Questions about Anthropic vs. the U.S. Government | Sharp Tech with Ben Thompson

0:00 Okay, that makes sense to me.

0:02 Um, Matt and many others asked a question along the same lines.

0:07 They said, "Why?" Matt says,

0:08 "Why would a private business not be able to set the terms

0:11 by which they do business with the US government?" This is freedom of contract,

0:15 which Ben has spoken about several times before.

0:18 Always in favor of businesses deciding who

0:21 they will deal with and on what terms.

0:24 How do you think about that particular question?

0:27 because I do think dealing with the government here is

0:31 different because of the implications of the technology at issue.

0:36 I mean again I I I'm pro contract.

0:39 I'm pro the government abiding by the contract.

0:42 the I I don't know what else to say like

0:46 like it's precisely because I agree with the emailers that I

0:51 would like Anthropic to back down here or to find

0:54 some face saving way or within the limit of the law.

0:57 I in I say this as someone

0:59 who agrees with their concerns about about surveillance.

1:02 You risk though by pursuing a principle that I agree is above

1:09 and beyond the law is both right and above and beyond the law.

1:12 I do think the current law is

1:13 insufficient in protecting the privacy of Americans.

1:18 Yeah,

1:17 I agree with anthropic there.

1:19 But everything is a trade-off.

1:21 And my concern is by imposing your standard,

1:25 which goes above and beyond the legal standard,

1:28 you imperil all legal standards, including property rights.

1:34 Fair.

1:34 Which I would like to preserve.

1:35 I'm very pro property rights.

1:36 I'm pro contract.

1:38 Uh I am the Remember, I'm the corporate overload bootlicker.

1:41 Uh you made this point, by the way, like you're the antitrust guy.

1:44 You're like, "Look, antitrust is about violating property rights,

1:48 right?" I'm [laughter] like, "Yes, it is.

1:48 Thank you.

1:49 50 years.

1:50 Um, yeah.

1:51 Well, and it's I just my understanding of companies

1:55 and their relationship to the government is probably

1:58 informed by my legal background and going to law

2:02 school and reading cases from the last 150 years.

2:05 In matters of national security,

2:07 there's a lot of difference given to the United States government.

2:10 Another good thing in your article,

2:11 you listed a whole bunch of examples of the US pretty

2:14 blatantly violating private company property rights

2:16 when it comes to national security.

2:18 So, it's not particularly new.

2:20 And when you talk about a technology that is

2:22 this important to the future of military technology,

2:26 it's not particularly surprising to me that the government

2:29 would put its thumb on the scale and expect

2:32 partnerships with all of the frontier labs and expect

2:35 that those are not going to be equal partnerships.

2:37 Um, so I not and look, a term a contract with the US government,

2:42 the government is going to retain the power

2:45 to change the terms of that contract.

2:46 that's just sort of the way it has always worked.

2:49 Um,

2:49 well, I think I mean I think it's very fair to say like I so as I understand it,

2:53 correct me on the timeline,

2:54 the initial anthropic contract was with the Biden administration,

2:57 but the Trump administration renewed the contract

2:59 last summer when if they had these concerns,

3:02 they should have brought them up then.

3:04 Um, so I think I think all those are sort of valid critiques.

3:07 Um, and I'm not defending the Trump administration.

3:11 I and I think the it just like it bothers

3:15 me that Anthropic was dishonest in the Super Bowl ads.

3:18 It bothers me that Secretary has it tweet was

3:21 asserting like things that are not powers he doesn't have.

3:25 Yeah, totally.

3:26 [laughter] Like like so and yes, I get the arguments, oh, you know,

3:31 Trump always does this, asks for the moon, gets a street light, whatever.

3:36 I I can acknowledge that and also not like it.

3:39 Um, and in this context, you know, is very damaging and and

3:43 you know, Gregory Al, it's just not surprising to me that they

3:46 that the government would reach for any lever

3:49 it has and marshall every tool it has

3:52 to ensure that these companies fall in line in

3:55 this all the stakes here are quite low and small,

3:57 which is actually I think a relief because I do think

4:00 like I followed up on Tuesday by pointing out why is it

4:05 that you know like oh the government can make their own AI

4:07 like no they can't just like they don't make their own chips.

4:11 One of the and this is about how technology works.

4:16 When the upfront costs are so substantial

4:19 and the marginal costs are basically zero,

4:22 which like chips is really the first one that's that that that's like

4:25 this to actually make money and to actually not just to like be profitable,

4:32 but to drive R&D, to drive innovation farther,

4:35 you need larger much larger markets than the government can provide.

4:38 And so, you know, Bob, I talk about Bob with wi with Fairchild semiconductor.

4:42 This started out with government contracts.

4:44 The government's like,

4:44 "We got all the IP and we're going to tell you what to do."

4:46 And like most companies did that because they need the money for R&D.

4:49 He's like, "No, we're going to sell chips

4:51 to the consumer market because it's going to be

4:53 much larger." The larger it is because

4:55 of this concept of zero marginal cost, high fixed costs.

4:58 The more you sell, the lower the price.

5:00 Uh like you make it up in volume.

5:02 Like that's the makeup in volume is sometimes a joke but it's rooted

5:06 in a very real concept of technology which is you can almost always make up

5:09 in volume if you have these dynamics of high high upfront costs and zero

5:13 marginal cost because the volume doesn't cost

5:15 you anything because it's zero marginal cost.

5:17 So you spread out your fixed cost over more and more

5:19 things and and within it took a very small amount of time.

5:23 By like the late 1960s there were more computer

5:27 chips being sold to non-government entities than to government entities.

5:31 And obviously it's not even remotely close today.

5:33 This is the late 1960s and Fairchild dominated that.

5:36 They sold like 80% of that market and uh and then they left formed

5:41 Intel because they wanted to actually make money.

5:42 Fairchild would give them stock options.

5:44 Uh and so uh but it pulled the innovation forward as well because

5:49 what the US government and so what happened was in the 1980s the US

5:53 government realized like we're actually our military is in in pretty poor shape.

5:56 we're way behind.

5:58 And they're like, we need to stop dictating and start buying.

6:03 And actually, there's been this transformation information technology.

6:07 We need to buy chips off the shelf.

6:10 And that was the foundation of of all the technology you saw like

6:13 the the last real demonstration of US military might was the Gulf War.

6:17 Uh and that was downstream from the 1980s.

6:21 uh where uh ba basically based on um you know

6:25 from the technological revolution in Silicon Valley in the 60s

6:29 and 70s and they were buying off the shelf

6:31 and and and AI takes this on steroids, right?

6:35 Like like the US government what's the DoD's like

6:38 the the amount of capex that uh what's the DoD's budget?

6:42 Uh

6:43 oh, I don't know.

6:44 I can look it up while we talk here, but all I know is it's

6:47 not the same as anthropics.

6:49 Uh, no.

6:51 Well, that's the thing.

6:52 Uh, DoD budget, this is great podcasting, but 2025, $850 billion.

7:00 Okay.

7:00 Um, the the tech industry is committed to $700 billion in capex for this year.

7:05 So, same ballpark.

7:07 Not just saying ballpark but on on the we're

7:10 on the verge of spending more every year on AI

7:15 than the department of defense which has always been held up war

7:18 whatever has always been held up as the biggest spender in the world.

7:24 Yeah.

7:24 And that's for they're they don't have zero marginal costs.

7:27 They're paying all the troops.

7:28 They're paying all the like all the staff like all that sort of thing.

7:31 Not all of $850 billion is going to technological development.

7:35 And I'm not sure.

7:36 This is where I think Washington needs to wake up.

7:41 Mhm.

7:39 I don't know to what extent do people in government understand

7:46 how little they matter to tech companies from an economic perspective.

7:51 Yeah.

7:51 I mean $200 million contract is a drop in the bucket for anthropic

7:55 which is currently getting into enterprise market

7:58 making all of gains on consumer side.

8:00 Yeah.

8:01 So, so the the problem is but this is why this issue is so important

8:06 because ideally like the government used to be

8:11 able to guide lots of companies with carrots.

8:14 It's like you want this government contract uh do do what we say.

8:20 Their carrots are not just baby carrots.

8:23 They're like slivers of a baby carrot that honestly

8:27 is just so hard to pick up with your fork.

8:29 it's probably not even worth the trouble to eat.

8:31 Like, and if it's not there yet,

8:33 it's getting there very rapidly with AI with these astronomical upfront costs

8:38 and not zero, but close to zero marginal

8:40 costs as far as serving in the long run.

8:42 And and so

8:44 a little baby carrot next.

8:46 This is why people have to be worried about this.

8:50 If the government doesn't have carrots, all they have are sticks.

8:54 Mhm.

8:54 I would prefer the government not getting

8:57 in a position where they feel we have no carrots,

9:00 all we have are sticks, so we're going to use that.

9:02 That's what happened last week.

9:03 That that tweet was not just a stick.

9:07 It was also like someone's like putting a light behind them so

9:11 the shadow looked very large so the stick looked bigger than it was.

9:15 Yeah.

9:14 Like it was kind of it was kind of pathetic actually in a certain

9:17 respect the extent to which Hexath overstated what he could even do.

9:21 Mhm.

9:23 Like this and these are the people

9:26 that are supposed to be worried about alignment.

9:28 That's what's so frustrating about this.

9:29 How can you not see this collision that's happening?

9:33 And what happened last week was not a oneoff.

9:37 Just wait till the administration changes.

9:39 It is a precursor of what's going

9:43 to happen as AI becomes more and more powerful.

9:49 Yeah.

9:49 I mean, I understood it as the government

9:51 setting a precedent for any of these AI labs.

9:54 Um, but I guess more will be revealed in terms of tension in the future.

9:59 Related to HEG, we got this email from Craig.

10:02 He says there are two actors here.

10:04 One which wishes to abide by the previously agreed to terms

10:08 of its contract or barring that help with the transition to a competitor.

10:12 The other expects the right to ad hoc change contracts

10:16 or barring that will ignore the laws to implement punishment.

10:19 Which one is pro-democracy?

10:23 So the answer the anthropic one is pro-democracy.

10:26 Again this is the this is the my my issue.

10:29 This is why some of the response were driving me bonkers.

10:32 Like yes the the following contracts

10:36 and or helping transition better that is pro-democracy.

10:40 That is pro property rights.

10:41 that's pro all the things I believe in where but there's this like assumption

10:48 that democracy is in the ether that property rights are these things that are

10:56 notable

10:56 and honestly I mean from my perspective

10:58 Pete Hexith is not undermining democracy by exercising

11:03 the government's own authority and leverage

11:06 in this area like the system of government

11:08 is a democracy and the leaders who are elected and abuse their power have

11:14 the ability to be voted out of office

11:16 if voters are unhappy with what they're doing.

11:19 But the government acts in undemocratic ways using democracy to stand

11:25 in for like the rule of law is actually super annoying.

11:28 You can be you can be pro you could be the democratic like you could

11:33 democratically elect someone who violates the law

11:36 and in line with what the people want.

11:38 All right.

11:39 the the point of law is to restrain in many respects the popular impulse.

11:43 So yes, there actually is more of a if you want to get into philosophy,

11:45 there's often a tension between democracy and the rule of law.

11:50 Uh and every executive branch in history is pushing the boundary of law.

11:55 Well, yeah.

11:56 And like if you're dealing with a government making

11:59 a product that's so powerful it's going to implicate government interests,

12:02 you have to be aware as an American company,

12:05 you're ultimately going to be answerable to the American government.

12:08 terms of any contract can and will change.

12:11 And like Sam Alman, he said earlier,

12:14 private businesses shouldn't be more powerful than the government.

12:18 And I'm sure a lot of people in the AI

12:20 community will call him a bootlicker for that particular comment.

12:24 Oh, they've called it worse, including in the last week.

12:26 [laughter] I know he's really taking it on the chin.

12:28 I look to me as an outsider, I'm not part of the EA community.

12:33 That's just true.

12:35 And it's certainly how the actual government sees things and that will

12:39 inform a lot of what plays out in this space

12:43 in the years to come including to anthropic if they

12:46 are on a different page on a lot of these questions.

12:49 I don't know here here's the question for you.

12:53 Okay.

12:52 If you were to redo your tech take rankings where does Anthroic now rank?

12:58 Well, it's very takable.

13:00 So I [laughter] actually I thought about that.

13:02 I was like, man, people are really just teeing off left and right.

13:06 Uh, the conversation has been going to crazy places for a week now.

13:11 I don't, however, think that Anthropic can unseat Open AI.

13:16 I think it's true.

13:17 [laughter] Open AI may lose AI.

13:18 They're not going to lose the take rankings.

13:20 And Sam, Sam just stuck himself right in there.

13:25 And you know what?

13:26 Speaking of vibes, he just he could not wait one day.

13:29 He's like,

13:30 "Oh my god,

13:31 no.

13:31 Same night.

13:32 I'm going [laughter] to stand by mythropic

13:33 Friday afternoon and sign a contract Friday evening.

13:36 Amazing.

13:37 Amazing stuff.

13:38 Oh, he never lets us down.

13:40 You know, I'm just I'm just trying to forstall controversy or whatever.

13:43 [laughter] It's like if you see a deal like like you think Trump loves deals.

13:50 Sam Alman loves deals.

13:51 Like he's like, I already signed a deal with Amazon today.

13:54 I'm going to sign one with the US government tonight.

13:56 It's going to be amazing.

13:57 [laughter] Um All right.

13:59 You want to know what's unaligned?

14:00 What's unaligned?

14:01 And it is the degree and you saw I think we might have talked about it that clip

14:06 what are they in South Korea or somewhere

14:09 where no India where they were all class pans

14:12 and like unity and Sam and Dario were standing

14:16 next to each other and refused to touch each other.

14:20 For people who didn't see it, you had 24 leaders on stage,

14:24 everybody raising their arms hand in hand

14:27 and Sam and Daario awkwardly next to one another.

14:30 hands held high, refusing to hold hands with everybody else on stage.

14:35 Um, in that moment I was like, "Wow,

14:37 they really hate each other." Then I read Daario's memo and I was like,

14:40 "They hate each other." At least one direction of that relationship.

14:44 Yeah.

14:44 I I think u I think the t like very stupid by OpenAI like like

14:50 even if you're whatever reasons you want to do

14:53 it to do it that night [laughter] was just ridiculous.

14:57 uh which evinces a certain degree of irrationality

15:01 which is very likely downstream from personal animus.

15:04 So I think it goes both ways.

15:06 Fair enough.

15:07 Uh well moving through the rundown here.

15:10 Billy says, "I was listening to Bradley Tusk's firewall after reading

15:14 your anthropic newsletter and he referenced

15:16 Apple's position during the San Bernardino case.

15:19 If I recall correctly,

15:20 you supported Apple's refusal to create a backdoor on the theory

15:23 that a backdoor for the US government is effectively a backdoor for everyone.

15:27 In your anthropic piece,

15:29 you argue that CEOs shouldn't dictate how generalpurpose software can

15:33 or cannot be used and that those decisions belong with elected officials.

15:36 In both cases, though,

15:38 the government is asserting authority over how a company's

15:40 technology should be deployed in the name of national security.

15:44 How do you think about the distinction between those two?

15:46 Uh I I thought about bringing this up because I knew someone would raise this.

15:49 I think it's it's quite straightforward.

15:51 It's the difference between deterministic and probabilistic.

15:53 Like like at the end of the day,

15:54 if you have a backdoor in encryption, it is it's a math issue.

15:59 Like it's it's definitionally not encrypted.

16:02 Like back it's just not possible.

16:05 And so what the FBI was asking was basically that phones not be encrypted.

16:10 Now just be less secure for every iPhone user.

16:14 Fundamentally flawed.

16:14 fundamentally flawed in a way that the user would never even know

16:18 if their if their stuff was just it's a math issue.

16:22 What the FBI was asking is that is that Apple Now

16:26 there's lots of you know ways that you can talk about this.

16:29 You know what if it's what if Apple diminishes like the password

16:33 reset so you can like try the code a bunch of times.

16:36 What actually happened is whether Apple was on board or not,

16:39 the FBI got uh a jailbreak.

16:42 Um yeah, you know, to to sort of they got on the phone sort of regardless.

16:47 But in this case, this is just this is Daario and anthropic

16:53 making a policy decision.

16:54 It's not like there's not math at question here.

16:58 Well, so the one question I had, I mean, so in terms of anthropic,

17:03 is it different because it's AI like this is constantly maintained by anthropic.

17:09 So it's not a piece of hardware

17:10 that they're just turning over to the government.

17:12 That's another thing here.

17:13 Yeah.

17:13 It's like, oh, why does the government just, you know, XYZ the problem is,

17:15 and this is going to be a big issue going forward,

17:17 is the issue with SpaceX, which I think some people also brought up.

17:20 Um, the you're not just dependent on it being a service,

17:25 but continually getting better and improving and all these sorts of things.

17:28 So, it does need to be some sort of partnership uh in this regard.

17:35 And I will say some of the tweets I

17:38 saw that I actually had the most appreciation for.

17:41 Some people are like I would rather have Dario Amade decide than and they were

17:45 I think my favorite one didn't even limit

17:46 it to Trump than like our democratic process.

17:49 And it's like points for honesty, [laughter] right?

17:53 Like that's what's implicated here.

17:55 Sure.

17:55 Well, I mean this goes back to though

17:57 and this kind of ties into the the Iran stuff.

18:00 Uh we when I think of the first podcast we did after Trump became president,

18:06 we talked about this like in the like as things get more

18:10 and more sclerotic and difficult to get things done, you're going to see,

18:18 speaking of democracy, more of a populist push for someone to just cut through

18:27 Yeah.

18:26 the tape and do what they want.

18:28 And uh and that

18:30 I would say a lot of that has happened over the last 20 years or so.

18:33 A lot of the authority and decision-m has been vested.

18:37 But but here's the deal.

18:38 All these all these quote unquote prodemocracy

18:40 folks critique the Trump administration for doing that.

18:47 Mhm.

18:46 And they're hoping Dario Amade does that.

18:49 [laughter] Yeah.

18:50 Well, that's a good point.

18:52 Yeah.

18:53 Now again, it's it's not quite Trump is as president

18:57 more constricted by laws than a private company executive.

19:00 So I'm not saying it's exactly the same,

19:03 but they are asking slashhoping that a private company executive

19:08 make the safety decisions that the government is either unwilling or unable to

19:13 dictate to the dictate to the government.

19:15 And um you know, again, I I mean trouble.

19:19 I'm not saying it's a perfect analogy, but I think you're going to see actually

19:23 on both sides or I predicted this a couple

19:26 years ago and I think you're seeing it happen

19:28 this desire and actually push for more centralized authoritarianism,

19:35 but it's sort of disguised by is it doing

19:37 the things that I want to happen or not?

19:40 And then you can label it whatever you like

19:41 to, but it's more this isn't a partisan thing.

19:44 It's a response to a society that feels

19:48 stuck and sclerotic and can't do what it wants.

19:53 And you got to see, you know,

19:54 the response like the like what is international law?

19:59 Who knows?

19:59 Do people like when the US goes in kick kicks people butts?

20:02 Yes, they seem to, right?

20:03 like like people like it.

20:05 And so maybe I'm the rube here by I made my case why I actually think the point

20:12 of my argument was to support property rights

20:17 and to support our system of government as it is.

20:20 And maybe I'm the one holding on to something that uh is going to be

20:25 look pretty antiquated 10 or 15 years from now.

20:29 Well, related to international law, one email that made me laugh.

20:32 Beef writes in all lowercase letters.

20:34 Your takes on international law and anthropics anti-democratic entitlement are

20:38 so good they make up for your take on advertisement.

20:42 Really, really good.

20:43 Best from your most out ofdistribution European the European.

20:46 No, I'm sorry.

20:47 I'm sorry, Beef.

20:48 Um, you are might be out of distribution in terms

20:51 of uh supporting my take on international law and entitlement.

20:55 You hating advertising is as European as could be.

20:57 So, [laughter] um, Exactly.

20:59 Just perfectly on brand from one of my favorite emailers.

21:02 Writes these long missives in all lowercase letters.

21:05 Uh, one more note here.

21:07 William says, "One thing that's bounced around my head this week has been how

21:11 terrible it would be for any large generative AI provider to be in a headline

21:16 like chat GPT responsible for targeting and bombing

21:20 school or similar for Claude or Gemini."

21:23 There isn't much a model company would be able to say if this happened.

21:27 Saying, "Yeah, we warned them not to do

21:29 that and that the technology isn't ready wouldn't prevent

21:33 the bad the wouldn't prevent the bad press nor

21:37 the lawsuits." I thought that was actually a fair concern.

21:41 And just as a passing observation,

21:43 I'm pretty surprised the companies in question here have not

21:48 rebranded the tools that they are giving to the US military.

21:52 Like Google I think has Maven um but it's not Oh my god.

21:56 Oh my god.

21:57 Project Maven was this project to provide

22:01 cloud computing for the Pentagon that Google

22:04 won and then their employees protested about it and so they abandoned it.

22:09 It was actually of Silicon Valley.

22:12 No, they like tore it all up.

22:13 It was like the peak of Silicon Valley.

22:16 Failing to appreciate that the rest of the world is not

22:20 the United States and actually supporting the the military is a good thing.

22:25 No, Amazon got the contract.

22:27 Uh and so and it actually Greg Allen explained

22:32 how it was that that Anthropic got classified.

22:35 uh in general is actually kind of an interesting story

22:37 like because they're worried about things like nuclear weapons they actually

22:40 got classified via the department of energy I believe to like

22:43 test different scenarios and then once you're they're the only

22:45 one that's classified because it's a lot of work and so

22:47 once you can handle classified information you're in um and uh

22:51 which is kind of funny but they're also like closely

22:53 allied with Amazon and so and they're also with with Palunteer

22:57 which are work closely with US government Amazon via this project Maven which

23:00 Google had and then got rid of I mean Google is sitting here

23:03 it's hilarious because actually I One of the contractual details of when they

23:07 acquired Deep Mine is that Deep Mind can never be used for military purposes.

23:11 Um, they're happy to be sitting out this right now,

23:15 I think, but that is looming uh as as a potential issue.

23:20 Uh, but uh and why is it a potential issue?

23:23 Is it a potential issue because

23:24 of the the researchers that are populating deep mind?

23:28 Is that ultimately like the sticking point?

23:30 It's a potential issue because uh

23:32 at what point does the government come knocking

23:34 and say give us AI uh or we we want to tell you what to do.

23:39 I guess my question is why would Google not play ball at that point?

23:43 But maybe it's a researcher issue and the same

23:46 employees who are pissed off about project.

23:48 I mean it turns out that everyone

23:49 can violate contracts and promises including Google.

23:51 So yeah, we'll see what happens.

23:53 Yeah.

23:53 Um well rebrand any of the models that are being deployed

23:57 in military combat would be my recommendation to some of these companies because

24:02 yeah this is our model this is not our model

24:05 from deep mind this is our model from shallow brain like

24:08 well and it was a it was a little odd to hear that like the US military

24:12 was relying on claude in the Maduro raid

24:16 and the same is true in the Iran operations.

24:20 like a very practical application right now is the there's

24:25 so much data out there already today that is collected

24:29 and sifting through that AI is very good at it and

24:33 no I understand it's super useful I mean what they're doing

24:36 in Iran right now still relying on claude because it does take time

24:41 to get some of this stuff classified phase it out but they're

24:45 analyzing different targets and um uh

24:49 collating coordinate data and prioritizing missions.

24:54 All of which would take much longer than it's taking with the assistance of AI.

25:00 I've seen a demo um I I don't know how much I I can

25:04 say of in broadstrokes this technology and that was a few years ago when

25:09 it was not nearly what it was today and it was mind-blowing like like

25:12 it like there's no question this like

25:14 this is use useful technology in a military

25:17 gamechanging interesting um okay Well,