TikTok's Last Stand + Google's Quantum Breakthrough + Cult of Claude

“The president was given a very clear and easy opportunity to say, ‘Yes, I am going to save TikTok,’ and he absolutely did not say that.”

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Your agents can talk to each other. They still can't think together. That's the gap Outshift by Cisco is closing. Outshift is Cisco's incubation engine, building the internet of cognition. Shared intent, shared memory, and guardrails that allow agents to work together from anywhere. Open source protocols and code. A system for intelligence that scales. Read the paper and experience the demo at outshift.com. That's O-U-T-S-H-I-F-T dot com.

• 0:32Kevin Roose?

  Casey, how are you?

• 0:34Casey Newton?

  Doing great, Kevin. I'm in full holiday cheer this evening. I had a chance to go to a holiday party with you last night even.

• 0:41Kevin Roose?

  That's true. Yes, this is the holiday party season where tech companies all invite reporters to nosh and drink eggnog and whatever. And it's a good time.

• 0:50Casey Newton?

  Yeah, it's a time where we sort of lay down our arms and we try to find common ground and scheme to get high-profile interviews in the coming year.

• 1:00Kevin Roose?

  Yeah.

• 1:00Casey Newton?

  Yeah. And I think we might have been successful.

• 1:02Kevin Roose?

  I think so, too.

• 1:03Casey Newton?

  Yeah, yeah.

• 1:04Kevin Roose?

  More coming on that suit. But we also got our caricature drawn at this holiday party, which was very fun.

• 1:10Casey Newton?

  If you've never been to a tech PR holiday party, they're structured basically like bar mitzvahs, where there are different stations for having your photo taken, caricature drawn, balloon animals. And we did get a caricature. And it was the first time we've ever sat for a portrait. A drawn portrait, I guess I should say.

• 1:27Kevin Roose?

  And in the nude. We were totally nude.

• 1:29Casey Newton?

  And let's just say, the world might not be ready to see what we have to offer.

• 1:34Kevin Roose?

  That's a joke. But Casey, show the people our caricature.

• 1:37Casey Newton?

  I'm going to show the caricature. And of course, if you're listening to the podcast, you may want to go to youtube.com/hardford just so that you can see this caricature and have the same reaction that my boyfriend did, which is, those guys look much hotter than you do.

• 1:52Kevin Roose?

  So. But usually caricature artists, they try to roast you a little bit, but this one was incredibly flattering.

• 1:58Casey Newton?

  Yeah, it's like, I mean.

• 1:59Kevin Roose?

  The chiseled jawlines on this caricature are really something.

• 2:03Casey Newton?

  Yeah, it is nice to have a caricature artist who's like, I can actually fix you. That's unusual, but I think it was a really nice sort of holiday spirited thing to do.

• 2:17Kevin Roose?

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

• 2:19Casey Newton?

  I'm Casey Noon from Platformer.

• 2:21Kevin Roose?

  And this is Hard Fork.

• 2:22Casey Newton?

  This week, TikTok may be entering its final days. For real this time. Then, Google's Julian Kelly joins us in the studio to describe the company's big breakthrough in quantum computing. And finally, Kevin investigates the cult of Claude.

• 2:36Kevin Roose?

  They're doing ritual sacrifices at dawn.

• 2:44Casey Newton?

  Well, Kevin, TikTok is back on the clock.

• 2:47Kevin Roose?

  Yes, we had some big news this week about TikTok. A three-judge panel on the DC Circuit of the US Court of Appeals has rejected TikTok's argument. Basically, they were trying to strike down this law that would force its sale.

• 3:02Casey Newton?

  That law, of course, the Protecting Americans From Foreign Adversary Controlled Applications Act, or PFACCA.

• 3:08Kevin Roose?

  PFACCA. So let's back up here a little bit and kind of walk up to this moment. Because I think it's been a pretty surprising turn of events. So the first Trump administration attempted to ban TikTok. That was not successful. Then the Biden administration took another stab at it and actually, Congress actually passed a law, PFACCA, that would force the sale or divestiture of TikTok by ByteDance. That law is later to come into effect in January. But TikTok has been frantically working in court to try to stop this law from going into effect, claiming in effect that it is an abridgement of First Amendment rights. Basically, you can't shut down a platform because there's speech taking place on it that you don't like. That argument has been made in court.

• 3:51Kevin Roose?

  And there was some hope among people who opposed the ban that the courts would sort of intervene and step in to stop this law from going into effect. But as of this week, it looks less likely that that will happen.

• 4:02Casey Newton?

  Yes, the court has ruled. And now TikTok, on Monday, asked that same Court of Appeals for an injunction while TikTok appeals all the way up to the Supreme Court. Now, the Supreme Court, Kevin, does not have to take this case. They could simply choose to let the lower court's decision stand. And if that is the case, then... TikTok either sells or gets banned.

• 4:27Kevin Roose?

  Yeah, so Casey, where do you stand on the question of whether TikTok is likely to exist in its current form in the new year?

• 4:34Casey Newton?

  Well, it seems increasingly unlikely, Kevin. I think at a minimum for TikTok to continue to exist next year, ByteDance is going to have to figure out some sort of sale. And the reason is that when scholars look at the court's decision, they find the court saying, look, we agree that there are very real national security issues here. And in the United States, courts typically defer to Congress on issues of national security. And in the case of PFACA, it was passed by a large bipartisan group in Congress. And so the thinking is the Supreme Court will be very reluctant to step in and challenge Congress on that point.

• 5:17Kevin Roose?

  And what did the judges who evaluated this case in the DC Circuit of the US Court of Appeals actually say when they denied TikTok's sort of attempt to prevent this law from going into effect?

• 5:30Casey Newton?

  Well, so there are two legal questions at the heart of this case, Kevin. One is about the First Amendment. The other is about national security. The First Amendment issue is the one that TikTok pushed really hard. And TikTok said, look, tens of millions of Americans are using this app to produce speech, to post speech on the internet. A lot of that is political speech. And members of Congress actually came out and they said, we don't like the speech on TikTok. We don't like it because it's too pro-Palestine. We don't like it because some of it feels like Chinese propaganda. And TikTok said, aha, Congress is passing a law because it does not like the content of speech on our app that is a very clear violation of the First Amendment. Okay?

• 6:14Casey Newton?

  So that's sort of legal argument one. The other argument which TikTok had to defend against was the national security argument. And here you have the government saying TikTok has incredible access to data of tens of millions of Americans, and we have no effective way to know how that data is being used or to regulate it. Furthermore, Because every consumer internet company in China is ultimately beholden to the Chinese Communist government, there's nothing to stop that government from intervening in TikTok and saying, you know what? We need to turn up all the dials on the Chinese propaganda in TikTok, so that when you open the app, all you see is good things about China and bad things about the United States.

• 6:57Casey Newton?

  Now, very little evidence, and really almost no evidence, was presented in public suggesting that that scenario that I just described had ever been the case. But when the court looked at it, they said, no, we agree with the government. That risk is simply too great, and everything in the national security and data privacy argument trumps everything in the First Amendment argument. And so therefore, we are going to agree that this law should be upheld.

• 7:21Kevin Roose?

  Hmm, and the deadline is fast approaching for this PAFaka forced sale of TikTok. We have about a month until January 19th, when the law would officially go into effect. So what are TikTok's options as of this point?

• 7:37Casey Newton?

  Yeah, well, I imagine that internally, they're having a big PAFAC and forth over what to do. Because as we've said before, Kevin, ByteDance has given no indication that they want to sell. The Chinese government had said that they essentially find it just completely unpalatable that ByteDance would be forced to sell. And so the public position of ByteDance has been, we're not going to sell. Now, tens of billions of dollars are at stake here. And you have to assume that if ByteDance can get 20, 30, 40 billion dollars for spinning out TikTok, it would rather take that than absolutely nothing, which is the scenario that it might be looking at if it is not able to divest. So the question then becomes, who might actually be able to take this thing off of ByteDance's hands?

• 8:20Kevin Roose?

  Yeah, so let's talk about the legal arguments being made here. Because I think a lot of people who opposed the TikTok ban were hopeful that the courts would see the First Amendment argument that TikTok was making as being very compelling and overriding any potential national security risks. But that did not happen. The courts don't appear to have bought that argument from TikTok. So Casey, what did you make? Of the court's reasoning about the First Amendment issue specifically.

• 8:49Casey Newton?

  Well, so in the spring, Kevin, I had talked with legal scholars who told me two essentially equally important things about these sorts of cases. One is that the court takes the First Amendment very, very seriously and provides really strict scrutiny when it comes to cases of political speech in particular. And they noted that there was a lot of political speech on TikTok. The second thing that they noted for me is that First Amendment cases tend to be really unpredictable and that the number one thing that usually sinks First Amendment arguments are usually national security claims. So heading into this decision, I did think that there was a good chance that the First Amendment

• 9:30Casey Newton?

  arguments could carry the day based on some of the precedents that I had looked at and that we've talked about on this show. But in the end, this did turn out to be the sort of First Amendment curve ball that those scholars had warned me about earlier this year.

• 9:43Kevin Roose?

  Yeah. And what was the judge's sort of rationale about that?

• 9:46Casey Newton?

  So one way that the court gets out of the First Amendment argument is by saying, look, as long as ByteDance sells it, there is no speech issue, right? Because if ByteDance offloads TikTok onto someone else, all the speech on the platform remains. So the quote from the ruling that I think is pertinent here is the court writes, "Content on the platform could in principle remain unchanged after divestiture, and people in the United States would remain free to read and share as much propaganda or any other content as they desire on TikTok or any other platform of their choosing, with the act's targets is the PRC's ability to manipulate that content covertly." The PRC, of course, the People's Republic of China.

• 10:28Casey Newton?

  So that is how they're able to sort of get out of it is by saying, this is really about a for sale. Now, when I talked to a First Amendment scholar earlier this year, Evelyn Duhak from Stanford, she had said, look, if in another situation, you can imagine a government of the United States saying, "We don't like the content being published by the New York Times, and we actually think it's a national security threat." But as long as the owners of the New York ...times, sell to someone else of our choosing, we will allow it to continue to publish. Well, you can see how there's an obvious speech chilling effect there, right? So that is why, you know, Evelyn, for example, thought that the court might reject that argument, but in the end that they did not.

• 11:10Kevin Roose?

  So can they appeal further? I mean, you said the Supreme Court might take this up, but presumably that wouldn't happen before this January 19 deadline. So we are like looking at a forced sale of TikTok, or else it will not be available? Is that what we're gonna see?

• 11:25Casey Newton?

  Well, you can imagine some sort of injunction being granted while this plays out. You can imagine Congress passing something to give TikTok an extension. There is some thinking that might be the case, particularly if incoming President Trump decides this is a priority for him. You can imagine, you know, Republicans will now control all three branches of government. They could presumably quickly pass something like that to give these folks a little bit more time. But if not, no, there is a process through which the president becomes able to ban TikTok if it is not sold. Now, that, the mechanics of that might themselves be subject to additional lawsuits. So I expect we will see sort of a lot of legal wrangling up until the very last minute. And I do think it makes exactly

• 12:09Casey Newton?

  what happens next month somewhat hard to predict.

• 12:11Kevin Roose?

  Yeah, but let's talk about the Trump factor in all this, because this is the part that I truly can't get my head around. So Donald Trump comes into office the first time, tries to ban TikTok, the ban doesn't work out. He then has a change of heart, perhaps motivated by some, you know, talks with major ByteDance investor Jeff Yass. We've talked about that on the show before. But during his 2024 campaign, he comes out and says, I will save TikTok. I do not want to ban it. And so people, including me, think, okay, when he wins the election, this is the end of the attempt to ban TikTok, at least for now. TikTok is saved. So what happened?

• 12:48Casey Newton?

  Yeah, well, so on Sunday, Trump went on Meet the Press, and a reporter asked him if he was gonna save TikTok. And Kevin, I think we might actually want to listen to his answer.

• 12:56Kevin Roose?

  Let's do it.

• 12:57Speaker 03

  The federal court upheld a law that could result in TikTok being banned. You said you're gonna rescue TikTok when you get into office. Are you going to take steps to protect it?

• 13:07Speaker 04

  Yeah. And as you know, I used TikTok very successfully in my campaign. I have a man named TikTok Jackie who was very effective, obviously, because I won youth by 30%. All Republicans lose youth. I don't know why. Maybe it's changing. And last time we were down 30% with youth. This time we're up 35% with youth. And I use TikTok, so I can't really, you know, I can't totally hate it. It was very effective. But I will say this. If you do do that, something else is going to come along and take its place. And maybe that's not fair. What they're doing, and really what the judge actually said was that you can't have Chinese companies. In other words, they have the right to ban it if you can prove that Chinese companies own it. That's what the judge answered.

• 13:51Speaker 03

  So are you going to try to protect TikTok just very quick, once you're in office?

• 13:55Speaker 04

  I'm going to try and make it so that other companies don't become an even bigger monopoly.

• 13:59Speaker 03

  Okay.

• 14:00Speaker 04

  Because that's what happens.

• 14:02Kevin Roose?

  So Casey, what did you make of that video?

• 14:05Casey Newton?

  Well, at first, we should say what NBC made of that video, which was that they overlaid a series of factual corrections over a bunch of stuff that President-elect Trump was saying. But look, I think the most important takeaway here was that the president was given a very clear and easy opportunity to say, yes, I am going to save TikTok. And he absolutely did not say that.

• 14:22Kevin Roose?

  Yeah, it's, I mean, it seems like such a straightforward one for him to answer. If he was staying consistent with what he said during the campaign, he could have just said, yes, I will do everything in my power to stop the forced sale of ByteDance and to either, you know, try to convince Congress to pass a new law repealing PAFACA or to do other things. But what can Trump actually do if he decides to try to prevent the forced sale of TikTok? What powers does he have as president?

• 14:49Casey Newton?

  Well, so the way that PAFACA is written, ByteDance has to divest TikTok if they want to keep their US users. But it is up to the president to determine whether it has been divested to his satisfaction. So I read one scholar proposing like, if Trump wanted to save it, one thing he could do is he could go to ByteDance or, you know, through intermediaries or whatever and say, look, why don't you just set up some kind of subsidiary that's maybe somewhat different from ByteDance and divest to that? But maybe ByteDance actually still has majority ownership of it, whatever. And then everything kind of gets to continue operating more or less as it already does. But because PAFACA gives the president the authority to determine,

• 15:32Casey Newton?

  okay, it has been divested to my satisfaction, then TikTok could skate under this sort of, you know, new subsidiary.

• 15:40Kevin Roose?

  Yeah. My other note on this video is that I'm absolutely desperate to know more about TikTok Jack, the character from the Trump campaign.

• 15:48Casey Newton?

  We would love to know more about TikTok.

• 15:49Kevin Roose?

  Please come on the show. And he would probably know what to do about this. But, you know, there are other things that it appears President Trump could do to stop the forced sale of TikTok. He could do things like direct the Justice Department not to sort of penalize ByteDance for failing to sell TikTok by the deadline of January 19th. So it does appear that there are still some things that Trump could do to step in and save TikTok. But as you point out, he had the opportunity to very clearly say that he would do that, and he didn't take it.

• 16:19Casey Newton?

  Yeah. So, you know, again, we will find out what that means. Nothing would surprise me less than if we came back after our Christmas break and President Trump said, "Oh, yeah, I'm going to save TikTok." You know, like, there's been so much back and forth here, Kevin, that it's very hard to predict.

• 16:33Kevin Roose?

  So what are people that you're talking to feeling about this TikTok ban? Because I know you said you've talked to some legal scholars who thought that it would be sort of struck down on First Amendment grounds. Now that that has not happened, or at least that the judges who have ruled on it so far have not defended TikTok on First Amendment grounds, what are they feeling about the strategies going forward?

• 16:55Casey Newton?

  Well, so there is a lot of concern here for this reason, Kevin. Congress knows that courts give them a lot of deference on national security issues. And so if the coming Congress wanted to really crack down on social media companies in America, you can imagine them saying, "Hey, your recommendation algorithms are now a national security threat." Or the kinds of political speech on your platform among Americans are a national security threat. They could pass a law. And if they're able to go to the Supreme Court and just say, "National security, national security," and not even have to show any evidence of harm, because remember, in the TikTok case, there was really no public demonstration that ByteDance had done any of the things that Congress was worried that it might do, well, then all of a sudden,

• 17:39Casey Newton?

  we just have a really easy way for Congress to ban speech online. And that is really scary. Now, what I suspect a lot of our listeners would say, and what people in Congress would say is, you You are forgetting the fact that this is about China. Okay, this is not about an American company. This is about a Chinese company. And so this is a completely different category of discussion. But the First Amendment scholars that I've talked to, they really take no comfort from that because they believe that this is essentially a slippery slope. And once you let people say national security, national security, and don't even have to show any evidence of harm, then we are moving into a world where all of us have fewer speech rights. Kevin, let me ask you. President Trump has been busy filling up his cabinet and naming a bunch of key advisers.

• 18:24Casey Newton?

  What do we know about what they think about TikTok?

• 18:27Kevin Roose?

  So I did some digging into this. It looks like most of the people that Donald Trump has sort of picked for prominent positions in his administration, who have come out talking about TikTok and the threats or lack of threats it presents, are firmly on the side of banning TikTok. So for example, Senator Marco Rubio, who Trump has picked to be Secretary of State, called a TikTok ban a, quote, "win for America" earlier this year. We've also talked about Brendan Carr, who is Donald Trump's pick to lead the Federal Communications Commission. Brendan Carr also advocated for a TikTok ban as part of Project 2025. So it seems like while Donald Trump may have mixed feelings about TikTok, may think that

• 19:09Kevin Roose?

  banning it is, you know, a bad idea, may believe TikTok Jack is sort of, you know, the reason that he got elected and wants to be nice to TikTok as a result, he does have people in his orbit who believe that this app presents a major national security threat and so are in favor of banning it. So it'll just be interesting to see which voices he's listening to and what he decides to do.

• 19:30Casey Newton?

  Well, what I would propose is that once they're, you know, back in the White House, we see an Oval Office debate between Marco Rubio and TikTok Jack. And they just really hash it out and preferably livestream that on TikTok for us all to watch.

• 19:42Kevin Roose?

  I think that'd be good.

• 19:42Casey Newton?

  Yeah.

• 19:43Kevin Roose?

  Let's do it.

• 19:43Casey Newton?

  Let's do it. Well, Casey, when we come back, we're going to do some new bits about Qubits. We'll talk to Julian Kelly from Google about their new quantum computing breakthrough.

• 20:23Speaker 05

  It's time to innovate at the speed of AI without compromising control. That's Outsystems. Take a look at what happened with one global retailer during a period of massive expansion. They used Outsystems to build AI agents to automate invoicing and translation processes, shifting 95% of their daily store operations to a new digital app, saving thousands of hours of manual work. Outsystems gives you the flexibility to deploy how you want and the governance to do it safely. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.

• 20:58Speaker 00

  AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.

• 21:30Speaker 05

  Stripe is built for whatever comes next. With Stripe Billing, you can easily test and iterate subscriptions, launch new usage-based and hybrid pricing models, and automate invoicing. Stripe Billing supports whatever business model you want to launch next, and Stripe powers all the Forbes AI 50 who have monetized. Learn how you can power any business model at stripe.com/billing.

• 21:56Casey Newton?

  Kevin, did you ever see the 2008 James Bond film Quantum of Solus?

• 22:01Kevin Roose?

  I think it was pronounced Solus, but I did see that.

• 22:04Casey Newton?

  Well, either way, today, we can take solace in the fact that you and I are going to learn something about quantum computing.

• 22:10Kevin Roose?

  Yes, so there was some big news this week from Google. They unveiled a new quantum computing chip called Willow. And in making this announcement, they described it as representing two major achievements in quantum computing. Experts are sort of arguing about how big a deal this really is, but some people think it's a, could be a turning point for the entire field of quantum computing and might bring this technology closer to real-world application.

• 22:38Casey Newton?

  So what were those two achievements?

• 22:39Kevin Roose?

  So in the blog post announcing this Willow chip, Google said that Willow can reduce errors exponentially as we scale up using more qubits. That is sort of the, the basic sort of unit of quantum computing. And that this is a key factor in cracking a challenge in quantum computing known as quantum error correction that has bedeviled quantum scientists for almost 30 years. And the second big thing that Willow did, according to Google, is that it performed a massive benchmark computation that would take one of today's fastest classical supercomputers, 10 septillion years, that is 10 to the 25th power years, which is much, much longer than the age of the universe.

• 23:24Kevin Roose?

  It did that computation Willow in under five minutes.

• 23:28Casey Newton?

  They're saying where there's a Willow, there's a way.

• 23:30Kevin Roose?

  That's, that's true. So I am on record on this very show as saying that I have sort of a mild and not very strongly held belief that quantum computing is all fake.

• 23:40Casey Newton?

  Yes, you predicted on a, on a very early episode of this show that this entire field was a fraud. And I thought, that's actually an extremely hot take, one of the hottest takes you've ever shared. But frankly, you know, in the, in the time since you've shared that prediction, there hasn't been a really big reason for us to dig in and try to get our hands around what quantum computing is. But then this news came out and you and I looked at each other and we said, well, maybe this is the moment, maybe this is the moment to bring in not only somebody who is smart about quantum computing, but somebody who is actually good at explaining it to two people who really want to learn, but let's just say, do not have a background in theoretical physics.

• 24:18Kevin Roose?

  Yes. So today on the show, we are going to be talking with Julian Kelly. Julian is the director of hardware at Google Quantum AI. He's one of the leaders of this new Willow chip project. And he's been working on quantum computing since 2008. He started doing research in the field as an undergraduate at UC Santa Barbara. And he's one of the best people in the world to explain what the heck Willow is and why it's such a big deal for people who are interested in quantum computing.

• 24:45Casey Newton?

  Yeah, and we asked him about the more practical applications of quantum computing, which appear to be a relatively long way out. But there is some hope that this technology, among other things, could help us to be much more energy efficient and could let us build other technologies that are perhaps not as bad for the environment as some of the ones we have today.

• 25:05Kevin Roose?

  Yeah, and in addition to responding to my accusation that the field that he's devoted his entire professional life to is a fraud, I was also curious to get his take on how long it will be before we actually start to see stuff materialize that you and I might use.

• 25:19Casey Newton?

  All right. Well, if all goes according to plan here, Kevin, by the end of us, you and I will be able to credibly explain what quantum computing is, what the breakthrough that happened this week was, and what we might expect in the future that speaks to how powerful this stuff could eventually be.

• 25:35Kevin Roose?

  Yep, let's do it. Let's bring in Julian Kelly.

• 25:37Casey Newton?

  Get him in here.

• 25:47Kevin Roose?

  Julian Kelly, welcome to Hard Fork.

• 25:48Speaker 06

  Thank you. Yeah, it's great to be here. Great to be with you guys.

• 25:51Kevin Roose?

  So you work on quantum computing, which is something that I would say Casey and I are both novices at. How often do you find yourself in conversation with someone about what you do? And you just tell them quantum computing and they just kind of stare back at you with a blank face?

• 26:06Speaker 06

  A lot, yeah, I would say. It's fun. Like, I love quantum computing, right? But the funny thing is, is that you kind of have to explain it so much. I kind of, to be honest, maybe avoid it in day-to-day conversation. "What do you do?" I'm like, "I'm a scientist."

• 26:23Casey Newton?

  We similarly avoid telling people that we do a podcast, you know? I think for the same, not because people don't understand it, they just don't want to hear about it.

• 26:32Speaker 06

  Not that they have feedback for you?

• 26:34Casey Newton?

  That's right. I mean, if they do, you know, they should email Kevin. But I want to start by trying to visualize a quantum computer because I've never seen one. And so my first question for you was just, what does this thing look like physically?

• 26:46Speaker 06

  Yeah, so this is great. I'm actually, it's funny, I don't know if we can get this on camera, but you guys have like a motherboard behind you, like a normal computer, right?

• 26:52Casey Newton?

  Yeah.

• 26:52Speaker 06

  And you can kind of think of it as like, in a quantum computer, there's the processor itself, right? And then there's a whole bunch of other things that go around to support it, right? The thing that is hard to visualize is for the quantum computers we build, the processor is sitting at the bottom of what's called a dilution refrigerator.

• 27:08Casey Newton?

  A dilution refrigerator?

• 27:10Speaker 06

  That's right. And so what that means is it's super, super cold. It's similar to a fridge that you have, only it's almost absolute zero, right? So you take your chip, you stick it down there, which makes it behave quantum mechanically. But then you have to stitch all the connections back up to room temperature. So what they look like is there's a chip sitting at the bottom, a bunch of wires that go up to like some racks, kind of conventional looking racks that sit at room temperature. And yeah, that's the kind of like the overall footprint of it.

• 27:36Casey Newton?

  And so like, how big is this thing? Like, once you have the whole package?

• 27:39Speaker 06

  It's kind of like as big as this room, you know, about as tall as a person. A couple of feet wide cylinder is the refrigerator and then a few racks next to it. Maybe the size of a normal fridge that the electronics sit in to run in.

• 27:53Casey Newton?

  Got it.

• 27:53Kevin Roose?

  So Julian, I have so many questions about this and I want to get into what Willow is and what this breakthrough represents, but I don't know much about quantum computing. And so I did, when I was getting ready for this interview, what I often do, which when I feel out of my depth, which is I went to an AI model and asked it to explain quantum computing to an idiot.

• 28:12Speaker 06

  Great.

• 28:13Kevin Roose?

  And I got back something that I want to just gut check with you so that I understand, in case he understands whether we are even in the right ballpark of thinking about quantum computing in the correct way.

• 28:22Speaker 06

  So let's workshop that.

• 28:24Kevin Roose?

  So the thing that this AI model told me about how to think about quantum computing is that in a regular computer, it's like flipping a coin that must land either heads or tails. Every bit that goes through a processor must be either one or zero. And this produces a sort of very clear Definite state in a quantum computer, the AI model told me it is more like a coin that is spinning on a table before it lands.

• 28:49Casey Newton?

  Like an inception.

• 28:50Kevin Roose?

  Yes, and while spinning, it's kind of both heads and tails at once. These things are called qubits and they can exist in multiple states simultaneously until you measure them and that because of this, you can kind of process certain types of problems much faster. Is that even close remotely close to a definition of what quantum computing is?

• 29:09Speaker 06

  Yeah, that's a reasonable kind of starting point to dive into it. So I've been listening to your guys' show and you had this top 100 technologies and I've I've got some feedback.

• 29:18Casey Newton?

  Let's hear it.

• 29:21Kevin Roose?

  You and everyone else.

• 29:22Speaker 06

  Yeah, that's what I hear. But yeah, you don't have qubits on there. You don't have quantum computing in there. So maybe next time.

• 29:27Casey Newton?

  Well, I think if and if I could give you some feedback, the reason why is as far as I can tell, a qubit has never actually done anything for me. Is that right?

• 29:34Speaker 06

  Not not yet. This is a really good thing to get into.

• 29:37Casey Newton?

  Yeah.

• 29:37Speaker 06

  The interesting thing about quantum computing is is that the fundamental physics, the way information is processed is based on quantum mechanics instead of classical. And the funny thing is is that often when people think about a computer, they very much think about a classical computer. So to know what a quantum computer is, you kind of have to unlearn what a classical computer is. So in a classical computer, we manipulate information using classical properties like digital gates of like and or or like not, you know, which you can think of as like flipping switches on and off.

• 30:08Casey Newton?

  Boolean logic?

• 30:09Speaker 06

  Boolean logic. Thank you.

• 30:10Kevin Roose?

  Look at you.

• 30:12Casey Newton?

  I've also talked to chatbot models about this.

• 30:15Speaker 06

  But quantum computing is based on the properties of quantum mechanics. So there's superposition, which is this being in multiple states at the same time. And there's entanglement where multiple qubits can come together to make even more complex states, so exponentially more complex and then kind of crunch on all of those at the same time. And then you can bring it back together to complete a computation.

• 30:36Casey Newton?

  You know, I went to Burning Man once and was in multiple states at the same time. But that's a different story. So let me ask you this at the sort of very highest level. So we've now learned a little bit about qubits and about, you know, what this quantum computer looks like and we're about to dive into your news. But before

• 30:50Kevin Roose?

  Do that, Julian, tell us, what is the sort of reason why we're doing all of this? What is the great dream of quantum computing that you are trying to bring into existence here?

• 31:00Speaker 06

  Yeah, so quantum computers can just do things that the computers we have today cannot do just ever fundamentally. And that's what's so exciting about them because like we know how powerful computers are. They're ubiquitous in everything that we do, but they all operate the same way. So like an abacus or a vacuum tube or a transistor computer. The principles are the same. They're basically on classical computation and that just fundamentally has has its limits. And so quantum computers will unlock the opportunity to solve problems that there's no other technology that can solve them.

• 31:31Kevin Roose?

  All right, very good. So this week you guys unveil Willow. This is a new chip. Tell us a little bit more about what Willow is and how you made it.

• 31:41Speaker 06

  Yeah, so Willow, it's so it's a new quantum computing chip. This is the thing that lives at the bottom of the refrigerator, but you can kind of think of the computer, it's the whole system, right? And the big breakthrough in Willow is that we made the the qubits themselves just fundamentally better, so they're a little bit less fragile, right? And this enables us to do two new things, right? So one is in what we call quantum error correction. And quantum error correction is basically it's the end game for quantum computing. It's how you take a quantum computer and you make it scalable. So you want to make it really big. You want to make it like industrial and do totally wild stuff with it. You have to overcome the fact that the qubits are fragile and quantum error correction is what lets you do that. So we show that we actually beat what's called the quantum error correction threshold,

• 32:25Speaker 06

  which has been this theoretical idea that's existed for about 30 years that everyone's kind of banking on, but no one's really shown yet, that you can put qubits together and actually reduce how fragile they are, reduce the number of errors by making it bigger. And so that enables a path towards scaling.

• 32:40Kevin Roose?

  So that's one explanation. I also got a second explanation from my chatbot that I instructed to talk to me like I was an idiot. And this is the analogy that it gave me, and I wanted to check this with you to make sure this is correct. It said, think of quantum computers like trying to conduct an orchestra where all the musicians are extremely caffeinated and jittery. The more musicians. Or qubits you add, the more chaos you typically get. Before Willow, adding more qubits was like adding more caffeinated musicians. It just made things messier. What Google did was figure out how to make the orchestra play better when you add more musicians. Is that more or less correct?

• 33:17Speaker 06

  Yeah, I think that's fair. And like, I've got an analogy that I kind of like because what's interesting in this is like, there's a threshold behavior, right? It's like, if it's not good enough, it gets worse. But if it's good enough, it gets better, right? So one way you might think about this is like, totally different technology. But suppose you want to, you want to make a rocket and you want to explore the universe, right? If you want to do that, you have to go faster than what's known as escape velocity of the earth, right? Otherwise, if you don't, you're just going to come back, no matter what. No matter what direction you come in, you're going to come right back into the planet. But if you can achieve escape velocity, go faster, then you actually can get out of the clutches of earth's gravity and go explore wherever you want. And so it's another just example of like, one of these thresholds you have to cross. You either cross it and the technology is going to work, or you don't cross it and it won't.

• 34:02Speaker 06

  And so that's why we're so excited.

• 34:03Kevin Roose?

  Hmm. Now, the big headline grabbing part of this announcement about Willow that got a lot of people's attention was this calculation that this chip was able to run. This was a calculation that would take a classical supercomputer to complete about 10 septillion years, which for comparison's sake is roughly the length of one Lex Friedman episode. But Willow, this quantum computing chip, was able to complete this giant calculation in five minutes. So tell us more about this calculation and how Willow was able to complete it in so much less time than a classical computer.

• 34:44Speaker 06

  Yeah, so this is an important benchmark problem that we like to use to understand what the capabilities of a quantum computer is, right? That's sort of, this is not like a useful problem, the way that we see it and understand it, it really is just like, where are we at, right? And so what's interesting is, we first started running this benchmark in 2019. And we're very excited because at that point, we could run this benchmark better than a classical computer could, which is an exciting moment. Because before that time, the field of quantum computing was in an uncomfortable position, which is, you know, someone could say, hey, you're building these quantum computers. Can you do anything better than a classical computer? Like, literally anything? It would be like, no, we can't. It's like-

• 35:23Kevin Roose?

  That's not a super position to be in.

• 35:25Speaker 06

  Yeah. This guy's got jokes.

• 35:27Casey Newton?

  Thank you.

• 35:28Speaker 06

  Yeah, you got the jokes.

• 35:29Casey Newton?

  Thank you.

• 35:29Speaker 06

  And so in 2019, we're like, well, actually, here's something like it's not useful, but it's something, right? It's a good starting point because like, once you cross that threshold, then you're like, OK, now, now we're talking.

• 35:39Casey Newton?

  Now we've got something.

• 35:40Speaker 06

  Now we can actually start to think about, let's apply it towards something useful.

• 35:43Casey Newton?

  Like running Doom.

• 35:44Speaker 06

  Right, like running Doom.

• 35:45Casey Newton?

  Yeah.

• 35:46Speaker 06

  Or Crisis or something like that. And so, as I alluded to, there's this fundamental advantage that quantum computers have for certain problems. This is evidence of that really kicking up.

• 35:58Kevin Roose?

  Can I ask a very dumb question? I mean, all of my questions today are going to strike you as very dumb, but this one may be particularly dumb. But it is something that I wondered about with this super calculation that Willa was able to do in five minutes. How do you know if it's right or not? I mean, do you use a quantum computer to check its work? Or can you even use a classical computer to say this was the correct solution to the problem?

• 36:20Speaker 06

  Yeah, this is such a good question because like-

• 36:23Kevin Roose?

  Oh, my God, I asked a good question.

• 36:24Speaker 06

  It's a great question. It's the right one to ask because basically, you have to kind of build a case to explore into the regime where you can't actually go. Like, the thing is, only the quantum computer can go into that regime. That's the whole point, right? So how do we know if it's if it's right or not, right? So we do, we do a lot of work. We make a bunch of smaller problems. Then we look at the scaling behavior and we're like, OK, it was right at a problem size of two and three and four and five, and it's following this trend line exactly how we expect. So our projection says that, therefore, it will continue to behave in the way that we expect. So you kind of have to look at it from multiple angles and build up an evidence-based case. But yeah, you can't directly verify it.

• 37:06Casey Newton?

  All right, so you've been able to solve a very complicated math problem. What, what do you do with that? Where do you go from there?

• 37:12Speaker 06

  Yeah, so of course, what we want to do is we want to actually solve problems that have real-world impact, right? Will is an important step forward, but it's not the end of the story. It definitely brings us closer to real-world applications. It's a step on our very long-term goal of making a large error-corrected computer. And at the same time, we're taking these machines and we're running them in like an unerror-corrected way to see like, hey, what can we get out of it today, right? And so, we're using that to explore what might be in the near term, while at the same time, looking at the very long-term as well.

• 37:44Kevin Roose?

  I want to ask you about a line in the blog post that was announcing Willow that just really made my head spin and that was pretty unusual in announcements from Google about new hardware or breakthroughs. And the line said that the success that you all had with this Willow chip, quote, "lends credence to the notion that quantum computation occurs in many parallel universes, in line with the idea that we live in a multiverse." Now, for people who haven't read a lot of science fiction, a multiverse is the idea that we are currently existing in one of many universes out there.

• 38:27Kevin Roose?

  What I don't understand is how the breakthrough that you made with this chip makes the notion of a multiverse more likely. Can you explain that?

• 38:36Speaker 06

  Yeah, so I think you look at like a little chip, right? That's fundamentally doing the computation. It's like this little thing. And then you look at like some warehouse-sized supercomputer. And then you imagine that running for longer than the age of the universe. You're like, how can there be such a difference between these things? And so there is some room for interpretation for exactly how that functions.

• 38:59Casey Newton?

  Yeah, and also, by the way, Kevin, I would say a lot of this is covered in Doctor Strange and the Multiverse of Madness, if you haven't seen that one yet.

• 39:06Kevin Roose?

  Well, I definitely understood the answers that you just gave in some other part of the multiverse.

• 39:13Casey Newton?

  It doesn't have to be like, my understanding is, you know, because the qubits are in all these multiple states at once. It's like the states themselves are the evidence of the multiverse, right? It's essentially all these states that they're in are all, not all, taking place in the same universe? Did I get that part right?

• 39:28Kevin Roose?

  I don't know, but I feel like I just took a huge bong rip.

• 39:34Kevin Roose?

  I'm sorry. I believe that what you are working on is very important. And I believe that because a lot of people who are smarter than me also believe that what you're working on matters. For example, there was a big back and forth after this Willow announcement was made between Elon Musk and Sundar Pichai, CEO of Google. Elon Musk responded to- The Willow News by just saying wow. And then Sundar said, we should do a quantum cluster in space with Starship one day. And Elon responded that, quote, "Any self-respecting civilization should at least reach Kardashian type two." What are these people talking about? What is going on here? And why are these people so excited about this breakthrough?

• 40:13Speaker 06

  I mean, I think at the end of the day, it just comes down to the fact that like quantum computing has been something that's been talked about for a long time, right? But like, it is a real technology and like, this is providing more evidence to that, right? It's like, there are computations that you can only do on a quantum computer. And there is a path towards taking these quantum computers and scaling them up to be really big and useful. And I think that's fundamentally why people are excited.

• 40:35Casey Newton?

  What are some like, problem areas that you're excited about? That if you could make a computer as powerful as the one we've been talking about today, we could have a shot at solving something big?

• 40:44Speaker 06

  So, yeah, in terms of application areas, there's some interesting ones. So one is on like, saving energy. So I don't know if you're familiar with the Haber-Bosch process for making fertilizer.

• 40:54Casey Newton?

  Not only am I familiar with it, it was one of our 100 most iconic technologies.

• 40:58Speaker 06

  Yes, right. So, and you know, something like 2% of the world's energy goes towards making fertilizer, right? And I think like, the fun history of this, you know, you guys probably know this. But like, this is a revolution, right? Because before that, people would go out and sail into the middle of Pacific and like, chip bird poop off of rocks and stuff like that to bring that home. And, you know, there, people were thinking like, I wonder if this is going to be sustainable for the entire planet for fertilizer production. Eventually, it come out with the Haber-Bosch process, which is great, right? But the issue is, is that it takes like, high pressures and temperatures. Like, this is industrial chemistry process for making fertilizer. And what's funny is, is that if you look at what happens with like, bacteria that just like, live in the ground with plants, like, they'll do this at ambient pressures and temperatures and things.

• 41:40Speaker 06

  But it's hard to know, like, how do you guys, how do you guys get that to work, right? And fundamentally, if you wanted to understand that, you'd need to really have a detailed understanding of what they're doing quantum mechanically. And a quantum computer would allow you to kind of like, peek inside and be like, hey, what's actually going on under the hood? And could we harness that to make something that's way more efficient? And, you know, potentially save 2% of humanity's energy, which is like, that's a lot if you think about it.

• 42:02Casey Newton?

  So you're saying this could... Knock the Haber-Bosch process just out of the top 100.

• 42:07Speaker 06

  Potentially. Hey, like, you know, it could still be like, we can, we can respect it even if we don't use it.

• 42:13Casey Newton?

  Yeah.

• 42:13Speaker 06

  So I might not need it anymore. Yeah.

• 42:14Kevin Roose?

  So my understanding of why people like Elon Musk and Sundar Pichai are very excited about the possibility of quantum computing has to do with what you just described, which is like, it is a means of unlocking more energy for things. If we can actually get quantum computers that work, that are useful, that are more reliable, less fragile, you could do just many more types of powerful computation without needing to like build these giant data centers, running supercomputers 24/7. Is that, is that part of why people are excited about this? Because it could shrink the energy needed to perform various tasks?

• 42:50Speaker 06

  Yeah, that's right. And in fact, I'd go one step further, which is like, some tasks are energy intensive and some tasks are just impossible, right? It's like, again, this one of 10 septillion years, that's just impossible. Like, we can talk about the energy. Okay, how much energy is it going to take to run this for like way, way longer than the age of the universe? But, you know, it really is just, it's, it's such a big problem that is impractical, right? And like, those are the ones that we're targeting. Like, it's, it's the impossible part that I think is so compelling to people. And of course, energy savings is great and, and that's very relevant, but like, it's, it's beyond that.

• 43:19Kevin Roose?

  So Julian, just summarize sort of where we are in the quantum computing story. Because, you know, I've been covering tech for like a long time. So has Casey. I know we've heard a lot of ideas and people getting excited about quantum computing. Every few years, there seems to be some sort of breakthrough that gets heralded as like a major step forward. But I still don't feel like I have a good grasp of what quantum computing will do for me. We still obviously don't have quantum computers out in the world in certain places where people can interact with them. So like, how close or far are we to, say, having a quantum computer that the average person will be able to use?

• 43:57Speaker 06

  I would imagine it'll take a while for the average person to interact with a quantum computer because, you know, right now, we have expert scientists and researchers and things working with them and they're pretty, you know, they're pretty like prototyp-y, to be honest, right? And where we want to get to with these error corrected computers in the next number of years is get it to the point where then you have experts that don't have to be as familiar with the underlying hardware and all the kind of quirky aspects of it, right? They can use it a bit more like a prototypical computer. That being said, we are also looking for the sort of near term potential applications, but those are going to be a bit more bespoke, is the expectation.

• 44:36Kevin Roose?

  Yeah. When do you think we'll hear about a major scientific discovery powered by quantum computing? Like, what's your best guess of when we'll hear that?

• 44:46Speaker 06

  It depends what you mean by major.

• 44:49Kevin Roose?

  Something that I could understand would be major.

• 44:54Speaker 06

  I did my best on this one.

• 44:56Kevin Roose?

  Something we would like conceivably talk about, you know, on this, on this show.

• 45:00Speaker 06

  I think there'll be some interesting things coming out in the near future, but I would say for like a huge, real world breakthrough, it still might be a couple of years out.

• 45:11Kevin Roose?

  Yeah. But like two or like five?

• 45:14Speaker 06

  Again, it depends on the caliber. It's tricky because it's like, it's research, right? So it's possible, you know, when we run the computers without error correction, you're like pretty, you're limited in terms of the computational space you have access to. If you happen to be able to fit a problem in there, you could do some really exciting stuff potentially, right? But the question is, can you fit the problem in there or not, which is a little bit hard to know, right?

• 45:40Kevin Roose?

  Right.

• 45:41Speaker 06

  But there's, you know, there's confidence on something like a five-year timescale, then you're starting to get into a pretty interesting territory and you're in less of a, you're kind of in a scalable regime where you will start to see more incremental progress of like year over year, you'll see the capacity expand and it'll be more sort of foreseeable and understandable. And I think that's going to be a very exciting point to get the field to get to from very abstract to kind of like planimal and engineering.

• 46:07Kevin Roose?

  I guess that's what I'm wondering about is sort of like, I mean, I am on the record, Julian, just for full transparency here, is saying that I am suspicious that quantum computing is fake and that it might actually just be like a jobs program for physicists, you know, who want to work at Google. But I am more persuaded after talking to you that this is actually real and that people who are smarter than me are working on it and that it may one day produce something that all of us benefit from.

• 46:33Casey Newton?

  Here's what I'll say. What I would love is the next time you guys have a breakthrough, it would be cool if Kevin and I could actually come out. I want to see this thing with my own eyes. You know what I mean?

• 46:41Kevin Roose?

  Yes. I want to be sucked into the multiverse by a quantum computer.

• 46:45Casey Newton?

  I want to touch something that is absolute zero and see what happens to my finger.

• 46:49Speaker 06

  Might get freezer burn, but, you know, that'd be fun. Yeah, maybe you guys could visit.

• 46:55Kevin Roose?

  I guess I'm being a little cheeky here, but I do think there is a problem that you all face in quantum mechanics and quantum computing that researchers in other fields do not necessarily face because the thing that you are building is very hard for the average person to understand and observe. And so I wonder if there is just a barrier there that, say, someone building a chatbot might not face.

• 47:16Speaker 06

  Yeah, I mean, so quantum computers, they're a real technology and we are making progress, as you mentioned. That being said, it can be hard for a layperson to understand the state of the field. There can be a lot of hype and kind of overselling. And what we're trying to do is be very clear about, like, when we get to these points, we've done this and provide the context around that. This is the things it can do, this is things it can't do. But what's nice about, for us, is that we have the opportunity to take a very long-term vision. Like, we're not trying to change the world tomorrow, right? Like, we have this longer-term vision of the thing that we want to build and allows us to be a bit more measured in our progress and reporting it as we head in that direction.

• 47:53Kevin Roose?

  Cool. Well, congratulations on Willow and thank you for explaining what you do to us. And I do feel actually like I understand what you're working on much better than I did at the start of this, so thank you for that.

• 48:03Casey Newton?

  I agree. It might not seem like it, but that actually was Kevin doing the best that he could.

• 48:07Kevin Roose?

  Julian, thanks so much.

• 48:10Speaker 06

  Thank you, Julian. It's been fun chatting with you guys.

• 48:12Kevin Roose?

  When we come back, we'll talk about the AI insiders who are very into Claude.

• 48:30Speaker 05

  Companies are pouring money into AI, but few leaders can explain how those AI investments work with the systems you already have. You know, the ones that actually run your business. Without Systems, you don't have to start over with AI. You can choose the projects where AI will have the most impact. And OutSystems gives you the platform to make it agentic on your terms with complete control every step of the way. So you can trust AI with the mission-critical systems that run your business. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.

• 49:05Speaker 00

  AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.

• 49:37Speaker 07

  Ripple has been powering modern finance for a decade. It's not our first trip around the blockchain, and it won't be our last. We're leading the charge to bridge the gap between existing financial systems and the future of finance. You can have it all: legacy, stability, and digital innovation. Systems that work with your other systems. Because with the right partner, the question isn't crypto or traditional finance. The question is simply, why choose? A new standard of modern finance. It's happening with Ripple. For informational purposes only. Availability of products can vary. Maybe provided by different Ripple entities.

• 50:13Casey Newton?

  All right, Kevin. Well, I got a note from you saying you wanted to talk about Claude. Yes. And I assume you mean the founder of the French impressionist Claude Monet.

• 50:23Kevin Roose?

  Not quite.

• 50:24Casey Newton?

  Who? Oh, so you mean the other Claude, the chatbot one?

• 50:26Kevin Roose?

  Yes, the chatbot Claude from Anthropic is something that I've been writing about this week because Casey, this chatbot is taking San Francisco by storm.

• 50:35Casey Newton?

  It really is, and I loved your column about it, and I think it is time to cast a little bit of light on a chatbot that is definitely much less used than some of the more popular ones, in particular ChatGPT, but as you note in your column, has to develop a really fervent following.

• 50:53Kevin Roose?

  Yeah, so I first started thinking about this column back at that AI conference that you and I went to together in Berkeley, where I was just talking with people about the AI tools that they're using in their daily lives. And over and over again, I kept hearing people who are in the AI community, many of whom work at AI companies, tell me that they were all using Claude not just for help with their work or for help writing emails, but really as kind of an emotional thought partner for themselves and a problem solver in their personal lives.

• 51:24Casey Newton?

  Yeah, so what kind of conversations are they having?

• 51:26Kevin Roose?

  So I've talked to a lot of people who are sort of Claude power users, and what I was hearing from people is that Claude has become a kind of social sidekick, that it helps them navigate relationship challenges, think through major life decisions that they have to make, maybe help them in an argument or a dispute with a friend or a loved one or even a romantic partner, that it has sort of become this all-purpose social helper for them. And, you know, this is something that you and I have talked about in the context of these AI companionship apps like Replica and Character AI, but this feels different to me for a couple of reasons. One is Claude is not built for AI companionship, right?

• 52:10Kevin Roose?

  It is a general-purpose AI model. Anthropic has not specifically designed it to become friends with people or to help them in their social lives. This appears to be kind of an off-label use of this model. But it also struck me as remarkable because the people who are falling for Claude the hardest, at least in the circles that I'm seeing, are people who work in AI, are people who know that These systems are not sentient creatures, that they are just sort of next word prediction machines who understand the way that they're built and trained, but who still seem to be having these sort of very emotional reactions to Claude.

• 52:44Casey Newton?

  Yeah, and I will say that I can confirm everything that you're saying. I also talk to people who have joined the cult of Claude and find it to be very useful for all of the things that you're talking about. For those people who have not yet had this experience, what is it about the way that Claude is responding, do you think, that is building this kind of cult-like following around it?

• 53:04Kevin Roose?

  So, part of it is just that I think Claude and some of these other frontier models have just gotten really smart. And this is something that I heard from people who are part of this sort of group of Claude super fans. It's just, it just sort of has a good grasp of a lot of different subjects. But the thing that really seems to be the X factor that is driving this frenzy around Claude among AI insiders is that it seems more emotionally intelligent than other models. And, you know, people will say, and I'm sure this is a fair criticism, like, this is not real emotional intelligence. This is, you know, this is just a sort of simulated version of emotional intelligence. But people who talk to Claude a lot, some of them feel like this model actually starts to understand them.

• 53:48Casey Newton?

  So, one of the reasons this is interesting to me, Kevin, is that when we think about the AI race that we're in, the way that we usually determine which model is the best is by looking at who is performing the highest on various benchmarks. What percentage are you scoring on these very difficult math problems, or word problems, or various other sorts of tests? And if you look at something like Chatbot Arena, which is one of the big systems that measures these things, Claude is often near the top of the list, but as we record this today, it's not on top of any of those lists. So, what is the difference between a chatbot that is the best and one that develops a cult following?

• 54:30Kevin Roose?

  Yeah, this is one of the other questions I was trying to answer in the course of reporting this column is, you know, what did Anthropic, the company that made Claude, do behind the scenes to make it feel more compelling? More emotionally intelligent, more insightful than other models. And so I talked to Amanda Askel, who is the sort of in-house philosopher at Anthropic. She's a trained philosopher who now is sort of in charge of Claude's personality, you could call it, like the way that it behaves in interactions with users. And we talked about sort of the steps that they had made trying to turn Claude into something very compelling. And before I tell you what she told me, I'll just say like, this is not always been Claude's default mode, right?

• 55:14Kevin Roose?

  I remember when Claude first came out, when it was released to the public, I could barely stand to talk to it because it was so preachy. It was so careful. You know, Anthropic as a company has always been very obsessed with AI safety. And so the first versions of Claude did not really seem to be fun to talk to at all because you'd ask it a question and it would give you like a moral lecture before responding.

• 55:36Casey Newton?

  Can I give you an example of that?

• 55:38Kevin Roose?

  Yes.

• 55:38Casey Newton?

  So one time, when this period of Claude was happening, I asked it to summarize some movie. And it said, I absolutely can't do that. And I said, well, why not? It said, the risk of me reproducing copyrighted material would simply be too high. And I said, well, don't reproduce the material. Just tell me what it's about. It says, I can't. It is simply too risky. And I said, you're being ridiculous. And it said, I know you might be frustrated, but this is what I have to do. And yeah, in that era, Claude, not super useful.

• 56:04Kevin Roose?

  Right. So they've done a lot of work to try to fix that problem with Claude. And part of the improvements that people are noticing is just a function of a better base model, right? We went from Claude 2 to Claude 3 to Claude 3.5. And now we have Sonnet and all these different models. And each time the model improves, it just, it gets a little bit better at understanding a user's intent and maybe giving them a response that's along the lines of what they're looking for. But what Amanda Askel told me is that there's been a lot of very careful work around defining a consistent yet adaptable persona for Claude.

• 56:43Kevin Roose?

  So, you know, one problem that these models have that we've talked about before is called sycophancy, basically. These models, they want to tell users what they want to hear. And so they flatter you, they don't challenge you, even if you have a sort of bad or potentially harmful idea. They'll just sort of tell you, oh, that's really fascinating. Let's talk more about that.

• 57:05Casey Newton?

  Yeah, these are the qualities that I look for in employees generally.

• 57:09Kevin Roose?

  Yes, you really like a yes man.

• 57:11Casey Newton?

  I do.

• 57:11Kevin Roose?

  At Platformer, but a lot of people don't like that about these AI models. They want to be challenged. They want something that feels more like it has, like it's actually evaluating what they say on the merits and not just agreeing with it. So Amanda told me that basically they have tried to make Claude less sycophantic. And the the personality type that she said that they were aiming for is something like a well-liked and well-respected traveler. Someone who can go around, who goes around to a lot of different places, who talks to a lot of different people from a lot of different walks of life and has a consistent personality, but is also able to adapt and communicate with people in different styles, depending on where they are, who they're talking to.

• 57:50Casey Newton?

  I have noticed this in my own interactions with the bot is I have the sense that it is giving me a feel for how many people think about this problem, right? It is sort of reporting back from the collective unconscious. And there is something kind of cool about it. Of course, it's limiting, like that is a certain kind of feedback that you're giving yourself. But I think a lot of what we want to know as we move through the day is, is this normal? And Claude is turning out to have really interesting answers to the question of, is this normal?

• 58:18Kevin Roose?

  Yeah, and it's also just a really interesting challenge for designing a chatbot that can behave in this way. And, you know, Amanda was telling me she's a philosopher by training, and she said that making Claude's personality and sort of fine tuning it has put her in touch with virtue ethics in a way that nothing else in her career ever has, because it goes to this question of like, what does it mean to be a good person? They want Claude to be a good person, and sometimes that means telling people that they're, you know, they're on the right track, and sometimes it means challenging them.

• 58:49Casey Newton?

  Yeah.

• 58:50Kevin Roose?

  And so the key for an advanced AI model is knowing sort of when to push back and when to go along with what a user is saying.

• 58:56Casey Newton?

  Yeah, and the way that I'm going to test that later is by trying to use Claude to plan a bank heist.

• 59:02Casey Newton?

  You know, I keep coming back to the idea of a therapist. And I have talked about this before, but really it comes up so much for me as we discuss this. Because what a therapist is trained to do is to just get you to talk and reflect yourself back to you in a positive way. And it does require an intelligence. Obviously, really great therapists have many more tools than that. But fundamentally, that's what it is. And most people don't go to therapy. But we now have this opportunity where for some price between free and 20 bucks a month, you can talk to something that will just sort of hold you, look at you, hear what you're going through, reflect it back to you, offer you a little bit of coaching and guidance. There are a million limitations around that.

• 59:47Casey Newton?

  There are a million dangers. But there is also something, I think, really potentially quite useful.

• 59:51Kevin Roose?

  Yeah, I agree. And I've been talking with people, as I'm sure you have, who are part of this new group of Claude obsessives. And it doesn't seem to be having sort of overnight transformative effects on their life. It's not like this chatbot suddenly solves all their problems. But it does seem to be kind of helping them make decisions, helping them think about things that happen to them in new ways, maybe help them understand some interpersonal dynamics that are going on in their lives. I was reading this thread by Nick Camerata, who's a former OpenAI employee and just someone who has a lot of friends in the AI scene who are early adopters of this stuff. And he described the way that Claude had become kind of a big part of how people that he knew

• 1:00:36Kevin Roose?

  were conducting their daily lives. And his observation was that the people who have really started using Claude all the time just seem to make fewer dumb mistakes, as he put it, like they just have what he called a computational guardian angel watching over them, where they can basically be consulting this chatbot, you know, maybe dozens of times a day on decisions, small and large. And it just seems to be kind of raising the floor for them psychologically.

• 1:01:04Casey Newton?

  Yeah, well, I'm sure a lot of people are hearing that and they're thinking, oh, my gosh, we're already letting the AIs make dozens of decisions a day for us. We're in a lot of trouble here.

• 1:01:11Kevin Roose?

  Totally. And there's some, I should say, like, there's something that terrifies me about this world. Like, I consider myself fairly optimistic about AI, but I do think it's worrisome, where some very smart people are walking around San Francisco, basically outsourcing their cognition to this chatbot. And that, to me, feels like maybe a worrisome indicator. But how do you feel when you meet people in San Francisco or elsewhere who have come to rely heavily on chatbots in the course of their daily lives?

• 1:01:41Casey Newton?

  Well, I mean, mostly just with curiosity, right? Like, I think this is a relatively new phenomenon, and it is just interesting to hear about what sorts of questions people are asking, what kind of answers they're getting back, if it is leading people to make any changes in their lives.

• 1:01:59Kevin Roose?

  And what are you hearing on that front? Like, what are people in your circles using this stuff for?

• 1:02:03Casey Newton?

  Well, for example, thinking through a very difficult situation at work, right? Like, making a career move, wanting to know how to communicate something internally to a manager, or to your manager's manager, and trying to sort of game it out with someone. That is something that you could do with a friend, but would you maybe want to tap into the collective unconscious and see what sort of, like, the statistical median of all people would do in that situation? Like, there's a lot of appeal there, right? You know, I was using it to prepare for an interview, and I had gotten a bunch of materials that I was going to be asking this company-

• 1:02:37Kevin Roose?

  Are you interviewing for jobs on other podcasts?

• 1:02:39Casey Newton?

  Yes, I am, actually. I want to co-host with Joe Rogan. No, but I had got some materials, and I had shared those with Claude, and I was, you know, trying to get a sense of what do I think the shape of the story is. And I was trying to understand one particular detail, and I was like, you know what, I think I've actually identified a tension in between, like, something the company is telling me, but like, something that they're telling me about this product. And so I asked Claude if he could sort of explain to me this discrepancy, and Claude said, "That's actually a really probing question." And I like, I felt good about myself. Now, that's silly, of course, to feel good about yourself because a chatbot is giving you a compliment. But I will note, that was really the only good question that I had that I asked of Claude, and Claude did go out of its way- To say, well, that was the good question.

• 1:03:23Casey Newton?

  So again, yeah, in some sense, like, we are just sort of like staring at like math and then we're developing feelings about the math that is staring back at us. I don't want to sort of continually underline the fact that there is some, you know, smoke and mirrors here and we are just sort of like falling for the smoke and mirrors. But I kind of just think that that's like how human beings learn and relate to each other. There's a lot of smoke and mirrors and therapy too. And guess what? It still works for people. So what I'm the most interested in is how do people find the help and support they need to lead happier lives? And if a way that you are leading a happier life is by talking to a free chatbot, I am not going to get in your way.

• 1:04:03Kevin Roose?

  Yeah, I do think it raises the possibility that there's essentially going to be like, you know, two groups of people in the very near future, one of whom are relying heavily on AI tools as a decision making assistant, as a productivity enhancer, as maybe a therapist or a relationship counselor, and then a group of people who just want nothing to do with them. And I'm curious if the group that uses AI will start to become more powerful as a result of having this sort of always on advisor sounding board therapist with them.

• 1:04:39Casey Newton?

  Well, this is why I think your column is important because, you know, some people might listen to this and say, oh, really? There's a group of San Francisco hipsters doing something weird with AI? Like, tell me something I didn't already know. But it made me think of this great quote by Chris Dixon, the investor and a one time guest on our show. And he once wrote, hobbies are what the smartest people spend time on when they aren't constrained by near term financial goals. I like to say that what the smartest people do on the weekends is what everyone else will do during the week in 10 years. And when I thought of that quote in relation to your column, I was like, this use of AI right now, it is kind of a smart person hobby that some people are really excited about. And I do think that there is a case to be

• 1:05:21Casey Newton?

  made that in 10 years, I don't know if this is necessarily going to be everyone's job, but I do think a lot of people, maybe even most people are going to be turning to some Something like a chatbot to help them make decisions.

• 1:05:34Kevin Roose?

  Yeah, I mean, another question I have more narrowly about the sort of cult of Claude is like whether it will actually drive mainstream adoption of this chatbot. Because I think, you know, there's an argument that the more people in San Francisco become obsessed with this chatbot, the more it might sort of become a worthy rival to ChatGPT. Right now Claude is just a lot less popular than ChatGPT.

• 1:05:59Kevin Roose?

  But I also think like this is the kind of thing that might, you know, not sit well with a corporate customer. For example, like they don't want their chatbots that they're using in their insurance company or their marketing firm to have a strong personality, right? They do want them to be more sort of generic seeming and productive, but not to have this kind of like carefully constructed personality. So I almost wonder if the world of AI chatbots is going to slowly separate into the the sort of work focused tools and the tools that are focused more on the more creative and relational parts of your life and whether those will actually just be separate products.

• 1:06:37Casey Newton?

  I mean, maybe my expectation is just that all the big AI labs will just offer different personas, right? And that if emotional Claude becomes super popular and drives tons of adoption of Claude, you can absolutely bet that there's going to be an emotional ChatGPT and an emotional Gemini, and you're just going to kind of get to pick. In fact, already in Claude, you can choose what style you would like it to respond to you in, and you can even create a custom style within Claude. So I think personalization, customization is sort of the future there. But I had a theory about this that I wanted to run by you.

• 1:07:12Kevin Roose?

  What's your theory?

• 1:07:13Casey Newton?

  I don't even know if I believe this, but I was like, why does Claude feel kind of better for some of these questions? And we have so far not talked about one of Claude's very real limitations, which is that when you ask it a question, it does not search the web. It does not go out and get updated real time information from you about anything. So, you know, its training data stops sometime earlier this year, I believe. But the result is when you ask it a question, it feels like it is really sitting with you and answering your questions. Whereas if you ask Gemini or ChatGPT a question, it often will go search the web. And I experienced this as something akin to when you ask a friend a question at a bar and they pull out their phone. And it feels like all of a sudden they're not in the conversation anymore. They're going to read a million things, right?

• 1:07:58Casey Newton?

  They've sort of stepped outside the conversation where it's like Claude is staying with you the whole time. Does that, is that kooky?

• 1:08:04Kevin Roose?

  No, I don't think that's kooky. I think there's something there. I mean, I think a lot of people would consider that a missing feature, right? It makes Claude less useful in some contexts than a chatbot that can go out and search the internet. And we should also say like Claude is not the most fully featured chatbot on the market, right? There's no like voice mode like there is on ChatGPT. It can't generate images like ChatGPT or now Gemini can. Like it doesn't have as many features as some of these other chatbots do. But I think in part because of those limitations, it does feel like a more sort of unmediated pure experience of like talking to the language model and having the language model talk back to you.

• 1:08:44Casey Newton?

  Yeah, so Kevin, let me sort of wrap up by saying, okay, we've done this sort of like, you know, cultural trend analysis, which I think is very interesting. Let's try to move it a bit into the business realm. If you're the head of product at Claude, which is actually Mike Krieger, the co-founder of Instagram, what are you doing to translate this enthusiasm that people have for the product into something that is actually running neck and neck with ChatGPT in terms of weekly users?

• 1:09:10Kevin Roose?

  I mean, I think it's a tall order. ChatGPT had a head start. It got to the market before Claude or any other sort of leading edge chatbots. It's now sort of synonymous with AI chatbots themselves. It's sort of become a household word in the way that Uber or Google did in previous generations. But I think, look, I think there's a fine line to walk here because if your model has too much personality, if it's too engaging, then it starts to feel like one of these AI companion apps. And there's a market for that, but it's not, it's not the right market for a company like Anthropic, which wants to, you know, sell through Amazon and make its chatbots very commercially useful to big companies. Companies as well as to individual users.

• 1:09:54Kevin Roose?

  So I think they've got to walk a fine line. I've noticed that Anthropic is leaning more into the the sort of social branding of Claude than I would have expected. I saw an ad in the airport the other day, basically talking about how Claude can be your AI friend. And I thought, well, that's interesting and maybe risky. But I think if I were the head of product at Claude, I would be playing up this sort of like hidden helper angle here, maybe making some ads showing people who sort of are, you know, going through their daily life and they run into a situation where they don't exactly know how to handle it and they, you know, consult Claude in their pocket. And then they go back and have the conversation with their with their spouse or something like that.

• 1:10:35Kevin Roose?

  Because I do think that is the use case that people in the AI scene are starting to find really valuable. What what would you do?

• 1:10:42Casey Newton?

  Well, so I do think there's a lot to be done on the marketing side. I think that Claude's marketing has actually been quite bad. There is a billboard that I saw recently that said something like, take your business up and to the Claude, which I think was supposed to be a play on like up and to the right. But yeah, that doesn't work. Also, the Claude logo, I'm not going to say what it looks like, but if you look at the Claude logo, you're going to know what it looks like. And I'm going to ask you, should it look like that? But to offer an even more constructive suggestion, Kevin, I would say, I think voice mode is the big opportunity here. Because as impressive as ChatGPT's voice mode is, technically, I do not actually find it enjoyable to talk to. It has that kind of nervous anticipation of an intern who's working for you.

• 1:11:27Casey Newton?

  It's like, and anything else, sir? Anything else that that I do not find warm or humane or comforting? If they can translate everything that we've been talking about into voice, then I can see a lot of people saying, no, that is actually the one that I'm going to be using instead. You know, this week I had a chance to use Gemini in voice mode on a pair of prototype glasses because I went down to Google for a demo. And again, it was that nervous intern. It was like, anything else? Can I help you with anything else? And it was, I almost found it very sort of like pushy and aggressive. It did not feel like a chill, cool friend. So if nothing else in this conversation, I think we have hit on a vibe that can be exploited and turned into money. And what higher purpose could a podcast have than that?

• 1:12:11Casey Newton?

  So we'll be interested to hear if anyone at Anthropic wants to make use of these ideas.

• 1:12:29Speaker 05

  Building and scaling AI today can feel pretty chaotic, mostly because teams are juggling too many fragmented tools at once. That's where Outsystems comes in. Outsystems gives you one unified agentic systems platform to manage your entire portfolio of agents, apps, and workflows. You get to innovate at the speed of AI, but with the security, governance, and control your enterprise actually needs. And the results are already there. A major manufacturer used Outsystems to build a multi-agent system that analyzes blueprints, and a global bank cut their development time by 60%. Own your agentic future today at outsystems.com/nyt.

• 1:13:05Speaker 00

  AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.

• 1:13:37Speaker 07

  Ripple has been powering modern finance for a decade. It's not our first trip around the blockchain, and it won't be our last. We're leading the charge to bridge the gap between existing financial systems and the future of finance. You can have it all: legacy, stability, and digital innovation. Systems that work with your other systems. Because with the right partner, the question isn't crypto or traditional finance. The question is simply, why choose? A new standard of modern finance. It's happening with Ripple. For informational purposes only, availability of products can vary. May be provided by different Ripple entities.

• 1:14:12Kevin Roose?

  So one more thing before we go. Last week, I got to go to New York and lead a fascinating task force discussion at the Dealbook Summit, including 10 of the leading voices on AI. These were basically a dream team of people that we assembled to talk about some of the biggest issues facing the AI world. These were people like Jack Clark of Anthropic, Eugenia Koita, the founder of Replica, and Ajay Kotra and Dan Hendricks, two of the leading experts in AI safety. It was a really, really fun conversation. And hearing these people talk to each other about issues where they agree and disagree was really striking. My goal in putting together this panel was to hear a bunch of different views from across the AI spectrum.

• 1:14:56Kevin Roose?

  And I think we accomplished it. I think it was a lot of fun, and I think we all learned a lot. The Dealbook podcast has put out the full conversation of our AI panel as a standalone episode. And I think if you listen to Hard Fork, you will find it very illuminating. And just to really give you a flavor of the conversation, I wanted to play one brief clip from the conversation. This was from one of my favorite moments from the discussion that was basically prompted by a question that I posed about what the US government's role should be in ensuring that US AI companies maintain their advantage. I asked about this thing that's been going around the AI world in recent weeks about the idea that we need a Manhattan Project for AI, basically a big federal coordinated

• 1:15:39Kevin Roose?

  effort to make sure that the US stays ahead of countries like China when it comes to building advanced AI. Here's some of what followed.

• 1:15:46Speaker 08

  It seems to me we have about three or four or five of them already. If you look at the big companies who are investing 10s or 20s or 30s of billions of dollars in it.

• 1:15:56Kevin Roose?

  Should the government be giving them more money to have them go faster? Eugenia?

• 1:16:00Speaker 09

  I actually have a little bit of a hot take here. I think the easiest way to win AI for the US would be to just make it extremely easy to come here on a visa for any computer scientist or mathematician or physicist who is working in AI, who has any inclination to work in AI. And the AI war will be over in one day, just if that immigration is solved. And I think this is actually a really simple thing to do, and unfortunately, I'm not seeing much of a progress.

• 1:16:30Speaker 10

  It's probably worth noting that economic security is national security and things like Manhattan Project are like at the end of a very long list of interventions government can do to gain a lead in a technology. And interventions you can do earlier are like sorting out immigration or, as Peter said, building the basic experimental infrastructure that would let scientists access the kind of computers that, you know, the companies assembled here have. And so when I hear terms like a Manhattan Project, I'm like, that's where like our imagination has got us. There are many other tools which are less expensive and scary that you can be using way earlier.

• 1:17:08Kevin Roose?

  If you want to hear the rest of this conversation, which ran about 90 minutes, you can search for Dealbook Summit Podcast and listen to the full episode titled "The AI Revolution." We'll also drop a link to it in our show notes. Thank you to all of the people who came and participated in this, as well as to the Dealbook team who put on an amazing summit.

• 1:17:29Kevin Roose?

  Hard Fork is produced by Whitney Jones and Rachel Cohn. We're edited by Jen Poyant. We're fact-checked by Caitlin Love. Today's show was engineered by Alyssa Moxley. Original music by Alicia Baetoupe, Marian Lozano, Sophia Landman, and Dan Powell. Our audience editor is Nell Galogly. Video production by Ryan Manning and Chris Schott. You can watch this full episode on YouTube at youtube.com/hardfork. Special thanks to Paula Schuman, Huwing Tam, Dahlia Hadad, and Jeffrey Miranda. As always, you can email us at [email protected]. Tell us what you would do with a quantum computer.

• 1:18:29Speaker 05

  Companies are pouring money into AI, but few leaders can explain how those AI investments work with the systems you already have. You know, the ones that actually run your business? Without systems... You don't have to start over with AI. You can choose the projects where AI will have the most impact. And Outsystems gives you the platform to make it agentic on your terms, with complete control every step of the way. So you can trust AI with the mission-critical systems that run your business. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.