Patrick McKenzie (patio11) is joined by returning guest Zvi Mowshowitz, AI watcher extraordinaire, to read Anthropic’s S-1 and what it shows about the AI industry. They cover a founder-controlled structure built on a single Class F share that holds 50.1% of the vote, an uncapped 3:1 employee donation match that may be the most expensive perk ever offered, and about $900 billion in forward compute commitments. They also discuss why softer Q3 revenue points to a shortage of GPUs rather than of customers, the bubble and circular-financing bear cases, and who should be blamed when AI-assisted hackers take $5 billion from Binance.

Presenting Sponsor: Mercury
Complex Systems is presented by Mercury. Mercury Payroll puts payroll, benefits, and compliance right inside your Mercury account. An in-house team backed by AI registers you in new states, files and remits your payroll taxes, and handles the notices, so you don’t have to copy from PDFs to Excel to web portals. Setup takes minutes, and Mercury Payroll is free for eligible Mercury customers through the end of 2026 at mercury.com/payroll

Timestamps:

(00:00) Intro
(01:35) How the S-1 talks about existential risk
(07:45) Anthropic has long had an attitude
(12:20) Anthropic’s uncapped donation matching
(14:18) Postscript: what the matching program will actually cost
(16:00) Sponsor: Mercury
(18:29) Revenue growth and cohort economics
(22:25) Postscript: Q3 was a supply story, not a demand story
(24:55) Is it a bubble?
(28:27) Sizing the market for tokens
(31:02) The commodification bear case
(31:56) Cash efficiency and $800 billion in compute commitments
(36:44) Postscript: what happens if the commits go bad
(39:20) Flexing compute between training, inference and subleasing
(43:37) Reading an S-1, with and without an LLM
(46:49) Circular financing and the dot-com comparison
(50:08) Wall Street versus the tech industry
(52:02) The three AI pills
(54:53) Industrial revolutions and how weird the future gets
(57:31) AI mathematics and the bitter lesson
(01:00:52) Capabilities nobody trained for
(01:04:40) Capabilities versus diffusion
(01:07:12) Token spend at the frontier
(01:09:09) Final thoughts on the S-1
(01:11:48) The hackening and who bears responsibility
(01:17:49) Trapdoors and irreversible decisions
(01:19:12) Wrap

Patrick: Welcome to Complex Systems, where we discuss the technical, organizational, and human factors underpinning why the world works the way it does.

Hideho, everyone. My name is Patrick McKenzie, better known as patio11 on the internet. I'm here with multi-time guest Zvi Mowshowitz. Thanks very much for coming on.

Zvi: Glad to be back. Yeah, I always enjoy it.

Patrick: So we're discussing the Anthropic S-1 today, and the larger window that it shares on the AI industry.

And because we are professionals discussing professionally adjacent topics, we have to start with the boring parts, similar to the boring disclaimers in an S-1. I don't believe I own stock in any of the labs. I don't invest in publicly traded tech stocks except through ETFs that are broadly market-tracking. [Patrick notes: You don’t come to me for investment advice, and that is a good thing, because this has been a really poor decision for the 20+ years I’ve been in tech.] 

I have a career in the tech industry, which is getting drawn into the gravitational well of the labs right now and the other parts of the AI supply chain, and you should assume I'm conflicted up the wazoo in that fashion. But I think I'm trying to be a reasonably responsible person talking about this company, and I think that's all for me.

[Patrick notes: To be slightly more explicit: I am socially connected to many people at the labs, including former coworkers and bosses, a few of whom are now quite senior. Both major American labs have had commercial discussions with me. Other non-lab companies in the AI industry have paid me substantial amounts of money for consulting and, separately, advertising spots on Complex Systems. You may reasonably model these and similar relationships recurring in the future. I believe that Stripe equity, which I own through having worked there, and which is the supermajority of my family’s assets, is materially exposed to the AI industry both mechanically and via investor sentiment.]

Do you want to add any for yourself?

Zvi: I do buy individual stocks. I own Google and Amazon, relevant to Anthropic, but I deliberately did not invest directly in Anthropic because of potential conflict of interest concerns.

Patrick: And nothing that you hear on this podcast is investment advice. Okay. So let's start it out.

You recently wrote something of your own perspective on the S-1, which had some key takeaways. Do you want to walk people through those key takeaways?

How the S-1 talks about existential risk

Zvi: Yeah. So as usual, when I look at anything from Anthropic, I want to see their attitude towards existential risk, alignment, whether or not they're going to potentially kill everyone.

So in this case, they mention existential risk only in passing, only in the passive voice. They're clearly trying to have it both ways. When they talk to the investors, they want to say, "Oh, look at our big mission, but our big mission is to build this super profitable hypercomputing thing." And then other times they want to say, "Our geniuses in the data center actually might kill everyone," and these two things have to coexist.

Similarly, pacing the frontier is something they do talk about here, but then they make clear to investors, "Don't worry. When we said pacing the frontier, we did not mean not making incredibly fast AI progress. We did not mean voluntarily let someone else go first. We just meant it would be great if everybody could collectively agree to go only at super speed and not at ludicrous speed." But they will do the embedded evaluations.

Similarly, we're all worried about recursive self-improvement. And here they call that "closing the loop" to be more business-friendly, but they say that their entire business model is they want to close the loop. They are worried that if someone else closes the loop before they close the loop, they will be out of the loop and they will lose.

So they intend to close the loop as fast as possible and as much as possible.

Patrick: I feel like there is an interesting tension around S-1s, because we largely have this culture in the tech industry of, necessarily, Zuck speaks for Facebook and Facebook speaks for Zuck. Tech has done complicated innovations with share classes that enable the founder/company relationship to be relatively durable over a period of years and decades.

[Patrick notes: Facebook was not the first company to issue super-voting founder shares but it popularized the model. Pattern matching goes both ways with Silicon Valley investors: you should believe your new investment is the second coming of Facebook. If you don’t, stop now; returns in venture are heavily weighted towards a handful of generational companies, and the entire game is spreading out your fund such that you are definitely in them and persuading the founder that you will be easy to work with (and so please don’t cut your allocation in any round). If you believe you are investing in the next Facebook, you affirmatively want the next Zuck at the helm durably regardless of how diluted their economic stake gets as they raise capital. Public market investors, seeing a company which grew from a dorm room to a global power, largely seem to accept precedent on provisions guaranteeing founder control. If founders have them at IPO, they keep them.] 

Patrick continues: Founder-led companies are something of a thing, and Anthropic is doing its own thing with respect to complicated share classes, which we can get into in a minute, if people are very interested in the mechanics of single shares that have more than 50% of the voting power, yada, yada, yada. But when you get the quote-unquote "responsible adults" in the room — S-1s are largely drafted by external counsel and counsel working for the banks that are the underwriters of the S-1.

Responsible adults say: "We all understand that everyone in the tech industry thinks that the CEO runs the company. However, managerial capitalism says that the CEO only runs the company to the extent that they're not fired by the board on behalf of the shareholders.

And while you've taken some amount of effort to make sure that that never happens, we have to pretend that in principle it is possible that it happens, and so we can't privilege the CEO's voice among many other voices." And there's a difference between when the CEO is stating their own opinion in their own spaces and when they're speaking on behalf of the company, et cetera, et cetera, et cetera.

And I think in some parts, there is the usual level of cynicism that prevails in corporate communications, and in some parts authors sincerely believe, "No, there is a ritual here. We deeply buy into the ritual, and we care about the boundaries of it." But there has been much mirth around the use of passive voice in talking about the AI risks to humanity, because "concerns have been raised" is an understatement, as has been pointed out by yourself and Matt Levine and many others.

The “concerns” call was coming from inside the house, at the executive level and essentially all other levels of Anthropic. So it is an interesting inversion of the normal form, where the normal form is not claiming that the company could potentially kill everyone, and that would be bad for business results.

Zvi: Yeah, I mean, obviously it's a compromise between — the bankers would rather not mention this at all, because if everyone is killed, you will not in fact be sued for securities fraud. There will be no one left to sue you. And Dario and others at the company presumably wanted to use not only the passive voice but the active voice, and be much louder and more extensive about it, and use it as an opportunity.

Patrick: And there's also liability-minimization happening here, where lawyers will always advise you, "Don't say anything that you don't need to say. Anything that you don't need to say to get the transaction, deal, testimony, et cetera done just creates risk needlessly." And so if you feel emotionally or aesthetically or morally committed to saying, "No, we could kill everyone," the lawyers will say, reasonably, "Well, when you say that, you're providing causes of action to people in the future for any of a variety of things that could happen, including nothing much different than what is happening in the status quo.”

[Patrick notes: I believe this factor is underrated. Consider a very prosaic lawsuit about an ordinary corporate matter. If the opposed side wants to demonstrate cavalier recklessness, which is frequently good legal strategy, one way to demonstrate recklessness is contrasting testimony that all normal steps were taken with “Really? Because at your firm, doing things that might end humanity is a normal Tuesday. Do you recall the following language from your S-1/senior executives/company culture guide/etc.”] 

The lawyer will continue: “And that will be quoted against you, and it will be worse for being quoted against you from an S-1 versus from a blog post, because the various people that look at these things, like regulators in Washington, DC generally and potentially courts in the future, care about things in an S-1 much more than they care about things that are in a blog post.

And we will be able to wave away things that are in the blog post in the future as, 'Oh, that was just his opinion,' et cetera, et cetera. But we can't really wave away things in the S-1.”

Zvi: Right. And from my perspective, the whole idea is, everything in terms of the financial incentive — the lawyers, the corporation people, the lobbyists, the PR guy, the usual suspects, all the business interests — will in fact want to downplay and ignore and sweep aside, and make everybody else sweep away and ignore, all of these concerns.

And my view essentially is that if we do too much of that, there's a very good chance that in fact AI will kill everyone, either from Anthropic or otherwise. And the argument that, "Oh no, this is the place where we have to say it in a way that matters, therefore we shouldn't say it the way that you want to say it" — well, I say that's exactly my reason why you should say it there, and why, if you choose not to make a costly signal, you've told me who you are. You've told me you're not willing to make a sacrifice that's meaningful. You are signaling to me that, in fact, you will prioritize other considerations higher. You just did so.

That's the whole point of the exercise. And I understand entirely why people would choose to prioritize messaging to other people in these situations, but I never treat it as an excuse or as a reason to not care in the other directions. Should you choose these other considerations over the considerations that I care about — and maybe I even think you're right to do so—I'm going to knock you for having done it.

Founder control and Anthropic's share classes

Patrick: Cool. Shall we go to the second part of your takeaways, which was on control, I think?

Zvi: Anthropic has long had an attitude, like many tech companies, as you point out, that the founders should continue to control the company, and that the worst thing that could happen to Anthropic would be to have the likes of Google and Amazon suddenly having them maximize shareholder value and then thinking about only corporate things. And then their company would probably go out there and increase the risk that the world is destroyed, and lose their unique company culture, and this would be extremely bad.

So to that end, Anthropic has a very weird structure that they made even weirder in advance of the IPO. They have the one Class F share that belongs to the founders, and the founders collectively vote on how to exercise that share. And that share has 50.1% of the voting power for a wide range of issues, which is another way of saying the founders will keep control of the company as long as at least three of the founders continue to show up to the company, which could last quite a long time.

Anthropic's anticipation is this will last until long after the fate of the world has been decided one way or another. And this seems rather likely, because there are something like nine of them, and only three of them have to keep showing up.

Patrick adds post-interview: patio11 here with a brief correction of the record. So formally, Anthropic has seven founders, and there is an additional sort of early employee who gets a footnote, Benjamin Mann, but he is not formally one of the founders and doesn't get this inclusion into the Class F share class.

Also, by the way, Zvi says at one point that it's Class H that might side with the founders. The employees that donate equity to charity are getting Class E voting rights rather than Class H voting rights. Okay, back to the flow.

Zvi: And similarly, they have set up a Long-Term Benefit Trust, which will appoint nominally four of the seven board members over the long term, which in theory would've given them control of the company.

But if you look at the actual details, the board appoints three, the Long-Term Benefit Trust appoints three, and then there's the seventh seat, which is suggested by the Long-Term Benefit Trust, but has to be effectively approved by the founders. And if the founders don't agree, then it goes to the vote of the employees who donated stock to charity, also known as the Class E shares.

And a default assumption would be that the Class E shares would side with the founders.

Patrick: I don't think I've ever heard of another company explicitly privileging employees with share classes. Generally, it's something of the reverse. 

[Patrick notes: Employees most typically get stock options or restricted stock units which are designed to convert into common stock, or into a second class of common that is economically identical and differs only in voting, with the distinction typically disappearing after IPO or sale. Companies sometimes offer employees inducements to impair their stock in some way, such as accepting a voting proxy agreement (ceding their votes to the founders) in return for e.g. better-than-market exercise terms for their options. 

Common shares vote like Republicans in Chicago: mathematically you have no say in outcomes, but you do have a right to have your opinion tallied.]

And it's interesting. It says you believe something strongly about your alignment, in the non-central sense of the word for AI, in that you have successfully recruited multiple cohorts of people who —

The traditional thing to do in the tech industry is to have the employees proxy their votes to the founders, and the founders literally get to vote them in whichever way they want. And you're, in doing this, suggesting, "No, I think the employees are sufficiently philosophically aligned with management that they will proxy without a proxy agreement being in place," which is an interesting point of view.

And I'm not saying "interesting" as a way of saying I don't think it is well-calibrated. In my many conversations with various people at Anthropic, they do seem to be quote-unquote "more aligned" than one would suggest for most tech companies at that stage, with that employee count, at that revenue growth rate, et cetera, et cetera.

[Patrick notes: Many companies experience, after hypergrowth, discontent among long-time employees (and sometimes founders) that the company has lost something special in the culture. Sometimes this is diagnosed as “we got lame and the cool people left.” I think it is more similar to reversion to the mean: you had a core of N people, distinguished in some way. You then intentionally, as a mechanical consequence of hypergrowth and with your eyes open, hired 10 people with impeccable credentials from AppAmaGooBookSoft for every one who previously worked with you. You will always find your culture drawn to the median culture of AppAmaGooBookSoft as a result of this choice. You will be warned about it in advance, and you will say “It will not happen here!” It will happen to you. It happened to AppAmaGooBookSoft, too.] 

Patrick continues: Not that there is another company at that revenue growth rate to point to, but we will get to that in a moment, I think.

Zvi: Well, so before we move on, I would say Anthropic has long cultivated a very special company culture that has a large awareness of safety, but also of altruistic goals and charity and making the world a better place.

The founders have famously pledged 80% of their shares to charity, and they've offered very, very generous matching donations for those who want to give their stock away. And in fact, how much stock you gave away determines how many Class E shares you have to vote. So they're, on several levels, determining, no, these are the preselected people who have very, very costly signaled that they want to influence the future in generally good ways, rather than build personal wealth and maximize profits.

Those are the people who will ultimately make the tie-breaking vote in this situation, and that is exactly what that was selected for.

Anthropic's uncapped donation matching

Patrick: I couldn't find if this was explicitly called out in the S-1, and I will edit the transcript if I missed it in my first two readings. But their matching program for employees might be one of the most expensive single corporate decisions ever made in the history of capitalism.

The matching program for employees was, I hear, calibrated to essentially allow employees to do something similar to what the founders were doing, and give away 80% of their economic stake. And to make those numbers work, they said that the corporation will basically match employee donations on a 3:1 basis, which is unheard of in the rest of capitalism.

And additionally, typically a corporate matching program is: okay, we will match on a one-to-one basis up to, let's say, $7,000 a year. And so you say, okay, well, optimally executed, that option is worth $7,000 to charity, and perhaps $10,000, $11,000 or so in pre-tax income. The Anthropic matching program is uncapped.

And so there are employees that will donate $20 million to charity and have it get backed by $60 million of Anthropic money in the wake of the IPO and their continued operation and similar. And when you sum that over extremely charitably minded employees, times thousands of them, times many of them that have presumptively eight- and nine-figure equity values, which they will quickly be giving away, that is just going to be an extremely expensive decision.

Zvi: Yeah, but they got what they wanted out of it. You want to strongly differentiate yourself as an employer and attract aligned employees.

If you're not committed to the mission, it sounds expensive, but if you are mission-aligned, the expense is a positive. And when you turn out to be the fastest-growing company in history, and the stock is worth close to $2 trillion, it will be a larger line item than you thought.

But at the same time, you'll be able to afford it conditional on success, so it works both ways.

Postscript: what the matching program will actually cost

Patrick injects via new recording: I failed at Ctrl+F prior to recording. The S-1 does indeed mention the donation matching program in a few places. For example, it says that the Class N shares are reserved against the need to make good on these commitments, and the company took a $2 billion non-cash charge in Q3 related to it.

So in terms of cash, the matching program will be close to free. Shareholders as a class, inclusive of donors, pay for it through dilution. In terms of accounting, though, when you promise people compensation worth billions of dollars, we want to record the fact that you have promised billions of dollars.

When you promise them compensation based on future world states, we update the cost of that promise as we learn more about the future. I suspect the Q3 charge is very far from the last non-cash charge attributable to the matching program. Future charges will happen as employees make new pledges, as past pledges not yet settled get remarked from a 409A valuation to the public price, and, probably the dominant factor, as currently pledged equity continues to vest and gets matched at public market prices rather than the private ones.

The S-1 suggests there is presently about $24 billion in unvested equity compensation, with that calculation being made as of the day the equity was granted. When the cost of the match is calculated, it is calculated the day the equity vests, which will probably be at a higher valuation.

And then you assume that much of the equity is going to employees eligible for three-to-one matching at the new mark, and much of the remainder to employees eligible for one-to-one matching at the new mark, and that probably tens of percent of it will be donated. And so, yep, that is almost certainly the most expensive perk in the history of capitalism.

I suspect the magnitude of the first post-IPO charges is going to surprise people. Much has been written, including by my former colleague Nan, of the coming tidal wave of charitable donations. A large portion of that tidal wave flows through Anthropic's P&L.

Revenue growth and cohort economics

Zvi: So the next section of questions is about money, of course.

So Anthropic is one of the fastest-growing companies in history. We're talking about 10x year over year for several years in a row. We're talking, in some places, even greater than that. First two quarters this year were completely nuts.

Patrick: About 15x in the first half, right?

[Patrick notes: $1.16 billion in H1 2025 to $16.3 billion in H1 2026, so about 14x, though what is one entire prior-year half between friends.]

Zvi: Right. Approximately, something like that, 15x in the first half, from a very large baseline to begin with, rather than from very small amounts.

I mean, in the beginning it was like, okay, you're 10x-ing, but this is not unheard of because it's a very small company. Now it's a relatively large company, and people are asking questions like, how many times can you 10x? The entire economy is only so big. There are only so many Fortune 500 companies.

It's just 500. At some point you're going to run out of targets. But that's not necessarily something we're getting close to. Quarter three, they did slow their growth. So their growth was still the envy of almost every company on Earth, but by their own standards, it was reasonably small.

We have a variety of explanations for this. Anthropic was relatively badly positioned in terms of the quality of their available models and their appeal to business, compared to how they had been in the previous period or in the later period that we're in now. So that's probably my guess as to the main reason.

There's also some talk about quarter three just being a slow period for growth, relatively speaking, due to the calendar. But my expectation is that quarter three was relatively slow, but that quarter four has picked up again, and you will see some very, very large growth numbers out of Anthropic.

Patrick: In analysis of securities, we have to pretend that the gross quarter-to-quarter revenue results matter. But in this case, I think Anthropic is properly modeled more or less as a software company, where what you are doing is stacking cohorts together. And for reasons, we do not require software companies to break out their cohort numbers in S-1s and similar.

[Patrick notes: This is extremely well-understood among the best managers in the world, and relatively well-understood among investors. One classic articulation of it is Bezos’ in 2017, referencing a much slower-moving business which was once the fastest growing business of all time: “Those quarterly results were fully baked three years ago. So today I'm working on a quarter that will happen in 2020, not next quarter. Next quarter is done already and it's probably been done for a couple years.”] 

If we did, that would be a fascinating window into the growth of the business. And so the cohorts are often heterogeneous with each other, and so the customers that you recruited in Q3 2026 are at a different point in their ramp of spend than the customers you recruited in Q3 2025, than in 2024.

And the very interesting thing that you would want to see on those graphs is: do they asymptote at any point? And finger to the wind, my guess is the answer is no. The S-1 says they have a net revenue retention of 900%, which is silly for any other SaaS business. If you were doing 110%, that would be wonderful. 

[Patrick notes: Slight exaggeration; that’s workmanlike execution for an IPOable SaaS company selling on annual contracts.]

But my guess is their "early adopter customers," the ones that were there in the ancient history of two years ago, have not figured the limits of cognition in their businesses yet, and are continuing to pour money at tokens and get business results out. And if you were hypothetically in a world where a cohort graph would show that, okay, after the first two years of adoption, the growth pace goes to that of a normal software business, then you would be much more bearish on Anthropic and labs generally as a business.

I suspect we are not in that world. I suspect the line for people that have been on two years continues to be vertical, and it will look less vertical in the first three months of a customer's life, as they do things like internal pilots, IT starts to get involved, yada, yada, yada, all that fun stuff for the absorption of this technology in the world.

And so I suspect quarter-to-quarter softness in the number is more an artifact of how the cohort graphs stack up than it is an artifact of the total corporate positioning relative to competitors in that quarter. But I do not have internal numbers on this. Obviously, if I had internal numbers, I would not be able to share them.

Postscript: Q3 was a supply story, not a demand story

Patrick: After recording the episode, I thought through this a bit more, including with Claude assisting, and have a better hypothesis.

As I mentioned, it is useful to have lenses available, like being a SaaS-y guy, because it lets you see the world in terms of cohorts. It sometimes impairs your vision, though, because you imagine every business as having arbitrarily high margins and the ability to scale to customer demand as quickly as DevOps can type into their terminal.

Anthropic is, importantly, less like a software business in those respects. Anthropic says in the S-1, "We have historically operated in a compute-constrained environment in which demand for compute has exceeded the supply available to us." Their ability to service revenue from fully bought-in customers is limited by the contracted, racked, powered GPUs available to them.

The S-1 says historically they make these arrangements 12 to 18 months in advance. And so, in roughly Q1 or Q2 of 2025, they had to make a guess as to how compute efficient a yet-to-be-released Opus 5.5 would be, and how much customers would be willing to use it. 

Anthropic seems to have underestimated or —in Zvi's parlance elsewhere—perhaps simply not to have been aggressive enough.

Their Q3 revenue softening isn't a demand story. It's a supply story. They're running all the GPUs they have, and it just isn't enough. Supporting evidence: they talk about hitting the spot market for GPUs for $5 billion, at substantially more than their multi-year commitments typically run, in such a way as to eat 10% of the margin just from the shift in the mix.

You'd assume low tens of percent of revenue is served by spot compute, but that spot compute made up 50% of their compute spend for the quarter. The S-1 says "short-term commitments," which might not line up one-to-one with, for example, how AWS will sell you a spot instance. Roll with me here on the lingo.

The important bit is that you're paying through the nose for availability as a trade-off for not contracting for that capacity in advance. And so why has the revenue softened? Either because you want to make the principled decision that you don't want to climb the cost curve in the spot market and sell at wafer-thin margins or for a loss just to inflate your top-line metrics, or perhaps because you just tapped all the spot that capitalism had available in Q3 2026.

I think analysts who read this as softening demand would need to explain tapping spot for 50% of one's spend. I frankly don't think there is an explanation available there. These analysts are just straightforwardly wrong. Even if one allocates all of the spot compute to training runs for Opus 6.x or whatever, that's still saying they fully saturated the long-term contracted chips, and their customers still had tokens waiting.

Is it a bubble?

Zvi: Yep, it's really hard to say. There's a lot of different factors. We have very low visibility, for a variety of obvious reasons. We have to endlessly hear the same people predict the imminent fall of OpenAI and Anthropic, and the imminent commodification of the models, and the imminent — everybody's going to stop spending so much on tokens.

And they say this every time the number 10xes, which has been a remarkably large number of times. And yes, at some point, number has gone up, and estimates of the future rate of number going up will go down, and therefore stock price will go down even as the token usage and the revenue numbers go up. And at that point, people will talk about the bubble bursting, and they will be wrong.

That is the default. Obviously, nothing I say is investment advice. As per usual, nothing you say is investment advice. But it's very easy to predict who will continuously predict these things happening and who will not.

Patrick: I think reasonably well-calibrated people could think, just on priors: I'm not paying much attention to what AI is doing. I've just heard this story before. "This time is different" as a meme. This feels in some ways like the classic examples from Byrne Hobart and Tobias Huber's book, Boom, where when a new input for industry gets developed. We overshoot the actual demand for it for a couple of years.

We dump hundreds of billions of dollars or more into investment that is initially unproductive. The early investors get shellacked at some point. Then, there is a period of wandering in the wilderness, and then we eventually figure out how to use all the infrastructure investment that we made.

Zvi: Yeah, there are so many differences with that model that exist here. One of which is that the initial investment is not initially idle or useless. There is actually overwhelming demand for everything that you can get right now, while the systems get clearly better at a very rapid pace. So even if you're not going to talk about AGI or superintelligence or all the things that we're going to do in a couple of years, this just seems transparently very different to me.

Patrick: Yep. Not to kick crypto while it's down — by the way, breaking news from the day before we recorded this: Binance was hacked for somewhere north of $5 billion, likely in a way that was accelerated by AI by North Korea.

[Patrick notes: This podcast was recorded immediately after the attack; Binance claims that partners successfully froze or intercepted approximately $2.85 billion of $6.34 billion taken. One should expect that number to evolve over time. Either way, North Korea’s Lazarus Group has stolen the crown for the largest robbery in history from… themselves. Their previous high-water mark was the $1.5 billion stolen from Bybit in 2025.]

At any rate, crypto boosters consistently said, for 15 years, "Next year is going to be the year of mass institutional adoption."

[Patrick notes: They have a reasonable argument that stablecoins for B2B payments are a scaled, novel, non-speculative use case for crypto. They also confidently said that Bitcoin would be used for payments because its security and fees were superior to banks’ offerings, that securities would move to blockchains out of technological superiority and to decrease counterparty risk, that NFTs would lead to a new flourishing of the arts, that novel VMs running on top of blockchains would be the primary application development platform for the financial industry, etc. And crypto advocates made a lot of money investing early and convincing others to buy them out of those investments using those claims, which makes them retrospectively effective claims if not true claims, to a certain strain of logic.]

And AI is like, "Well, no, this year is the year of mass institutional adoption, except next year we will 10x this number." A billion dollars in revenue going to $16 billion in revenue, going to who knows what it'll be next year, for a part-year period, is just wild. So I think I'm with you.

I think folks that are doing the "this is just a bubble because it pattern-matches to every other bubble" thing.

As dangerous as it is to say "this time it is actually different," the actual object-level reality matters. This time is indeed discontinuously different compared to other bubbles we have seen in our lifetimes and historical examples.

[Patrick notes: “This time is different” is a phrase which notoriously ages poorly after it is said during bubble conditions. “This time is actually different” ages even more poorly. We will see how “This time is actually different and I believe I have a compelling vision for why this statement will not blow up in my face, enough to explicitly acknowledge the track record of generations of investment analysts who have said the same thing” does.]

Zvi: Yep. We'll keep talking about the money, obviously, but I really do not think this shows particular evidence that this is a bubble. I mean, obviously the number might go down. Number might always go down. But I definitely would not want to bet against any of this. That would seem like a very, very poor idea.

Sizing the market for tokens

Patrick: Five years ago, I thought there is a reasonable chance that LLMs asymptote “just a bit higher than where we are today” on the capabilities curve. And after being chastened by thinking that many times over the past five years, I now think, well, it's extremely unlikely that it will asymptote just a bit over where we are today. We'd have to hit an incredibly narrow bullseye for that to happen. 

So I'm trying to update on evidence. There are some other people that have not really updated on that evidence. But even if we never got better than Opus 5.5 and Fable 5.1 — if this was the best it ever got, just waiting for this to roll out for the next couple of years in capitalism would make these some of the largest businesses by revenue in the economy.

Now, obviously they're structurally predicated on that not happening. But John Collison of Stripe — which previously employed me, and I'm not speaking on their behalf — had Sam Altman on John’s podcast Cheeky Pint a few days ago, and John said—and all of the following is my gloss rather than a quote—, plus or minus, round numbers, 1% of the entire operating expense of Stripe is AI tokens right now.

Asked what would the correct number be, I would describe John as hemming and hawing a bit, and then he suggested it would be in the tens of percent of OpEx.

If you have white-collar workers doing intellectual labor, then clearly tens of percent-ish of your entire opex is the quote-unquote correct number, and it will take you more or less time to get to that. And when you multiply tens of percent times the amount of intellectual labor in the economy, you end up with a total addressable market in the tens of trillions of dollars.

[Patrick notes: the above is my verbal gloss; I believe it to be fair to the perspectives of both speakers. You can find it in Cheeky Pint’s transcript. Ctrl-F for “1%-ish order of magnitude of Stripe’s OpEx”.] 

Which implies 10, 20x plus, plus, plus that number as the aggregate value of the labs and the supply chain from them, et cetera, et cetera. Yeah, that's a wild number to even put in one's head and think about.

[Patrick notes: Implicitly here, I believe on their public statements my general sense of industry trends that Stripe is a leading, scaled early adopter of tokens-for-all-the-things. Given Stripe seems to see a path to an already-heavy-user 20X+ing their usage, the total value ascribed by customers to tokens would be enormous. That doesn’t flow directly into 20X+ increase in the value of the equity since investors are already making predictions about future growth, but I think John Collison is more bullish on the check he will write for tokens than AI investors are regarding the check he will write for tokens. As always, though I was once an employee for Stripe and am currently an advisor there, Stripe (and John) do not necessarily endorse comments I make in my own spaces.]

Zvi: Yeah. And that even discounts the idea that there will be completely new forms of labor and production enabled by AI, which there almost certainly will be, even if, again, we don't advance much.

So the bear case is not a bear case of "if we don't improve the AI models more, there won't be that much for the AI models to do." There'll be a tremendous amount for the AI models to do.

The commodification bear case

Zvi: The bear case is essentially the expectation by many of commodification. It's the idea that, well, the cheaper, faster, open-weight, competitive models are going to be good enough for whatever you want to do.

You will choose to use them instead. If Anthropic's models don't get better, then a year from now, they would think, well, okay, at that point, the DeepSeek model or the Qwen model will be almost as good, and it will be cheaper and faster. So if all that Anthropic does is make their own model cheaper and faster, this will destroy their revenue margins, and their business will no longer function.

But there's no longer a case of, oh, AI doesn't do much, and the economy will adopt almost no AI. Again, even if AI is the minimum it could possibly be, and we never see another release that's any good again, and Carbon turns out to be terrible, and Fable 5.5 turns out to be a disappointment, and we go from there, and so on and so forth.

Cash efficiency and $800 billion in compute commitments

Zvi: The other thing to know about Anthropic's financials is that Anthropic's financials are actually incredibly good for a startup that is growing at this kind of rate, especially one that just eats capital in various ways. So you would think that Anthropic would have spent a tremendous amount of money to get this far, and it turns out they haven't.

Not on net. So they have raised a total of $141 billion, according to the S-1, and they have $131 billion of cash, cash equivalents, and investments on hand, which means their true loss, in the sense of how much cash flow loss they have in total, has been only $10 billion getting here. And they are directly profitable in the first half: cash flow of $3.9 billion positive.

Now, if you adjust for the convertible notes not having been converted yet, and therefore they have to be considered as losses every time the stock price goes up, they post an $80 billion loss from that. And therefore, they are shown as having a seventy-plus-billion-dollar loss. But this is entirely just an artifact of accounting, as far as I'm concerned.

They've sold that stock. Amazon has not converted to stock, but for all practical purposes, Amazon just has stock in Anthropic. So we should be thinking of this in terms of, okay, that's just part of the company's cap table, and they shouldn't be charged for losses any more than we should charge for Dario's stock going up in value, and therefore having been a more valuable thing to give to Dario.

That is a very silly way to do accounting. And then the actual expenses that we have to look forward to are — Anthropic has compute commitments of roughly $800 billion.

Patrick: The one thing I would say with regards to them being incredibly cash efficient — and they have been incredibly cash efficient — is that there are ways to trade off cash and future promises in ways that matter, and they've done some excellent corporate finance.

And I'm not saying that as a criticism at all. This is just someone who enjoys the game looking at a team of people that are very, very good at the game. They have locked up compute for the next couple of years in a trade where both parties get something they want. Anthropic says, "We want a price, and we want availability," and the other party says, "Okay, we want a guaranteed spend over a number of years."

And so a way in which they keep their cash expenses in the current period low is to say, "Okay, we're going to smear out a very large commitment over the next couple of years," and they're up to, I think the number is about $800 billion in commitments going forward. I think a midwit way to look at that is, well, that's equivalent to having $800 billion of debt, just like it would be if you had $800 billion of real estate leases. And that could cause, quote-unquote, "the bankruptcy of the company" if demand growth slows, or yada, yada, yada.

[Patrick notes: The conventional way to round $895 billion is to $900 billion but what is $95 billion between friends.]

I have reasons for not thinking that, but what's your perspective on the forward commits?

Zvi: So Anthropic has been incredibly conservative, in my opinion, on its forward commitments and its acquisition of compute. OpenAI has been much, much more aggressive. And still, OpenAI proved to be more conservative than it would have been optimal to have been, as things played out.

But you are trying to guard the downside option. And Anthropic has basically been saying, "Oh, we don't want to go broke and suddenly be in this position if growth is only very good and not insane." I think that they basically underestimated what the floor should have been on their potential growth, slash how dead they would be as a company in terms of their plausible amount of quality future if growth had in fact stalled for various reasons — that they were now falling out of the top tier and were not able to make these sales.

That didn't happen. But if it had, it is not clear to me that not having bought enough compute would have saved things that were that valuable. Whereas the cost of not having the compute available when things go really well is very high. Obviously, there are reasons to want to be risk-averse, but I do think they played it too conservatively, even though from an outside perspective, oh, $800 billion of forward commits, that's completely insane.

That's over a number of years. Compute is growing rapidly. AI is growing rapidly. The only way this works out badly for them is if the price of compute actively goes down substantially, so they can't resell it, and there's no demand for it. And that is a very strange world that I very much do not expect.

Remember that SpaceX, right, bought a ton of compute, right? They built Colossus, they built Colossus II, they did all these things, and then they used some of it to train Grok models, but nobody wanted the Grok models. So they didn't have any inference customers. So they couldn't sell the inference — they couldn't sell the compute to make money, and they didn't really have enough use for the compute to train models.

But now that is the main way that SpaceX makes money, because they sell that compute to Anthropic.

Postscript: what happens if the commits go bad

Patrick: Hideho, folks. So, a bit of a post-interview riff on this topic. There's an old saw: if you owe the bank $900, you have a problem. If you owe the bank almost $900 billion, the bank has a problem.

What does a future look like in which the buyer is unhappy about the $900 billion in forward commits? The most likely future is one with globally soft compute prices. You get there via some combination of: A, a supply glut of powered and racked GPUs because of capital overshooting demand.

B, major improvements in inference-time compute efficiency. C, a saturation of current approaches to intelligence, such that new training runs don't improve economic impact and are therefore shelved. D, policy restrictions on the amount of compute that can be put into training, or perhaps, less likely, inference.

Or E, a large decrease in global demand for cognition. I think you need to imagine some improbable world states to get there, but perhaps one gets there. Okay. What does one do with the commits then? Does one proceed straight to bankruptcy, like, for example, WeWork eventually did to get out from under many billions of dollars of now-uneconomic leases? No.

[Patrick notes: I had WeWork bonds, believing that they could refinance their way through unattractive leases using excellent access to the Japanese capital markets. Many investors would say, for career risk reasons, that the trade was blown up by the pandemic, but I think it was probably mistaken the day I made it. Anyhow, when you owe Softbank twenty billion dollars, that is SoftBank’s problem.

SoftBank is, of course, financing a large slug of the AI infrastructure buildout. They once owned about 5% of a company which made gamer-focused GPUs. SoftBank exited that niche hardware business, keeping the large stake in the CPU designer, because obviously CPUs are more economically important than GPUs. SoftBank tried to sell the CPU company to the GPU people, but the deal died in anti-trust, because the U.S. authorities were worried about that the GPU company might become too important if it controlled CPUs. (CPUs are, of course, the chip that gets racked by the hundreds of thousands in data centers worldwide.) Chastened by the feds, the GPU company returned to focusing on their niche business, which is now worth approximately $4 trillion. Whether this accomplished the goals of advocates of a robust anti-trust policy, I will let them articulate.

Some might read SoftBank’s checkered investment history as a reason to be skeptical of anything they try to put tens of billions of dollars into. Personally, I think the broadest outline of their strategy is simply correct. Collecting the Internet tax in a developed country a) throws off a huge amount of cash b) requires ongoing access to capital markets for nationwide infrastructure refreshes every decade or so and c) is done as a business. You can either choose to operate a utility, or do something interesting with your life. Funding industrial revolutions is a high-variance game but you only have to be right once.]

First, you start using some of the contractual features of those commits. For the $120 billion, or $3 billion a month, commit to SpaceX, you point to the line in the contract that says cancelable on 90 days' notice. This buys you some breathing room, if for some reason your crystal ball says compute is now hyperabundant at better prices relative to your anticipated needs, and this is likely to continue.

Not enough breathing room? You break out the other contracts. Nscale, Lambda: we contracted to pay you once the compute was delivered. It turns out the future we've arrived at sucks for us. Bad news: it sucks for you a whole lot worse. We will of course pay when you deliver compute as agreed upon, but you need tens of billions of dollars of capital to successfully deliver that compute.

And we anticipate that if you look around right now, you may possibly be able to afford a ham sandwich. So our commitment to you is going to evaporate when you fail to deliver. Not enough yet? Okay. Then you start triggering the prearranged contractual provisions which allow you to pay rent with convertible notes rather than dollars.

That will get you a few tens of billions. Small beer here, but tens of billions of dollars can be real money. Is that still not enough? Then: Google, Amazon, Microsoft, our beloved suppliers that are also major shareholders, it turns out we want to play Let's Make a Deal. Here's the first sketch of the new deal.

You take an accounting hit against contracted revenue for a few quarters. In return, you are a larger chunk of our cap table than you expected to be. This structure is definitely not robust against all possible world states. Again, they contemplate material risk of the world ending. But it is clearly anticipated that perhaps we come to regret these deals in the future.

Flexing compute between training, inference and subleasing

Patrick: And this is true within the labs as well. They have a couple of dials on — I think the three classic buckets a lab breaks down its own compute into are training for their own models, inference for the customers' use of the models, and then what is number three? Post-training or alignment or whatever. I think it gets described as different things in different places.

Zvi: Yeah. I think of lab compute spend as roughly divided in half between a combination of internal use, pre-training and post-training on the one hand, and serving the models on the other.

Patrick: Yeah. And so hypothetically in the future, if demand for inference goes up, keeping compute constant, you can adjust for that by saying, "Okay, we'll accept some slowdown with regards to our next model release," or shoot for less in terms of top-end capabilities, or yada yada.

And if there is a slowdown in customer demand for the model, then you think, "Well, pack more features into the next one and spend more on training it." Or, you know, the third option is essentially: sublease your compute to other people that are making more valuable use of it.

This is not a new phenomenon in human experience. One of the reasons Amazon and Google, et cetera got so good at racking servers in the early days of quote-unquote cloud  was they were basically the best in the economy at turning servers into revenue. But they got so good at racking them that then they discovered, wait, it is in fact not the case that you have to be the best in the economy at monetizing a server to afford an almost infinite number of servers.

[Patrick notes: Amazon and Google specialized in racking servers because they had problems nobody else did. Google had a high-margin worldwide software business. Amazon had a predictable demand spike on Black Friday/Cyber Monday so large it required inventing an engineering discipline.

They then found, essentially as an industrial byproduct, that they were the world’s best marginal operator of servers. They informed the rest of capitalism of this fact, enterprises shifted their workloads off-premises over the course of about a decade and a half, and the three hyperscalers (Amazon, Google, and Microsoft) printed money hats.

As one of their hundreds of cloud infrastructure product lines, they got into GPU hosting. The motivating use cases at the time of this decision included video codec conversion at global scale. This turned out to not be the most interesting pipeline you could run on a GPU.] 

If we just clone-stamp these boxes out into the universe, Fortune 500 companies will shift their workloads over to here at the prices we can offer, and it would be a ludicrously high margin business for us. And so AWS is probably the most important business inside of Amazon. Google's compute business is reasonably important, et cetera, et cetera.

And so there is a possibility that all that is old is new again, where, even if hypothetically the direct use of the API is soft for a few quarters, then they can say, "Well, we're sitting on a mountain of compute right now, and other people will need to use that for their own purposes. We will sublet 10 or 20% of our compute hoard for the quarter at some price that is mutually amenable," and then see what the global diffusion of this technology is next quarter.

Zvi: Yeah, it's certainly an option, and I don't expect them to ever have to take it. But we spent a significant period of time where everyone was talking about the depreciation schedules on all the GPUs, and how, oh my God, this will only last for so many years, the value will go down, the business model doesn't work, they're counting on such thin margins, how could this possibly work?

And now, of course, we have the situation where there's so much demand for compute that these things are staying in business indefinitely. The prices, if anything, are going up over time rather than down, because as much as we are producing more compute every year than we did the year before, and therefore have much more compute in the ecosystem, it doesn't matter. The demand is growing much faster than supply.

Patrick: Yeah. My understanding of the last couple of years of history — which is imperfect, because being a semiconductor analyst is not actually my job — but we were acutely compute constrained for a while. And so one of the reasons to make these forward-looking commits was that if you didn't make the commit, you didn't get an allocation.

[Patrick notes: Anthropic says in the S-1 that it has been routinely constrained, for the entirety of their corporate existence, by lack of sufficient compute.]

And since everyone was constrained on the number of powered H100s that were available in the world, which is a large number, but not an infinite number, to stay in the game, you kind of had to commit to the future. Having made that commitment to the future, we are trying to produce and power and rack, et cetera, H100s and H100 equivalents as fast as humanly possible, with trillions of dollars of industrial build-out happening in the United States to enable very powerful gamer GPUs in all the data centers.

[Patrick notes: There are not factually H100s or equivalents in all data centers, but data centers are basically “powered boxes.” You have to have a much more fine-grained model if you are an infrastructure engineer or a data center investor, but you can round them to “essentially fungible between workloads until you turn the racks on.” (Give or take a cooling refit. Crikey are GPUs heat-dense.) A new data center will tend to have a mix of racks corresponding to the new marginal demand, which in 2026 is very few Bitcoin ASICs and very many GPUs.] 

But we will probably run all these chips until they burn out in place, and then replace them with new chips, and just continue doing that, because we are not close to discovering the maximum amount of economic value of substitute intelligence, or substitute cognition, maybe.

Reading an S-1, with and without an LLM

Zvi: Yep, strongly agree.

And then the other point that I came away with from the S-1 — it was my first S-1 — was just how much of the S-1, while it is in fact written by very expensive lawyers and bankers that run around and go over every sentence with a fine-tooth comb because a mistake would be so expensive, there are also quite a lot of areas that seemed very, very boilerplate, repeated about five times.

Very much no one was intended to ever read this. Very much you just have to check a bunch of boxes. And you wouldn't have trusted GPT-4 with this, but I think you absolutely would trust GPT-6.1 or Opus 5.5.

Patrick: So, open secret in the industry that many S-1s are copy-pasted, or large portions of them are copy-pasted.

Get out the usual boilerplate about cyber incidents. We're a software company. We need the cyber incidents paragraph in there, and the cost of it is amortized over every other software company that the investment bank has done. And it's basically designed to be sort of maximally — well, if we disclosed all the risks to you, we can't be sued.

But if we forgot one of them, then you can say, "Oh, I wasn't informed about the risk that your executives could die, and that would impair your business results." And then an executive dies, and you have to say, "Oh, shoot. Update the standard document for 'sometimes people die,'" et cetera, et cetera.

That is not one that is in this document. I'm making up a hypothetical joking example to make a point. Although substantially every S-1 says, "We might not retain key members of our staff."

Zvi: This one does.

[Patrick notes: In one of several mentions: “Achieving [our goals] and maintaining this position depends on our ability to … retain the continued service of our co-founders, senior management team, and key technical employees, and attract, retain, and motivate leading researchers, engineers, and technical talent in an intensely competitive and evolving market[.]”

Patrick: Yeah.

Zvi: Yeah, no, I remember reading that, and to me this is just such a stupid outcome of capitalism. But also, if it's actually literal copy-paste, then who cares?

It's just a matter of, everyone gets their first S-1 where they don't know this yet, or they don't understand it yet, so they have to go look at all this stuff. And after that, presumably, you would sit down with the S-1 and you would go, "Claude, remove all the stuff that was copy-pasted that I don't need to read, so that I can read the actual S-1."

Patrick: Yeah. I saw on your Substack that this was your first S-1, and I was briefly surprised, and then I remembered that, oh yeah, the finance industry contains multitudes. So this is not quite my first S-1, and there's a rhythm and a language to these things, and after you've read 10-plus of them, you get good at, okay, there will be nothing interesting in this part.

You can just skim over it or skip it entirely and get to the numbers. And in particular — not in this S-1, but speaking generally — you're looking for interesting choices made, and differences from how things are usually presented is often where the intellectually meaty content of an S-1 lives.

And so, go straight to the footnotes, et cetera, et cetera, is a technique of people who read these on a recurrent basis. Of course, state-of-the-art models are pretty good at — the prompt can be as quick as, "Hey, read this S-1 and tell me what you find surprising."

And they get a lot of gold out of just that.

Zvi: Yeah. I am continuously learning how to easily, without having to think about it, ask the right questions of LLMs to get them to point out the things that I would want them to point out.

Circular financing and the dot-com comparison

Patrick: So we've talked about their sort of capital stack issues.

So they had convertible debt. Morally speaking, that will be converted. You should just think of it as stock, yada, yada. But the convertible debt is currently convertible. Obviously, they're raising additional money through issuance of common stock in the IPO. There's also capital raising happening essentially across the tech industry to support this build-out, because we are in the midst of what is frankly an industrial revolution.

To the extent that Wall Street is worried about things going badly, I think they're worried about a symptom that we have seen before, where, through complicated shenanigans — either with a central actor, or through just individual participants in the market making locally rational decisions and ending up in a globally irrational place —

If you find a way to turn capital into revenue, and you loop that very quickly, sometimes things get to crazy places. And so a classic example of this is the interplay between Yahoo and the other startups shortly prior to the dot-com bust, where Yahoo's share price was increasing rapidly.

People treated that as the index price for the future value of the Internet. And startups would raise, saying, "Look at Yahoo's share price rising. You should bet on us as being part of this Internet land grab, and so assign a high valuation to us. We're going to take in that money. We will spend it on, of course, Yahoo ads, because that is the way to get users."

Spending money on Yahoo ads increases Yahoo’s revenue… which increases Yahoo’s share price… which increases investor interest in funding Internet startups… who promptly spend the money on Yahoo ads.

And then this cycle happens a few times, and we get a bit ahead of our skis. And so there is a complicated industrial apparatus happening right now in, again, a build-out for the ages. We are living through history in this moment. And one bit of that complexity is that the shareholders, Amazon and Google, are also major suppliers, are also major customers in some fashions, et cetera, et cetera.

And I think there is a sense in some people on Wall Street that there is a circularity here that could end badly. I don't think I share that point of view, but to the extent that there are credible bear cases, I think that is probably the intellectual scaffolding for the credible bear case.

Zvi: Yeah, I think it's a reasonable thing to be worried about from the outside view.

I think that, based on my understanding of the technology and what's happening, I am much less worried about it than I would be if I hadn't invested a lot of intellectual effort in this question, but it's certainly possible. It's certainly possible that something like a debt crisis could affect the industry.

The world is being asked to provide quite a lot of debt in ways that are really unprecedented. I certainly understand why you would do that. These companies are increasing in value rapidly. You don't really want to sell more stock than you have to. So it's certainly reasonable. 

And another way of thinking about it is, if you are trying to efficiently raise capital and efficiently allocate capital and efficiently do an industrial revolution, the proper amount of risk of a crisis situation due to being a little bit over your skis on the revenue assumptions and the circularity is very much not zero.

Wall Street versus the tech industry

Patrick: I also think that the Wall Street side of the financing community and the tech side of the financing community do not really see eye to eye as to what is happening, where Wall Street is looking at it like, wow, these are incredible software companies. And the software side of the industry is basically shaking Wall Street by the neck and saying, "No, you don't understand."

It's possible that the entire business history of Google to this point was a bootloader for this business. 

[Patrick notes: Tip of the hat to Elon Musk for saying that humanity might be a bootloader for the AIs, which is not my argument. The way both of us are using this metaphor, though, is that the program you are currently using to read this essay runs on an operating system, which is an incredibly complex and valuable bit of software. That operating system needs some amount of tying to the specific hardware to be viable. This is historically done by a bootloader, a very minimal operating system which turns on when the device turns on, and then bootstraps the rest of the operating system. Bootloaders are important software, but are not particularly complex or economically valued relative to all software sitting on top of them.]

Google is willing to spend all of its free cash flow and then some, and max out its own credit lines, et cetera, et cetera, to get as much of this business as they can at various points. This is not simply because LLMs threaten Search and therefore the ads business. I think, bluntly, that's oversold. (And Google is moving aggressively into making Search a thin layer of paint over their own LLM, Gemini.)

But this will make Search irrelevant because it is such a better business than the best business in the history of capitalism. And Wall Street, they don't buy that yet, for various reasons. One is, we underestimate how small the tech industry is, and underestimate that we all read the same blogs, et cetera, et cetera, at different levels.

[Patrick notes: Several centuries of AI time ago, or about 12 months ago in human time, I had a chat with some people who work at the Bank of England. One topic for that chat was a recent talk of the town: responsible adults in London believed that it was unlikely the capital required for the next 12 months of infrastructure buildout was available. This had been reported by responsible adults in responsible publications after consultation with responsible adults at financial institutions, organs of the state, and other places which have power. 

I told some responsible adults that they were welcome to their opinions but that the tech industry was welcome to its money, and piddling sums like a trillion dollars were easily within our capability to commit without asking anyone else’s blessing. And, in fact, we had done so. (Responsible adults may ask a junior analyst when capital needs to be raised to fund a transformer that will be delivered six months from now. Spoiler: it was fully committed years ago.)

I told my hosts that my model was the most senior executives of AppAmaGooBookSoft were more AI-pilled than literally anyone they had ever met, and that many people at the labs considered those executives radically underinformed about research results the labs had achieved twelve months ago.]

Patrick continues: And also, the engineers drive the bus in the tech industry, and you can point at the adoption of Claude Code, et cetera, and say, "This completely disrupted the entire profession of engineering in less than 12 months." That will not be the last time this happens. Obviously, this is tens of percent of the economy going forward.

And investment banks — JPMorgan, without loss of generality — are not yet seeing, "Oh, I could really see a world in which we're spending billions of dollars a year on token costs." And when they get there, they will have a much higher implicit valuation and discount the debt less than they do currently, but they're not there yet.

The three AI pills

Zvi: Yeah, it's weird to have these people refuse to believe, as far as I'm concerned, even what AI can currently do and what AI's current impact is. I talk about three AI pills, right? There's the AI pill, the AGI pill, and the ASI pill. So the AI pill is just: do you believe that AI exists, and are you vaguely aware of what it can do right now?

And what we would be able to do with that AI if we just got to keep it around and diffuse it throughout the economy. And I feel like pretty much everyone in the tech industry, at these companies, is at least AI-pilled, but most people in places like banks are not AI-pilled. And that's separate from the question of artificial general intelligence, AGI, and then artificial superintelligence, ASI.

These economists are going around making these financial projections and saying, "Maybe this will boost GDP growth by half a percent." And then, if you really, really want to go crazy, 1%. And people who are actually thinking about the object level of what it means to have a tremendously larger amount of intelligence are like, "No, no, no, you're off by an order of magnitude, in the worlds that are relatively normal."

And then there are the worlds where everything completely transforms and people may all die.

Patrick: Yeah. Checking in on other corridors of power: I have a bunch of friends in DC, and I'm working in a non-profit (Sparrow Infrastructure) to help DC and people from the San Francisco Bay Area get together more often. And my sense of DC is that in the last year we have flipped, where, if you pick a representative, countable group of influential people, like, say, congressional staffers, we have flipped to more than 50% of congressional staffers now believing that you can get useful work out of AI tools.

Awesome. They have taken the first pill. And some very small number are saying, "Okay, not that I'm saying it, about there being existential and/or catastrophic risks to the United States, but I'm hearing other people say it, and they don't sound insane. So can you walk me through what it would look like if, hypothetically, these got very, very good at computer hacking, and hypothetically maybe they were as good as a nation state at computer hacking, and hypothetically if they were better than the nation states at computer hacking?"

And so the leading edge of DC is updating to the world of the Mythos release in — what was that? February of this year? Non-ironically, I give massive props to people for being able to risk their reputations and professional lives, et cetera, to catch up to February only eight months delayed. [Patrick notes: Public release was April. Mea maxima culpa.]

By DC standards, that is incredible forward-thinking. But taking the third pill — I don't think I've taken the third pill yet, by your standards.

Zvi: Yeah, you haven't. You're insufficiently ASI-pilled.

Patrick: I'm getting more worried about if and when I will have to swallow this.

[Patrick notes: Dwarkesh Patel had an arresting footnote in his book: “[LLMs are] already plainly superhuman on a few dimensions: reading and writing speed, breadth of knowledge, number of languages spoken fluently, and accuracy at predicting the next word in a sentence. I encourage you to confirm these by inspection; I think it’s that obvious.” 

I cannot confirm by inspection that LLMs are plainly superhuman in research mathematics, but this seems to be correct and also new in the last few months. Plausibly there are domains where being the smartest doesn’t mean being the best but after research mathematics falls to the machines I think the handwriting is on the wall: to the extent a field is g-loaded, the frontier of it will shortly be performed by machines, not by grad students.]

Industrial revolutions and how weird the future gets

Patrick: I will say there is a bit of an intermediate, where, even if one doesn't credit fundamental transformation of the economy, the way industrial revolutions work is you redevelop entire supply chains to take advantage of the new thing. And so the kind of fundamental impact of the automobile versus horse-drawn carriages is not: people will be able to take a lot of long road trips across the United States — although that is one impact of it.

But it's like, no, substantially all transportation of goods and services will no longer go by horse. And we will take advantage of this, over a period of decades, to redraw the interlinkages of the economy. And there will be impacts of that in agriculture, in finance, in et cetera, et cetera, et cetera.

And these will be iterated and looped and interconnected in ways that are very difficult to understand when you see the first Model T roll off the assembly line and say, "Yeah, that is a really cool thing. I could see how that is going to play out." And we have very limited ability to see how it is going to play out right now.

[Patrick notes: The Model T doesn’t by itself imply the interstate highway system. Trucks carry the majority of America’s goods transportation, substantially enabled by that system. Long-haul trucking is one of the most common professions in the country; long-haul horsing is close to physiologically impossible. Long-haul trucking makes nationwide fast food chains viable, because it delivers bulk food system inputs year-round. And so, in a real way, the Model T changed what children grew up eating.] 

The thing I keep going back to is, the future is going to be really, really weird. I think almost everybody is underestimating how weird the future is going to get. And certified Internet weirdos like you and I and some buddies of ours — I think we are probably underestimating how weird the future is going to get. Okay, I will say I am probably underestimating how weird the future is going to get.

Maybe taking the third pill helps you get to that. Maybe not.

Zvi: Yeah, I have successfully overshot on how weird it will get, how fast, in some ways. But the future is going to be really, really weird. And the present is in fact really, really weird. It's just that I managed to overshoot modestly on some of those questions.

And it is really disappointing, because it's just a microcosm of trying to warn about future dangers: even when we're talking about just future industrial processes, or future economic activity, future ordinary things, people measure AI almost always as, how do I slot the existing AI that exists today into the existing action that a human or a machine is currently doing, and have it substitute directly for the exact thing that's currently happening?

How good is it at that thing? And that is obviously a very small portion of what is going to happen. And that alone would in fact make the world really weird, but that is a very small fraction of the amount of weirdness that will happen, again, even if AI is relatively normal from here.

AI mathematics and the bitter lesson

Patrick: And it's frustrating, because we've already seen this happen enough that we can point to the recent past and say: if you thought the way AI was going to transform engineering was it was going to be the world's best autocomplete in existing IDEs — there are people that thought that, big bets placed on AI is going to make IDEs 20% more efficient than previously, and that will be a 10% increase in the aggregate output per engineer.

Twelve months after having that belief, you have, I would say, probably tens of percent of the engineering industry no longer booting up an IDE to do software development, because why would you? That would imply that you were looking at computer code. And the leading edge of the cool kids are just so far beyond that. Why would a cool kid have one’s main agent look at computer code when the swarm can look at computer code for him and save his context window?

We, over very short periods of time, get fundamental rethinkings of what the task looks like, done in very, very hard domains, and then it starts cranking through.

The field of mathematics is weird because the individual dollar value associated with mathematical breakthroughs is very low, but they're known to be pretty hard. And so, essentially as a marketing stunt, both labs have been just knocking down mathematical theorems, or producing proofs of hypotheses, et cetera, et cetera, with the same aplomb that SaaS businesses use to write blog posts.

And the math community feels very emotionally conflicted about this. And I think all of us are going to be emotionally conflicted about something coming in the next couple of years, to put it mildly. But the thing that generated that outcome was not that the chatbots are better at talking to a mathematician and driving hypotheses forward than any previous Wolfram Alpha has ever been.

It's like, no, this is the RSI, the recursive self-improvement. The agent swarm, with some amount of direction at problems, will just chat about it within themselves and essentially autonomously write out proofs that all the king's horses and all the king's men have not been able to come up with in the last several millennia, and then crank them out at such a rate that literally the entire field of research mathematics cannot check them as fast as they are being produced.

But there's good news. We're pretty confident that they're correct, because they come with formalizations, which is a thing that research mathematics doesn't normally ship with proofs, because it is some hundred-plus times more expensive than the proof itself. My mind is kind of blown. And this future is weird.

We're going to get our minds blown in so many things over the course of the next couple of years.

Zvi: Yeah, it's the bitter lesson. If you just make the things more capable, they will suddenly be able to do things, pretty much kind of automagically, kind of on their own, that you didn't expect them to be able to do.

You're suddenly like, "Oh, actually, the LLMs can drive." It's like, "The LLMs can drive?" Yeah, the LLMs can drive. What did you expect?

[Patrick notes: A long time ago, in a place far far away, I was attached to a computer vision lab. Computer vision was a thriving academic subfield for decades. It had journals, conferences, and PhD programs, grinding out incremental improvements to practical, useful problems like “Find the face in this JPEG” and “Does this photo contain a bird?”

LLMs solve many computer vision problems by accident. Does this photo contain a bird?

Claude Fable 5.1 high, asked the question and told to elaborate in a pedantic level of detail, responds:Yes. There is one bird: an adult male Indian peafowl (Pavo cristatus) in full display, roughly center-frame. … 

Claude then went on to write several thousand words, including estimating the peacock as being approximately 6 meters from the camera, locating the photo to the Barcelona City Zoo with 35% confidence (“It sits inside Parc de la Ciutadella, which is wall-to-wall plane trees; its older exhibits use exactly this log-and-faux-rock barrier style; its peacocks roam free; and that pale sandy surface looks like Barcelona’s sauló. Runners-up: Lisbon Zoo, Rome’s Bioparco.”)

Claude is correct; I took this photo at the Barcelona City Zoo in 2015. It is unpublished on the Internet and I stripped the EXIF data prior to asking the question. Claude objects pedantically to that statement: The file you sent is not the camera’s JPEG. It is a 1076 × 1404 RGBA PNG (alpha fully opaque) carrying an ICC profile named “Display” and a 62-byte EXIF stub whose only tags are PixelXDimension=1076 and PixelYDimension=1404. That signature is what macOS writes for a screenshot or a Preview export. Corroborating: the left 18 columns are a perfectly uniform gray (150,150,150) that is not part of the photograph, i.e. the viewer’s background caught by the selection rectangle. So “EXIF stripped” is true in effect, though it looks like it happened by screenshotting rather than by an explicit strip.]

Capabilities nobody trained for

Patrick: A few years ago, I visited one of the labs, and they showed me a pre-release model and said, "By the way, we did not intend for this model to speak Japanese.

We don't have anyone in the building that does. We did not intentionally have Japanese in the training data, et cetera, et cetera. We think the model has intuited the structure of the Japanese language from sufficient pages on the English-language internet. And so here's our demo, and you tell me, Patrick, whether this model speaks Japanese.

So I'm going to go to CNN right now, and I'm going to say, 'Hey, translate this CNN.com news article,'" which is clearly not in anybody's training set yet, "and just translate this into Japanese." And I read it, and I was like, "Okay, as a one-time professional translator, I'm not the optimal person to do a quality check on this translation, but finger to the wind, yeah, your model speaks Japanese.

It probably speaks better Japanese than I do, after 20 years of devoting a significant portion of all my cycles to getting good at that. You achieved it by accident. Wonderful."

So you mentioned the other day on Twitter that Anthropic, for reasons of its own, had decided never to release — well, they did not prioritize releasing an image generation model. And there's been a sort of side stream of AI research done on that topic, with Stable Diffusion. OpenAI has a quite capable model. There's Midjourney, yada, yada, yada. And it turns out Anthropic kind of invented one, quote-unquote, "by accident," by just focusing on Claude Code enough, where they are strikingly capable at doing various image-related tasks, by, I don't know, being really intelligent, and having image generation be one of the things that you can do if you write enough computer code.

Zvi: Yeah, I think basically everything can be done if you can write enough computer code and you have reasonable ability to think and analyze data. And you should expect AIs to figure out everything that is inherent in the structure of all the data that comes into them. And pretty much everything on Earth is inherent in the structure of theIinternet.

Patrick: Which is a wonderful, terrifying lesson. Something of the opposite of the bitter pill lesson. I don't have a great sense for how many neurons are in the weights right now, but I do remember when the first version of Stable Diffusion came out, and by current standards, Stable Diffusion is laughably terrible at image generation.

But you could reliably get something that looked like art out of it, and it was a two-gigabyte download. And it was a little bit wild that you could compress the artistic history of humanity into two gigabytes. It seems like information theory says that that's impossible. And yet Stable Diffusion was an artifact that you could download and coax into getting some quality of outputs out of.

And a huge portion of the human experience is compressed into the weights. It's just a statement of the world we live in, which is wild. And then the fact that it can be harnessed in this agentic fashion to do things that are not within human experience, but just bootstrapped on it, is its own flavor of wild.

And for all the flavors of wild and weirdness that you can see visible in the physical universe right now, reflected in the S-1, et cetera, et cetera, we haven't seen anything yet. I think that might be a good thing to say. We mutually expect to be more surprised in the future than we have been surprised in the last five years of being surprised, right?

That is the correct way to be calibrated.

Capabilities versus diffusion

Zvi: Absolutely. I think so far, I would summarize it as: I have been surprised the models have gained capability as quickly as they have so far. And, like everybody else, I have been surprised, conditional on how much capability they have gained, at how little the world around us looks different as a result of that.

Diffusion has been slower than I would have expected, but ultimately that doesn't matter. And as the models themselves are capable of driving the diffusion, the diffusion will speed up. So yes, expect high weirdness.

Patrick: I think I was far worse calibrated than yourself and other people on the rate of advances in capabilities, and better calibrated with respect to diffusion.

They are diffusing faster than my model for diffusion. (“Diffusion” here is the term of art, folks, for how fast capitalism is adopting these things.)

[Patrick notes: My naive model for diffusion was “It takes capitalism about 10-15 years to get this into All The Things beginning at GPT 3 levels of capabilities.” You’re welcome to compare that to e.g. AI 2027. Importantly, I think one thing I believe which is not widely believed was that GPT 3/3.5 was already sufficient to kick off an industrial revolution, given sufficient effort to harness it.] 

Zvi: Yeah, I don't think I was crazy stupid high on rate of diffusion, but I do think I was substantively high on that. And it seems so crazy to me. You look around and you're like, "Well, they have these fantastic things that can do all these cool things for you for very cheap.

Why are people not investing in it?" And people don't do things. It takes a while. They need to see other people do them. They need to have it kind of just naturally there, step by step, without having to make any leaps or make any investments.

Patrick: Yeah. And to be clear, when we are saying that capitalism is not doing anything with these yet, it's like there's only tens of billions of dollars a year spent on inference.

I mean, clearly, if we were serious about things, we'd already be at trillions.

And nothing is investment advice, et cetera, et cetera, et cetera. But you'd have to assume crazy things about the world to assume that we don't get to trillions of dollars in spend on inference: either a complete collapse of inference costs, or, I suppose, the death of all humans, or massive contraction in the world economy, et cetera, et cetera.

These things provide cognition on tap. There is not an obvious ceiling for how valuable that is.

Zvi: Yeah, I mean, obviously it's possible that five years from now, it will turn out that model capabilities de facto kind of peter out for most practical purposes, and the model everyone uses costs a tenth or a hundredth of what we currently pay for our frontier models, and it's just very, very compute efficient. And then we'll just use that much more of it, right?

In that situation, because it'll be so good and so valuable, and we'll have had time to figure out what to do with it.

Token spend at the frontier

Patrick: Do you have any anecdotes for — if you think of the people in your social/professional set and similar — what the peak is for token usage per month at the moment?

Zvi: So I'm in a weird spot where I'm a writer, so I don't really code massive things, and I don't hit limits, right? I have the max subscriptions, because it's very important to me to never hit a barrier, and also it's very important for me to be able to have access to everything on the first day, so that I can evaluate it and report back.

Even if most people can wait a week and it doesn't really matter.

Patrick: I am not at liberty to say who it is yet, but I can say I'm aware of at least one person who spends more than a million dollars a month in inference for the acceleration of their professional job, and whose employers look at their inference spend every month and are like, "Great.

That continues to grow. Wonderful." And there don't have to be all that many people in the world hitting a million dollars a month in spend that they can personally authorize for these to be wild, wild businesses. And again, the job of this person in the status quo resembles the job that they were doing two years ago, except they're now doing just a whole lot more of it, because they have teams of minions. Virtual minions.

Zvi: Yeah. My guess is that the fact that I'm not spending at least ten times as much as I am on inference, and probably a hundred times, is mostly a skill issue. I have chosen not to prioritize figuring out how to do that. But it just seems absurdly like I should be doing that.

Patrick: Yep.

I feel just the same way. I'm sometimes thinking of hiring for this podcast, and if I was being strictly rational about it, it would just be spending four weeks on "automate all the things." But I have not done that yet, even though I make extensive use of AI in various parts of the production function here.

Final thoughts on the S-1

Patrick: So do you have any more thoughts on the Anthropic S-1 that we have not already covered?

Zvi: So I think those are the main ones. It's just sort of — they chose mostly a standard corporate document to write, and they are presenting as if they might have very, very fast growth, and they gave us very, very attractive, very good financials and prospects.

But aside from the corporate structuring that we got more detail on, and the choice of how they communicated, I felt like, sort of — it wasn't like the central issues were news, right? We knew vaguely where Anthropic was at. They'd already shared a bunch of details in advance. And so we're all looking at it to sort of get a slightly better feel, get a bunch of details that we didn't have.

But yeah, I think that Anthropic's S-1 was the S-1 we always knew was coming, to a large extent. I would've been very surprised — very pleasantly — if they had stepped up to the plate and issued a much more robust warning, but I wasn't expecting it.

Patrick: The one micro note I would have that we haven't already discussed is, the S-1 at one point explicitly contemplates capabilities growth and product improvement, and Anthropic creating products which economically dwarf their products to date.

And we've indeed seen a couple of iterations of this, where they had the API, and then they had the chatbot on top of the API, and then they have Claude Code, which is something much different than a single-threaded chatbot, and are continuing to build on from there. I've always perceived the chatbots for the labs as being essentially a sales engineer for the real product, which is the API.

But the agents and harnesses seem to be, quote-unquote, "the correct way" to use the API, versus single-threaded calls. And we will see if that is still true two years from now. But they are certainly where the economic interest is at the moment. And it is possible that, as they get tighter and tighter into this self-improvement loop, there are more tiers of the business that don't cause the previous business to go to zero, but they cause it to be economically irrelevant, because they're just worth so much more.

Which is a wild thing to say, when the thing that will be economically irrelevant is the one that was doing 15x revenue gain on a base of more than a billion dollars. But we are priced into a wild, wild few years, regardless of what happens to stock prices or individual returns to securities, I think.

The hackening and who bears responsibility

Zvi: Yeah, I mean, I think that's pretty much guaranteed, and the pace of things is accelerating. The Hackening is only getting worse, it looks like, over the last few days, so.

Patrick: Can we discuss that briefly? I think there's going to be a bit of backlash against the labs for this, because — I don't think it's true that the labs made us more vulnerable to it.

Software has been insecure for the entire lifetime of software. That's the default way to create it. We are lately getting exposed to the bill via AI-uplifted hack incidents, where, in most cases, it seems like the actual tool used to uplift the hackers was not directly one of the labs' products, but it was one of the open-weights models that were created with substantial reliance on distilling the labs' products.

(Although there have been cases of documented misuse of the labs' products directly against various targets.)

And so yesterday we had two major incidents, where I started the morning thinking, "Wow, Polymarket had a loss which was large and inconsequential." And then Binance had a hack loss north of $5 billion and counting, to North Korea. Which, speaking of AI impacts on the global economy — we have already passed the point at which AI is more than 5% of the GDP of a nation.

It's more than 5% of the GDP of North Korea, by enabling them to innovate in their business model of stealing money from crypto people. And I shouldn't kick crypto too much. They're only stealing the money from crypto people preferentially because that is the lowest-hanging fruit. After they have taken many more billions of dollars from the crypto industry, they will start hitting the rest of us more aggressively than they currently do.

But I think people are going to get to a period where, wow, it is now obvious that we need — my estimate is like $500 billion of remediation, at least at pre-LLM prices. And this $500 billion feels, to some, like a ransom demand from the labs, because the only way we can reasonably pay it is by adopting AI for the blue team, for defenders, et cetera, et cetera.

And "you created the problem, and now you're selling us the solution to it for tens of billions of dollars" is going to discomfit some people. What are your thoughts on this?

Zvi: So I feel like you're painting kind of the optimistic case, where there's this fixed set of technical debt that we incurred, and mostly it's just the bill is coming due much faster than it would've otherwise come due, which is forcing you to deal with this via AI, which is going to cost you a lot of money, but is ultimately a one-time bill that you can pay, and then the problem is solved.

I don't think of it that way.

Patrick: I also don't think of it as a one-time bill, but just the backlog is already priced into $500 billion.

Zvi: Right. The issue is that the moment we get down that $500 billion, I don't think that that results in something that is now robust. I think that's like, now the next level of things, the next set of dominoes, start to fall.

I think it was very reasonable, obviously, to build the entire internet and the entire modern world around things that were not fully secure, because being fully secure is next to impossible. And previously, there was a very fixed pool of people who were, one, trying to cause trouble; two, had capacity and knowledge to cause trouble; and then three, targeting you and your thing specifically, because there are so many things in the world and not that many people left to target them.

And now you can probe and exploit at scale, in parallel, and any number of people can do it, and those capabilities are going up, and they can concentrate their firepower on whatever they decide the weak spot might be at any time, including figuring out where the weak spot might be in order to concentrate their firepower.

Yeah, it's going to be dramatically, dramatically harder. And I am not one of those who has the maximalist view of "the internet will break," or you will assume that any phone that activates its Wi-Fi will immediately get rooted. I don't think these are likely scenarios, but I do think that we're going to lose a lot of the internet, probably permanently, on the long tail.

A lot of things are going to go down reasonably often. A lot of things are going to get stolen. A lot of things are going to get compromised. A lot of information's going to leak. We're going to have a continuous set of problems. Will the labs be blamed for this? I mean, somewhat, sure, but they should be blamed for this, right?

There would not be these Chinese models that are capable of doing this without the American models.

Patrick: Huh. I think there's complicated notions of desert there, where I agree that but for the labs having state-of-the-art models, this state would not be occurring. I don't know whether it is reasonable to lay at their door consequences of decisions made by other people using their technology. I similarly don’t know whether it is reasonable to argue they simply didn't get distillation locked down quickly enough, and now they are in some way morally responsible for the consequences of that.

Zvi: I mean, I think that's a good point to close on: it's just sort of, it's not the incentives, it's you. And this goes to existential risk, and potentially building things that will lead to an AI race that gets everybody killed when we don't know how to control things that are smarter and more capable than ourselves, or when this creates equilibria that have no good solutions that we know of, and where all the natural outcomes are extremely bad.

And it also applies to hacks, where any individual decision makes sense, but inevitably goes down this road, and you didn't actually figure out how to mitigate this sufficiently. You didn't actually figure out how to fix it. And they've done some very commendable things, like Glasswing and Daybreak and so on, to try and mitigate the situation.

I'm sure it'd be vastly worse if they hadn't done those things. But at the end of the day, yeah, it's not the incentives, it's you. Don't pretend that just because someone else would've done it, that gives you the okay to do it, because then someone else is pointing at you.

Trapdoors and irreversible decisions

Patrick: I will say that there is an element of trapdoor-ness to — we are making decisions that are not fully reversible.

And that has been obvious for a few years, but there were, I think, many people involved who had reasonable points of view in a world where you think you would be able to have take-backs. And so there's many people in the tech industry, including many people who have dumped a lot of money into lobbying on this issue, who are quite committed to open source models being available as a thing.

And once the open source models and the open weights achieve a capability, that is the worst capabilities will ever be going forward, unless you successfully do a worldwide surveillance regime over all basic computing, which is hopefully impossible. It would be an incredibly dystopian world if we were able to blacklist mathematics.

So once it's out there, it's out there, and it is now out there with the capability of uplifting North Korea and other bad actors. And we will never be able to yank that back. We can only go forward from where we are right now. And I hope we learn the lesson of that with respect to things like — by the standards of things that AI could potentially do, stealing $5 billion from Binance is a prosaic risk.

There will be other trapdoors in the future, and we do not want to fall down those trapdoors as well. And I guess that is as good a place as any to call it. Well, Zvi, thanks very much for discussing. These days, your newsletter continues to be kind of essential for people trying to keep up on the frontier here.

Where can people find that?

Zvi: You can find my newsletter at thezvi.substack.com.

Patrick: Thanks very much for being on the program. And to everybody else out there, thanks very much for listening, and we'll see you next week on Complex Systems.