（预告）Anthropic 传奇仍在继续、Fox 与流媒体的未来、ChatGPT 问答、智能体购物、自动驾驶

Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp, and on the other line, Ben Thompson.

Ben, how are you doing?

Sad, Andrew.

Why?

Sad because I've been forced over the last week to use these really crappy AI models. Only 5.5 in Opus. Where's Fable?

Oh, boy.

I told you Fable was good.

Somebody needs to jailbreak Fable back to life for us, man.

We did need the 48 hours of truly leap forward in terms of model capabilities just to be truly—

Poor Ben. His wings have been clipped here. Yeah, now here we are, living in the Stone Age. But look, it's only a matter of days until the rule is reversed and Fable is brought back to life, would be my guess.

I was going to say—

We're going to start with Anthropic.

Was that Andrew Sharp Reports, or—

No, look. It occurred to me as we came on to record and talk about the export controls and everything else that by Friday night, I'm sure there will be a twist that renders half of this obsolete. We're not going to start with Anthropic, though, actually, because you informed me that you have a World Cup take related to the hydration breaks.

Oh, you want to bring—

And—

No, I don't know if we can do this at the top of the show because—

Okay.

I'm concerned I'm going to make people so mad that—

They will hit stop and unsubscribe.

Right. Well, if it was at the end of the show, then they might go and unsubscribe, but at least they've listened to the whole show. But if we put it up front, they might not listen to the rest of the episode. Although maybe if we make the rest of the episode really good, they'll decide to continue to be subscribers.

A bit of grace after 90 minutes. Fine.

Okay.

We can save it to the end.

No, that's fine. I'll drop it. Okay, here you go. Here's my World Cup hydration take. You remember previously, people sort of got annoyed at this, but I still stand by it.

The U.S. implements these policies in the stupidest way possible, right? So, as I said, we have universal health care: everyone has to be served by an emergency room. That made a lot of people mad, but I stand by it. The fact of the matter is, if you're in an emergency, you're going to get health care. That is universal health care in the stupidest way possible. “Oh, it doesn't cover preventive care.” Yes, I know. It's dumb.

Yeah, in a very narrowly defined way, it's universal health care.

Right. The point of making that analogy is that it's a really stupid way to accomplish policy. Or, if that was in the context of AI data centers, they're just going to end up paying people to build them, and that's how we're going to backdoor into AI putting money in people's pockets.

UBI.

Right?

Right.

Yes.

Yeah.

So, this is the stupidest way possible. This is the context for my hydration-break take.

Soccer is—and it is called soccer, by the way. Do you know where the word “soccer” comes from?

I have no idea. Where?

It comes from England, to be clear.

Okay, great.

Football was a sport played on your feet, as opposed to on horses, and then there was association football. S-O-C-C was soccer, came from England. I'm sorry that we preserved English culture better than they did and used the word soccer.

And we've been getting it wrong for 150 years. What the hell happened, England?

That's right. Okay, no, wait, that's not the take. Don't unsubscribe yet. Give me a couple more minutes and I'll really make you mad. Soccer has all these traditions, the way it's always been played.

One of them is that you have 45-minute halves, the clock doesn't stop, and that's just part of it. The whole thing with the hydration break is, well, the U.S. is really hot. Number one, come on. Get over yourselves.

It's not that hot.

Seriously. This is tied into the whole thing. Go to Taiwan in the summer. That's hot. I would come back to Wisconsin every year, and people would say, “How do you go to Wisconsin in the summer? It's so humid.” I'm like, “It's not humid at all.”

But it's not that hot. Whatever. Everyone's like, “It's so hot. We need to have some games where it's going to be too hot. Players need to stop for their health and safety and get water. If we do it in some games, it won't be fair. We have to do it for all the games.” What it's actually doing is giving us an extra commercial break in the middle of the game.

Right.

It's like a commercial break, but you're not going to change the channel because you're in the middle of the game. It's a huge moneymaker. Everyone's upset about it: “You're destroying soccer just to make a few bucks,” blah, blah, blah.

Here's my take: we're actually fixing soccer. One of the complaints about the hydration break is that the momentum changes. A team is on its heels—

The purists hate it because it's actually changing some of the outcomes of these games. Players come out of it—I think Haaland scored right after a hydration break. That's Erling Haaland, the player.

Well, it happened in the England-Croatia game yesterday. Croatia was on its heels. They came out, reset, and scored a few goals. And guess what? England was definitely the better team than Croatia. England was incredibly impressive in that game.

That game was so much more entertaining because the hydration break allowed Croatia to reset and score a couple of goals, which made it very gripping. England ended up winning 4–2. It made the game better. Guess what? Corners are a good idea.

Yeah.

Hydration breaks—a blatant sort of “grow up, it's not that hot out” combined with a pure money grab. We're backing into making soccer better, and everyone should thank us for it.

Fewer exhausted players is a good thing, I think. That's the kernel of the take.

Exactly. I like to see people performing at their best, uh, Suby.

Don't unsubscribe, but that's my take.

And don't sue Ben.

That's right.

My take on hydration breaks is that they should just be called commercial breaks, so maybe we'll get there by the time we get to the next World Cup. I do understand the instinctive rage at branding this profit-optimization strategy as a player-safety initiative.

That's right. There's a lot to be said for just being honest, okay? And guess what? You know what's good for the game? If everyone's making more money.

We're all adults here.

Yeah.

If everybody else has commercial breaks, it's fine. I hadn't considered the idea that it's actually going to make the second halves more entertaining than they would be otherwise. So, as far as I'm concerned, not a bad hydration-break take, soccer fans—

You know what?

—of the world, or football fans—

And if you're super excited for the game, imagine that the U.S. is playing Australia tomorrow. Maybe you've already watched it by the time you listen to this. Maybe you're pre-partying; you're getting ready. You know what you might need? A dehydration break. In the middle of a half, run to the bathroom.

There we go.

All good from my perspective.

Well, my number-one World Cup take—

Yeah.

You created a giant World Cup group chat, and there are 15 different people in there. You added me without asking me, so I've been in there for the past week.

Because you were dropping World Cup takes in other group chats, so I thought you wanted in.

I was dropping Champions League final takes a couple of weeks ago, not World Cup takes. But being in the group chat, I'm fairly busy. I've got 2 toddlers and a bunch of podcasts. You are busier than I am, though, and I am flabbergasted by how many World Cup games you've consumed over the past week or so. You continue to be the biggest sports fan I've ever met, so congratulations.

Well, thank you. I love the World Cup. It's great. Games are on all day. Did you see Xu Xi Jin, the Chinese propagandist?

No.

Oh yeah, he put out a tweet complaining about the game times, that they should be catering to China as far as when the games were played. The games are being played from 11 in the morning until 1 in the morning.

Okay.

Relax, okay?

9:00 PM Pacific.

Yeah.

Some of the games are clearly being catered to China, so China’s being taken care of here. And you’re watching almost every single one. I’m getting texts from you throughout the day about different World Cup games I didn’t even know were happening. So I look forward to joining you probably by the time we get to the knockout round.

Well, just in general, what’s great about the World Cup, I’m not actually a soccer watcher. Every World Cup, I think, “Oh, I should watch more soccer.” No, it doesn’t happen. The whole week, I’m sure it’s great. I appreciate the quality of play with teams that play together all year. I get it in theory.

Whatever.

I am all in on the eventization of sports. I love big events. It’s like tennis: I watch the Big Four. I don’t watch your regular tennis match.

Mm-hmm.

I will watch the big golf tournaments. I’m not watching it every week. I want events. I’m in on the events. I enjoy the atmosphere around the events and the social media around events, and I don’t want to deal with it the rest of the year, so it’s perfect.

There you go. Well, I would say I’m halfway in on the World Cup but all the way in on America, and I had a great time watching USA’s opening match and will be watching on Friday night in Seattle. But for now, Ben, we have news to get to. We did talk Anthropic and Fable last week, but as I’m sure 98% of the audience knows by now, late Friday afternoon the Commerce Department issued an export control directive to suspend all access to Fable 5 and Mythos by any foreign national, whether inside or outside the United States. So Fable was—

Or inside or outside of Anthropic.

Right, indeed. That was part of the problem as well. Claude was pulled from the market altogether and remains off the market, as you noted at the top. As we record this Thursday afternoon, that directive is in place. We’ll see how much longer it lasts. What do you think of where we are right now? It’s kind of a frustrating story to talk and write about because, on one hand, it’s the biggest story in technology. It also seems like there’s a good bit of nonpublic information that’s informing behavior here. So how do you wrap your arms around where we are?

It’s hard to wrap your arms around it, in part because of what you said there. What do we not know?

Mm-hmm.

It’s also hard because everyone in this situation is incredibly frustrating to me on a personal level.

Great.

So I actually feel like, almost this week, I took the opportunity to just write about Anthropic in general. We don’t need to rehash it because we talked about it on Sharp Tech last week.

Mm-hmm.

And that was one of those times when I actually got a fair number of requests like, “You need to write this out”—this idea of the sort of fusion of the safety culture with lots of choices that they make that are great for their business.

Business incentives.

Yeah, what a magical alchemy that is. So I wrote that on Monday. We already covered it last week. You can go listen to that. But Fable was a way into that, and I said it there. I’m not actually going to really talk about this because who knows what the heck is going on? But implicit in that—and anyone who’s been listening to these podcasts for a while knows this—is an overall frustration with Anthropic and a general concern about this company, not because they’re purposely being bad actors.

Mm-hmm.

But my overarching concern about people who think they are the only ones who are trustworthy and ought to make decisions for the rest of humanity is that it’s just bad.

The track record is not great.

Yeah, I don’t think that works. And that underlies so much about their posture in general. There was a write-up on this AI summit, or whatever, a day or two ago.

Yeah, G7.

And you could just see in Altman’s comments that he clearly doesn’t trust OpenAI. He kind of trusts Google. But it’s like, “We should decide.”

Mm-hmm.

And I think Altman took a much better approach, which is actually, “We shouldn’t be deciding. Someone else needs to be deciding.” That’s closer to my sensibilities, I would say, for sure. So in general, I’m frustrated with Anthropic, not just in this case.

Yeah.

At the same time, from what we’ve been able to garner, there’s a very good chance that the administration is just being incredibly stupid here.

Hmm.

And so I almost feel like I’ve been a little too hard on Anthropic this week because, from what I can gather—and again, who knows? We don’t have enough details about what’s happening or what’s not happening—there’s sort of an Occam’s razor explanation here, which is the US administration is like, “Stop it from doing this thing.”

Yeah.

And Anthropic is like, “You can’t stop it from doing this thing.”

That’s not how AI works.

That’s not how it works. The example here is, what’s the difference between “Find the vulnerability in this software” versus “Help me patch vulnerabilities in this software”? It’s the exact same request, and the reality is Mythos is clearly—if you used Fable, you could just feel it—it’s clearly ahead. It’s something new. No one has anything to match it. But if the history of the last few months and years continues, we’re going to have Mythos-level models in the near future.

Yeah.

And so what is the best possible thing we can do knowing that’s coming? Everyone needs to be working overtime to patch their vulnerabilities.

Mm-hmm.

And we’ve been talking about this for several months on Sharp Tech. It’s going to be a messy few years, and it feels like there’s this sense in the White House of trying to patch things, put your thumb in the dike against this wave that’s coming.

Yeah.

And the reality is, the wave is going to go over the whole wall. It’s going to be a mess, and the best thing we can do is—offense is the best defense. I don’t know if that’s the right analogy here. But go out there and work feverishly for the next little bit to fix as many vulnerabilities as we can before this is truly broadly available. And every day that this conflict is going on is a day when people aren’t able to do this.

Yeah.

And I think it comes back to this—and this is where I do bring Anthropic back into the conversation—how many people in the White House understand how AI works?

There are some people in the White House, for sure, who understand how AI works.

Mm-hmm.

But at the highest levels, this distinction between a probabilistic piece of software versus a deterministic one—

With a deterministic system, you could say, “You can’t make this available,” and, “This isn’t possible,” and you could hit some buttons to turn off different features that create unreasonable risks. You can’t really do that with AI.

Right. That’s like the magic of AI, right? The strengths and weaknesses apply to everything. The reason AI can do that—is it truly fully generalizable? I think not as much as the advocates claim, but it’s pretty darn generalizable—is precisely because it’s not deterministically designed to handle every sort of thing.

We see this in self-driving cars, right? The reason why self-driving cars took forever to actually be viable is because people spent years trying to hard-code every single scenario that a car might encounter.

Yeah.

There are so many variables in a car that you can’t do that. What you need to do is have an end-to-end model that is basically—not exactly the same as an LM, but the same concept of a neural net—that generalizes from being exposed to tons and tons and tons and tons of scenarios and develops its own sort of internal heuristics and probability distributions, such that when something new happens, it kind of knows how to react—

It can adapt.

—without being—

Yep.

—told what to do. And you think about this: it’s not just that it’s not deterministic, it’s that putting deterministic rules into that functionality makes it worse in many respects because you’re actually limiting it artificially in its capability to understand different situations. We’re just moving into this fundamentally new paradigm that’s totally different, and it’s really easy for people to look at AI as a fancy search engine.

Yeah.

We had this conversation a few years ago.

You used AI for the greatest-of-all-time talk and looked up some stats.

Oh, yeah. Which, by the way, it’s gotten a lot better at pulling recent stats.

For sure.

And everything else—it’s a lot more accurate than it was 2.5 years ago. But indeed, it’s a different sort of tool than traditional Google.

Right. And what I told you at the time was like, “No, that’s the exact wrong function that you’re trying to do.”

Yeah.

Right?

It’s the worst thing you could use ChatGPT for.

Well, then this gets into so many issues, or so many concepts that flow out of this. One point that I made again and again is: What is the shape of the crash going to look like, or the bubble popping?

Mm-hmm.

What it’s going to look like is people attempting to plaster AI onto existing workflows that were built around a certain way of computing, being frustrated that it continually doesn’t work. And then what’s actually going to happen is people are going to completely rethink their approaches.

This idea of moving from putting advertising next to an article to a feed—like the—

Yeah.

Like AI—

Create new companies built atop this technology to take—

That’s right.

—advantage of it.

That’s right. And so the challenge is, this takes years. It takes a long time for people to even reset their expectations of what’s possible and what this means. And in the meantime, what we do have is this ability to look at code bases, find vulnerabilities, and exploit them—basically autonomously.

Hack the planet. Just like Hackers. Great movie. So, to anchor people in this take, I think a lot of this springs from the report that Amazon was able to jailbreak Fable and raise some concerns with the White House—

Right, and no one knows for sure what this jailbreak is, right? There’s this analyst-researcher who saw a report from Anthropic, so this is all Anthropic spin, to be clear—

Mm-hmm.

—who characterized it as very basic: just slightly change your wording, and now it’s looking for vulnerabilities. Which I think, again, is an Occam’s razor explanation. You could believe that would happen.

Mm-hmm.

And then Anthropic is saying, “Well, look, that’s very minor. You can’t stop that.” And the White House is saying, “You guys are doing it again. We’re just asking you to stop it, and you won’t stop it.” And that’s where Anthropic’s at fault. They’ve created the conditions of mistrust with the administration.

And also, though, is the core issue for Anthropic that the way they discuss AI risk then creates unreasonable expectations for how it can be managed?

Of course. The whole Mythos rollout about this is like, they go on for years talking about AI being like nuclear weapons. Then they have this huge Project Glasswing: “This is way too dangerous to release. We’re going to secure the banks and secure the big tech companies.” And then 2 months later, “Okay, everyone can use it now.” And then you turn around and you’re creating the conditions for this reaction.

They create the expectation of perfect security, given how dangerous this weapon is for the whole world, and we’re going to be very careful. And then when it’s revealed that you can’t possibly paper over every possible risk that arises from this technology, other people who have been paying attention freak out and start to try to take back control. It seems like that’s some of what’s happened here.

Yeah. Well, the problem is that—I’m not minimizing the risk—it’s actually quite scary. Almost all software on the web is likely insecure.

Mm-hmm.

But the problem is that you’re not going to foreclose that forever.

Entirely, no.

Yeah. What is the administration buying with this? Weeks or months, maybe a year.

Okay.

The only way to solve the security problem is to go through it—

Mm-hmm.

—is to actually work and patch stuff like crazy. And by the way, the defenders have an advantage. They have the code.

Yeah.

They can examine their code bases. They can fix the problems. Even if you’re attacking something from the outside, it’s a little bit more of a challenge. It’s more obfuscated. You have to sort of figure it out.

The Manhattan Project we need is a whole-scale effort: let’s go through all this code on the web, all these open-source projects. Which, to Anthropic’s credit, they did do. That’s the sort of aspect of Project Glasswing. But we’re just not there. There’s just not an acceptance of this reality.

Yeah. Well, Michael says, “If Ben’s going to get so outraged about Anthropic’s actions even when he believes the safety threat is real, I’d love to hear a positive vision for what they should be doing to address safety concerns that wouldn’t be self-serving. I agree with Ben’s analysis about this being a convenient source of cultural alignment with the business incentives, but I just don’t have a problem with it. What am I missing?”

So, if you were running policy at Anthropic, what would you advise them to do to handle some of the safety risks?

The problem I have with Anthropic is the implicit and sometimes explicit message: “Only we can handle this. Only we are responsible. Only we are properly concerned about the dangers.”

And what’s implicit in that is, only we should control AI, combined with the rhetoric that AI is going to run the world. People got mad at the sentence in my article, but it’s just a logical construction. If A equals B and B equals C, then A equals C. They want to run the world. That’s the implication of saying, “We should run AI, and AI is going to run the world.”

Mm-hmm.

I mean, how explicitly have they ever said, “We should run AI”? Is it more an inference from their behavior?

Yeah. Their behavior and, I think, their rhetoric. They clearly don’t think Sam Altman should run AI, that’s for sure—which, by the way, I completely agree with, to be clear.

That’s true. Yeah.

And it sounds like, to Sam’s credit, even Sam agrees on that point at this stage.

Well, he’s cleaned up his rhetoric a lot. I think there was a lot of bad rhetoric that came out. And this is the challenge, because people are like, “Look, why shouldn’t they talk about job upheaval?” No, there’s the opposite mistake, which is to say nothing is going to happen, when you’re basically lying through your teeth.

Mm-hmm.

And I think there’s an aspect of that in Silicon Valley that says nothing is going to happen. No, it is going to be disruptive. Now, we might not know exactly how it’s going to be disruptive. You go back to the internet. The internet was incredibly disruptive.

Right.

The entire structure of society—

It took 25 years. We were the frogs in the boiling water for a while there.

This administration is the Trump administration, which is directly downstream from the structural changes wrought by the internet and what it did to media. Imagine going back to 1993 or whatever, when the World Wide Web comes out, and saying, “This is going to lead to Donald Trump being president.” It would have blown people’s minds. AI is going to be arguably even bigger.

But part of the problem is the certainty and prescriptiveness of what’s going to happen.

Mm-hmm.

There’s an aspect of this: things are going to change, and I think everyone could do with a little more doubt. What’s the word I’m looking for? Epistemic doubt.

Humility?

Humility about the fact that we don’t know what’s going to happen. This idea of just explaining why this is different—

Mm-hmm.

—appreciating that just because you understand AI and other people don’t doesn’t necessarily, one, mean you’re smarter than them, although often you are, to be fair, or two, that you know everything about everything because of this factor, right?

Yeah.

You’ve been in conversations with people where they’re explaining something to you, and their tone and approach is so patronizing, and it just is infuriating, and you’re like, “Screw this guy. I don’t care.”

Are you describing conversations that you and I have had about various consumer technology that you demand that I adopt and then I stubbornly refuse to?

Right, and you want to punch me in the face.

Yeah.

Maybe I can identify this attitude because I can personify it at times. That’s a fair pushback.

I think the emailer does have his finger on something. When I was at Microsoft, I remember I never got promoted, and I was very frustrated about it.

Mm-hmm.

I ended up on this team where I was actually a level 61, and everyone else on the team was level 63. I was doing all the work, and I thought I should get promoted. But the one piece of feedback I remember getting at my year review or something like that was, “Your what is excellent. Your how needs a lot of work.”

There was this idea that it wasn’t just my manager that had to be involved—the other managers on the promotion committee, the manager above the manager, and all those people.

That’s amazing.

There’s an aspect of corporate politics that is just putting on a happy face, talking to people, and selling yourself and your ideas. This is why I didn’t last long in corporate America. I’m much better off, I think, on my own, to say the least.

Yeah.

There’s an aspect here to my feedback to Anthropic: your what is fine. It’s the how that needs work.

Right.

You need to figure out how to communicate these things.

And to describe it, there needs to be a capacity for risk tolerance over the next 4 or 5 years, and there’s not much of that coming from Anthropic in the course of these conversations. They create the illusion of perfect security as long as they’re the ones in charge of that security, and that’s going to be a problem long term.

I’m open to the possibility that members of the Trump administration simply don’t understand that you can’t keep these models from being jailbroken, and you can’t reduce the risks on that front. I will also note, though, that there was that report in The Washington Post about Anthropic giving the government a list of 111 organizations that had advanced access to Mythos, and then there were an additional 50 or so entities that were added and had already received access.

Anthropic was slow to get that list of companies to the government, as the government was asking, “Who actually has access to Mythos?” One of the entities was a South Korean company. What’s interesting in that report is it seems like Anthropic still doesn’t understand that the government is ultimately in charge, and it’s really not that complicated. Dealing with the government shouldn’t be as impossible as Anthropic is making it look.

Yeah.

So again, blame belongs on both sides of the equation here.

This goes back to the Department of War debate, which ultimately is a philosophical question of who’s in charge. The fundamental problem with Anthropic’s how is actually not just the how; it’s the substance.

In that debate, their red line was, “We get to decide how this is used.” So, as you asked earlier for evidence of them saying they should be in charge of AI, there’s your example.

That’s the evidence. Yeah.

They’ve already demonstrated it. And by the way, we talked about this last week: they’ve also demonstrated the capability and willingness to silently nerf their models to stop behavior that they don’t think is acceptable.

Yeah.

This is the core issue. There’s an assertion of authority that is reserved solely for them, and it’s not just with their customers, it’s not just with other companies, it’s with the federal government. That is the fundamental issue.

Mm-hmm.

It actually does go back to that episode.

Yeah.

That’s the issue here: they’re trying to dictate terms, and the government is probably being dumb. The export-control thing—who knows if it’s even legal or whatever it might be? That’s for the courts to decide. My guess is probably not, right?

Mm-hmm.

What’s actually happening is the exact same battle that was happening before.

Yeah.

Who gets final say?

And when the government says jump, Anthropic should say, “How high? You want the additional 50 entities? Here are the additional 50 entities. You have a problem with this South Korean entity?”

How’s that boot taste, Andrew?

Well, no, that’s just the way it’s going to work.

Right.

And so—

No, that’s the thing.

It’s inevitable. If you’re talking about a superweapon that poses dangers to the entire world, this is the inevitable conclusion of that conversation. But they still have not fully grokked that—

And that’s what’s so frustrating: the people who are implicitly saying they should decide everything seem incapable of understanding the reality of the world as it actually is. If you can’t understand the reality of the world as it actually is, on what basis should I have faith that you get to decide how the world is going to be?

Mm-hmm.

All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at email@sharptech.fm.