特朗普上任100天、关税冲击贸易、AI Agent与Amazon退让

We're back, and what an amazing panel we have today. With us today, Ryan Petersen, friend of the pod, is back on the show. He's the CEO of Flexport. How are you doing, Ryan? Did you get any skiing in this year? I know you like to ski in the deep powder, like I tried, but it was a busy year for work and I got 2 little kids.

I did a few days.

Okay, so you're—oh, yes. We all forgot you gave control of your company to somebody. It got a little shaky, got a little contentious, and then you took the reins back. How's it been being back in the pilot seat?

That was a year and a half ago, so it's a distant memory. In Flexport time, that's like a decade. We've really had an amazing run. Although, these tariffs—I guess that's why you guys invited me on—have created a lot of new uncertainty in the Flexport world.

Okay, so we'll definitely get into that. And, of course, a fan favorite is back for his 4th appearance on the pod.

I saw the comments last time I was on. I'm officially not a fan favorite, but glad to be back on, and I will be representing free markets in this version.

What did the comments say about you?

It was like, "Loves Biden," "totally beta," "soy boy."

You're filling in for me, though.

I think so. Yes. I was trying to represent libertarian values at the time, but I love that the leftists are embracing Milton Friedman. I think it's all worth it if that's what comes out of all of this. AOC is going to be a complete free-market person soon.

Yeah, free-market months are coming soon. Embracing free-market values and the stock market, right?

Yes, exactly. Because any decline in the stock market is Trump's fault, so now they're embracing the stock market. Unfortunately, the one day that he said it was Biden's market, it was a green day, so that didn't help the case.

I mean, gosh. Well, listen, it's another massive green day already. All that matters to me is that Uber is the anti-tariff stock. It just does great. It's not impacted by tariffs. So here we go with Chamath.

It started already. We have TDS on both sides. We've got Trump Derangement Syndrome from Aaron.

Wait, no, no, no. I want to be Trump Defender Syndrome from Sacks. We've got both.

Defender and derangement.

Trump bias syndrome on your part.

Who, me? I call balls and strikes. What are you talking about? Let's get started. It's starting already, folks. It's going to be a great episode.

Jason has the self-sabotage—find every way to not get rich—syndrome.

I do. What are you talking about? You guys said you'd buy me out of this thing so I could get the hell out of here. You know how much my shares in All-In are worth? For the love of God, write a check. Get me the hell out of here.

I just may.

Oh, God. I mean, I'm going to be a terror if Uber breaks 88. That's my number. 88 is the number.

All right, let's get started. We have so many topics to get through. With us again, David Sacks. Hey, David, you're doing more episodes now. The audience wants to know. I don't know if we're allowed to make any initial announcements, but people are asking me on the streets, in the airports, and in the comment threads: Is Sacks back?

Well, the ratings are back ever since I came back to the show, that's for sure.

The ratings are back, sure, but is the show back? Is Sacks back?

I'm back as much as I can.

And you are a partial employee of the government. You can do 130 days a year or something. Is that still the status?

Yeah, it's roughly half of the workdays.

Got it. And so what do you do? Do you have a punch clock there? When you get to the White House, do you punch in and punch out like Fred Flintstone, or what? Are you keeping track of these days?

I know why you don't know this: You have yet to be invited to the White House. But that's not how it works. Normal people just badge in and badge out.

Like that? Badge in and badge out?

It's a natural place, Jason. I mean, literally, it's interesting.

Interestingly, I don't know if you gentlemen know this, Ryan and Aaron, but there's a new private club in D.C. that Don Jr. is doing, and Sacks is a member. Chamath's a member. I just checked my Gmail. I checked all 3 of my Gmail accounts—everything. No invite.

You must have gotten lost again. Did you send a paper one? Was it like—you sent a gold card or something?

Sacks, how do I get invited to this private club? What is this private club? Everybody wants to know.

We'll be happy to have you as a guest.

Okay. Do I have to wear a MAGA hat? Are there courtesy MAGA hats at the door?

If you want to be a member, obviously there are dues and a membership fee—

Okay. I just didn't want to waste your time with an offer that I knew you wouldn't be willing to accept. It's only $500,000, is what I read. Is that true?

That's true for founding members who have additional benefits, but there's also a lower level that's the more reasonable membership level. I think people are getting a little bit carried away with that number.

Got it.

Yeah, there's like 10 founding members who have that level, and then there's a lower level for the more average member.

Chamath, are you one of those 10?

Yes.

Do you pay more if you have TDS, or how does that work?

TDS premium.

What are you talking about? J.D. specifically, or what are we talking about?

The TDS surcharge. Asking for a friend.

It's a place to hang out in D.C. All of us have been to clubs like the Battery—I don't know if you go to L.A.—like those kinds of places. There's Malibu Beach House. There's Bird Streets Club. There are places in Palm Beach that are really cool.

In any event, we wanted a place to hang out. The clubs that exist in Washington today have been around for decades. They're kind of old and stuffy. To the extent there are Republican clubs, they tend to be more Bush-era Republicans as opposed to Trump-era Republicans.

So we wanted to create something new, hipper, and Trump-aligned. Since I'm in the government, I can't be an owner, but I told him I'd be happy to be member number 1. And so I said, "Great, let's do it." We're creating a place for us to hang out. That's basically it.

We want a place to go where you don't have to worry that the next person over at the bar is a fake-news reporter or even a lobbyist or something like that who we don't know and don't trust.

Got it. So it's like any private club. You want to go somewhere that's highly curated.

This private-club movement is happening all over the country, not just Washington. But we're creating something that didn't exist before in D.C., which, again, is younger, hip, Trump-aligned, Republican.

I actually started a Kamala club in the Bay Area.

Yep. I don't think anyone would pay to join that, though. That's the problem, right?

I mean, it's an open bar, that's for sure. Where do you guys meet up? In, like, Redwood City?

We actually meet up at the trade ports.

We're 100 days into Trump 2.0. It's just a random 100-day thing, but everybody's talking about it. What has it been like for these first 100 days? How does it compare to Biden? How does it compare to Trump 1.0?

There have been 143 executive orders, the most ever in the first 100 days, and they're moving at a different pace, to be generous. Major indices are down 7% to 10%. Obviously, there's the trade war and tariffs. The yield on the 10-year is down about 40 basis points. There's a lot going on.

Let's go around the horn. Ryan, Aaron, you're our guests. What's your take on the first 100 days? Is it what you expected—good, bad, and otherwise, wins and fails, everything?

I'll go first. I think it's a whirlwind. If you look at the John Boyd, the fighter pilot, he has this concept of the OODA loop, which is observe, orient, decide, and act. The concept is that if you're in a dogfight, if you're able to maneuver through those OODA loops at a faster pace than your competition, they get disoriented and don't know what to do.

I think that's got to be how Democrats in Washington and maybe mainstream Republicans in Washington—and certainly journalists—are all feeling this. The Trump administration takes action, and before anybody can respond to that, they've already done 4 more things. You're like, "Wait, I forgot to actually follow up on the other thing that they did that I didn't like."

So it's pretty disorienting. They can't find a line to fall back to and say, "Hey, we're going to push back against this policy," because they're already moving on to the next one, and the next one.

Obviously, I come at it from a trade angle. I think everybody knew that Trump was going to—he told us during the campaign that the most beautiful word in the English language is “tariff.” Don’t tell them it’s an Arabic word, but the most beautiful word in the English language. And so we knew that was coming.

I think the suddenness of it all caught people by surprise. They told us April 1st, April 2nd would be Liberation Day. They didn’t tell us that it would go live the next week. In effect, you’ve already ordered these goods. So, that’s one aspect that people are kind of disoriented about, and we’re going to unpack that.

Aaron, your thoughts on the first 100 days? Obviously, you are a Democrat, and you were pretty vocally not in support of Trump. So what’s your take on the first 100 days? Any bright spots for you, things you support?

Actually, Sacks’ world, I’d say, has been a bright spot. I think we have a very clear message on AI, and that has been a huge net positive. If you look at the past few months, out of all the AI push from the administration, it’s unmistakably pro-open-source and pro-bringing as much AI innovation to the US. Obviously, the tariffs add a little bit of a headwind to that.

I have some very strong asks around high-skilled immigration because I think that AI talent is going to be super critical to actually win the AI war. So I’d say that directionally has had some positive momentum.

From my perspective, this is kind of playing out almost exactly how I thought it would 6 months ago. And then 3 months ago, I think there were some signs that maybe it wouldn’t play out this way, just based on some of the early groups that were coming to the White House and the sort of deep business-centricity of the White House.

I think it was day 1 or 2 that Stargate was announced at the White House: We’re going to go build massive infrastructure. The case I’d like to make once we talk about tariffs is I think there’s an alternative universe where you just lean into acceleration as opposed to adding headwinds. So that would be the case of what maybe could have been: We just keep doubling down on what’s working while fixing the parts that aren’t working. But that would be my judgment so far.

Chamath, I mean, you’ve been talking about it here every week. You and I have been talking about it pretty consistently, so I don’t think there’ll be many surprises here. But take a second and maybe assess what you think. If you had to pick a singular thing that’s gone really well and a singular thing you think could be improved, what do you got?

Let me give you my overall grade, and then I’ll tell you how I get to that. I think the first 100 days have been a B+.

There have been 2 things where I think Trump has frankly hit a home run. The first is all of the direct investment, and specifically the foreign direct investment, into the United States. I think it’s approaching, if not already exceeding, $1 trillion from corporations and organizations and individuals from around the world who have committed to bringing money into the United States. And I think strategically that’s a legacy that will live past him. So, I think that’s been an A+.

The second is we had a very unsafe border situation, and he ran on shutting it down. I’m not talking about the execution of the deportations. I’m just saying getting the illegal crossings to zero, and he’s done that. So, that’s been an A+.

I think what’s going to be more controversial are these next 3 things, though. But in my interpretation, I think the tariffs have been an A, and I think that the market reaction—the stock market is only down 4%—and the interest-rate markets are 4.25%. I think those have been an A.

Now, the reason I think tariffs have been an A is because it has uncovered, in my opinion, how beholden we are to a brittle supply chain, and specifically to China, who is a friend but who’s also an enemy. I think that’s going to really severely complicate our flexibility and optionality in the future as they do what is in their best interests.

Okay. So where have they then not done so well?

I think the documents have frankly been a D. We were supposed to get the Epstein files. We haven’t yet. We were supposed to get the Martin Luther King files. We haven’t. We did get the redacted JFK files. I don’t think there’s been very good communication about why it’s taking so long. So, I think it’s a very small, narrow thing, but I think it had a lot of attention on the way in.

I think the communications of the tariffs and the back-and-forth have been a C. I think the markets were not led in enough of a way where they could absorb the volatility. But if you take it all in its totality, I would give it a B+.

I think it’s been a very productive 100 days. And when you look back in 3 years, 4 years, 5 years, okay, we’ve made some important progress.

Sacks, obviously you’re part of the administration, so I’m not sure exactly how to ask you this, but you heard some nice compliments about AI from Aaron. I happen to agree with those. I actually agree with a good portion of the crypto stuff, too. I think actually getting those tightened up—which are your 2 zones of excellence and the areas that you’re focused on—I think you’ve done a great job there.

So, just bestie to bestie, great job there.

Thank you.

What’s your take overall? It’s kind of hard, I guess, to ask somebody in the administration to criticize the administration, but hearing everybody else’s take, what’s your response, maybe?

I would highlight 3 main areas that I think are big accomplishments for the Trump administration in the first 100 days. Number one has to be the border. Like Chamath said, I think you have to give the administration an A+ on this. They’ve completely stopped the border crisis.

I think we all knew that Trump would take action on this because it’s one of the main issues he campaigned on. I think if you had asked any of us 4 months ago, would this problem be completely solved—meaning border apprehensions completely stopped, the border completely sealed—within the first 100 days, I don’t think we would have believed necessarily that it would get done so quickly, but it has.

Recall that for 4 years during the Biden years, we were told for the first 3 years that the problem didn’t even exist. Whenever the videos were published of caravans coming or throngs of people running across the border, we were told that these were cherry-picked videos on Fox News. It wasn’t real.

Finally, in the last year of the Biden administration, they said, “Okay, we’re finally going to do something about it.” They took some limited actions, and they said that doing more than that would require new legislation. Well, all of that was just gaslighting. It turns out Trump came in, he restored “Remain in Mexico” and other policies, and completely stopped it. He had this line at the State of the Union, which I think is exactly right: We didn’t need a new law; we just needed a new president.

So, I think that’s area number 1.

Area number 2, I would say, would be the vibe shift in the culture around wokeism and DEI. You know how quickly we forget about this, but wokeism has completely collapsed. I don’t know that anyone is endorsing it in a full-throated way.

Moreover, beyond just the cultural aspect of it, I think we’ve had significant policy changes on DEI. Trump has basically ended DEI at the government level. He also signed an executive order ending the use of disparate impact for affirmative action.

This is the policy that said that even if you have a policy that’s applied in a completely neutral and objective way, if it results in a disparate impact, where different groups are represented in a different way in the outcomes, then somehow that must be racist. And that led to essentially engineering the results of various populations to basically fit quotas.

I think all of that now has fallen by the wayside, and I think that meritocracy and colorblindness are back. The only holdout really has been these universities, where Trump is now taking action against Harvard, and I think that ultimately we will win that battle.

You see that even in relatively liberal companies, the DEI departments have been canceled, and they’re moving back toward more of a meritocracy. So, I would say that’s big shift number 2. And I think if any of us had tried to predict that 100 days ago, we would have thought, yes, Trump will do something about it, but I don’t think we would have predicted the total collapse of wokeism and DEI so quickly.

Then I’d say the third area, which is still in flight, is the reprivatization of the economy. That’s a term that Scott Bessent used. I think that the Trump administration needs to reprivatize the economy, and I like that framing of it. There are a bunch of different pieces under that.

I’d say number 1 is DOGE, again, ending this wasteful spending. I do think that Trump has come into office inheriting a very weak Biden economy that was being propped up by massive amounts of government spending that was not only stimulating the public sector, but it was also goosing the employment numbers as well.

And we knew that that spending was unsustainable. We have to do something about it. So, I think for the first time in decades, we’ve actually started to make real cuts in government, real cuts in the federal workforce. And look, we’d like to do more, but that is a huge shift in the conversation. There are other pieces of it as well.

I mean, President Trump has signed a significant number of executive orders on deregulation. There’s also been an unleashing of energy. He ended Biden’s EV mandate and a lot of these “Green New Scam” projects, offshore wind, and he’s been encouraging oil and gas exploration.

I think there’s that. And then I appreciate what Aaron said about tech innovation. We did repeal Biden’s executive order on AI, which was 100 pages of unnecessary regulation on AI. We’ve ended the war on crypto, and I think we’re trying to stop the regulatory capture that benefits large incumbents.

So, you have all these things, and there have been other things done on the economy as well, but I do think that this sets us up for a Trump boom in the future. It’s just that a lot of these changes take time to play out.

Okay, great. Well done. I think we knew Sacks would be very pro. Chamath seems really pro, other than he wants the alien-conspiracy files released, which we’ll get soon. What’s the view from JCal, when you’re the left-leaning guy in the room?

I’m kind of independent, but, yeah, social liberal. I look at what all Americans believe and try to build some consensus here. It’s one of the things I’ve been trying to do on the pod: look for where we agree.

Americans universally want the border secured. They don’t want illegal immigration, and they don’t want fentanyl. This is the biggest win, I think, for Trump, which I think everybody on the panel pointed out. Sacks, you were dead right: when we were seeing those videos, some of them were 5 years old, some of them were recent. Biden really covered up what was going on at the border, and it took years to figure out exactly what was going on there.

So, that’s the biggest win possible. Overall, just to be brief, I give Trump a B for these first 100 days, and I give Biden a C-minus. The second thing that everybody agrees on is that they want to downsize the government. They don’t want waste and fraud, so I think DOGE is the other huge win.

The things I think could be improved are really just 3 simple things. Economic uncertainty is really terrible for running a business. I’m seeing a lot of folks in my circle, on This Week in Startups and here, telling me, “I don’t know how to plan for the future.” We’re going to get into that with this tariff stuff and the trade war.

I think economic uncertainty means we have to slow down and maybe make it easier for people to understand what the administration is trying to do. I think the rule of law really matters to people. People didn’t like Biden’s pardons. They didn’t like the cover-up of his mental acuity, and I don’t think people like the deportations without due process. We talked about that on a previous episode.

Overwhelmingly, people want Trump and the administration to obey what the Supreme Court says. They really want the rule of law. The third-term talk—8 out of 10 Americans don’t like that kind of talk. And then there are conflicts of interest. Obviously, people hated the Hunter Biden stuff, and they hate the memecoin stuff.

That’s where it could improve: crisper communications, more thoughtful execution, maybe less trolling. I don’t like the White House Twitter account trolling. And then focus on what got Trump here. You all said the same thing: what got Trump in here was the economy.

One thing that wasn’t mentioned by everybody is the peace dividend. Trump is making massive progress in Ukraine, apparently. I don’t know if it’s on the docket today or not, but stopping the wars and making the economy boom—those are the 2 most important things that he could do.

Build on that.

I totally missed that. You’re absolutely right. That’s another one where I would give Trump an A-plus. Nat and I had dinner with POTUS 2 weeks ago.

Wait, you had dinner with Trump? This is breaking news.

Well, okay, whatever. Yes.

I think it’s remarkable how much of a Putin apologist JCal has become. I mean, you want to end the war in Ukraine now? You’re going to give it to Putin? You’re not going to stop Putin?

I’m totally in favor of what Trump’s doing in negotiating a deal to get more money.

Oh, you want to talk to Putin now?

I’ve always wanted to talk to Putin. I just don’t trust him.

But you can trust him. Let me tell you what Trump said.

There were a handful of us at dinner, and then he got up to say a few words at the end. He reminded me why I was so inclined to vote for him, which is that he talked about his uncle and how his uncle taught him about the severity of nuclear war—how people don’t understand how intense and destructive it is, and the power of these weapons.

He ended that speech by saying, “This is why I’m so fundamentally against this thing.” It reminded me, to your point, Jason, that it is so easy to forget that there’s only 1 existential risk, short of aliens coming from the heavens, right? There’s only 1 existential risk where all these issues become fringe issues.

You mention the rule of law, border security, foreign direct investment, tariffs—it all goes out the window in a nuclear war. I was like, “I am so glad this guy’s in charge,” because on this 1 issue, he never wavers.

Yeah, and I think there are all kinds of complicated moments that could make this an issue. This was where my biggest issue with Biden was: I did not know who was in control.

I think that Trump, in the first 100 days, to your point, has completely reinforced that there are no conditions under which he’ll go to war. He has time and time again shown that he’ll find the off-ramp. I think that’s really healthy for Americans to see.

Let me build on that point with respect to Ukraine. We were on a glide path before the Trump presidency that Biden had put us on. Kamala Harris gave every indication she would have continued it.

What was that path? It was a path of continued escalation and doubling down in Ukraine. Recall that it was Biden himself, at the beginning of the war, who said that if we give Ukraine Abrams tanks, F-16s, ATACMS, or HIMARS, or if we allow them to hit targets inside Russia, it would lead to World War III. He actually used the word Armageddon.

At the beginning of that administration, they were very concerned about how an escalatory path could lead us into direct conflict with Russia and World War III. And yet, despite that, at every fork in the road where they had a choice, they ended up doubling down.

They gave the Abrams tanks. They gave the F-16s. They gave the HIMARS. They gave the ATACMS. Finally, when Biden was a lame duck in his last couple of months in office, they did the most reckless and irresponsible thing, which was to allow American weapons to be used to strike targets on Russian soil.

It wasn’t just fighting in Ukraine; it was striking targets on Russian soil. Moreover, we now know from a New York Times article that came out in the last few weeks that it was American generals and American intelligence who were planning this war.

When you’re talking about striking Russian targets on Russian soil, it’s not just the Ukrainians using our weapons. They’re using our targeting, our guidance, and our satellites. We are deeply integrated in the kill chain. This is the United States being a co-belligerent in the war, hitting Russian soil. That is incredibly reckless and dangerous.

I have no doubt that if the Democrats were still in office, we would be in an escalatory spiral right now, with the destination being World War III. I do think that Trump has pulled us back from the brink there. There’s obviously still more work to do.

But I really appreciate the efforts that Steve Witkoff has undertaken, where, for the first time in 3 years, we’ve at least had direct diplomacy with the Russians. We weren’t even talking before.

I mean, talking is a great thing, and apparently we’re going to keep supplying them with weapons as long as they pay for them. So, it’s going to be very interesting to see how this all hashes out over the next 100 days or so.

Let’s keep moving. I don’t think we know that yet. Let’s wait and see on that.

Okay, yeah. I mean, I think that’s what was reported, but you’re right. We should wait and see.

Downstream tariff impacts. We have to talk about this, and this is why we have you here, Ryan, since you’re in the thick of it. You tweeted a thread last week about the lag time of shipments from China, and when you were on, I guess during COVID, you really educated us on how the supply chain works.

According to the thread that you shared, somewhere around early June we’re going to expect warehouses, trucking—the entire supply chain—maybe to start to seize up, or layoffs. I don’t know how you would frame it, Ryan, but are we at the point of no return with regard to the supply-chain tariff conflict? Is there an off-ramp for this tariff conflict, this war, this negotiation with China, in your mind? What are you seeing on the ground and in the purchase orders and invoices at Flexport?

Definitely not past the point of no return. I think we’re still right in the middle of “don’t judge the cook while he’s cooking.” Let’s see what it tastes like at the end. I think that’s a starting point here, and they’re still in active negotiations, so I don’t think today’s situation is static.

The world does want a lot more certainty, and that’s a big cause of what’s happened here. What has happened is a 60% decline in bookings of ocean freight from China to the U.S. That’s really dramatic, probably exceeding what was expected.

I don’t think, when they rolled out the initial reciprocal tariff plans on April 2, that it was meant to be a 54% tariff on China.

Then there are multiple cycles of escalation. We ended up at what’s now a 154% tariff. So this is a lot higher than anybody planned for, and therefore, I don’t think anyone’s planning for a 60% decline in ocean freight.

Ryan, let me ask you a question about that. Are people actually paying that 154%? There’s been this discussion online, and it’s sort of unclear from the administration and from retailers, with stuff that’s landing, whether people who ordered before April 2 are actually paying the 154% on top of what’s landing.

It’s live now. It was based on departure date. Goods that departed China after midnight Eastern time on April 9 are subject to the tariffs upon arrival. Now enough time has passed that pretty much all the ships that are arriving left China after April 9, when that started. So yes.

So what happens? Are people paying it, or are people saying, “I won’t take delivery”? You have to pay it, Ryan—correct me if I’m wrong—but you have to pay it at the dock in order to get the goods released.

More or less, that’s true. They allow you to have a bond in place, so you can pull the goods out before you pay, but the money is owed at that time, and then you get about a 2-week timeframe to actually make the payment. But there are strategies here. A lot of people are doing that. You can use what’s called a bonded warehouse and move cargo into this warehouse, and then you only owe the duties when the cargo leaves.

That’s what I was asking: Is there a hack here to do that?

It’s not a hack; it lets you defer things, and it’s very common right now. People are searching everywhere for bonded-warehouse capacity because, in a bonded warehouse, not only do you defer payment until the cargo leaves the warehouse, but you only owe the duty amount based on that date. So if the duties come back down, which a lot of people are betting they will on the China-specific duties, you’ll actually lower your tariff burden.

And then there’s another hack for this, which is to effectively use a Mexican or Canadian bonded warehouse. You move the goods into Mexico, and then you technically only import them into the U.S. at a future date, when tariffs are lower. I understand a lot of companies are doing that right now, too. We’re helping some people with that type of strategy.

Do you think the government will view that kind of hack as okay? If you look at the GDP numbers, one of the craziest things was the inventory pull-forward that people did—to your point, trying to get as much stuff into the United States before April 9, for example.

Yeah, I mean, it’s not a hack. Bonded warehouses have been around for decades, and they’re very commonly used. I don’t know that it’ll be that material in the scheme of things that it would cause a change in the law around bonded warehouses.

So you don’t think, for example, the Department of Commerce will have an issue with the strategy of sending inventory into Mexico? Essentially, isn’t it a workaround? Instead of paying the China tariff, now you pay a Mexico tariff, which should be less. Is that the idea?

Well, you can move it into a bonded warehouse in Mexico and not pay Mexican tariffs either. You just wait until it imports. But what is the Department of Commerce or Customs to do? You just delayed importing the goods. You’ve imported them in the future, and it doesn’t—I wouldn’t even call it a hack. It’s just people are going to get creative here. That’s the job. Actually, the government should set the rules, and the rest of us have to figure out, all right, how are we going to compete and make money in this environment that they’ve created?

Ryan, in that tweet you wrote, which was a pretty dramatic tweet painting a very—I don’t know—a pretty dire situation, where are we in terms of how dire this will get or how resolvable it is? Paint us the best-case scenario and what you expect could happen in that case. Or, if this gets extended, are we going to see, as people are hand-wringing, empty store shelves, Christmas getting ruined, and all these layoffs starting to happen in the supply chain? Take us through the 2 scenarios that people are debating.

Yeah, I mean, the bleak scenario—which I don’t really think is going to happen—is that the administration doesn’t want this to be their legacy, that they created a policy that just kind of tanked small business and the supply chain. So I don’t actually think this is going to happen, but the bleak scenario is that tariffs stay at this level: 145% on China. The 10% goes way back up to what was originally announced in the reciprocal tariffs, so there’s no safe haven for tariffs, trade just falls off a cliff, and a lot of companies go bankrupt—especially small companies that are importing from China.

The reality is that tariffs have been high on China for a long time. Labor costs in China are not low; you’re not there for cheap labor. You’re there for quality manufacturing at this point. There’s much cheaper labor in Southeast Asia and other parts of the world than there is in China. So you’re in China because of the manufacturing capabilities and the ecosystem, not just for cheap labor. If you could have moved, you would have already, with the 25% tariffs from Trump’s first term, which were high enough incentive.

And so that’s the bleak scenario: small businesses start getting wiped out, especially the ones that are buying from China. It’s a lot of brands. It’s not just Amazon sellers selling stuff that you don’t need; it’s all the brands that you know, like fashion brands and apparel brands. I had Cuts Clothing on This Week in Startups last week, and he said that if this doesn’t get resolved in, let’s say, 2 to 4 weeks, people are going to start layoffs in his group chats, and they can’t physically restart the supply chain in Vietnam or wherever to make T-shirts.

Aaron, what’s your thought on this as well? I’m just bringing you in.

Sure. Well, first of all, Ryan has supplied me with a high degree of doom-scrolling, and it’s just a horror show reading his tweets. I would feel better if the messages out of the administration were either more consistent or had a logical connection. Do we either want to raise the tariff revenue stream, or do we want free trade? Those things are working against each other because, depending on who you talk to, they say this is a mechanism to bring down income tax, which obviously, by definition, means that they expect the tariffs to persist. That’s totally different from, “Let’s go negotiate deals that just allow for free trade to actually increase.” So are we worried about reciprocity, or are we worried about a revenue stream? That’s a whole issue.

You also have this issue, which is the messaging from the government—and this is the meta point I’ll make in a second—about how we could have actually accelerated into the transformation of the economy. But you have folks like Lutnick, et cetera, going on TV talking about the end state of our economy, which are actually probably fine messages, but we haven’t seen what that vision looks like.

So everybody is kind of confused. Does this mean that we literally go into manufacturing plants and we’re the ones literally doing the screws on an iPhone? Or is it a bunch of next-generation jobs, where we’re managing robots, and shipping and logistics grow as a result of this, and all of the surrounding supply chains start to grow?

Most people on this call have managed teams. You do change management; you lead people to the end state that you want them to see the potential in. Something that gets missed—and the part that confuses me—is, exactly to Ryan’s point, people are in China because of the ecosystem of manufacturing, yet the messages you get out of the administration are like, “Oh, we’re going to have fewer toys at Christmas time.” It’s like, no, that’s not the big picture. The big picture is that this is supplying the parts that go into building a manufacturing plant and building a car that allows us to actually be even remotely competitive in car manufacturing.

So where should, in your mind, this all lead? You have some thoughts on American exceptionalism and maybe skating to where the puck is going. If you were to become an adviser on this as a technology expert and somebody who spent their whole career in it, what would you advise them to do?

I’d get rid of Navarro immediately, and you would basically say, “Mea culpa,” like, “Oops.” Obviously, you need to land that with some really cool trade deals that make everybody feel happy. You basically say, “You know what? Let’s go back to the first 2 days of Trump,” which is: let’s announce massive deals. We’re bringing manufacturing here with Stargate. We’re doing TSMC. We’re building NVIDIA chips. We’re going to do a deal where you get a 5% tax rate if you build in America.

So you just stimulate a manufacturing boom in the country. You incentivize automation across manufacturing, and you use that as a competitive weapon to go and compete with the lower-cost labor that happens internationally.

We find every incentive and tool we can. We deregulate. You allow people to build these plants, so you don't have to go through the 3-year EPA process. You just accelerate from this position, and you see it all as upside.

If you go talk to the Fortune 500 company that actually has to build anything right now, you give them a path to say, “Listen, we're going to help you transition away from your current supply chain, and we're going to make it even more competitive and compelling to do that in America.”

There's a reason Elon builds in America. He has actually made it more effective to bring automation to manufacturing and to build locally, but he wasn't forced to do that. I would argue that you use as many carrots as possible in some surgical areas.

Chamath, I've heard your points about chips, pharma, and AI. In those surgical areas, we get tough where necessary. If we have to do a couple of very surgical tariffs to make people move in the direction that we want, that's totally fine.

But even arguing the premise is hard because we act like countries are screwing us, when actually businesses are independently making decisions about where they want their supply chains to exist. In a free market, they've made that decision. They don't need the government to tell them where they're supposed to, or where they're allowed to, have their supply chains operate.

That ends up with lots of economic distortions that everybody on the right would have called the left socialists for trying to implement central planning around supply chains. So that's my piece.

What do you think, Chamath, about this sort of reframing and off-ramp, and maybe the positive spin on it? If you want to make T-shirts, you want to make commodity items, have at it—free trade, reciprocal tariffs, great, check that box. But here is a series of incentives and a path forward to do the advanced stuff, to do robotics, et cetera.

Let me answer this in a different way. A lot of those things he's actually doing. I think this is where we are: we're beyond TDS. There's something that comes after it, and I think that the mainstream media has just lost their mind to a degree that they hadn't even lost their mind in Trump 1.

I'll give you one example. A shout-out to our friend: completely brazen, ridiculous, shitty reporting by The Wall Street Journal last night. When they were told that this whole Tesla thing was a total farce, they continued to publish it. They were referring to Elon—the board starting a search to replace Elon—and then the board said, “Wait. We told you we weren't doing that.” They didn't even mention that the board had communicated that directly to The Wall Street Journal. The Wall Street Journal said, “I don't care. I have an axe to grind.”

I think that Trump has a strength, which is that he shapes these potholes for the mainstream media to fall into. The downside, though, is that the mainstream media then doesn't do the other part of the job, which is to tell us the things that are important.

For example, we spent a lot of time breathlessly talking about the MS-13 knuckle tattoos of the guy. Then we spent a bunch of time talking about how MSNBC blurred out the names of the placards on the lawn. But here's the other part, where they get so tilted: they don't report that, for example, when Trump took a shot at Harvard, he also reinforced and strengthened historically Black colleges and universities. That totally did not get written.

I'll give you another example. This past week, Scott Bessent said that the tax bill will allow you to fully deduct all the PP&E and all of the incidental costs of building a factory. I heard that and immediately went to my wife. She runs a pharma business, and this is exactly what she's trying to figure out. We now are asking, “How do you build a business case if this actually gets effectuated?”

The point is that this thing would create an absolute economic bonanza if it were passed. Other than people hearing it on this pod or randomly finding it in a direct clip that Bessent puts out on X, there has been zero coverage by the mainstream media.

Yeah. First of all, the MSM, and whatever we want to call it, aside, that is still on the administration for driving a change-management process that causes people to build on momentum and not causes boards to basically say, “Are we going to pivot our entire supply chain this week because Trump didn't get a call back from Xi?” That is really not a TDS-MSM issue.

If you talk to Fortune 500 CEOs, did you know about the PP&E thing?

No, but that's not the point. I don't need to know about it, but there are many other CEOs who do. They're controlling trillions of dollars of capital allocation. It's an important thing.

If we had Mary Barra on this call and we said, “Mary, has Trump increased your ability to execute and operate and accelerate the transition to the U.S., or has he created headwinds that make it tougher to navigate right now?” Which way do you think she'd go?

I think that she would give you a calculated answer that is neither pro nor con.

I think that answer changes by the day. If you talked to her last week, she would say this has been a major headwind. Then yesterday, to Chamath's point, they did this thing where you can depreciate or fully expense in year 1 capital improvements or building out factories.

But earlier this week, they made it so that auto parts are not subject to the tariffs. They created a huge exemption that wasn't there. They should have had all of this in the beginning, by the way. If this was planned, it should have been there in the beginning because these auto companies were saying, “This is going to bankrupt us if we have to pay taxes on this.”

Circling back to communication and making a crisper and clearer expectation, you are right: it's my job to stay informed. As a CEO of my company, I try to stay informed, and you're right, it is hard because sometimes I find myself hunting and pecking to find the things that matter.

But I do put a bunch of that responsibility into the lap of the people who are supposed to actually report the facts. They can choose. They didn't have to run that article about Elon, which turned out to be total horseshit, on the front page of The Wall Street Journal.

They could have talked about what Ryan just mentioned as the first article and said, “This completely changes your ROIC and ROE calculations for 90% of the S&P 500.” That was not the article they chose to write and publish.

I also think it comes back to my original point around OODA loops. The administration is running these very tight loops: let's take an action, let's see what happens, let's see the reaction, and then take another action.

Washington is used to having all these committees that plan everything for 5 years, or whatever—18 months—and then roll it out slowly. They're going, “Let's roll it out. Oh, crap, we're about to cause this huge problem in the auto manufacturers, and they're all telling us they're going to go bankrupt.” Three days later, they push an update. It feels chaotic.

To summarize, Ryan and Aaron, your position, so we can keep going through the docket: a little less shock and awe, maybe a little more predictability, and a little crisper communication. And Chamath, I think your position is maybe the mainstream media can play a better role here in focusing us on what matters.

That wouldn't be my takeaway.

What's your takeaway?

Zero shock and awe—not a little less. My strategy would be 100% different.

Scott Bessent has an incredible podcast from September of last year, and he basically said, “Biden got it all wrong.” I was listening to it and thinking, “Okay, this is kind of cool.” He basically says: deregulate the U.S., make it easier to build manufacturing in the U.S., increase GDP, and then you'll be able to take in less tax revenue and spend less in the government.

It was, “This is actually a glide path.” We could take the fact that we had a soft landing relative to the rest of the globe. We're winning in AI. We're winning in a number of categories. We obviously need more energy, and we need to bring manufacturing into the U.S.

You have this great momentum. We are the tech leader in the world. Let's pour fuel on that. To pour fuel on that, you do a series of carrots and the wins that build a flywheel of positive energy.

The reason why I take a little bit of exception to Chamath's MSM point is that I think, to some extent, Fortune 500 CEOs are not thinking, “Rachel Maddow said this, so I'm going to go worry about this topic now.” The information coming at them is not the issue. I'm not talking about the information that's presented. I'm talking about the information that's excluded.

How do you get the information that's not published and shared broadly? Goldman Sachs and JPMorgan are not writing reports on the fact that we might enter a recession because of MSNBC's reporting on this topic.

Again, that's not what I'm talking about. I'm saying glad-handing some high-level prognostication, which nobody ever gets right, is, in my opinion, worthless. What I'm talking about is the details. So when you talk about something as narrow and specific as excluding PP&E or allowing you to double- or triple-depreciate something in a given calendar year, fantastic. That is narrow, precise, specific, and actionable.

And what I'm saying is, if I surveyed the 500 CEOs of the S&P 500, dollars to donuts, the overwhelming majority would not have known. Had they brushed up against that somehow in their normal media consumption and then asked their teams, the odds of that would have been zero as well. So I guess then, who do we blame for this, Cham? Is it the administration's job or mainstream media? I bet you everybody knows about the blurring out of the stupid pictures on the lawn and the MS-13 knuckle tattoos.

Yeah, okay. So let's wrap up on this, just a really lightning round here. Amazon flip-flopped on a new tariff notification on its websites. Trump said he had a great discussion with Bezos: “He solved the problem very quickly. He did the right thing. Good guy,” et cetera.

If you haven't seen this, it's something that Temu is doing. Here's what Temu does today. Nick, you have that image. If you could pull it up, it shows that when there is a tariff, they explain the tariff coming into the country. They put it in as a line item. I thought this was actually kind of cool. I don't know why people take offense, Ryan. This is pretty standard stuff.

Amazon's competitor, Temu, is putting in the import charges. They don't say tariffs. They don't say taxes. They say “import charges.” This is a standard thing. This happens in other countries, too. What is this? What is Temu?

It's like a dollar store. It's basically a dollar store.

Is it the last place you would ever buy jeans?

Yeah, you can buy $12 jeans. Basically, your left sock from Loro Piana costs less than Temu's entire inventory of jeans.

The point being, I thought this was actually a plus. I think they totally misplayed this. They rolled it out, then they got criticized. I think they were called a treasonous company from the White House press secretary. They totally misplayed this because they should have leaned into it and said, “Yeah, we're showing you all these tariffs when you buy from China. If you buy from America, you don't have to pay any tariff. And look at all these other products.”

Come on. That would have lasted 3.5 seconds. This is exactly consistent with the other issue, which is that they're playing whack-a-mole. “Okay, we're going to do something with the automakers. We're going to try and solve some problem with Amazon.” This is a sign that it's not a good strategy if you have to do this much whack-a-mole. They can't cover up what Amazon is going to end up dealing with, because there are going to be 500 other retailers that don't get the call with Trump.

To me, that's evidence that they clearly didn't think through the entire downstream set of conditions that are going to change as a result of this.

Sure. Yeah. I thought this was a big win, Chamath, because they could then have Amazon—here's a mockup somebody made. I'll pull it up here. It was interesting. They could, to Ryan's point—I'm sorry, Ryan's point—just show, “Hey, here's a bunch of American companies. Buy American when you do a search.” Here's what it might look like. Pull that Oral-B toothbrush one up, Nick, if you have it right there.

Somebody mocked this up. I think this could be the hugest win. You could have the retailers do “Buy American.” Buy it once. Buy a high-quality product from America.

If you look here, we don't have the products. It wouldn't work. We don't have the products.

Well, I mean, we do have American-made products for some products. I buy my boots from Danner, and those are all American.

Yeah, so we should just go back to communism and we're all going to make our shoes. It's like, we're in a global market. We buy from everywhere.

I got a bunch of emails from people, and a bunch of them were Amazon sellers. I don't know, Nick, if you can find it, but I posted their comments and reshared them just to highlight the issues that they were going through. At the core of it was a feeling by them that Amazon had abandoned them as American purveyors and sellers of goods, and that Amazon, on the margins, had tended to help competitors from abroad come stand themselves up and compete, and essentially cannibalize on price and margin.

This is my view completely: This is the biggest opportunity that I think the Trump administration is flying at 40,000 feet, doing macro-level negotiations and, look, failing to see some of these micro-optimizations that are really, really real.

In the United States, you can import goods as a foreign company. You do not have to create an LLC or any sort of registered entity in the United States to import goods. Sometimes they say, “Americans pay the tariff.” That is not true. In many, many cases, the foreign company just imports this stuff and sells it on Amazon.

When they get caught cheating, they can lie about the valuation and pay a lower tariff. They can change the classification and pay a lower tariff. They can import stuff that is harmful to children, has lead paint, or whatever else. There is no enforcement at all.

So you're saying Amazon's third-party marketplace is a bit of a backdoor to abuse the system, right? I mean, Amazon is sort of just playing the game that's on the field, but this is legal in the United States? These companies import stuff, and I think it's 60% of all the sellers on Amazon that are Chinese-registered. They're not registered in the United States at all—just Chinese companies.

Which sounds profoundly unfair in terms of the playing field.

Can we just take a step back and also acknowledge that we are talking about a level of detailed issues that we would never have talked about 6 months ago? There was no interest in even bringing this up. If Ryan wanted to bring up the hollowing out of American salesmanship, let's say, because of this arbitrage that Amazon does for GMV, that would have been a snoozefest.

Except today, it can actually get a lot of awareness. Friedberg mentioned this, and I've mentioned this before, but I think there are 4 things that really matter: batteries, AI, pharma APIs, and rare earths. That is now on the agenda.

The positive way to look at this is that the American economy is too complicated. If you had waited, Friedberg, for a study of all of the implications, we would have been waiting forever and nothing would have happened. I think that we've made macro-level moves, you're right, and now we are finding what the implications are and course-correcting in real time.

I hope what happens, though, is that when we find these big, thorny issues—I think the Amazon thing is a pretty interesting issue about American competitiveness—we follow through and get to the root cause of it and fix it.

Those feedback loops are there. The Trump administration is going to act on this, and there's an act coming out of Congress as well to shut down the foreign importers of record. So those feedback loops are there in ways that I don't know if they were there in the past.

Go ahead, Friedberg.

We can totally chaos-monkey the economy and just see what breaks. You sort of phrased the idea that we could do the research paper and do the Aspen Institute as a bad thing, but it can also be a bad thing if you're the small-business owner right now who has 30 employees and you literally don't know what you're going to do next month.

That’s the argument to counterbalance. That's why you do have some bureaucracy, and why you don't chaos-monkey the economy, and why Rand Paul is literally saying we shouldn't actually let you have unilateral control over tariffs. Interesting dynamic there.

All right. You want to wrap us up here, or do you want to pass? Oh, am I still on the pod?

Yeah. Well, I spent 80 minutes debating this topic with Larry Summers 3 weeks ago. The point I made then is that we had, in this city, for 25 years, a globalist consensus on trade that distorted a lot of outcomes. I don't need to rehash that debate, but I'll just recall that Larry Summers's main argument for why this would not work out is that the market was down. Do you remember that? That was his evidence—that this wasn't going to work.

It was all about the market not pricing in lower expectations. Well, guess what? The market is actually up since Liberation Day on April 2nd. So what happened 3 weeks ago was basically a panic in the market over this policy, and the media has been trying to fuel that panic.

What I said as well is that we do have to stick the landing on this. President Trump shifted the conversation away from this globalist consensus, and he's now redefined the debate. But it is now up to Scott Bessent, the Treasury secretary; Howard Lutnick, the commerce secretary; Jamieson Greer, the U.S. trade representative; and the rest of the Trump trade team to negotiate these deals and stick the landing.

I agree with you to the extent that the sooner that is done, the better, because it is good to provide business certainty.

But the idea that this hasn’t worked so far—I think, again, the main argument against that was the market reaction, that now the market’s not positive. So I think my point is just we need to give this time to work. I think it’s too soon to be judging this policy as if it hasn’t worked yet. It needs to be executed properly.

And quite frankly, Ryan, I remember the last time you were on this pod, you were coming on about—wasn’t there some union deal that was supposed to shut down all the ports and all the shelves would be empty?

That never happened either.

It did happen.

It did. They shut down for 3 days.

I don’t remember the shelves being empty, which is now the new panic the media is trying to create. So, look, there’s a lot of pants-wetting that’s occurring here that’s being fueled by the media.

I do want to bookmark one thing. The only thing I was more frustrated by was listening to the Larry Summers and your conversation, because I was like, “Why did Larry make this point? Come on. Don’t go down the WTO rabbit hole. That’s not relevant.”

Of course it’s relevant. It’s how we got here.

No, that’s 25 years ago. Let’s worry about literally today and what we do going forward today.

China, over the last 25 years, has been able to strategically annihilate our rare-earth processing capability and our ability to cast rare-earth magnets. We just sat back and watched as the market basically went to the lowest bidder, which was being subsidized by the Chinese government, which the WTO allowed them to do.

And now we have a critical dependency in our supply chain on China for basically every electric motor in every product, including cars. That was crazy. We should not have allowed that to happen. How are you going to change that?

So we needed to shift the political conversation to recognize the ways in which free trade led to unfair trade and created unacceptable dependencies on the American economy.

Wait, wait. I just want to make one point. It’s more than that.

Make your point, and then I’ll—but this is the national security of the United States that’s at stake. Let’s take your favorite pet issue: China invades Taiwan. Okay? And we have to take a side.

Jason, my point is—hold on. Let me just finish. The Chinese say, “Here are the implications of supporting Taiwan: You don’t get any pharma APIs, you don’t get any rare earths, you don’t get any batteries.” Okay? It’ll send life back 50 years.

Or let’s say China and India get into a fight and we’re forced to pick a side. Same situation. The point is, there are all these scenarios that we never even considered, where we would be able to have strategic optionality to make the decision that’s morally and ethically right for the United States.

And I think that we have learned through this lens that these are huge issues. The thing that the Chinese did that was so brilliant, which we still don’t have an answer for, is they have these national champions. Being a national champion allows you—and we’ll talk about this in AI—it allows you to blur the lines between the public and private partnership. It allows you to blur the law. It allows you to blur capital.

I’m not saying we have to do that, but what I am saying is we need to have our own answer to it. And that was never on the table until April 9.

So, yeah, 100%—do that strategy, and then don’t have a mad rush. What is the strategy?

No, no. Because if you have a—you have, you know, Ryan, what is the number? I don’t know, a trillion of imports or whatever. You don’t need everybody jamming the system to build their supply chain in the U.S. to solve that problem immediately.

What Howard Lutnick said last week—and again, we may have all gotten caught up in the knuckle tattoos and missed this—but he was very clear: “We have one country. A deal is already done. We’re convening Parliament. It’s going to be the first of many.”

For all we know, there are 30 deals waiting in the wings, and the first one will set the tone. I think Sacks is right here: It’s way too early to declare defeat and say that it was “chaos.” I think if we’re sitting here in 9 months and foreign direct investment has shriveled up and domestic investment has shriveled up because there’s just no continuity, you have a claim.

But that’s—

No, no, because I don’t think that’ll happen. I think we’ll end up in a good spot. I’m with Ryan—we’ll end up in a good spot because we’ll iterate through this. My only point is there’s an alternative path that could have occurred. It could have been done in a more thoughtful, well-communicated pattern instead of, “Hey, let’s do barrel rolls with the airplane.”

I don’t disagree with you, Aaron, and I can tell you we’ve already started to see layoffs. Nobody wanted to even initiate the barrel roll, guys.

Yeah, listen, we get it. It’s like, “Hey, I don’t want anything to change.” I think we agree to disagree on this. We’ve got to move on to the next one.

Hold on. This is one last point. Excuse me. You didn’t call me for 40 minutes. I just want to make one final point.

Here he goes. Aaron, where were you with this perfect plan?

Where were you with this perfect plan before Liberation Day? I was telling Chamath about it.

You were telling Chamath. Okay, great. They were talking about this specific issue. All the people who suddenly know what the perfect plan is and how to perfectly execute it—no barrel rolls—had nothing to say about this topic for 25 years. Now all of a sudden, they’ve come forward with their perfect plans.

I would say that’s a victory for Trump.

Finish.

Look at this. The best thing of all of this is you’ve got the liberals embracing Milton Friedman and backing down on it. I love it.

Yes.

All right. Listen, we’re going to agree to disagree on this one. Sacks, Kyla is coming on next week. We’re going to make some cocktails. It’s going to be wonderful. We’ll ask her some direct questions about it.

But I want to talk about AI agents. 2025 is shaping up to be the year of AI agents. There’s tons to talk about here. OpenAI is planning to charge between $2,000 and $20,000 a month for different levels of AI agents that would basically be cron jobs they would run in the background and do things for your company that humans are doing right now.

You may have heard of this agentic tool. Again, “agentic” is just a fancy word for “agent,” which is a fancy word for a cron job that just runs perpetually.

Yep.

Manus is the company in China that Benchmark invested in, and it’s created a whole buzz on the side. Manus’s website has a really good visualization of what these agents would look like.

First of all, I think you’re giving a little too much credit to Manus. They didn’t come up with agents, but I do think that they have a very good demo, and it’s hard to know exactly how real it is because not everyone’s used it and it’s from China.

It’s from China.

I’ll get to that in a second. If you go to their website, you can see a bunch of their demos. I do think that what they deserve credit for is advancing the ball on the UI paradigm. It’s not that other people weren’t doing this. Notably, Anthropic was doing this with its Operator product.

The basic idea is that you’ve got this 2-pane view. In one window, you’ve got the standard chatbot interface, and in the other view, you can see what the agent is doing. That agent has the ability to toggle between currently 4 apps: search, browser, code, terminal, and document editor.

When you give Manus a task, the first thing it does is create a to-do list in the document editor. You can see it there. Then it works sequentially to achieve each of those tasks and puts an X on them. You can see it working.

What’s cool about the demo is just the way that it seamlessly toggles between those 4 apps. You can see what the AI agent is doing: It’s browsing the internet, searching for things, writing documents, and crossing things off its to-do list.

Now, I think it’s pretty easy to imagine where this goes, which is that you’ll be able to connect an agent to all of your SaaS apps. It won’t just be 4 applications. It’ll now be connected to dozens of applications, including ones that already have your data. It’s going to know what actions it’s possible to take in those apps.

So when it creates its to-do list, there’s a much wider range of things that it can accomplish. In fact, there’s a new standard called MCP that’s taking off like wildfire, which is built specifically to enable agents to connect with applications, understand the data, and understand the actions that are possible in those SaaS applications.

Manus is just at the tip of the iceberg here. I think this will become a very standard UI paradigm. That’s the reason why I mention it—not because I’m predeclaring them to be the winner in the space, but because I think there’s a lot of talk about agents, and it’s hard to conceptualize what that means without just seeing it visually.

A great summary there, Sacks. And Aaron, I want to get your thoughts on it because, obviously, you're running Box and have your finger on the pulse of this. We actually started building one of these in our venture firm. We have 20,000 applications a year, and we have updates coming in from investments. We are now taking those stacks, Aaron, and having an agent sort them, then look for competitors and compare them to the last update.

We're looking into our Notion, our Coda, and asking what other communications we've had and what questions we should ask about the startup and its strategy. Then we're presenting that in Slack to our team. This is coming fast and furious, and we spend, I don't know, probably 15 minutes on each of those incoming applications. You start doing the math on that, and we're talking about 5,000 hours of work. Aaron, what are you seeing on the street? What are you doing at Box in terms of agents landing right now in Q2 of 2025?

Yeah. I think Sacks represented it well: you have to now think about AI as effectively being able to do anything on a computer or another piece of software that a human can do. The little distraction that I think happened 2 years ago after the ChatGPT moment was that we thought about it as, “We're just going to do typing and information retrieval, and that's a new paradigm for user interfaces.” You just talk to your software and search Zillow via chat. That was a little bit of a distraction. It's super helpful when you want basic information lookup or whatnot.

The big breakthrough was starting to think through these as full, effectively agentic systems that operate on any amount of data and any amount of tools for as long as you want to complete any task that you want. This is the big year where agents are starting to enter the vocabulary of enterprises, IT people, larger organizations, and certainly small organizations. It requires you to have a little bit of a reset moment on how you think about AI. It's not just a copilot that you talk back and forth with. It's actually something running behind the scenes that's starting to deliver real, automated work for you.

There are lots of implications, including massive implications for what the software business model is in the future. I would argue strongly that it's a massive TAM increase because software starts to go after labor spend. It completely changes the dynamics of how you build a moat in a world of AI agents.

But unpack that piece there, Aaron. You said something very interesting: how software is going to go after human spend. Explain that concept. Unpack it for a second.

David and I go way back in SaaS land, but you used to basically build a piece of software and sell it for the number of people in the organization. A company has 500 employees, and you sell that thing for, let's say, $10 a user a month—$120 a year—and you make $60,000. That's the business model.

Now, when your software actually brings the underlying workflow or outcome to the customer, that company might have 10 lawyers. Previously, if you were selling software for lawyers, you had a maximum of 10 seats that you could sell. Now, all of a sudden, if your AI agents are doing the equivalent of, let's say, paralegal work or some form of professional services, you might be able to sell a multiple of the initial 10 seats that you would have sold previously.

So you see it as a huge opportunity because now you're not enabling a human to be 5% more productive. You're replacing a human, or you're replacing 1 out of 10.

Yeah. I'm going to underscore this point, though. I don't like the word “replace” because I think most of the upside is actually going to be for companies that now deploy labor on things that they wouldn't have deployed labor on before. Maybe I'm biased from the view we have, but most of our conversations with customers are about how, when they have AI agents, they can now deliver work in areas that would have been unaffordable previously. They weren't doing the work.

True in logistics. We're making thousands of phone calls a day using AI, calling truck drivers. We have 400,000 truck drivers using the Flexport mobile app. I don't have enough loads to keep them all checking it every day to see if there's a load that matches them. If they don't check it, they're useless to me now.

It was too expensive to call the truck driver and have a human talk to them, even if it was a human in a call center in the Philippines. Whereas with AI, it's almost free. We're calling thousands of them a day, going, “Hey, this load looks like it's a good match for you. Are you interested?” Then we activate them on the platform. That's new work that wasn't going to happen before, not just a replacement.

Yeah, I think probably we have, unfortunately, been co-opted a little bit in the Valley by a doomer mindset in some areas, and we think of AI as, “It's all fixed-pie. It's going to replace things.” On the ground with large enterprises, the vast majority of the use cases are the ability to finally review the contracts that we never got around to reviewing, automate an invoice process that we never did, or create marketing campaigns in every language that we never got around to.

I think that probably 90% of the usage of AI in the future will be things that, if we look back and draw the line right now and say, “This is what knowledge work is today,” we don't do today. Ten percent will replace what we're doing in some areas.

I think that's the right take because, Chamath, I can tell you, in our firm, we would never have associates, researchers, or analysts—Sacks, you also were in this line of work, venture capital—review legal documents. That's something lawyers would do in the legal department.

But now, because of AI, we can say, “Here's the SAFE. Here's the term sheet. Here's the edited version. Dump it all in. Find out what the changes are, what the deltas are, and then let's have a discussion about what the founder changed in a standard document.” We don't have to bother with an attorney, and maybe you wouldn't have even checked those documents if you were a seed fund or an angel fund. You would just go along for the ride because you're the 10th person signing the documents.

What do you think, Chamath, in terms of the premise that maybe it's 10% replacing work that's happening, but this is blue ocean and we're going to do 90% of new stuff that we just never got to?

Yeah, I tend to believe that's true. I think the customers that we sell into at 80/90 are largely large enterprises as well, so not dissimilar to Aaron's customer base.

What I would say is that what they are encountering is the trough of disillusionment. I don't know if, Aaron, you're seeing this as well, but every single CIO ran around signing up for some sort of AI product, in large part because their CEO would say to them, “Hey, what's your AI strategy?” The reason the CEO asked them that is that at some point somebody on the board said, “What are we doing about AI?” That's the cascade that we went through in the last 2 years.

I think what has happened now is people have spent billions and billions of dollars. You can see it in the revenue traction of the AI companies. But I think where we are today is that there are some real technical complexities that have not been solved. We have a lot of customers in regulated industries, which is to say that if you make a mistake, you will get fined or shut down. Life sciences, healthcare, and financial services are 3 examples.

People still don't seem to appreciate that when you replace software that is deterministic with software that is probabilistic—meaning software that somebody wrote for you: do A, then do B, then do C—with an LLM that can hallucinate, you'll have errors. What used to be a throwaway thing, which is quality assurance and QA, right? Unit testing and integration testing are now the only things that matter.

Why? Because if you're a financial services institution and you're supposed to do KYC and hit OFAC, and now all of a sudden you send a wire somewhere in Syria, guess what? You're in trouble. If you're a healthcare company and you're supposed to do some clinical diagnosis to send out a drug on time and you don't do that because the model hallucinates, that's a real problem.

I'm guaranteeing you, we have not seen the class-action lawsuits that will come when those errors are eventually made. They're guaranteed to be made. We just don't know the scope and scale of them. That's why I'm of the posture that I think we've sold in a ton of promise, but the reality is much more tactical. It's a little bit more banal. I think we're sorting through the exact use cases where you can put guardrails around these error rates, where it's okay and tolerable.

Like Ryan will probably tell you, there are some phone calls that just sound totally fucked, but he’s okay with that because the broader thing is okay. I don’t want it talking to my customers, though. I have it calling truck drivers to offer them loads, but I’m not having it talk to my customers. Sorry, I meant your truck drivers, but my point is that I think agents are real, but I think we are far away from that because we’re still at the phase of how to build reliable software in production for an enterprise versus the toy apps that you see on the internet, which are like, “Let me vibe-code something.” I think these things are worlds apart still.

Okay, so let me get Sacks in on here. Just to inform the audience, you heard “trough of disillusionment.” This comes from the hype cycle; it’s something Gartner has been talking about for a long time. In case you’re taking it for granted, if you’re watching, you have some sort of technology trigger, like agents. You have this peak of inflated expectations. Now we’re in the trough of disillusionment: “Hey, this stuff doesn’t work. It’s hallucinating.” But we’re kind of going up that—

I don’t see the disillusionment. I don’t know where this is coming from. I don’t even think we’re at the peak yet.

Oh, okay. So you think we’re still going up? Because a lot of people, to Chamath’s point, were buying stuff and saying, “Hey, it doesn’t work,” and now we’re in the mess.

Let me say it differently, Sacks. I think we have not yet figured out how to move the budgets from experimentation to mainline production, meaning where large chunks of the U.S. economy are comfortable enough with the ways in which hallucinations are managed that they will replace legacy deterministic code with this new probabilistic, model-generated code—model-enabled code, let’s just put it that way. Where are we on the slope here?

Yeah, look, I would separate change-management issues, which are always going to be important, and there are always going to be big ones whenever there’s a big disruption, especially in enterprise and especially around compliance and legal and all that kind of stuff. I would separate that from the impact of the underlying technology trend, and I don’t think the impact has come anywhere close to peaking yet.

In fact, I would say the rate of progress is exponential right now on at least 3 key dimensions. Number 1 is the algorithms themselves. The models are improving at a rate of 3 to 4 times a year. They’re not just getting faster and better; qualitatively, they’re different.

Remember, we started with pure LLM chatbots. Then we went to reasoning models. The difference there is that with a chatbot, it’s like a smart Ph.D. or college student giving you an answer off the top of their heads. With the reasoning models, it’s more like the Ph.D. saying, “Okay, let me go off and think about that. Let me do a project on that.” It could work for 30 seconds or a couple of minutes—as much compute as you want to throw at it—and it will break down your complicated question into a bunch of subquestions. Then it’ll try different approaches, validate some of those approaches, and come back to you with a much more impressive answer.

If you’ve been using Grok 3 DeepSearch or the new ChatGPT o3 to do these types of new reasoning tasks, it’s pretty mind-blowing what they’re capable of. Have we even come close to figuring out how to tap the potential there, especially in an enterprise context? No. But my point is that the rate of progress on the algorithms is, again, 3 or 4 times a year.

Okay, go finish up, Sacks, and then I’ll take it and pass it. Go ahead.

Well, I was trying to lay out the dimensions in which progress is proceeding exponentially. Number 1 is the algorithms, which is not just quantitative; it’s also qualitative. We didn’t even get to the agents part of it yet, but that’s the next big leap after reasoning models. We’re just starting to scratch the surface there.

Then you’ve got the chips. The chips are getting better at, I don’t know, 3 to 4 times a year. We’ve gone from the H100 to the H200. Now we’re on the GB200. We’ll be on the GB300 soon.

No, no, no. The chips themselves, depending on how you measure it, are getting 3 or 4 times better with each generation. NVIDIA is back to rolling out a new generation of products roughly annually, and I’m just using them as 1 example. Obviously, there are other companies as well.

Basically, the leap from Hopper to Blackwell to Rubin—I guess Rubin will be out next year—and then I think Feynman is coming after that. It’s really an astounding rate of progress. It’s not just that the individual chips are getting better; they’re figuring out how to network them together, like with NVL72. It’s a rack system to create much better performance at the data-center level.

That would be the 3rd area where you’re seeing basically exponential progress. Just look at the number of GPUs being deployed in data centers. When Elon first started training Grok, I think they had maybe 100,000 GPUs.

Colossus was 100,000, correct?

Right. Now they’re up to 300,000, and they’re on the way to 1 million. Same thing with OpenAI’s data center, Stargate. Within a couple of years, they’ll be at, I don’t know, 5 million GPUs, 10 million GPUs.

You see that on the power side, right? You’re going from 100-megawatt data centers to 300 megawatts. We’re just starting to see the first gigawatt-power data centers. I don’t even think they’re live yet, but this is where they’re trying to get to. I don’t think it’s beyond the realm of possibility that we could be at 5- or 10-gigawatt data centers in the next several years.

My point is, look, the algorithms, the chips, and the data centers are all improving or scaling at a rate of 3 to 4 times a year. That’s 10 times every 2 years. People don’t understand exponential progress. If you’re getting better at 10 times every 2 years, that doesn’t mean you’ll be at 20 times in 4 years. It means you’ll be at 100 times.

The models, the chips, and the data centers will all be 100 times more powerful in 4 years, let’s say at the end of this presidential term. You multiply those things together—the algorithms, the chips, and then the raw compute that’s available—and you’re talking about a 1,000,000× increase. Some of that will be captured in price reductions, some of it will be in the performance ceiling, and then some of it will just be in the overall amount of AI compute that’s available to the economy.

The impact of this thing is going to be absolutely massive, and I think people still don’t appreciate that fact because they don’t understand exponential progress.

Yeah. I think maybe, just to square the circle, everything that you just said, Sacks, is what I think is propelling the industry, and then the reality on Chamath’s side—just to connect the dots.

We have an eval test that we do where we run enterprise data through every model to figure out its accuracy rate and how much data it misses when we ask for facts. The best model in the world—interestingly, it was Grok 3 on this particular test. We send it 500 documents and ask for 40 data fields back from the documents, so it has to get every single data field correct, and we only do a single pass. We send the document to the model and get a single pass back.

Right now, the best score is about 90%. You can imagine a number of industries where you can’t have 90% accuracy if you give it a question on 40 data fields. There are ways to solve it: you rerun it multiple times, or you chunk up the document into smaller parts so it doesn’t get confused by the large context window.

A lot of the people who were deploying AI a year or a year and a half ago weren’t doing that. They had a pilot run of something, and it kind of worked okay. What they have to realize, back to your point, Sacks, is that this space is literally changing exponentially. You have to run the data through the model multiple times, chunk up the data into smaller parts, use a reasoning model, and make sure your prompt is hyper-tuned for the particular use case. If you haven’t done those 4 things, then you probably will end up with a project that fails.

Even when you do all those things for harder problems, you’re still going to run into issues. I think the challenge is that everybody’s running a million miles an hour right now and trying a lot of things. Some work, some don’t, at the same time that the space is changing at a pretty crazy rate.

What is the dimension? You know, we use Gemini, so for many tasks at 80% to 90%, we use Gemini. It’s incredible, but for most of our code generation, we use Anthropic, and Claude kicks ass. It’s exceptional.

This is based on the best scores in the Chatbot Arena, which just became a for-profit company. That is slightly different because people have gamed tests, so that is a rub there. People are now building their AI model for the evals, right? All the models are way overfit for these evals.

But if you had to pick, who’s your— I mean, so I guess, Chamath, you’re saying you have to take it task by task.

It depends on the task. I agree that what Sacks said is right. It’s kind of like, what problem are you trying to solve? Then you have to ride this technology wave that’s compounding very quickly.

All I was trying to get across is that the error rates have been diminishing, but not nearly as fast as you need for some sectors of the economy. You can use a model to generate deterministic code, and that’s great. As long as you unit-test it and integration-test it, it’ll be fine. But I’m saying if you’re going to use a model in production in an environment where there are consequences, we’re not there yet. You could use it for writing or writing jokes, maybe, but—

That’s too binary. It’s already used right now in health care, but it’s just the doctors’ meeting notes that would normally take 30 minutes to transcribe.

Yeah, so you can’t be too black-and-white on that one. What’s happening right now, and the reason why the progress is so rapid in coding assistance—I think you’re right that Anthropic, with Claude 3.7—

3.7, yeah.

I think they’re the leader. In fact, I think the Manus demo that we showed isn’t entirely a wrapper on Claude, because they actually do a number of different things, but I think they are significantly using Anthropic for the code-assistant part of it.

In any event, the reason why the progress is so rapid with coding is because code compiles, and you can determine objectively whether it works or not. You can validate it.

And so that makes it a perfect area for AI to get better at through reinforcement learning and test-time compute: AI tries a bunch of things. It sees what works, sees what compiles, sees what the user then accepts, and is able to learn and iterate based on that. That's why coding right now is really the big breakthrough application and use case. But it's not going to be the only one. Math is another good area where I think AI is improving rapidly, again, because in math you have proofs and you can look at the results and see if they validate.

Now, I think one of the big questions in terms of AI progress is: How extensible is the progress to other areas that don't easily validate that way? For example, legal work is a really good area for AI, but how do you validate that it's correct? You would have to go to a court, right? The court is the compiler. A lawsuit is the compiler. Or maybe the law is the compiler. You could hire 1,000 lawyers or experts in an area to basically do reinforcement learning. People are doing that, but it's not like a compiler, to your point. The progress isn't going to be as rapid because it's harder to validate.

Absolutely, but my guess is that once they figure out how to nail coding, math, and the things that are easily validated, they can move to the things that are harder to validate. I think this is one of the big questions, because people just kind of assume that AI progress will be equally fast in all areas. I think it's possible that AI gets really good in some areas—better than humans—but it's sort of childlike in other areas.

A narrow possible outcome is, to make your point, a finite answer or an answer we know is definitive.

Well, this is the important thing about the agent framework or architecture: Instead of just saying, “Okay, we're going to do a single pass through the model, and whatever it comes back with, we're going to be satisfied,” the legal work might be reviewed by another agent whose job is to review legal work. We can just throw more and more compute at the problem, and we're just early in figuring out how to architect those, or multiple models, right? You could have Anthropic check GPT and Gemini. When you have some anomaly, it spits back out to the user, so a human in the loop still matters in this type of process.

In the early days of OCR, you would have a computer say, “Here are the characters in this legal document.” Then you'd have 2 humans type it in, and then you would get a certain level of certainty.

You'll quickly find that when you layer these models on top of each other, the test-time compute costs are astronomical. Ryan and Aaron have probably dealt with this. I get a bill from AWS, and it's like, “Oh, wait, hold on a second. I just spent $100,000 this month. What's going on?” So we have to get to the bottom of that.

That, by the way, is another major trend line: The new applications that we talked about are all much more token-intensive. We went from basic LLMs, which don't require that many tokens to give you an answer, to the reasoning model, where you can spend 1,000 times more tokens just getting 1 answer to a question. Now the agents are going to be even more token-intensive than that. The amount of compute required to serve all these new applications is going to be massive, which is why I think the capex buildout actually makes sense.

When you do deep research, to your point, David, you're firing off maybe 200 queries, and it's asking—the AI is saying, “Hey, what query should I ask on behalf of the user?” Then you go down that rabbit hole. It's basically like doing 200 of them at once.

Ryan, your thoughts here on AI-first companies and agentic computing?

The one I really wanted to tie back to was actually our earlier conversation on tariffs. There's a very real use case for LLMs: How do you classify a product? What we see today is that when we did our first machine-learning-based natural-language classification of a product, you take a product URL, a listing page, a Shopify page, or an Amazon page, and say, “What classification code is this? What duty is owed?” Six years ago, in a hackathon, we got to about 70% accuracy. We're now in the high 90s accuracy versus what a human-trained expert will get to, but what you actually get to is not good enough. You're wrong 3% of the time. You might have committed a violation of the law, for sure.

But actually, what is truth in that regard? There's a lot of gray area in this. And truth ultimately is: What does Customs say? What does the CBP determine is correct? Those guys are using software, right? That's a very simple algorithm, and it's a decision tree that's going, “Okay, is it a shoe? Yes. Is the top made of leather? Yes. Is the bottom made of rubber?” They just go through a very simple decision tree, and that outputs it.

On some level, if you convince the government to use your LLM, it becomes true whether it's true or not. I think there's going to be some interesting cases like that that we haven't really thought through, like when does the government adopt these to be the source of truth?

All of this speaks to something that's going to sound totally esoteric, but we all used to shit on QA, right? The least talented engineers were allocated to QA. I think in the world of AI it'll end up being the most talented. We internally at 8090 call it improvement engineering, and it's a total specialty. It's similar to when I kind of coined the growth team at Facebook. I feel it's the same kind of moment where improvement engineering is really the skill that translates toy apps and vibe coding into something that's very practical and real.

My team and the leader of this team—he's steeped in things like Japanese kata management from Toyota and quality systems—and these are all the things that matter when you're trying to shrink the error rate down to 0 so that you can use it in a reliable way, and also to document it so that if people want to question what happened or have recourse, or some way to come back and say, “Hey, that really harmed me,” how do you even do that? These are all very complicated issues that will get sorted out.

Super interesting.