（预览）Anthropic 乱局仍在继续：前沿 AI 与法律的不确定未来，Netflix、约会应用、F1 问答

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, it's great to see you.

I'm doing okay. It looks like you just leave, go on vacation—leave me to my own devices, and I just go out and get canceled, apparently.

Can I tell you what? Nothing on my vacation was more relaxing than not recording a Friday night emergency episode about Anthropic versus the Trump administration.

Oh, yeah, we're out of emergency episodes.

Yeah.

We're not doing those anymore.

We may be post-emergency episode on this podcast, but we are going to talk about Anthropic and the Trump administration on this podcast. One week ago, for anybody who's missed it, the Trump administration canceled a defense contract with Anthropic and announced a plan to phase Anthropic out of all government institutions within the next 6 months or so. That will take some time.

If it happens.

Yeah, TBD on all of this. It's changing by the day here. Pete Hegseth also threatened to label the company a supply-chain risk to national security, potentially implicating many more Anthropic contracts.

Well, he went beyond that. He said that no one who has contracts with the U.S. can work with Anthropic, which goes beyond even supply-chain risk. It's a very extreme measure.

Serious designation, yeah.

It also does not go that far. So there have been some definitional issues all along the way.

That was broadening the scope of the potential implications, at least in Hegseth's tweet, and they have not followed through on that threat 1 week later. So again, TBD on a lot of this. For now, we will talk about everything you wrote earlier in the week. I wrote about it for Friday as well.

We do have to start with the updates as we record Thursday afternoon here. I'll read from The Information, which republished an internal memo from Dario Amodei to Anthropic employees. It was a long memo. This is not the whole memo, but it did culminate with this:

"The real reasons DoW and the Trump administration do not like us is we haven't donated to Trump, while OpenAI/Greg have donated a lot." That's Greg Brockman, for anyone keeping track. Dario continues and says, "We haven't given dictator-style praise to Trump, while Sam has. We have supported AI regulation, which is against their agenda. We've told the truth about a number of AI policy issues, like job displacement, and we've actually held our red lines with integrity rather than colluding with them to produce 'safety theater' for the benefit of employees."

In parentheses, he says, "Which I absolutely swear to you is literally what everyone at DoW, Palantir, our political consultants, et cetera, assumed was the problem we were trying to solve. Sam is now, with the help of DoW, trying to spin this as we were unreasonable, we didn't engage in a good way, we were less flexible, et cetera. I want people to recognize this as the gaslighting it is."

Then, just to keep all of us on our toes, about 4 hours after that memo leaked, we got a headline from the Financial Times: "Anthropic chief back in talks with the Pentagon about an AI deal." So what do you make of that?

One important detail about this memo is that this was a memo he wrote last Friday. It was not a memo he wrote yesterday.

In the heat of the moment.

Yes. The memo was published on Wednesday, but it was written in the heat of the moment. Is the memo a good look? No, it's not. Does the memo evince an accurate theory of mind for everyone that Amodei is interacting with? I don't think so.

I found this memo quite validating, actually, in terms of what I wrote. I have a principle: assume everyone is smart and figure out why they're doing dumb things. No one is denying that Dario Amodei is smart and brilliant, and everyone at Anthropic is.

Mm-hmm.

And yet I've felt for ages that they and this broader movement have had a significant blind spot, which we talked about in our first podcast of the year when we were going to the dorm room.

Mm-hmm.

And that blind spot is this entire article we talked about that Dorish Patel wrote about the future, when AI does everything and humans have no worth. Everything in the world was different except for modern conceptions of property rights, which were still exactly the same.

Yeah.

That was one of the pushbacks we had. You almost feel silly pointing this out, and we're going to get to a couple of emailers. I know I've already gone in on you for similar reasons here.

But there's just this weird blind spot that is consistently there, and I feel like it was at play in this case. It's like there's an assumption that certain rules and laws are immutable, with no conception of how laws come to exist.

It's kind of funny because the first article I wrote when ChatGPT came out was an article called "AI Homework."

Yeah.

It opened with a question about Hobbes versus Locke, and ChatGPT got it wrong.

Mm-hmm.

I thought that was interesting. It sort of started the initial hallucinations—obviously a bad thing, but also amazing. You can understand how that particular hallucination happened, because Hobbes and Locke are always referenced together in textbooks and it basically confused their 2 views. This was ChatGPT 3.5, right?

Yeah.

They would do fine on that today.

This was very early on, and it's one of my favorite articles of yours.

Well, thank you. I think that's a compliment.

Maybe we could all go back to Hobbes and Locke, or just political philosophy in general: Where do rights come from? What is the foundation of authority and government? Hobbes was sort of on the side of the monarch, and Locke was more on the democratic will of the people, but even then it goes back to: What's the foundation of rights?

Totally. I noted that in my article. The idea of property rights, which are supposedly implicated by some of what the U.S. government is trying to do to Anthropic, are themselves a legal construct that is given force by the American government that allegedly upholds those property rights.

Exactly. When I read this memo, it's interesting because I think one of the frustrations I had when I wrote what I wrote—I didn't expect the degree of blowback that it got, but that's fine. I stand by it.

Actually, I'm going to bring up the blowback a little bit because there's an analogy to a previous article I wrote when we talk about your piece in a little bit. You certainly have people who only see things through a partisan lens, which is very frustrating.

The whole point here is that this is not a partisan issue. I think it wouldn't matter who's in the White House. Maybe today it matters, but in 10 years, if AI is as powerful as it potentially is, it's not going to matter. It's going to be the government versus the AI companies.

There's a bit where I'm actually fairly grateful this came up in the interview I had with Gregory Allen this week, which I highly suggest everyone listen to. I thought it was really, really good.

Yeah.

It fleshed out some of these issues, which he doesn't completely agree with me on. We went back and forth on a few different things.

I'm quite grateful this debate is happening right now in 2026, when, all things considered, the stakes are pretty low. The stakes have the potential to be drastically higher in a very short amount of time, given the exponential increase in the power of AI.

There are actual, real, fundamental questions at play here that need to be discussed. I don't think it's a partisan issue, and yet, of course, there are people for whom it's a partisan issue, right?

Sure.

What's frustrating is they don't even attempt to grapple with the article. It's just, "Is this broadly in line with the Trump administration or not?" Then they get insulting and call me a fascist and a bootlicker, or whatever it might be. It's sort of easy to dismiss.

You're only a bootlicker when we're talking about antitrust, just to clarify for the audience out there.

That's right. I'm a bootlicker for our corporate overlords—

That's right.

—not our political ones. Someone's going to clip that.

Yeah.

It was a risk on my part, but here we are.

What I was frustrated about with the response was, again—for just categorization purposes, we already have an issue with the Department of War versus the Department of Defense.

Mm-hmm.

Anthropic used Department of War, so we’ll use Department of War.

And for everybody out there, I believe there is some sort of congressional threshold that needs to be cleared before it’s officially the Department of Defense. But everyone involved in this mess is referring to it as the Department of War. It is a bit of a minefield in D.C., depending on who you’re talking to.

Right.

You have to choose your nomenclature carefully, but Department of War for the next hour here as we talk through all this.

Yeah, I know. I was just using Department of War because Anthropic did. I wasn’t really thinking about it, but I just assumed that was the case. And then some people are coming at me: “Oh, you’re using State Department.” I’m like, I honestly did not give two shits about it.

Relax.

So I’ve just been saying “the U.S. government” ever since, but whatever it might be. I’ve tried the first sentence in my article—the quote, “You might not care about politics, but politics cares about you,” or whatever it is.

True for Sun Tech as well.

That wasn’t just about Anthropic. That was also about me, right? To be clear, what I found disappointing was—

The other categorization that I’m going to use in this podcast for ease of use is EA, the effective altruists, from whom a lot of the AI researchers come. Anthropic is kind of the standard-bearer among AI companies for that. There are a gazillion definitions of who it is or whatever it might be. This is a very blunt instrument. I’m just using it for terminology purposes.

Okay.

But I think one of the frustrations I had with the EA critiques is that they sounded a lot like the anti-Trump critiques.

Mm-hmm.

It was like, “Did you actually read the article?” Instead, there was just this visceral reaction.

I’m like, I addressed that. That was in there.

So how so? What were they objecting to in terms of the EA side of things?

Oh, well, no, actually, a lot of them just fell into Trump-like things. Like, “Oh, he’s trying to back into a conclusion here because he wants to support X, Y, Z.”

A lot of them really fell into this “property rights will last forever” sort of categorization. My whole point here is that laws are ultimately human constructs, and they come back, at the end of the day, to this: laws are formed, nations are founded on the tip of a bayonet.

Mm-hmm. I was going to say, Hobbes and Locke. Here we are.

That’s right. Oh, and Hobbes in particular. Given that, a lot of people—this is where I brought up Palmer Luckey: “We should let Democratic leaders decide,” because ultimately they have the legitimacy. I was not making that argument.

Okay.

Because what I was talking about was the risk that we face here: AI becomes so powerful that it becomes a source of power in its own right. It becomes the ultimate gun.

Mm-hmm.

The ultimate—the sharpest of the bayonets. And when that’s the case, the problem is that the temptation—or call it a felt necessity—of the government is going to be to violate property rights, to leverage its monopoly on violence, not just because it wants the technology, but because it actually fears it’s losing its monopoly.

Yeah.

And that’s antidemocratic. My point was not a democratic argument. My point is that the risk we are running if we don’t get alignment here between AI companies and the government is that we’re creating the conditions for the government to turn antidemocratic.

Mm.

To violate the laws. The laws exist and are enforceable because the government enforces them. What happens when the enforcer is the one that decides the laws don’t matter and need to be violated?

Right. We’re seizing—

Who do you call?

—your property under—

Right? Are you going to call the international police?

—the DPA or something like that.

There are no—right. And so this is actually something I wish I’d made clearer. I think I did in that Gregory Allen interview, sort of fleshing this out.

I almost regret even bringing in the Palmer Luckey bit, because I did say in there—this is the thing about reading it—I’m like, “This is not my argument.”

Yeah.

What I should have added is that it’s not just not my argument; it’s that my concern—and it’s so funny that people are calling me a fascist—my concern is fascism. We’re creating the conditions for the government to violate its own laws—

Mm-hmm.

—to obliterate the concept of property rights precisely because it fears the power of AI more than it fears the ultimately human constructs of laws and the self-control to abide by them.

So let me tease out what you’re saying there. Are you saying that your fear is that if there isn’t close coordination with all of the frontier AI labs, these labs will then be nationalized and weaponized by the state without any sort of input from those labs?

Yeah, and that’s arguably an optimistic outcome.

Mm. Okay. So there’s sort of a techno-dystopian hellscape that could emerge if the government—

Yeah.

Yeah.

At a very fundamental level, the government has a monopoly on violence.

Mm-hmm.

Gruber brought this up in Dithering, this sort of idea. I think he said it was P. J. O’Rourke. I wasn’t familiar with the article, but the idea is that, at the end of the day, you pay taxes because the government has guns.

Yeah.

None of us think that way. We just pay taxes because that’s what you’re supposed to do. But ultimately, what happens if you don’t pay taxes?

They come and take your house, and you go to jail.

Right. And who comes to take you to jail?

Probably someone with a gun. Yeah.

People with guns.

And this is why I thought the Iran— that’s why I started with the bit about international law and Iran as a useful analogy, I thought.

That was a sensitive issue on Monday morning, as everyone was processing the world.

Yeah, I know. And maybe there’s an element there too—

It probably didn’t do you any favors as far as the backlash is concerned, but yeah.

But the issue isn’t just that the government has a monopoly on violence. The problem is that if AI is what it is promised to be, the government doesn’t just want it for its own purposes; it will feel threatened by it.

Yeah.

Again, is it threatened by it today, in 2026? No. If these exponential curves actually play out, is this potentially going to happen sooner than we think? Yes. And so, if we want to retain our system of government, if we want to have democratic control, we ought to, all of us as an industry, be thinking about how not to create the conditions that make the government feel oppositional to and threatened by AI.

Because last week the government felt oppositional and threatened by AI, and fortunately we’re still in a place where we can laugh at it and say they overreacted, they overdid it.

Yeah.

But what if AI was 100 times or 1,000 times more capable and powerful than it is today? And that might be here pretty soon.

Indeed. Well, before we get to more about what you wrote last week—or this week, excuse me—two questions on Anthropic’s next moves here.

First, Scott sent this in after the memo leaked. The subject line was “Strategy credit overdraft.” He writes, “Hello, gents.” One of Ben’s hottest takes recently was that his concept of a strategy credit is his greatest idea, even better than aggregation theory. I can’t disagree. Is it possible, though, to exhaust one’s credit or overplay the strategy-credit hand?

“I wonder if Anthropic has a strategy credit in saying no to the Department of War to further sell itself as the ethical AI shop, better to attract talent and consumers. But could this strategy credit lead it so far in one direction that it undermines its core business? It feels a bit like a politician winning the primary but going so far left or right that they can’t win the general election.”

What do you think about that possibility?

Yes. But I actually don’t think it undermines their core business. I think it probably, by and large, helps their core business.

1. You have the Pentagon saying, “We’re going to ban you because we want you so much, and you won’t help us, and your technology’s so awesome.”

Mm-hmm.

Mm-hmm.

You have the talent wars, where I'm sure, by and large, across AI talent, people are much more sympathetic to and aligned with Anthropic's views.

Yeah.

I would guess they'll gain more talent than lose anything in the next couple of years.

Right. You have this—they've been number 1 in the App Store, I think, all week, right?

Mm-hmm.

People deleting OpenAI, going with Claude, and maybe, for the first time—and maybe it turned out those Super Bowl ads, which I thought were somewhat deceptive, but whatever—all ties together, and they're building this great brand. The problem here isn't that they're undermining their core business.

Mm.

The problem I have is they're becoming unaligned with our democratic government. Right? The problem is—

You know, that does undermine their core business. That's the risk that I see as far as the next year or 2 for Anthropic here.

No, my worries are much more basic than that. I'm worried about them undermining our democratic government and our way of living. The means of that isn't necessarily Anthropic acting badly; it's prompting, putting in the incentives for the government to act badly.

Hmm.

Again, the great irony about this is, particularly on the domestic surveillance point, I have tremendous sympathy and, by and large, am aligned with Anthropic. I think the holes, the loopholes we have for domestic surveillance, are fairly atrocious. I think a lot of our privacy policies were predicated on a world where there was friction—it was hard to spy on everybody—and we've already blown past them to a great degree because computers can wiretap everyone.

Mm-hmm.

Just one of those secret closets in the AT&T building. Don't ask. And now we can analyze all that at tremendous scale. I think it's a totally legitimate concern, but everything is a trade-off.

Yeah.

I just don't think this was a fight worth picking, particularly. This is where we don't know what the details were. You can't have a situation where, on a case-by-case basis, you get to have a say.

Right.

Let's get as clear of guidelines as we can, so that the lawyers in the DoD/DoW can decide go or no-go, and they don't have to ask you for permission. Let's do that not just because we support national security, but because we are developing potentially the most powerful technology ever.

Mm-hmm.

We need to appreciate the threat that is going to pose to not just people's jobs, not just SaaS companies, but to the very fundamental legitimacy of the United States of America.

Indeed.

The other funny criticism I got was, “Oh, yeah, Ben just wrote that because he doesn't take seriously AGI or its power.” I'm like, that's the exact opposite. It's because I'm taking you seriously.

Right.

If I was convinced—I could see the arguments—that this is just another technology, just like computers, and it's actually not going to be that big of a deal, in which case, let's unapologetically support them, because we saw the abuse that happened with computers. But I am taking seriously the possibility that this is unbelievably powerful and destabilizing. For that exact reason, it's super important that we be aligned with our democratic system of government.

Well, related to this, one more question on Anthropic's calculus in the short term here. Adam says, “Stripped down, the analysis in your post on Monday is pretty banal. Anthropic can't and shouldn't be able to dictate its terms of service with the U.S. government because of obvious power dynamics. Here's the thing: No firm in Anthropic's position would be unaware of the U.S. government's ability to force Anthropic's hand. Forget the supply chain risk designation for a minute. The real hammer the U.S. lords over Anthropic is the Defense Production Act. It's not quite all the marbles, but it's most of them. So why, then, has Anthropic performed such very public line-drawing over the last week? What's the endgame?” Any guesses on the endgame and/or why this has all played out in public the way it has? I don't think that has been strategically optimal for Anthropic.

I don't know. I can speculate, but I'm hesitant to. I think one of the critiques that can be leveled at maybe the EA movement—and where I do feel a certain sort of allergic reaction to them—is this idea: “We are the smartest people in the world, and we're the only ones that truly think through costs and benefits.”

Mm-hmm.

“Everyone else is kind of stupid.” That's one thing, but it sort of follows on: “Therefore, we should make decisions for everyone.”

Yeah.

One of the big issues I've had with the AI labs in general—and this goes back to OpenAI being a nonprofit, and certainly applies to Anthropic—is this assumption that AI must be used for the good of everyone.

Mm.

The question is: Who decides? When you push on it, it's always, “We do.”

Mm-hmm.

I am very allergic to that point of view. I think that leads to hubris. It leads to very dystopian outcomes. It leads to the worst abuses we've seen in human history. The worst actors are not the ones who are attempting to be evil; they're the ones who are so convinced that they are good—

Yeah.

—that everyone who opposes them ought to be eliminated for the greater good.

Yeah. Well, and it's funny—

I'm not saying they're there. I'm just saying it's shades of that.

That's the risk. Yeah.

It's shades of that.

You can see shades of it in the memo. There's a fervor permeating that memo that made me upset.

Yes, and that's why I said the memo felt—maybe felt—justified. This is not fully grappling with reality.

Yeah.

Is Sam a tricky guy and an operator? Yeah, I think we've figured that out. But there's a fervor here, and I find fervor fairly terrifying, particularly when it's being evinced by the leader of, again, potentially the most powerful technology ever.

Mm-hmm.

That's just unsettling. I find it very unsettling. I do want to note, I thought the analysis was pretty banal also, right? This didn't rise nearly to the level of—but I would say there was at least a shadow of the out-of-pocket response I got to net neutrality in the first year of the Trump administration.

When I've told this story, I wrote about net neutrality in the context of Netflix versus Comcast in, like, 2014 or something. I basically said Netflix is full of BS here in terms of their peering agreements and stuff like that. I basically made the same argument in 2017 when Obama, in I think it was 2016, the previous year, imposed Title II of the Telecommunications Act on internet data.

Yep.

It imposed net neutrality, but it came with this massive amount of bureaucratic paperwork that was basically designed to apply to AT&T back when they did long distance. My point in that article was that I'm definitely pro-net neutrality, but this is the worst possible way to do it, and it's going to restrict investment in broadband. It was a very technocratic article that I thought was quite banal.

Mm-hmm.

For which I got not only a few hundred replies here, but thousands—tens of thousands of replies. I got death threats. I got the whole kit and caboodle. I had people like—take Nilay Patel, who this week called me a fascist on Bluesky—at that time writing entire screeds about how I don't care about democracy, this whole sort of thing.

Hmm.

It's funny: If you go back to that article, I quoted several tweets. They're all deleted because every single person who lost their minds about that was totally wrong. Not just that, the vast majority of them changed their position. They didn't want neutrality. Suddenly, they wanted platforms making decisions about what was and wasn't allowed.

Yeah.

I appreciated your piece this week on Sharp Text, which didn't get into the issues as much as saying, “These reactions are out of control.”

It's pretty exhausting. You know what the common thread was between your piece on net neutrality—which, as I have said in the past, I cited in law school, got an A+ in the class on the final exam, and cited Ben Thompson before I even knew you at that point—

That’s right.

But the common thread, I believe—didn’t Ajit Pai share that piece on Twitter?

Pai, yes. Yes.

Yes. A member of the Trump administration—

Well, no, so I deserve blame—

—shared this week’s piece as well.

—because I originally entitled it, “Why Ajit Pai Is Right.”

Oh, great.

That title was a mistake. The title is no longer that. It is “Pro-Net Neutrality, Anti-Title II,” or something very dry like that.

What a fun, spicy title there.

You wonder why my titles tend to be fairly dry.

I’m getting death threats. Yeah.

Yeah. But yes, and then Emma Michael retweets my article on Monday.

Yep.

Not helpful. I’m not on a side here. I’m just trying to make a point, but yes.

That’s right. Yeah. Well, and as far as Anthropic’s endgame, all I will say is that I wrote in my piece I would expect them to broker some sort of compromise in the midst of all this.

And I think that’s going to happen. It is possible this happens even before this podcast publishes.

Okay.

I think he was very—I think he settled down. He was at the Morgan Stanley TMT Conference, and I think was fairly conciliatory in his tone.

Yeah.

I’m hopeful that maybe, in some small way, what I wrote is a little bit of, “Let’s wake up here. This is—let’s not go down this route.”

Mm-hmm.