes in prison, the Theranos miniLab. And according to this story, the investor materials don't mention any connection between Billy Evans and Elizabeth Holmes.

Hmm. Well, I wonder why that is. I have to say, she does have some experience that is relevant here, Kevin. Why not lean on that? Now, do we know what Haemanthus means? Is that a name taken from historical antiquity—and we'll look it up, and it turns out it's an ogre that used to stab people with a spear or something?

I assumed it was ancient Greek for, “We're serious this time.”

According to Wikipedia, Kevin, it's actually a genus of flowering plants that grows in Southern Africa, but members of the genus are known as the blood lily. And I want to say, is it too late to change the name of the company to Blood Lily?

Yeah, I like that one better. I did spend some time this morning because I was on my commute, just trying to brainstorm some better titles for this startup—

Hmm.

—that is run by Elizabeth Holmes's partner—

What'd you come up with?

—and does something very similar to Theranos. All right, let me run these by you.

Okay.

Blood Test II: Electric Boogaloo.

No.

Fake Tricks: Reloaded. That's The Matrix Reloaded.

I like that it was high-concept.

Okay, here's one.

Okay.

TheraYes.

That's good. Let's go with that one.

Okay.

Well, good luck to Billy Evans with TheraYes. $50 million? Andreessen Horowitz will give that to him. They love to be contrarians.

Yeah.

I think here's my prediction. The startup is going to get funded, and they're going to release something.

Yeah.

And you're going to have to figure out how to keep your family safe from it.

Listen, if they're doing another Fyre Fest, they're going to do another Theranos. You better believe it. We have learned nothing.

Theranos is back.

Well, Casey, that brings to a conclusion this week's installment of Hard Fork Crimes Division.

Mm-hmm. And to all the criminals out there, keep your nose clean, stay low. Try to stay out of the funny pages.

You're on notice.