An A.I. Assistant That’s Too Good to Be True?

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, phone-to-phone combat in the streets. We'll talk about the tech industry's response to ICE killings in Minneapolis. Then it's the reason Mac minis are sold out across San Francisco. We'll tell you about the sensation now known as Moltbot. And finally, it's time once again for Hat GPT.

So today we want to start by turning our attention to Minneapolis and the ongoing confrontation between ICE officers and protesters. Kevin and I have both been horrified by the fatal shooting of Alex Pretti, the second fatal shooting of a U.S. citizen by federal agents this month. And while we don't often turn our attention directly to matters of politics and protests on the show, we both believe that the tech industry is playing a role in this story and that we would be remiss if we did not look at the elements of what is going on in Minneapolis that are ours.

Yeah, I've been really horrified by what's been going on in Minneapolis. I have family in Minneapolis. They're scared. Some of them are leaving. It is a really tense and hard time in American civic life right now, and I think we should talk today about not just how the big tech companies and their leaders are acting right now, but what role technology itself is playing in some of the reaction to these shootings.

Yeah. And I think the tech industry is playing a significant role in what is happening right now. It is an infrastructure provider behind many of the key tools that are being used in this fight. I would say primarily a lot of the surveillance tech that ICE is relying on to identify and detain migrants, but also the social platforms where once again we're seeing violence being turned into content and influencers on both sides trying to make it go viral.

And then you just have the dimension of the fact that there is actually now state power—the government—that is using these tools in ways that we have not seen before in America. So there's a lot to talk about here.

1. CEOs Navigate the Crisis

Yeah. So let's start with the CEOs. At the start of this week, Sam Altman of OpenAI, Dario Amodei of Anthropic, and Tim Cook of Apple each made statements or sent internal memos to their employees lamenting what was going on. These were slightly different in the way they were worded, but they all came after employees at some of these companies had been urging their CEOs to speak up, to say something.

Sam Altman had an internal Slack message to OpenAI employees that said that he had spent some time talking with Trump officials, and he described ICE as going too far. He also said that he thought that part of loving the country is the American duty to push back against overreach. Dario also discussed it in a post he made on X, where he called the events in Minnesota “horror,” and Tim Cook sent an internal message to Apple employees saying, quote, “This is a time for de-escalation.”

So we're not seeing the kind of full-throated denunciations that we might have seen with the first Trump presidency from some of these leaders, but they are saying, at least to their own employees, that they are not happy about what's going on.

Yeah. And I would say that in several of these cases, we are seeing these CEOs say about the least that they possibly could. You can tell, particularly in the Altman and Cook statements, that they are very afraid of irritating the White House. At the same time, they're facing a lot of pressure from their employees to say something, and so I think this represents their best efforts to find a middle path. Personally, I found it a little weak.

Yeah. I don't like it when CEOs or other business leaders talk insincerely out of obligation. I think we saw some of that several years ago from the left. I think we saw some of that during the first months of the Trump administration aimed at appeasing the right. I just don't like when CEOs talk in ways that are not their own views. In some of these cases, I'm sure these are sincerely held views, but it just seemed to me like, why are we doing this?

Well, I think the reason that we're doing this is because these guys are politicians, too. They are, in their own way, heads of state. They represent hundreds of millions of users and giant employee bases, and so they have to get in there and play politics. I'm sure they hate it, but the fact is that in these cases they have to.

And I think we should say it is risky to do this in this current environment. There was a post from Chris Olah, who's one of the co-founders of Anthropic, who posted a very heartfelt and sincere message about how horrified he was by what was happening in Minneapolis. And Katie Miller, the wife of Stephen Miller, the White House adviser, reposted it and basically said, “Look, this is the AI companies. They're building this bias into their models. If they talk like this, this is how their AI models are going to talk.”

So there is a cost politically to saying something right now, and so I don't want to be totally dismissive of the people who are doing that. I think they are displaying some backbone. But as you said, in some of these cases, they're saying about as little as they could.

2. X Becomes State Media

Mm-hmm. So, Casey, one of the last times we talked about political violence on the show was after the assassination of Charlie Kirk, when we talked about some of the online commentary and how weaponized it was feeling out there—how these platforms were becoming places full of rage bait and incendiary, provocative content, and how that was bad for us as a society. I'm curious: Does this feel like another case of that to you, or what are you seeing out there?

I think there are definitely ways where it's similar, Kevin. People are drawn to conflict, and so are social networks, right? This is why Charlie Kirk was so successful in his way: He would create these conflicts by visiting college campuses and staging debates, and those conflicts would then go viral in ways that served him.

We're seeing a similar thing here, where the Trump administration uses spectacle as one of its main weapons. And if it can create a crisis that begets a bunch of video, that can serve its interest, too.

I think the real difference here is that Charlie Kirk did not have the power of the state behind him, and Trump does, right? Trump has an army behind him. He has thousands of ICE agents and other federal agents. And so we have just seen, after he returned to power, the way that various Trump administration officials would bring along influencers and their entourages when they would go about their business.

So last year, Kristi Noem brought a bunch of right-wing influencers with her when she visited an ICE operation happening in Portland. And another important thing here, Kevin, is the way that the Minneapolis operation began after a video alleging fraud in daycare centers went viral on X. In this very real way, a social platform is setting the president's policy agenda. So that's just very different from the Kirk stuff. The stakes are a lot higher here.

Right? I think in a lot of public spaces, what's happening in Minneapolis is not a 50/50 issue. I think it's a pretty widely held belief that federal agents should not be shooting protesters in the streets. But on X, you see people defending it. You see people spreading these conspiracy theories. You see people circulating this manipulated footage and these deepfake images that purport to show more of a violent standoff than there was.

And it just seems like a platform that has either decided to put its thumb on the scale in favor of the administration in its responses, or where just a lot of this is going on unchecked.

Yeah. You predicted after Trump won that X was going to become state media, and I think that it basically has. This is where the Trump administration goes to get its policy ideas and to advance whatever operation they're working on at the moment.

And so you have just seen, all across X, administration officials making their case that this ICE campaign is justified or necessary.

Yeah. And I think another difference between what's going on now and the Charlie Kirk assassination is that the government itself is getting involved in actually producing some of this content. There was a great story by 2 reporters at The Washington Post last year about ICE's media strategy, which was based on a bunch of internal agency documents and messages they got.

What I found remarkable about this is that ICE, like some other federal agencies, now essentially has a team of content creators who work at the agency and who are spending every day trying to steer the narrative using the same techniques that a brand or a celebrity would. They're out there using paid social media tools. They have a team of producers and editors and video makers who are out there making clips.

It seems like these agencies are not just thinking that their jobs are enforcing the law. It seems like they are also trying to control the narrative around how they enforce the law.

And I would say almost to the point where winning on social media has become almost the entire point, right? Yes, there are policy objectives here, but in a very real way, they seem secondary to getting the most retweets.

Yes.

Yeah. So that feels new.

3. AI Blurs the Evidence

Right? So in addition to the government actively working to seize and control the narrative online about these events, we also have the rise of AI tools that are making it very hard for people to know what's real and what's not.

Mhm.

So my colleagues at The New York Times had some reports on the way that AI was being used to alter images of Alex Pretti, the victim of this shooting. One was doctored to make him look like he was pointing a gun at an agent, although it was actually a phone. Another was used to enhance a blurry freeze-frame from a video of this shooting.

And we should put “enhance” in air quotes because it didn't actually enhance it. It just made a bunch of stuff up.

Right. Exactly. So it was sort of upscaled in this way that introduced some errors into the image. Casey, I can't imagine this was surprising to you. We've seen this kind of thing after most major news events in the past couple of years, but did anything stick out to you about the actual media that was being circulated?

Well, when it comes to online detectives who want to participate in the story of the day, that is not new. That's been going on for well over a decade. But Kevin, the White House released this AI-altered image of a civil rights attorney named Nakeema Levy Armstrong, who went to an anti-ICE protest last week and was arrested. In the aftermath of her arrest, the White House put out this image to make her look like she was crying her eyes out, when in fact she was not.

And it's kind of a weird story because DHS had already released the original, unaltered image before this altered one came out. But once the one of her crying—which, again, she wasn't doing—came out, J.D. Vance, the vice president, retweeted the image.

So the thing this has me thinking about is this phrase that got associated with Russia in the 2010s: “Nothing is true and everything is possible.” It referred to the way that Putin and his lieutenants would try to shape a world where you just couldn't trust anything that you were seeing anymore. The state invested huge resources into just blurring the line between truth and fiction. And if you wanted to know what was really happening, you were just going to have to invest enormous amounts of resources.

And the way that this serves the administration is this concept we've talked about a few times on the show that I love, which is called the liar's dividend. It means that because people know that evidence can be fabricated, no matter what evidence you see now, you're always wondering, “Was this maybe fabricated?” And that just erodes trust in our society more broadly.

So I think it's hugely irresponsible for the vice president to be sharing these sorts of images. But when the administration has been asked, “What are you doing here? Why are you sharing these obviously doctored images?” a spokesman for the White House just said, “The memes will continue,” which I honestly found chilling.

Yeah. What do you think the platforms should be doing about this? Should they be taking down doctored images? Should they be applying some sort of label? What is the correct response here?

Well, I go back to the way platforms handled this during the first Trump administration, where they were much likelier to act. During the 2020 election, when President Trump was sharing misinformation about mail-in voting ballots, Twitter put a label on his post that directed people to high-quality information about how mail-in voting actually worked.

Now, of course, Twitter doesn't exist. Elon Musk owns X. And if anybody is going to correct anything, it's going to be in this Community Notes feature, which is just this volunteer crowdsourcing effort. And I'm glad that it exists. I'd rather have it than not have it. But, of course, it's unpredictable and doesn't always appear in a timely fashion.

It's my view that platforms have a strong interest in labeling obviously AI-generated content. I think that if you want people to trust your platform, you should take steps to help them understand what they're seeing is real and what is false. But, of course, in many cases, they're going to be terrified to do that because there might be a mean tweet about them.

Right. Exactly. And I think, too, for me, what I've been thinking about is less the platform response and more the AI companies. Obviously, I have AI on the brain sort of permanently. But I think this is actually why these companies should support regulation for the technology that they're building, because right now they're in this fragile state where they can't really speak up against the administration in very forceful terms without risking some kind of retaliation: some new law, just a bunch of mean tweets. There are actual political risks to taking a stand on something.

Even if they were to come out now and say, “Oh, we're going to apply a label to AI-generated images that are misleading, or we're going to have some kind of watermarking system,” I think some people in the Trump administration would interpret that as a response to what they were doing, and they would get mad about it.

Yeah. They call it censorship.

Exactly. So I think this is actually the case for regulation that is passed by Congress, signed into law, and persists across administrations, so that you don't have to operate in this weird limbo where you're trying to do the right thing but you're scared that the administration will punish you for that, and anything that you do is going to be seen, correctly or not, as a reactive countermeasure instead of just having one set of rules that apply across all administrations.

I think one set of rules is a great idea. I think we should pursue this. I also think that if you find yourself in a situation where there's a presidential administration that decides it wants to lie about what is happening and is willing to fabricate evidence to support those lies, platform policy is only going to help you so much.

I want it to exist. I want to be able to trust in it. I just want to also point out there are limits to it.

4. Phones Become the Battlefield

Exactly. I want to talk about this other dynamic here, Kevin, which in my head I've been calling the battle of phone versus phone. Smartphones are playing a central role in this conflict. And while, of course, phones have long been part of protests, I feel like we're seeing a direct confrontation between the state and the protesters involving the phone in a way that we haven't quite seen before.

Yeah, this stuck out to me, too, as I was looking at some of the awful videos of the shooting of Alex Pretti. Everyone's holding phones, right? This is now a standard part of any confrontation with law enforcement in this country: if there's anyone around to witness it, there are going to be phones pointed at it. Someone is going to be filming that.

Absolutely. Alex Pretti was holding a phone when he was shot. Renee Good's partner was filming when Renee was shot, and videos of their killings have spread on social networks and I think have significantly shifted public opinion against the Trump administration. Governor Tim Walz of Minnesota said, “Carry your phone with you at all times in order to help us create a database of the atrocities against Minnesotans.”

This is a dynamic I think we have seen before. You can go all the way back to the Arab Spring and look at the way that people were posting on Twitter images of state violence. Here's, I think, the twist. The Trump administration has been doing 2 things.

One, they've been putting pressure on American citizens not to film. DHS Secretary Kristi Noem said in a press briefing in July, “Violence is anything that threatens them and their safety. So, it is doxing them. It's videotaping them where they're at when they're out on operations.”

Also, Tricia McLaughlin, who's the DHS assistant secretary for public affairs, said in a statement, quote, “Videotaping ICE law enforcement and posting photos and videos of them online is doxing our agents. We will prosecute those who illegally harass ICE agents to the fullest extent of the law.”

So there's this recognition in the Trump administration that being filmed and having your video put on social media is dangerous to you. And we've started to see a lot of threats against people who are doing this.

And we should just say it is not illegal to film law enforcement. It is not doxing to post the name or the image of a federal agent who is operating on behalf of the state. These people wear badges, or at least they're supposed to, because we want this sort of public accountability. If they are acting in ways that are illegal or unethical, we want to be able to hold those individuals accountable.

But I agree that there's this feeling among certain members of the administration that these phones present a danger.

Absolutely. And when you consider the fact that all of these ICE agents are just permanently masked when they're out in the world, it makes you wonder, Kevin: What are they more scared of? Is it the guns that some of these protesters may be legally carrying, as Alex Pretti was? Or is it the phones that are in their hands? Is it the threat that someone will find out who they are and what they're doing and link it back to their actual identity?

There's this final thing that we're seeing in this phone-versus-phone battle, though, Kevin, and this is where I also think that we're in some new territory, which is that the Trump administration is bringing their own phones to the battle. They're bringing along those conservative influencers. They're giving them talking points and ideas for what sorts of videos they might post. There's a long tradition of the government using celebrities and influencers to promote their worldview or various things that they're doing.

But the thing in Minneapolis that seems really different to me is that it is almost like phone-to-phone combat between Americans, right? You have protesters who are trying to document what is happening, and you have the administration trying to shoot photos and videos to suggest that these are very effective anti-immigration raids.

Yeah. And I think for me, the question is what happens when the videos—the artifacts that these phones are producing themselves—become untrustworthy? As you mentioned, people were filming the Arab Spring to try to bring accountability to their leaders. When people started taping their encounters with law enforcement 10 or 15 years ago, it was seen as verifiable proof that what you saw filmed from these cameras was what actually happened.

I don't think we can make that assumption anymore. I think we are no longer in the era of video being proof that something happened, or pictures or any other form of media. I think we have entered this kind of postmodern state where things need to be interpreted in context, and there is no one canonical source of truth.

That is true. And yet, at the same time, I have a more optimistic view of where we are, at least in this moment, because this particular shooting, the Alex Pretti shooting, was filmed from many different angles. Journalists put a lot of work into verifying the videos that were posted. Interestingly, as they spread on social media, there was not a big movement to say, “These videos are fake.” There just wasn't.

And in fact, I think the video of Pretti's killing shocked the conscience of many Americans, some of whom had previously been fully aligned with the Trump administration in these anti-immigration raids, right? We saw the NRA criticize the administration's portrayal of Alex Pretti's killing, which I think really shocked a lot of people.

In a very grim moment, the silver lining for me was that Americans do still trust the video that they are seeing, including the video being shared on social media, and they are not immediately leaping to, “I am being lied to. This is just manufactured evidence from partisans.” Now, of course, the question is how long that state of affairs will hold as AI tools improve. Does that bargain break?

But I don't know. There's something happening here where, if enough people see it, if the video is captured from enough different angles, and if the administration isn't willing to just lie at an ever-greater scale and manufacture ever more evidence to support its version, maybe the truth survives here. Yeah. Well, that's a hopeful note to end this discussion on, and I'm glad we talked about this.

Casey, did you watch Alex Honnold's free solo of Taipei 101 on Netflix?

I was able to watch about 60 seconds of it, and my palms started sweating so much I thought I was in “8 Mile.”

Yes, it was very nerve-racking. But speaking of extreme risk-takers, daredevils, people who take their lives into their own hands, I hear you have actually been experimenting with some new technology this week.

5. Moltbot Enters the Danger Zone

I have, Kevin. I have installed Moltbot, formerly known as Clawdbot, onto my laptop, and I've said, “Damn the consequences.”

Yes. And this is kind of the AI version of a free solo up Taipei 101. My understanding is that this is insane behavior. So please tell me about Clawdbot, which we should say is not affiliated with the Anthropic Claude chatbot. It is spelled—or was spelled—C-L-A-W-D.

So, Clawd, like a lobster, is clawed.

Yes. And it has since changed its name to Moltbot, I guess because of some copyright concerns.

Well, when lobsters shed their shells, they're molting. And so now the former Clawdbot is Moltbot. This thing was the talk of the town over the past weekend. I saw many people who work in tech in San Francisco becoming obsessed with this tool, talking about how it was running their entire lives.

I'm glad that you experimented with it so that I didn't have to, and now you can tell me about it. But let's do our disclosures real quick, because we are going to talk about Anthropic and other companies in this.

My boyfriend works at Anthropic.

And I work for The New York Times, which is suing OpenAI, Microsoft, and Perplexity over alleged copyright violations.

Yeah. Well, I first saw Clawdbot in a post on MacStories by the great Federico Viticci. Federico, like me, is always ready at the drop of a hat to make some new piece of software his entire personality. Clawdbot looked like one of those things.

It is a free, open-source personal AI agent that you can run on your computer, plug into various systems, services, and AI tools, and, if all goes according to plan, have something resembling a little genie inside your computer who can work for you.

So, Casey, who created this bot?

Clawdbot, now Moltbot, was created by a developer named Peter Steinberger. He has worked on many cool, hacky, tinkery projects. You can go to his GitHub and see dozens of projects that he has worked on that are all in this do-it-yourself vein.

He started Clawdbot to scratch his own itch. Basically, he observed that there was no great personal assistant agent yet on the market, and he wondered if he could figure out a way to use Claude Code through WhatsApp. That's how the whole project began.

Okay. So you saw this and thought, “I should try this.” How easy or hard was it to set up?

I would say the setup is pretty easy. If you've installed Claude Code, it will feel quite familiar. You go to the Moltbot website, grab a little one-line piece of code, put it in your terminal, and it will begin the setup process.

I saw a lot of people over the weekend talking about how they were buying Mac Minis to install this thing, sort of giving it its own computer. What is the deal with that? Why are people doing this on separate computers rather than their main computer?

Yeah. Well, like when you free solo a giant skyscraper, Kevin, there is some risk involved with installing an AI agent onto your main machine. A few of the risks I would name are that somebody could just hack into your computer. Clawdbot can connect to a messaging app like Telegram. If somebody hacks into your Telegram account, theoretically, they could take over your entire computer and do anything on it that you could.

Yeah. So we should just say: Don't do this. Do as Casey says, not as he does. This is a very bad idea, and you like to live on the edge a little bit. What were you thinking?

Yeah. Well, what I was thinking was, first of all, I have disabled the Telegram integration. So I have gotten rid of that particular attack surface, and I'm also being careful.

How are you buying your drugs?

No comment. So, yeah. Anyway, that's one of the risks. You have the prompt-injection attack risk, which is that you will visit a website that has some hidden instructions buried in it, and the agent will see them and follow those instructions. That's very hard to protect against.

Basically, if you're going to use something like this, you would be better off putting it in what they call a sandbox, some sort of contained environment where it only has access to a narrow range of tools and nothing that has access to, for example, your bank account. So that's the right way to do it. I tried to do it in a way that is undeniably risky, but I've tried to take the precautions I can.

6. Moltbot Builds a Daily Briefing

I'm only connecting it to services that are not life-or-death for me. As I'm sure we'll get into, the main thing that I've been doing with Moltbot is trying to build a personal daily briefing for myself. I do not have it working for me as a full-time AI employee that is out browsing the web and doing a bunch of errands on my behalf every day.

Right, which is the kind of thing that I saw people experimenting with, or at least posting about, this weekend. We've been talking a lot about agentic AI systems. How is this different? Is it just that it can connect to stuff that Claude Code, Codex, or all these other coding tools can't?

The first thing to say is that, yes, this is actually a lot like Claude Code or maybe Codex from OpenAI. You can get a lot of the same benefits out of one of those tools as you can with Clawdbot.

Clawdbot, though, tries to offer a few things to distinguish it. No. 1, it runs locally on your computer. That's important to a lot of people. There are actually ways to restrict the amount of data that it is sending back into the cloud. No. 2, it has a better memory system than Claude Code currently does.

In Claude Code, you're always trying to cram everything into a context window, and often you'll run out of context window and it has to compact the conversation. You often feel like you're starting from scratch, and Claude can't remember what you did yesterday.

What Moltbot does is just write memories to a Markdown file and continuously revisit it. In my experience, it has been a little bit better at understanding. For example, if you built a tool with it the previous day and want to make a tweak to it, it knows where to look to find that project and edit it, which sounds very simple, but it's actually something that Claude Code is very bad at and has frustrated me constantly.

Like if North Korean state hackers are infiltrating your machine and stealing your money, it will sort of leave them a nice little Markdown file that they can return to the next time they want to steal from you.

That’s right. It’ll say, “Here’s exactly how to break into the vault.” Okay, yeah.

Okay, good. You’re really reassuring me here. So, give me a sense of what Moltbot is doing for you. Give me 1 example of something that it has done that other AI tools have not been able to do.

Well, yeah. So, again, danger zone, but I wired it up to my email and my calendar.

Oh my God. And I thought—I didn’t know I was in the presence of Evel Knievel. You are truly doing the most out here. What happened to you?

I’m always on the hunt for something interesting to write about, is the honest answer to that question, and I want to see if maybe this would be one of those things.

As 1 of my mentors used to say, it’ll either be a good time or a good story.

There you go. There you go. And I would say this is probably more in the realm of a good story than a good time.

Yes. But I connected it to my email, my calendar, and a few other services, and I thought it would be great if, in the morning, when I visited my Clawdbot and just wrote “good morning,” I would get a nice briefing that would have things like: Here’s the weather. Here are some important emails you should answer. Here are the calendar events that you have going on today.

And then I thought, okay, now I want to make it extremely specific to me, right? Because this is maybe where AI can help me get beyond the very basic stuff, like email and calendar. So I said, “I want you to tell me whenever there is a new pro wrestling pay-per-view on TV. I want you to tell me when there’s a new episode of RuPaul’s Drag Race. Every Thursday, I want you to tell me what movies are coming out this week, who stars in them, and give me a 1-sentence synopsis.”

And so I just started typing into the box. You’re saying, “Here’s what I think the best daily briefing would look like for me.” By hook and by crook, over a handful of days, I got something, Kevin, that works about 70% of the time.

This is why you can’t buy RAM, because you need your RuPaul’s Drag Race briefing.

Yeah, it’s definitely me and not the 800 data centers that are being built all over the world.

I have not seen this in action yet. Can you just show me, or tell me how you actually use this thing?

Yeah, I can show it to you. Let’s see. I’ll open up my Clawdbot here, and this is running in—you’re going to see this in a browser window, but this is running locally. So, let’s see that—

Can I talk to it?

No. I have it trained on your voice not to respond to anything you say.

Okay.

But I can just type “good morning” in here, and it’ll think for a second and then pull up a briefing. But you can see it’s a plain beige chat window. Nothing fancy.

A lot of people are messaging Moltbot through third-party messaging apps. There are ways to hook it up to WhatsApp, Discord, Telegram, and various other messaging apps. That’s very convenient. It means you can easily access it from your phone, but of course then there’s a lot of risk involved.

What can it do? Show me 1 thing it can do.

Yeah. It’s going to pull up the briefing I’ve been putting together. You’re going to see it come up here. It’s going to show you the weather, some emails you might want to look at, and some overdue tasks that you have in your to-do app.

You’re also probably noticing that it’s taking a really long time. Now, I want to say something about this thing that I did. I wanted to be able to get this sort of daily briefing that I’ve described just by typing the words “good morning” into Clawdbot.

I’ve already done this a few times this morning as I was testing and preparing for this segment, so that might be why this broke. But I just want to point out that, in this moment, the thing that I built did not work.

One reason why I was excited to talk about Claude Code on a recent episode was that I felt like this had gotten to a point where, if you told Claude Code to do something, it basically would just work. And I’m telling you, I do not personally believe, in my own experience, that Clawdbot is yet at that level.

I now run into this every day that I use it, where I will tell it to do something and something will break along the way and it won’t quite work. So, what is going on under the hood? Here’s the great part of vibe coding: I have no idea, and I never will.

Yes. This thing is wild, and I think we should just make a little caveat that you should not install this on your machine unless you know what you are doing and are willing to take some risks. Put it on your ex-boyfriend’s machine.

So, I mean, look, here’s the thing. I entered into this thinking this might be something really fantastic that could be a new way that I use my computer, and over the first few days of using it, my expectations have been managed, I would say, very far downward.

The security risks, as you mentioned, are not at all resolved. This is not a safe tool to use for a lot of things. But honestly, in some ways, just as important to me, it does not enable that much new stuff, right?

You can spend a lot of time hacking around, wiring up various services, having that feeling that you’re being productive, and then when you look back on it a day later, you’re like, “Okay, well, I have a very complicated new widget that’s telling me the weather.”

So why are people so worked up about this? The way that people were talking about Clawdbot, Moltbot, over the weekend, it was like the rapture had happened and Jesus had come back. So why are people so excited about this?

So here’s my best guess. I think that Clawdbot is a very compelling vision of the future, right? We have lived in a world where, to do anything, basically since the dawn of personal computing, you’ve needed a dedicated piece of software that had a discrete set of capabilities.

And Clawdbot and AI come along and they start to chip away at that vision of the world, and they say, “What if instead of having a bunch of apps on your computer, Kevin, there was just a genie who lived inside your computer, and every time you had a wish, you could go to the genie and say, ‘Genie, I wish for you to make me a website,’ and then, abracadabra, to quote Lady Gaga, it exists?”

That is something that is really interesting. And while it is only barely beginning to become possible, I can imagine what a good version of Siri would look like in the year 2036.

Yeah. I mean, yes, it may arrive a little sooner than that. We’ll see.

But that was my first thought: This is actually the thing that Apple said they were building into Siri. It seems like this 1 random developer has actually gone out and built it. Maybe it doesn’t work all the time. Maybe it’s got all these risks, but it does feel like people are figuring out how to stitch these tools together.

I’ll show you—or tell you about—1 example of an interaction with Moltbot that got me thinking somewhat excitedly and also nervously about the future. This is a user named Alex Finn who posted this screenshot of his Moltbot, and he says, “Can you make a reservation for me at this restaurant?”

Then the bot says, “I can’t book directly through OpenTable right now, but if you want me to just call them, I can call and book under your name.” And the bot actually did use a synthetic voice from ElevenLabs to call the restaurant and place this reservation.

So that, to me, was—I have not verified that this example works. I have not tried this on my own machine. I will not be installing this bot, because I value my security and my data. But this is the kind of thing where, if it does work in a sustained way, this would actually be a very useful tool for a lot of people.

Yeah. And this is honestly just what it is often like to cover technology: You are writing about stuff that is directionally correct, way too early, and barely works.

And I think a lot of people’s relationship to stuff like that is they don’t want to know about it because it’s irrelevant to them. But I think for you and me, it’s really important to pay attention to that frontier because it’s been our observation, particularly in the AI era, that stuff that barely works in January might be pretty freaking good by November.

So that’s why I have my eyes on this one. I think you’re seeing it starting to capture people’s imaginations and say, “Well, what if there was a safe version of this? What if this was just natively part of the operating system? What if I could trust it with all of my personal information? Maybe I would never have to visit OpenTable ever again, right? And what a world that would be.”

7. The AI Adoption Divide

Yeah. Casey, I want to talk a little bit about this tweet that you got a lot of heat for over the weekend.

A heat tweet. I was watching all this happen, with everyone experimenting and tinkering with Clawdbot, Moltbot, and I started feeling like the gap we are seeing between the people who are early adopters on this stuff is really diverging from the people who are more cautious, who are more skeptical, who really are only using chatbots in the way people were using them in 2023.

People got very mad at me for making this. They thought I was endorsing this early-adoption mindset where you turn over all of your decision-making to an AI chatbot and you let these Claude swarms run your life.

And I guess I’m curious whether you think I was right there or wrong. Do you think there is something notable going on here?

Well, let’s go to the text. Listen, if you ever want to know what’s a gift you can give any man, ask him to read one of his tweets out loud. So, Kevin, why don’t you read us this tweet?

Okay. I wrote, “I follow AI adoption pretty closely, and I’ve never seen such a yawning inside-outside gap. People in San Francisco are putting multi-agent Claude swarms in charge of their lives, consulting chatbots before every decision, wireheading to a degree only sci-fi writers dared to imagine. People elsewhere are still trying to get approval to use Copilot in Teams, if they’re using AI at all.”

The backstory here is that I had just been meeting with some groups about their AI policy. These are schools, nonprofits, people who want to know: How should we be using this stuff? And I’ve been struck recently that so much of what people are struggling with is the same thing they were struggling with a couple of years ago.

It’s like, what is my enterprise IT policy about letting people use AI? Should people be allowed to use AI to take notes in meetings or write performance reviews? It’s these institutional rules for the last generation of AI products. And what’s really been striking to me is how little people seem to believe that the future and the past are going to look very different from each other.

We are heading into something new and strange, something that requires us to shift the way we think about these things. So I think that what you’re hitting on is a dynamic that has long been predicted among those who say that the diffusion of AI technologies will be slower than the accelerationists think, because it will just run into a lot of institutional roadblocks and bottlenecks along the way.

Here in the real world, we do have IT policies. We do have companies that aren’t going to let you install whatever AI tool you want on your laptop. And we have a startup ecosystem full of wireheads that doesn’t have any IT policies. The minute that some crazy new insecure open-source project comes out, they put it on their laptop to see what they can do.

And so that dynamic, I think, is going to persist. And I think the question is, is there any real alpha to be gained by these wireheads who have access to Clawdbot, who can try any new AI tools, and who can lean all the way in? Or do they wind up just spinning their wheels a lot, running into a bunch of security problems, getting their GitHub accounts compromised, and having phony crypto coins associated with their names?

There’s going to be some variation. People are going to have different outcomes, but I think that’s what we want to watch.

Yeah. And I’m worried about this sort of new axis of polarization that I think is starting to happen right now. It’s like, do you think this technology is real and important, or do you think it is fake and overhyped? And I think we are starting to see the results of that polarization, where the minute something new comes out, you have thousands of people rushing to try it on GitHub and install it on their machines. And you have other people saying, “Stop this train. You’re all huge nerds and losers, and we should focus on the problems that people really have.”

Right now, I think it’s debatable whether these tools are actually supercharging people outside of maybe coding, which I think is its own special domain. Maybe this is all productivity theater, and people are going to regret installing these things and go back to using Google Calendar, the Notes app, and whatever else they’re using, like the rest of us.

But I worry that in the world where this stuff does actually continue to improve, where it’s not just productivity theater, the people who are using things like whatever the 17th version of Clawdbot that actually works are may be way more productive than the people who are still doing things by hand. In the same way that I think if you’re a coder today who writes your code by hand and doesn’t use any of these AI-assisted coding tools, you’re actually just much less competitive and much less well-suited to succeed than the people who are using the tools.

Maybe I think that if these tools wind up being as productive as you are suggesting, they will sell themselves to the mainstream, right? You don’t have to convince anyone to get an email account or buy a smartphone. They know that’s going to be helpful in their lives.

I think the concern that you’re seeing among people who are more skeptical—and part of it is just reflexive skepticism and maybe a resistance to change—but part of it is that it’s actually not clear to me that this is going to help me in my life. Maybe this is just a tool being built by people who do not have my best interests at heart, who are mostly just trying to take my job away for their own purposes.

To the extent that these tools can help empower people, help them make money, help them get and keep their jobs, and help them find work that they like to do, I think you’re going to see very enthusiastic adoption. But if it just remains, “This is getting better and better at doing the thing that I’m currently being paid to do,” you better believe we’re going to continue to see a lot of resistance.

Yeah. I’m not saying that people should be excited or enthusiastic or adopt every new tool the minute it’s out. I think there are very good reasons to be worried about and skeptical of some of the hype that’s out there.

At the same time, I do want people to understand what the tools are capable of doing. Right now, we have a situation where if you are at a big company or a more traditional organization, you are probably not able to use this stuff, at least not at work, on your work devices. If people are experimenting with this stuff and deciding this is not for me, that feels like a totally fine outcome to me.

But if people aren’t even aware that these tools are out there, or that people who are using them are getting some benefits from them, I just worry that they’re going to get left behind. And I was thinking about this in part because of something else that happened this weekend, which is that Andrej Karpathy, the legendary AI researcher, came down off Mount Olympus to give one of his regular pronouncements about the state of these AI coding tools.

And he said that programs like Claude Code and Codex are “easily the biggest change to my basic work coding workflow in 2 decades of programming,” and that it happened over the course of a few weeks. This is one of the most accomplished programmers in the world, someone who has built and maintained huge, complex systems, saying that he is trying to adjust to this new era.

His job—the old job, the job that he spent his whole career doing—has essentially been taken over. And if that is true for him, it is also true for other programmers. And if that is true for programmers, it is probably true for people in other areas, or will be soon.

Yeah. Well, the good news is everyone on the Hard Fork show is now fully aware of all of this. So if you listen to the show, you’re caught up. And if you don’t, once again, you’re out of luck.

Yep. That’s the moral of the story.

Mhm.

8. HatGPT Draws the News

Casey, it’s time for HatGPT. HatGPT, of course, is the game we play where we put a bunch of the week’s top stories into a hat, draw them at random, and discuss them until the other person gets bored and says, “Stop generating.”

Should we use our Hard Fork hat?

Please produce the Hard Fork hat. This is an official piece of Hard Fork merchandise.

Very nice.

Now, is this still available for people to buy?

I don’t think so.

Good. I want to gatekeep it. You can’t have it. But there’s a different Hard Fork hat that you can buy at The New York Times shop.

Oh, well, there you go.

Yeah. All proceeds go to supporting journalism. It’s about time somebody supported journalism. Okay, here we go. You’re up first.

All right. Well, Kevin, this story comes to us from Business Insider: Some Amazon employees got a calendar invitation titled “Project Dawn” discussing upcoming job cuts.

On Wednesday, Amazon announced 16,000 layoffs. The day before that, Amazon seemed to mistakenly send employees a calendar invite about those very layoffs. It was for an event at 5:00 a.m. Pacific time on Wednesday, but was canceled shortly after it was sent. And so, the layoffs were called Project Dawn.

Yeah, that’s bad. We shouldn’t do that. I’m pretty sure I’ve played a sci-fi video game where Project Dawn was the name of an effort to wipe out all of humanity.

God, just say Project Layoffs, Project Hatchet, Project Costing—

Project Bye-Bye—

Project Efficiency.

Much better names. Let’s not call this something that sounds like a science-fiction movie.

Now, here’s what I hope. I hope that reporters stay on the case and figure out which AI tool is responsible for that calendar invitation being sent. They’re deploying Clawdbot, and it’s sending people random notices about their layoffs. Okay, stop generating.

This one comes to us from The Guardian. Former FTX crypto executive Caroline Ellison has been released from federal custody. She served about 14 months for her involvement in the multibillion-dollar FTX fraud scandal.

She was originally sentenced to 2 years. Her release comes as the whole FTX saga is set to be turned into a Netflix series this fall called “The Altruists,” starring Julia Garner and Anthony Boyle.

Here’s what I will say about the icon Caroline Ellison: Kevin, never have I been more confident in my life that someone was about to start a Substack. So, we want to wish Caroline well on her upcoming journey into content creation. I would be very excited to read her Substack because she is actually a good writer.

I enjoyed the Tumblr posts of hers that I read during the sort of conflagration at FTX, and I think we should officially extend Caroline Ellison the invite to come on Hard Fork.

Absolutely. In the gay community, we love a problematic queen, and we believe everyone deserves a second chance. So, Caroline, get in touch. We’d love that.

Now, this one really has been the talk of the town, Kevin. A TikTok data center outage triggered a trust crisis for its new U.S. owners. This comes to us from Wired.

Over the past week, ByteDance officially transferred control of its American business to a new subsidiary with a group of majority-American investors, and things immediately went haywire. We’ve seen celebrity after celebrity, along with a bunch of politicians, raising claims of censorship, saying that as they have tried to post in protest of ICE, their posts are getting zero views or not appearing at all.

Gavin Newsom, the governor of California, has said he is launching an investigation because he tried to send the word “Epstein” in a direct message and it did not appear, which was something that other users reported as well. So, what do we think is going on with this new TikTok?

I have tuned out most TikTok-related news because I think I made the logical conclusion that nothing was ever going to happen to TikTok and just said, “Wake me up when it happens.”

But I’m awake now.

Wake up, sheeple. There are 2 things that have happened. One, TikTok was sold. And then the second thing is this theory that these new owners might be trying to put their thumbs on the scale and suppress certain content they don’t like.

I don’t know if this is happening or not. I always go back to the line, “Never attribute to malice what could be attributed to stupidity.” I think probably during the transfer, some stuff got shuffled and some cords got pulled. But do you think there’s anything real here?

I would be very surprised if there is. It pains me to say this because this controversy has roped in some of my favorite celebrities. Billie Eilish and Meg Stalter are out there posting about this. Meg Stalter said she’s going to delete her TikTok account.

I would just urge a little bit of calm here. If you’ve ever moved apartments, you know that some of the dishware gets shattered in the move. I think something like that is happening here. We know there was an outage at one of the company’s data centers.

We’ve seen stories in the past where, for some limited period of time, YouTube videos were showing zero views. It didn’t actually mean no one was watching them; it meant that the view counter was broken. So, I’m not saying trust TikTok U.S. forever. I am saying that in this particular case, which did admittedly happen at a very bad time for them, it probably was some sort of technical bug.

Yeah, they installed Moltbot on their servers, and it’s just been ordering stuff.

Okay, next up we have “Anthropic CEO’s Grave Warning: AI Will Test Us as a Species.” This comes to us from Axios, and it is about a new 38-page, 19,000-word essay by Dario Amodei, the CEO of Anthropic, titled “The Adolescence of Technology.”

This is a follow-up to his hit essay “Machines of Loving Grace,” which we talked about with him on the show. This one is kind of the flip side of that. That one was all about how AI might create this amazing acceleration of scientific progress—all the optimistic things. This is Dario saying, “Let’s not get too excited here,” because there are also all these scary things that could happen as a result of AI.

Yeah, I think he should have titled it “Machines of Violent Domination,” since that seemed to be what most of the essay was about. Of course, we always love to read essays like this from the CEOs of the companies that are actually building the Torment Nexus. It’s like we’re all trying to find the guy who built the country of geniuses in a data center.

I think his name was Dario something. Anyway, we’ll look that up later. Might be wearing a hot dog costume.

What I don’t understand is how these people have time to write. I would find that writing a 19,000-word essay would take me several weeks. I don’t run a large, multibillion-dollar AI company, so whatever kind of productivity-enhancing, focus-enhancing drugs Dario is on, I would like to get some of those.

On a bit more serious note, I feel like the last essay got a lot more attention.

No one seems to want to read about any of the potential downsides of AI.

Well, I think there are a couple things here. One is that it was somewhat out of character for Dario to put out an optimistic essay about AI. He’s known within the AI community as something of a worrier about the risks. He’s been writing papers about risks from AI for many years.

Anthropic sort of billed itself as the AI safety company. So, I think it just was a little more notable when he was saying, “Well, actually, people think I’m this doomer, but I’m actually quite excited about this technology.”

I think this latest essay is just more in line with what he’s been saying all along, which is that these systems are getting better quite quickly, they present all these new dangers, and we don’t know where it’s all heading.

Here’s an essay that would have gotten more attention: “The 5 Worst Decisions Made by My Rivals.” Something to think about.

All right. All right. You hate to see this one, Kevin. An app for quitting porn leaked users’ masturbation habits. This comes from 404 Media.

The name of that app? Salesforce. The name of this app is being withheld to protect the privacy of the masturbator, Kevin. This was an app that purported to help people stop consuming pornography, and it exposed highly sensitive data. Some of the data included users’ ages, how often they masturbate, and how viewing pornography makes them feel.

I have to say, this is a sticky situation, and I think it’s gotten entirely out of hand.

What? You don’t agree?

I wondered why you wanted this one in the segment, and now I know it was to make that joke.

I didn’t even put it in here. But here’s what I will say: I’m serious. I’m calling for a full investigation because I think it’s important that, at the end of this, we are able to finger the culprit.

What? Why are you laughing? This is serious.

These people had their data leaked.

All right. Okay, this one’s a doozy, Casey. This one comes to us from Dig Watch, and it is titled “Alaska Student Arrested After Eating AI-Generated Art in Protest.”

On January 13, Graham Granger, a film and performing arts major at the University of Alaska Fairbanks, was arrested and charged with criminal mischief after ripping AI-assisted artwork from a campus gallery wall and eating around 57 of the images as part of what he describes as a protest and performance piece against the use of AI in art.

Now, I’ve heard of consuming content, but this is taking it too far. I love this so much. Move over, Banksy. I maybe will reflect further on this and not feel this way, but here’s how I feel right now: I think it’s a great protest.

Honestly, no, because this was another student’s artwork.

Now, let me be clear: I don’t actually know anything about this art or how it was made. But what I can tell you is that people often save a backup to their desktop. So maybe that could come into play here.

Maybe it could, maybe it couldn’t. Maybe this was great art. We’ll never know because now it’s sitting in Graham’s intestines.

And it was delicious.

They make them different up there in Alaska, Kevin.

All right. Oh, now, here’s a story that I think we all saw coming, and now it’s here and we’re like, “Oh, finally.” Steak ’n Shake has added $5 million in Bitcoin exposure, deepening its commitment to Bitcoin. You might be surprised to learn that this one comes from Bitcoin Magazine.

Steak ’n Shake, which is one of the greatest triumphs of the Midwest, I think we could say, is a place where you can get a delicious hamburger and a milkshake. It’s continuing what it calls its burger-to-Bitcoin transformation.

In a post on X, the restaurant chain wrote that it had increased its Bitcoin exposure by $5 million in notional value, or as it’s sometimes called, value, and that all Bitcoin sales go into our strategic Bitcoin reserve. So, Kevin, anything you could tell us about what you think Steak ’n Shake should do with its Bitcoin holdings?

I think they should expand to the West Coast, right? Obviously, In-N-Out has a monopoly on that kind of experience out here, but I grew up eating Steak ’n Shake in my Midwestern town, and it slaps. It’s really good. They’ve got a good steakburger.

Unfortunately, many years ago now, the chain was acquired by a private equity guy named Sardar Biglari, who has made some very strange decisions, including putting his own picture up at every Steak ’n Shake.

Oh my God. Like a tinpot chairman now.

Yes. He’s opening versions of Steak ’n Shake in European cities. There was one in Cannes the last time I was there. I was like, “What’s a Steak ’n Shake doing in Cannes?” This guy, for some context, was also the owner of Maxim magazine. This is part of his lifestyle brand.

And now he’s become a big crypto guy and wants to increase Steak ’n Shake’s profits by investing some of its money in Bitcoin.

Well, let me say that I hope that someday we get a big sorry from Mr. Biglari.

This one comes to us from The Information. Apple is developing an AI-powered wearable pin the size of the AirTag that is equipped with multiple cameras, a speaker, microphones, and wireless charging, according to people with direct knowledge of the project. The device, this report claims, could be released as early as 2027. I think this is really interesting because this shows Apple pursuing what I like to call its Vision Pro approach to hardware, which is making things that cost $3,500 and you’re not sure what they do—

—but they are available. This is total vindication for the Humane Pin.

Really is.

They died too young.

Are they going to be upset that they let the Humane AI Pin go to work at HP and didn’t snap it up when they had the chance?

No, I think they’ll be okay.

But I think they are trying to compete with OpenAI, which is also reportedly developing some sort of pin-like hardware thing. And I, for one, cannot wait to buy and then immediately lose this.

Yeah. Well, look, I’m just excited that AirTags are going to have a second use beyond illegally stalking your ex.

You’ve got to stop doing that.

Yeah.

Now that you want to talk about the social event of the season, Kevin, let’s go inside the White House’s screening for Amazon’s “Melania” doc. This report comes to us from The Hollywood Reporter. On Saturday, there was a black-tie event that was not promoted or advertised, but it took place in the East Room of the White House and attracted about 70 VIP guests, including Queen Rania of Jordan, Zoom CEO Eric Yuan, Apple CEO Tim Cook, the CEO of the New York Stock Exchange, the CEO of AMD, Lisa Su, and Mike Tyson. Kevin, is that giving nightmare blunt rotation or what?

Yes. I’m sure there are many perks associated with sucking up to a sitting president and groveling at his feet. One of the downsides is that you have to go to the White House to dress up in black tie and sit through a screening of what I can only imagine is a terrible documentary. And you probably can’t even look at your phone while you’re doing it.

You really can’t. I have still not seen any reviews. I guess Tim Cook is still working on his Letterboxd post, so keep an eye out for that.

But here’s what I’ll say. When I think of all of those people, all of these power players in one place, it has all the makings of a great reality TV show. But unfortunately, we already have one called “The Traitors,” so we’ll have to come up with something new.

Hope the screening was good. SpaceX is weighing a June IPO, Kevin, that is being timed to a planetary alignment and Elon Musk’s birthday. This comes to us from the Financial Times. The rocket maker is targeting mid-June for its IPO, when Jupiter and Venus will appear very close together, known as a conjunction, for the first time in more than 3 years.

Now, I heard that boys go to Jupiter to get more stupider. Is that a factor in this as well?

I’m honestly surprised he didn’t plan some sort of Uranus joke here.

Honestly, don’t give him any ideas. Edit that out of the show.

We don’t want him to hear that.

So, what stage of capitalism is it when the world’s richest man starts factoring planetary alignment and his own birthday into his IPO plans?

It’s just so dumb. I’m sure the SpaceX IPO will be a huge whatever. It’s a company. They make lots of money. I’m sure investors will be excited about it. Can we just stop doing the “4/20/69, Jupiter’s in retrograde” thing, lol? It just feels like we should have left that one behind in 2013.

Yeah, your IPO is not a Tumblr post.

Knock it off.

Last one. Last but not least, LinkedIn will let you show off your vibe-coding expertise. This one comes to us from Engadget. LinkedIn is partnering with Replit, Lovable, Descript, and Relay.app on the feature and is working on integrations with fellow Microsoft-owned GitHub as well as Zapier. LinkedIn is allowing the companies behind the AI tools to assess an individual’s relative skill and assign a level of proficiency that goes directly to their profile. We’re getting vibe-coding badges on LinkedIn. Casey, what do you think?

I have to give it up to LinkedIn because every year they invent levels of B.S. that have never been seen before in corporate life. The idea that there’s a badge on your profile that says you can type in a box—there might as well be a little trophy you can get for having computer access for this. What are we even doing here?

Yeah, LinkedIn, I think, is the sneaky first all-AI social network. I recently went on there. God knows why. I think I had blocked access to my other social networking apps, and every post on my feed was clearly generated by ChatGPT. It was like, “This isn’t just a transformation, it’s a revolution.” The vibes over there are extremely strange. This is a little bit unrelated, but am I right that anyone who asks to connect with you on LinkedIn, you accept the connection?

Correct.

I do not.

No, because when I grew up, you would connect with people on LinkedIn because you knew who they were.

Oh, no.

And it seems like that norm has just collapsed.

Well, I don’t really use it, but I like having a big network for unspecified reasons.

Is it weird to you when you’re just seeing a feed of utterly random posts from people you’ve never heard of?

No. It’s fine.

Do you ever congratulate people for getting a new job when you have no idea who they are?

I don’t actually type that much into LinkedIn. But I will say that one of my tests that I give to new AI models is: can they autonomously respond to my LinkedIn messages for me?

That’s very good.

Listen, just to wrap this up, here’s what I’m going to say. If you find yourself tempted to put a vibe-coding badge on your LinkedIn profile, here’s what you can do instead: put on a rainbow wig and some white makeup and a big red nose, and look in the mirror. There’s your vibe-coding badge. Congratulations.

Congratulations.

All right, Casey. That’s Hat GPT. That’s Hat GPT.