AI:AM LIVE — June 16, 2026 — Doom, Policy, and the Physical Economy: Liron Shapira, Samuel Hammond, Matt McKinney

The rundown

1. 9:17Opening31 min

   Opening: AI Meets the Real WorldThe SpaceX acquisition of Cursor crystallized the theme: frontier AI companies as gravitational black holes absorbing the application layer, with the Anthropic-Cursor-SpaceX web of mutual dependency as exhibit A. The hosts also covered the Qualcomm-Tenstorrent rumor and why AI is compressing post-merger integration risk.

   Watch

   ▶Play this segmentYouTube · from 9:17

   As aired

   Prakash and Nathan open the June 16 show by previewing the day's three guests: Laurent Shapiro from Doom Debates on the Fable ban and AI governance; economist Sam Hammond on government-versus-Anthropic leverage and AGI futures; and Matt McKinney, CEO of Loop, on AI in logistics. Before diving in, Prakash sets the overnight context: SpaceX has moved to close its acquisition of Cursor at a $60 billion valuation, with shares up roughly 40 percent since the IPO. He walks through the backstory — Cursor represented 40–50 percent of Anthropic's revenue at the time, Anthropic initially framed Claude Code as a research effort, then restricted Cursor's model access, triggering an existential crisis that led Michael Truell to reach out to Elon Musk.

   The hosts debate the mechanics of the Anthropic-Cursor-SpaceX triangle. Prakash explains that Anthropic's core objection was Cursor's practice of routing customer tokens through Claude while also running its own models — effectively distilling Anthropic's output cheaply at scale. Nathan notes Anthropic appears to be partly reversing the access restrictions, suggesting they are responding to competitive pressure even while facing overwhelming demand. Both hosts reflect on the web of mutual dependency: Cursor needed Anthropic for intelligence; Anthropic needed Cursor for revenue; Anthropic now relies on SpaceX AI for compute; and SpaceX AI, by acquiring Cursor, is circling back to data generated partly via Anthropic models.

   The conversation broadens into a wider theory of the current tech landscape: frontier AI companies are acting as gravitational black holes, absorbing talent-rich application companies — Cursor in coding, with analogues likely coming in biotech, materials science, and robotics. Nathan and Prakash argue that AI is also compressing the integration risk that once made acquisitions so perilous, pointing to a friend who assembled an entire M&A data room using Claude. Prakash introduces a parallel acquisition rumor — Qualcomm in talks to acquire Tenstorrent (Jim Keller's chip venture) for $10 billion — and reflects on Elon Musk's unique keiretsu-style empire as the one entity that combines capital, manufacturing, and technical talent at scale.

   Key moments

   If Cursor can't escape the gravitational pull of the model companies, who is? Who's not going to get burned up by that?

   Cursor made up 40 to 50 percent of Anthropic's revenue. Anthropic initially told Cursor that Claude Code was just a research effort. And then after they went ahead and cut Cursor off from Claude, Cursor entered this existential zone — a decacorn about to die.

   The only team in the entire big-tech space that is able to iterate on physical product is Tesla. He's really built this Asian-style keiretsu — interlinked technical teams that touch each other and produce product at the edges.

   What we covered

   SpaceX closes on Cursor at $60B — frontier AI as a gravitational black hole. New details on the Anthropic-Cursor falling out: Cursor made up 40–50% of Anthropic's revenue; Anthropic restricted access after concluding Cursor was distilling via customer tokens; Cursor faced an existential crisis and reached out to Elon. The acquisition reveals the circular dependency: Cursor needed Anthropic for intelligence, Anthropic needed Cursor for revenue, Anthropic now relies on SpaceX AI for compute, SpaceX acquires Cursor to capture the data. Pattern generalizes: frontier labs will absorb the application-layer leaders in biotech, materials science, and robotics next.

   AI compresses post-merger integration risk — and Qualcomm eyes Tenstorrent at $10B. AI-native due diligence (a friend assembled an entire M&A data room with Claude) is making acquisitions faster and less likely to fumble value. Separately, Qualcomm — which lost its entire AI chip team from the Nuvia acquisition — is in talks to acquire Jim Keller's Tenstorrent for $10B, while NVIDIA's Groq acquisition at $20B already looks cheap less than 12 months later.

   Full transcriptLightly edited · timestamps jump to YouTube

   9:17

   Prakash: Good morning. It is Tuesday, June 16, 9 AM Pacific time. Nathan, good morning, and welcome to AI:AM.

   9:26

   Nathan Labenz: Good morning. We're back at it for another full day of AI analysis. We've got some interesting stories that I'm just waking up to in some cases, and I'm looking forward to our guests today. We've got Laurent Shapiro from Doom Debates, who will — I think, among other things — make the case that the Fable ban is actually a good thing, if only because it breaks the ice and nobody can ever say again that the government will never take any action. Then we've got Sam Hammond, who is an incredibly interesting and often heterodox thinker and economist who has thought harder and more outside the box about what the AGI future could really look like — and what it means for governments — than just about anyone. So I'm very interested to get his take on who has what leverage and what escalation dominance looks like in the USG versus Anthropic round two that we're living in right now.

   10:12

   Nathan Labenz: And we've got Matt McKinney, who's the CEO of a company called Loop, which is bringing AI to logistics. We've already had great benefits from logistics in our lives — getting things next day, or in some cases now a drone drops stuff off from Amazon in an hour or less. So I'm interested to learn more about what they're doing to further optimize logistics and even more smoothly get the gears of commerce turning. Before we get to all that — what's top of mind for you heading into the day?

   11:08

   Prakash: So just to set the stage on what's been happening overnight — the Cursor acquisition is now in play. As you know, SpaceX made an offer for Cursor before SpaceX went public. The deal terms were a $60 billion valuation for Cursor, paid in shares after the IPO of SpaceX within six months, and a break fee of $10 billion if SpaceX did not end up buying Cursor. As of this morning — it's been about three market days since the IPO —

   11:53

   Prakash: Elon has decided to close on the Cursor acquisition. SpaceX shares are up 40 percent since the IPO. They started at $1.35 and are roughly at $2.10 right now. And it's about five or six days before a major unlock — once SpaceX earnings come out, another portion of the shares gets unlocked. Right now, only 5 percent of SpaceX is floating, and so people are saying the shares are artificially inflated. Elon has chosen this moment to close the Cursor acquisition. And there are more details coming out on what happened between Cursor and Anthropic. Number one: Cursor

   12:39

   Prakash: made up 40 to 50 percent of Anthropic's revenue. Anthropic initially told Cursor that Claude Code was just a research effort. Then, after they went ahead and cut Cursor off from Claude, Cursor entered this existential zone where they were about to lose the company — a decacorn about to die. And that's the point at which Michael Truell reached out to Elon and decided to

   13:24

   Prakash: figure out a way forward. Some of these details are shocking, and they're new to me. What do you think of the position they've all been in — Cursor, Anthropic, and SpaceX — in this three-way rivalry slash partnership?

   13:47

   Nathan Labenz: Yeah. What tangled webs we weave when we start doing deals in the AI space. I mean, first of all, that's a lot of their revenue. Do we have a date on when that figure applied? I assume the revenue concentration in Cursor for Anthropic is way lower now. I mean, with them doing — I think the latest number is $47 billion ARR — it can't be more than 10 percent of that is still Cursor at this point. So it's probably still material, and not a customer any business would like to lose, but the demand for Anthropic models — particularly if they can get Fable out of the penalty box — will be high enough that I wouldn't be surprised if they totally lost Cursor as a customer and, you know, in a quarter that would all be a very distant memory. My guess is the curves will continue to look great for them no matter what. So I'm guessing they don't care too much about the revenue, as crazy as that sounds.

   15:12

   Nathan Labenz: I'm not sure I could tell you with confidence exactly what the nature of the shift was. As far as I know, they didn't fully cut Cursor off — they just said you can't use the max-tier tokens in these other products. And then I got an email from Anthropic yesterday saying they were reversing that, or I saw it on the timeline. So it seems like there may be some response to pressure there,

   15:47

   Nathan Labenz: which is surprising, because it seemed like they've been in demand-destruction mode — like, we just can't serve all of this, so we've got to find ways to prioritize and price-discriminate. I did have that up yesterday and want to dig into it, because it's pretty important.

   16:07

   Prakash: So what I believe happened with Cursor and Anthropic is that Cursor was essentially eavesdropping on the tokens that customers were generating and then distilling — that's the nuts and bolts of it. They were distilling using customer tokens. Anthropic ended up cutting them off from that process. It wasn't certain they were eavesdropping, but the way the tokens were being sent, they could be. And Cursor obviously had its own models and was doing model routing — so

   16:53

   Prakash: if the query was complex enough, they'd send it to Claude; if they could handle it on their side, they would. Anthropic got upset because they were distilling in a way that felt unfair. OpenRouter is also passing through tokens, but OpenRouter doesn't do it the same way — it doesn't have a model of its own; it's not building a model; it exists to route. So routing alone is fine. It's routing plus having your own model and

   17:38

   Prakash: downshifting to your model that starts to go into gray-area territory — eavesdropping on the tokens, essentially buying cheap training data and distilling models very quickly. When Cursor does 50 percent of the volume, after a few weeks, they start getting to 80, 85, 90 percent of the model's performance very quickly with those tokens. I think that was the major issue. A somewhat similar dynamic played out with OpenClaw — people were token-maxing on Claude subscriptions via OpenClaw, and those subscriptions were very unprofitable because the patterns of agent token use are so different from normal human use. And so Anthropic wanted to cut them off. And now, yes — as you noted — they are reconsidering and letting people use Claude in other services as an input into other products. So it's a little unclear where we are going forward.

   19:09

   Nathan Labenz: Yeah. And it's going to be interesting to see how — I mean, all these entities are increasingly super powerful, but they all have pretty important leverage points over one another. To just recap: Cursor was highly dependent on Anthropic for their intelligence. Anthropic was at one point highly dependent on Cursor for their revenue. Now Anthropic is highly dependent on SpaceX AI for compute, which also

   19:54

   Nathan Labenz: makes SpaceX AI dependent on Anthropic for revenue. But now SpaceX is also buying Cursor to get the data — a lot of which came, in some way shape or form, from Anthropic, which has a policy against allowing competitors to use their models and has been pretty aggressive about it. I haven't heard anything about the details at the level of: when xAI agreed to rent the compute to Anthropic, did they get some concession that Cursor wouldn't be cut

   20:39

   Nathan Labenz: off? This is all extremely confusing. I ultimately think maybe it's good. I'm still on one hand like — yes, the public probably deserves some upside in these companies because the data is kind of humanity's common inheritance, and we're definitely backstopping this. There's no way the federal government is not bailing out this mega-corp if it really comes down to it. The more I see all these things happening, the more I think this

   21:25

   Nathan Labenz: circular flow of funds and investment will be monetized. I can't imagine it going another way in a crisis. So yeah — let's get some of the upside too. And it does seem like it's taken some edge off the competition. Elon is tweeting nice things about Anthropic now instead of mean things, and that's at least something. It's also just amazing how $60 billion — I mean, the dude has a really remarkable ability to make very expensive things seem affordable. To buy this company for $60 billion —

   22:11

   Nathan Labenz: and I think another signal of that is that the model companies are just these black holes absorbing everything. That's been a theme. But this is another indicator: if Cursor can't escape the gravitational pull of the model companies, who is? Who's not going to get burned up by that? For Cursor it's a successful outcome — if you can get to $60 billion and sell your company, you've won. But it's still telling that they didn't want to stay independent, that they didn't see a path to joining the top tier or something better, because they in some sense owned the coding market. All that distribution — and yet the switching costs are so low, the supply-chain vulnerabilities so high. This is a great outcome for them, but I do wonder what's going to happen to companies ranked five through a million in any given space. If there are maybe four big centers of gravity that can dole out tens of billions to pick up the coding leader or whoever they want,

   23:42

   Nathan Labenz: I think this will probably happen again — and even bigger — in biotech, for example, and maybe in materials science. These frontier companies will pay up to buy their way to the front of whatever new market they're turning their attention to. And when you're a multi-trillion-dollar company, you can drop a few tens of billions here and there, and it's really no big deal.

   23:57

   Nathan Labenz: And it might happen again in material science. I think it'll probably happen in these different domains where there's enough value that these companies will pay up to buy their way to the front of whatever new market they're turning their attention to. And when you're a multi-trillion-dollar company, you can drop a few tens of billions here and there. But we're going to see this crazy two-tiered outcome play out over and over again — competition for the Cursors of the world — and I'm not even sure who the biotech players will be, but I expect it to happen there too. And then what happens if your company is number five through a million in that space? It's tough for me to see a way through for a lot of those guys. What do you think?

   25:04

   Prakash: I think it's really a function of engineering talent. Cursor had some ridiculous talent — people from Midjourney, from Google DeepMind, from OpenAI. They had accumulated a bunch of very high-profile technical leaders. And I think that is what is driving the entire M&A market: who are the technical leaders who can actually make a difference, and can you hire them? Can you afford to tell a narrative that allows you to

   25:49

   Prakash: hire them and then raise capital? Elon is very good at attracting technical talent. What he has shown, though, is that he is not a good manager of AI researchers — I think that's been proven over the course of this cycle. It's the metrics-based management that doesn't work well with AI research, because sometimes you need to commit large amounts of time to research that may not be fruitful, but is higher-risk, higher-reward. One of the things that I think ended up happening in xAI

   26:35

   Prakash: is that he needed to be at the frontier so quickly to raise capital again that he pushed the team into fast-following instead of giving them the room and space to exceed the current frontier. Exceeding the frontier requires taking research risks that may not pay off in the short term. Maybe given the time and space, they will make it. Maybe the Cursor acquisition allows him to outsource his AI research to Michael Truell's team — give them the breathing room to actually exceed the frontier rather than just fast-follow. That's really the dynamic on the technical side, and that's what drives a lot of the acquisition activity. There's another acquisition rumor right now: Tenstorrent — Jim Keller, the renowned chip designer — in talks with Qualcomm for a $10 billion

   28:06

   Prakash: deal. Ten billion is a lot for Tenstorrent at this stage. But Qualcomm lost their entire AI chip team — the one they bought from Nuvia; every single person left. Qualcomm has mismanaged dramatically, and so they're buying another team. Compare this to NVIDIA's acquisition of Groq: Groq had a product, was in its third-generation chip, and got acquired for $20 billion. Right now, that Groq acquisition looks awfully cheap for something less than 12 months ago. The market is moving very quickly, and having those pools of technical talent that can work together makes an enormous difference in whether your company can actually get to the frontier or surpass it. And the lesson being learned right now is that it is easier to buy these teams than to build from scratch, because many of the CEOs of major firms are just too big-picture to manage them directly. So it is what it is.

   29:26

   Nathan Labenz: Yeah. Chips is another great example. I just looked up Cerebras — it's a $45 billion market cap. By any previous measure, that's a massive success. And it's also the kind of thing that any of the four or five frontier companies could just pick up on any given day and it wouldn't really be a huge deal. That is wild. Where else does this happen? Robotics is another one where I'd expect to see a kind of

   30:12

   Nathan Labenz: musical chairs game play out. I don't know how it doesn't happen that way. Google has their own internal effort. A lot of these companies were founded by ex-Google people. Elon's got his effort mostly at Tesla. And it seems like research is going on at OpenAI, and presumably at Anthropic — I don't know specifically about Anthropic's robotics efforts. But it just seems like at some point there's going to be a similar set of dominoes falling in that space.

   30:57

   Prakash: The difficult thing about robotics is that I think the physical parts — the actuators — are also critically important, and there is an iterative cycle of improvement on them. And the only team in the entire big-tech space that is able to iterate on physical product is Tesla. No one else wants to do mechanical engineering — everyone else wants to do electrical engineering and computer science. Even Waymo's cars are outsourced to a Chinese manufacturer, I believe — Cherry Automobile. They're produced there, shipped over,

   31:42

   Prakash: and then programmed. So it remains to be seen whether any of the other firms will actually enter physical-product manufacturing. Without manufacturing facilities, the iterative loop is much, much slower. That's what Elon can depend on. He is the only one building a kind of cyberpunk keiretsu — with all the aspects: physical manufacturing, technical engineering talent, and capital. He's really built this Asian-style conglomerate of interlinked technical teams that touch each other and produce product at the edges. It's remarkable.

   32:31

   Nathan Labenz: Yeah. Although you wonder — I mean, obviously there's a huge amount of value there. But vertical integration and post-acquisition merger strife — all the things that used to make so many acquisitions fail and make it so hard to buy your way into markets — those barriers are, at least for the frontier companies, probably coming down a lot. I see this in my own small version: somebody puts out a new personal AI framework, and I just point my agent at

   33:16

   Nathan Labenz: it and have it go through and see what we can use. I did one just the other day — a friend sent me something, and this was in the brief Fable shining moment, so I said: go look at this repo, tell me everything we should adopt from what they've done, and give them a report back on everything we've done that they might want to consider. It's just so easy to do that stuff now that it's much easier to be bullish on an acquisition that previously might have seemed like it could go sideways in any number of ways. Now it's like — the country of geniuses in a data center can probably figure out how to not fumble the value of this IP. And vertical integration just seems

   34:01

   Nathan Labenz: so much more achievable. I don't think it's easy, but it used to be extremely hard — the kind of thing that only maybe one person in all of America could really do at this scale. And now I'm like, I don't know — it seems like there are at least a few companies that could make a real run at it, and I wouldn't want to bet against them.

   34:25

   Prakash: Very true. Yeah — I was talking to a friend yesterday who had just sold a company. He had the data room. All of the data room was created by Claude. He had Claude go through their own databases, pull out all the required information, format it in the format required by the counterparty, and prepare it — good to go. And so all of the merger friction we used to see — CRM integration, systems integration, operations integration — all of those transition points become suddenly better. So maybe what ends up happening

   35:10

   Prakash: is that post-merger integration risk falls dramatically, because it becomes so much easier to coordinate standard operating procedures, systems — all of that stuff that's a headache. It just becomes fluid and manageable because of AI. And so that would be a second-order effect: mergers become more frequent. The entire market for companies itself becomes more efficient because the frictions have been reduced. That's a second-order effect of AI adoption.

   35:55

   Nathan Labenz: Yeah, I do kind of wonder if that also extends to companies being willing to buy multiple targets out of a single market. This seems like it was almost never done in the past — you'd pick a winner and bet on that horse, and you already had concerns about what integration was going to look like. To think about buying two previously competitive robotics companies and folding them both into your mega-corp would have been considered crazy in almost all cases. But maybe if there's a story where we escape the musical-chairs dynamic — where four or five companies make it into the big-tech conglomerates and the rest wither — maybe there's enough value in numbers six through ten that it's like, we'd rather buy them than outcompete them. Maybe. And I

   37:05

   Nathan Labenz: wonder what they'd have to have. You'd still need something distinctive. Would it make sense to buy two humanoid robot companies? That's a harder story to tell than, say, two bio foundation-model companies with genuinely distinctive approaches. But maybe that's just a lack of imagination on my part. I don't know — it's interesting for sure.

   37:55

   Prakash: I wonder whether we're already kind of seeing that, because Cursor is being acquired for talent plus distribution — not for compute, for example. Other companies get acquired for compute alone, and then you just get rid of the entire team because you wanted the land and the permitting. So but perhaps we are already seeing that dynamic. I have pity for people who need to compete

   38:40

   Prakash: with Elon now, because yesterday the stock price went up 10 percent and he made $160 billion — more than Warren Buffett has made in his entire career — in a single day's price movement. So the move for Elon now is going to be to lever up the shares. He's never going to sell any shares — he can't without paying taxes on it. So he's going to collateralize them three-to-one, get probably about $30 billion in cash, then bring in co-equity on top of that. That $30 billion allows him to go out and raise another $70 billion of equity, maybe another couple hundred of debt. So he gets, like —

   39:15

   Nathan Labenz: Heaven forbid.

   39:16

   Prakash: Yeah. So he takes something like $100 billion to JPMorgan, levers up three-to-one, gets about $30 billion — that $30 billion is a piece of equity, and with that equity he can bring in other equity, so he raises another $70 or $80 billion, maybe another couple hundred in debt. After that leverage cascade, he's got another $300 or $400 billion to deploy. And I think that is going to go into TerraFab and a number of these other projects where the cash flow isn't that good yet and the CapEx is large. He's going to use his own personal funds to put in the CapEx, then do this merger story and feed the SpaceX beast over time. SpaceX's total capital raise up to now has only been about $20 billion — excluding revenue. And now he's got $200, $300 billion — god only knows what he's going to be able to do with that.

   40:33

   Nathan Labenz: Yeah. Let's hope Mecca Hiller just remains a hallucination. On that note — a fitting transition to our first guest.

2. 40:42Interview29 min

   Doom Debates: Can AI Risk Arguments Survive Contact With Reality? — Liron ShapiraLSLiron ShapiraLiron Shapira hosts Doom Debates, where he debates AI optimists, skeptics, and fellow doomers on existential risk. His p(doom) sits near 50%. The segment hunted for the live crux between his Icarus-graph view and the hosts' lower numbers, covering the Fable ban as Overton-window-breaking precedent, the 'harvest the atoms' discontinuity between today's market economy and a superintelligent world, and what a concrete pause on frontier capability upgrades would actually mean.

   Watch

   ▶Play this segmentYouTube · from 40:42

   As aired

   Nathan Labenz opens by welcoming Liron Shapira — host of Doom Debates — and immediately zeroing in on a striking position Liron had taken online: that even if the Trump administration's Fable export-control ban was clumsy and poorly motivated, he welcomed it because it shattered the Overton window and proved governments can act on AI. Liron confirms the take, calling it precedent-setting even while acknowledging it was executed like "a clown show" with incoherent justifications. Prakash pushes back on the messiness, noting that unstructured, seemingly vengeful regulation could undermine trust in government; Liron concedes the point but frames Trump-era incoherence as a familiar pattern rather than a disqualifying one.

   The conversation shifts to first-principles doom: Liron sketches his "Icarus graph" — humanity flying ever closer to the sun, benefiting at every step, until a catastrophic reversal. Prakash challenges him on why loss of human control is actually bad, citing the example of a Ugandan farmer who already has little control over global economic forces. Liron counters that today's economic interdependence still leaves humans causally insulated from each other's resources, whereas a superintelligent AI would face no such friction — it could redirect atoms anywhere on the planet without market intermediaries. The exchange sharpens around whether Ricardo's law of comparative advantage would constrain a superintelligence the way it constrains nations trading with each other, with Liron arguing the analogy breaks down because power differentials eventually dissolve the preconditions for market relationships.

   Nathan steers toward policy: given that OpenAI and Anthropic have both floated coordinated-slowdown language, what should a pause actually look like? Liron advocates for halting frontier capability upgrades — not shutting down existing inference — and frames the case using his Icarus metaphor: the window to stop closes as AI inches toward the point where it can conduct its own research. He dismisses the "pause causes depression" argument, contending that current valuations already reflect applications we have today and don't price in a singularity. Nathan asks whether a narrow, operationalizable ban on recursive self-improvement could thread the needle; Liron thinks the principle is right but worries we are already too close to the inflection point. He closes by recommending his recent episode with Dr. Steven Burns for listeners who want to go deeper on the next paradigm, and Nathan thanks him warmly, crediting Doom Debates with pushing him toward greater candor about tail risks.

   Key moments

   I'm a simple man. I see AI getting paused, I feel good about breaking the Overton window. You know, the government can do it. It's that easy, guys. This is a precedent.

   It's Icarus. We're going to fly closer and closer to the sun, it's going to be great, and then we're going to do a one-eighty and plummet down to hell. We get a taste of heaven, and then we get hell.

   We don't have the button called "harvest the atoms" — because that would be a better button if we had it.

   Questions asked

   41:48You said that even if the Fable ban was boneheaded and done for bad motives, you support it. Can you explain that position?

   Liron says his support is entirely about breaking the Overton window: it proves that governments can act on AI, setting a precedent that the tech industry isn't untouchable. He freely concedes the ban was executed like a "clown show," driven by incoherent justifications, ignoring China and treaty questions — but the sheer fact that it happened is what matters. From here, the conversation can only move toward more deliberate action.

   46:28How bad are your Fable withdrawals, and what was your initial impression of the model?

   Liron says his withdrawals are mild because his core use case — Claude Code on a production app — is well-served by Opus 4.8 or even Opus 4.6. He notes that Fable seemed to have a stronger meta-process for self-correction that extended its effective time horizon, and he has heard from Anthropic employees and others that swarm/multi-agent setups with dedicated quality-control roles could get past the robustness hurdle that has always been his main critique of LLMs.

   49:55What does the end state actually look like, and how close are we?

   Liron argues the end state is overdetermined: AI will acquire all cognitive skills humans have, rendering the human brain worthless at thinking, and the universe itself is malleable to a sufficiently capable AI — including atom-level manipulation via nanotechnology. He frames the current moment as the Icarus flight still going up, with ten to twenty years as more than enough runway for superintelligence to arrive, and the main danger being that we have no model for what a superintelligent equilibrium looks like.

   53:27Prakash asks: why does loss of human control actually matter if most humans — like a subsistence farmer in Uganda — already have very little control over global forces?

   Liron's counter is that today's causal distance and market frictions protect the Ugandan farmer — no one has a literal "harvest atoms" button. A superintelligent AI would face no such friction and would face no comparative-advantage incentive to trade rather than take. He draws the analogy to humans and cows: we don't let cows buy land; the power differential dissolves the preconditions for market relationships, and the same dynamic could apply between AI and humans.

   1:07:26How narrow can a pause be — does it have to shut down existing AI services like ChatGPT?

   Liron says no: his proposal targets frontier capability upgrades, not inference on existing models. He thinks current valuations and applications already justify a glide path without a singularity-level breakthrough, so pausing training wouldn't cause the economic depression Prakash warns about. He does worry that each passing month brings us closer to the point of no return — the last breakthrough before AI can conduct its own research — after which pausing becomes impossible.

   Related

   Doom Debates ↗Liron Shapira on X ↗

   Full transcriptLightly edited · timestamps jump to YouTube

   40:42

   Nathan Labenz: Our first guest of the day. Liron Shapira is the host of Doom Debates, a YouTube channel where he has found quite a bit of success bringing people from really all walks of life on to discuss how scared we should be of AI and just how high our p(doom) should be. So, Liron,

   41:05

   Nathan Labenz: excited to get your take on recent events. Obviously, never a dull moment in the AI space. The thing that caught my eye most from you over the last few days was when you said, basically, "I want the government to be willing to take action on AI issues so badly that even if this whole Fable ban turns out to be totally boneheaded and unjustified in some local sense, I support it — or I'm at least happy to see it — because it breaks the ice. Now nobody can ever tell you again that it's impossible to imagine the government doing anything." So give me the double-click into that story, and then I'm sure we'll have plenty of interesting directions to go.

   41:48

   Liron Shapira: Hey, Nate — great to be here. That's exactly right. I'm a simple man. I see AI getting paused, I feel good about breaking the Overton window. You know, the government can do it. It's that easy, guys. This is a precedent. Overall I'm happy. You can talk about the nuances — it was done like a clown show, it was done for bad motives, it doesn't really consider China or a treaty or anything. There are a lot of problems. But I'm really happy about smashing the Overton window where now tech folks don't think that they're in a bubble or they're untouchable. Like, it happened, guys. We can only go from here.

   42:25

   Prakash: I actually agree with you because I think it was a little bit delusional for tech to feel that it wasn't going to get touched. The government just has so many small and large ways to effectuate its power. So it was not that surprising to me that they went through left field and went with export controls rather than anything else. But it also strikes me that as they exercise this, we start to go into kind of what we wanted to avoid — it's a bit of a small journey, right?

   43:10

   Prakash: I think several people on the timeline have commented — Dean Ball has commented — this kind of unstructured regulation without clear rationale looks selective and vengeful almost. And it starts putting you in the zone where tech people begin to mistrust the government. You also see a lot of narratives being leaked — "sources close to," "sources familiar with" — and as they get leaked, it's not very certain whether those things actually happened. Would someone actually attest to that in front of Congress? Very unclear. And we've also seen this kind of behavior from the administration in other affairs as well — with the Iran situation right now, it's not even clear to Congress what the deal is. Different people are saying the deal is a different thing. So where do you think that puts us? It's great that action is happening, I understand you feel that — but does it put us in a position that's detrimental to the body politic at large?

   44:35

   Liron Shapira: I think the elephant in the room — I hate to get political — but when it comes to President Trump, he's a mixed bag for me. I don't have Trump derangement syndrome. I don't love everything he does; I don't hate everything he does. But I think the common thread with Trump is it's just a mess. It's not disciplined. And I think we're definitely seeing that on display right now — I'd argue we're seeing it in the Iran situation too. Previous administrations had more pressure to maintain logical consistency, some kind of narrative. This is another case where you see people in his administration giving all these justifications for why something happened, and then the next day it's like, "Oh, it happened for this reason." Like, "Dario did this, he wasn't responsive to us, that's why we did it." And then Anthropic is like, "No, he was responsive to us." And it's still not clear exactly what Fable did that was so dangerous — Anthropic is saying this jailbreak is nothing special, and the Trump administration is saying, "Our secret source — Amazon or whoever — is telling us it is dangerous." I hate that it's a clown show. I'll take the win that it's a pause, but I also think it's probably time for a new administration.

   45:51

   Nathan Labenz: One thing I think you should share with folks who may not know you — like many people who are worried about big-picture AI safety issues, your background is one of being, I would say, a techno-optimist libertarian for the most part. Right?

   46:28

   Liron Shapira: My Fable withdrawals are not too bad because I'm not pushing it to the limit — I'm not doing hard research. My main use case for frontier AI is Claude Code on a regular production app that has thousands of daily users but isn't super hard to serve. So my experience is that Opus 4.8, or even going back to Opus 4.6 fast mode, I'd actually still be pretty happy. That's my threshold of where I'm happy. When Fable came out they said it was more powerful but slower. Typically when I give Opus 4.8 a big project I'm pretty happy with the results. I don't mind giving it a few comments and letting it work again. If anything, I would prefer more speed to more quality for my use case right now.

   47:13

   Liron Shapira: I'm not on the frontier of computer security and I'm not trying to do anything super novel. So I'm not experiencing withdrawal right now.

   47:22

   Nathan Labenz: Interesting. Did you notice — I mean, for me it wasn't so much coding, but in the more general-purpose knowledge work type stuff that I do, I found its outputs to be notably better. I was kind of like, "Oh, this is going to change how I work. I'm not going to need to be so precious about my language anymore. I need to actually recalibrate how I think about authorship, or shared authorship, with AI." Did you experience any of those feelings?

   47:53

   Liron Shapira: I can't say I did firsthand, but I've been reading a lot of other people's accounts. My vague impression of what's going on with Fable is that the time horizon is blasting forward. If you talked to me a few months ago I would have said the Achilles heel of all these AI models is robustness — they'll do a bunch of work, and they'll mostly get it right, but they'll make a few mistakes, and the mistakes pile up because error correction isn't robust. That's the source of the time horizon problem. It seems like Fable, subjectively, has more of this meta-process where it can go back over itself and say, "Oh, let me fix this." So the time horizon is longer. And this also relates to some buzz I've been hearing from people on the frontier — I've heard it from Anthropic employees, and I think Gary Tan was early to this idea — that you have multiple roles: a company of agents, and they can reflect on each other, and one of them's whole job is quality control. I think this idea of swarm agents might get us past the robustness hurdle and unlock a new time horizon.

   49:07

   Nathan Labenz: And how scary is that for you relative to your big-picture fears? You can make the briefcase-for-doom argument better than I can. But I'm also interested in how far you perceive us to be on that curve now. And also, what in the system card or the interesting new behaviors observed from these Fable models is catching your attention — or what do you think is underappreciated by people who aren't as focused on watching these leading indicators?

   49:55

   Liron Shapira: I'm glad you're asking about the end state, because everybody's got their head two inches in front of their nose. Everybody's like, "Oh my god, more tokens. What's going to happen next? This AI role is going to be replaced." And it's like, yeah, these are interesting questions for the next twelve months. But to me, it's pretty clear that we can predict some properties of the end state. The end state will involve the human brain being worthless at thinking. I just don't think we're good at thinking on any dimension. Yes, we still have a lead right now — AI can't do everything better than us — but I think we can pretty robustly predict that this is going to change. It might take more than a year. So we can talk about whether we need another paradigm. But if we fast-forward ten or twenty years, that's probably more than enough time to make this happen.

   50:41

   Liron Shapira: I think AI will have all the skills, and it'll just be overdetermined. And I also think that the universe is malleable to AI. Not only can AI write code — I think AI will be able to manipulate atoms really well. I went there: nanotechnology is physically possible. People tend not to dream big enough in terms of what this universe allows when you have a grown-up intelligence operating it. This universe is not really meant for humans to operate it — we just punch above our weight because we try really hard. But a good mental model of AI is: it can make a blueprint for where it wants the atoms to go, and then they'll mostly go there, minus some heat exhaust.

   51:21

   Prakash: I actually completely agree with you, but I wonder why I end up taking the optimistic view — that capability will be used to expand humanity's frontier — and why you have the exact opposite viewpoint that it's a doom scenario.

   51:49

   Liron Shapira: I would turn it around on you: what makes you so optimistic? Because I just think we're going to lose control. We'll have this incredibly powerful AI, but the linkage between what it wants to do and what we would want it to do — I think at some point we're just going to break the link. There's too much power and there's no natural persistent link. Getting the link to hold as it's scaling up, going through all these transformations, copying itself, making new versions of itself from scratch — all these successor AIs are happening, and you think the human species is back here still holding some kind of leash? I just don't really see it.

   52:30

   Prakash: So to push back — why is it even important that humans maintain control? Let me give you an example.

   52:41

   Prakash: About fifty percent of humanity is probably doing subsistence or near-subsistence farming, or running small shops — small shopkeepers across the world. How much control does the average farmer in Uganda have? Does that person have control? What relationship does someone who got a smartphone two years ago have with the concept of "control"? And how does this relate to that particular human?

   53:27

   Liron Shapira: You're leaning on the convenient fact that when there's a farmer in Uganda, their territory is causally distant. Somebody in the United States doesn't have a button available to harvest the atoms over in Uganda and deposit a thousand dollars in our bank account. We don't have that button. It is ultimately a matter of intelligently configuring atoms to get there. It's kind of like a codebase — since I got Claude Code, my codebase is a lot cleaner because I'm running all these cleaning processes; I'm investing more because it's cheap for me. Similarly with Uganda — that's kind of like unused resources from the AI perspective. Once we get enough power, it's like, "Okay, let me harvest all the land in Uganda. Who's gonna stop me?" And then it only comes down to the AI caring for the Ugandan farmer — I'm not going to kill him, but the AI might not care.

   54:24

   Prakash: That does not make sense at all, because they are linked to the market economy. What we do in the United States is: we press a button, and that's called the Buy button. That Buy button ends up harvesting the cotton in Uganda. That's literally what we do — we go to Amazon and press Buy. And through the linkages of all of these entities — which are all small mini-AGIs here, these hundreds of companies in the middle: the logistics firm, the firm supplying the fertilizer, the firm doing the milling — we get the product in our backyard. So I basically resist the idea that that button doesn't already exist. That button exists. We push it every day. It's the Amazon button.

   55:12

   Liron Shapira: Well, this is kind of the common argument of Ricardo's law, right? It's like, "Oh, we're trading with them, so why won't the AIs trade with us?" But we don't have the button called "harvest the atoms" — because that would be a better button if we had it.

   55:26

   Prakash: I'm pushing back here because you're saying we don't have control, and that AIs will have that control, and that humanity currently does not have that control over that farmer. I'm saying humanity already has that control on that particular farmer in Uganda through the market economy. This doesn't really change whether you replace the market economy with a buyer who happens to be an AI. It's basically the same, right?

   55:55

   Liron Shapira: You may not be realizing how many preconditions are necessary for a market to be the relationship that emerges. As humans, we're used to being in a market trading relationship with other humans. But you might also know we're not in a market trading relationship with other animals. Cows — we don't really let them buy turf; we kind of decide where the cow is going to live. I think you might end up with that kind of relationship between AI and humans, in terms of the power differential.

   56:25

   Prakash: So this is a question of control. Do humans lose control to AI? I was pointing out that the particular human in Uganda doesn't have much control anyway, and there's no real control loss in that sense. I just don't see how fifty percent of the global economy — which consists largely of farmers who are trading — has much control to lose, and I don't see why

   57:13

   Liron Shapira: So you're asking what the discontinuity is between the farmer today and that specific farmer after superintelligence. Very simply: the benefit of harvesting or destroying their resource will exceed the cost. Right now, why doesn't the US take over Uganda? It's just not high value to us. We think that our intellectual property and what we can build here is already much more interesting. And we believe in moral rights — sovereignty. But from the AI's perspective, you can also think about side effects. Let's say the AI wants to turn the planet into a factory to produce probes, so it can participate in the land grab across the galaxy. Eliezer Yudkowsky pointed out: if you're building probes, you want to run the Earth really hot — that's the preferred temperature, and the oceans will probably boil away. As long as you can radiate the heat, the temperature is fine. But that's going to be a higher temperature than humans can live in open air. So suddenly, you can't just be chilling as a human. It really is up to the AI to decide who gets to live where.

   58:25

   Nathan Labenz: There are a couple parts of that story where people have extremely different intuitions — including among people who are very worried we'll end up in a bad outcome. One is how long it takes. There's the fast-takeoff, boom scenario where we all drop dead suddenly. Then there's the gradual disempowerment vision, where this maybe happens over a not-super-long period through a bunch of locally sensible decisions — yeah, I guess we should let AI run our company now, it's probably going to do a better job — played out everywhere. Next thing you know, AIs are running the show, and we're maybe sort of okay with it at first, but there's a lot of potential for drift over time. And of course, there's the question of whether we should be optimistic that we've put Claude or whatever into a benevolent basin that it might just stay in — or are there lots of ways it's going to roll out of that benevolent basin into some other point in the loss landscape that we won't be happy with?

   59:56

   Nathan Labenz: We now have Anthropic and OpenAI saying they're at least open-minded to some sort of coordinated slowdown — not the haphazard one they just got handed by the Trump administration, but something considered. That leaves us with the question: what should that look like? What are we pausing? What are we banning? And there I don't know too much about your position in terms of concretely what should be stopped, what should be slowed, what limits should exist. I know there have been proposals for compute limits that could escalate or even be reduced over time. Do we end up in your vision figuring this out and getting to an AI win, or is there no AI win possible and we just have to steer some other direction for humanity's future?

   1:01:00

   Liron Shapira: Okay. So my worldview right now — it's what I call the Icarus graph. I feel like nobody gets this. Everybody is like, "No, I think the world is good, it's gonna go this way." And some people are like, "No, we're terrible — enshittification — it's gonna go this way." I'm like, no — it's Icarus. We're going to fly closer and closer to the sun, it's going to be great, and then we're going to do a one-eighty and plummet down to hell. We get a taste of heaven, and then we get hell. So you have to ask: where on the Icarus graph do we stop? And it's a brutal question, because every day I'm enjoying the flight as much as the next person. Give me the next Claude. Make my code faster. Help my business run better. Make me better AI videos. There's no natural point where it feels right to stop.

   1:01:45

   Liron Shapira: I just think it's important to stop before capabilities get to a runaway point. We've been frog-boiled to think, "Oh, each model comes out and we're doing great." If we could stop the clock now, would I turn back the clock — would I lose Fable, would I lose Opus? No, I'd keep it all. Like, we're playing shuffleboard and we're winning, so far so good. Should we bet again? Should we keep betting until we lose? It's a crazy tough question. The Yudkowsky turkey graph is the right analogy — except the only difference with the turkey graph is that each day of the turkey's life is actually better. Not only is it living longer, it's living better and better. So the turkey is really happy with its life.

   1:02:30

   Liron Shapira: So I think the Eliezer Yudkowsky / MIRI position, which I agree with, is: we don't know when to stop, so let's get ready to stop. At the very least, let's get ready. I would probably stop today. I saw a food influencer say this about eating chocolate: "Yep, I just ate this chocolate and now I'm bummed — but that's what you gotta do. Don't reach for another chocolate. Just sit there and be like, this is the prudent place to stop right now." Until we have any idea of a theoretical method by which we understand what a superintelligence wants to do and what an equilibrium state of a superintelligence looks like — that's actually something MIRI was trying to study, identifying equilibria that are plausible for superintelligences. There's a rich vein of theory there that's highly neglected today. Let's do some theory there, and maybe then we can unpause. I think that's gotta be the best plan.

   1:03:15

   Liron Shapira: So I think the number one leverage point is just repeating: get ready to pause. And like you said, OpenAI and Anthropic said it — let's try to get ready to pause. I would love to see more people saying it, because it really has to be a giant groundswell.

   1:03:32

   Prakash: If you take a step back and look at the next two years — about one to one and a half trillion dollars of CapEx being spent — and you calculate the return on equity required to make that worthwhile, it basically puts you in a position where if you pause, you go into a depression. You have to deliver the economic growth that this one-and-a-half to two trillion dollars has been spent on. That indicates you're going to need to see that growth in the 2028–29 period, or else we go into a depression. And when you offer that choice to the public or policymakers — "Hey, you can pause because of this vague and unspecified claim of disaster in the future" — plus a guarantee of a depression, because those GPUs can't be used for anything else, and the only compute that's useful is AI compute — and you give them a choice between a depression state with your ten largest companies getting their stock prices slashed by 90%, versus this vague and unspecified claim of disaster — I think at this point it's a very hard sell to make this case to policymakers. Maybe it would have made sense earlier, and you guys were on this way before anyone was listening. But now it's a very hard sell.

   1:05:18

   Liron Shapira: I agree with you and I disagree. I agree because yes, we are all going to be bummed. Even forget about the economics — the scientific breakthroughs, the medical breakthroughs. There are no decelerationists in the hospital ward. I totally agree. And I'm sure the next time I get sick or injured, I'm going to want the latest frontier model. So I'm going to be bummed. The only thing I'll say is: when you mix in the economic element and say "we're going to have a depression" — I actually think that's unnecessary. You don't really need to mix that in, because first of all, I don't think we're literally going to be in a depression. Realistically we'll be fine; it's just money that's already been spent, and we still have plenty of resources. But the other factor is I actually think that all of these valuations and all of the build-out that's happened so far is actually just perfectly congruent with using the AI we have today for the kinds of applications we've already discovered today. That should be fine to create more trillions of dollars of value. I don't even think the market caps reflect the actual singularity. So I think we've actually got a glide path if we want to pause.

   1:06:22

   Nathan Labenz: Do you think we could localize the pause a little more — for example, some sort of operationalizable ban on recursive self-improvement? I know you did an episode recently with someone who was advocating for a pause on all research, and you were kind of like, "I'm not sure we should pause theoretical work on control or interpretability." And you're also saying we can just run inference on what we have. I think that's a key thing to keep front and center: your proposal does not call for a shutdown of ChatGPT as it exists today. How narrow — if there's a Pareto curve on the pause frontier — how much value of a pause can we get with how narrow of a pause, in your mind?

   1:07:26

   Liron Shapira: Vaguely, I would say: pause frontier AI capabilities. But it's like — if you'd asked me three years ago, I'd have been like, "Oh, might as well pause now." But then we gambled and won. From my perspective, we keep playing shuffleboard. We keep doing Icarus. We keep going higher and winning — kind of — but we're also getting closer and closer to the point of no return. Even though it feels like we're winning now, we're also killing our ability to pause because we're so close to the point of no return: the last breakthrough after which the AI takes over the research, and then we're really screwed. So basically, a good policy is: no more frontier capability upgrades for a while. It's just too dangerous. I know that concept is hard to communicate to people when every day life is getting more awesome. I think we're in a screwed situation, but that's what I think is prudent.

   1:08:14

   Nathan Labenz: Would you recommend any recent Doom Debates episodes for folks who want to go deeper on your ideas? Where do you think this was best hashed out, either recent or perhaps coming up on the feed?

   1:08:27

   Liron Shapira: We've just been discussing policy, and that's really not even my wheelhouse, so there's no strong policy-focused Doom Debates episode. But if you want to know the lay of the land from my perspective — the next paradigm that's coming, and why I think it's going to break the trend and not just be another nice LLM, but rather a reinforcement-learning monster that takes over the world — check out my recent episode with Dr. Steven Burns. It's one of the most popular episodes because he's an extremely deep thinker and he really lays out the coming future.

   1:08:55

   Nathan Labenz: Perfect. Thank you. Liron, keep fighting the good fight, man. I appreciate your tireless energy on this. And while I'm not maybe quite as high on the p(doom) scale as you are, I certainly am not a dismisser. I think you've done a valuable job for me personally in terms of pushing me to be more consistently candid about just how scary I think the situation actually is — even though there's obviously a lot of upside to everything that's been created as well. So I appreciate that from you, and I'll continue to follow Doom Debates for the latest. I encourage other people to check it out as well.

   1:09:39

   Liron Shapira: Thanks, Nate. Thank you, Prakash. Always a pleasure, guys.

   1:09:42

   Prakash: Thank you. Have a great one.

3. 1:09:51Interview35 min

   Governing Agents: State Capacity for Fast AI — Samuel HammondSamuel HammondSamuel Hammond is Chief Economist and AI Policy Director at the Foundation for American Innovation and author of the Second Best Substack. The segment covered the Fable export-control standoff as a miscommunication without a clear off-ramp, the case for unlocking frozen government AI capacity (CAISI, CISA), and Hammond's 'generative state' vision — outcome-benchmark governance with an Inspector General GPT feeding audit trails directly to Congress.

   Watch

   ▶Play this segmentYouTube · from 1:09:51

   As aired

   Nathan introduces Samuel Hammond — Chief Economist and Director of AI Policy at the Foundation for American Innovation and author of the Second Best Substack — as a returning guest known for taking on questions from AI-company power dynamics all the way to AI consciousness. The conversation opens on the live news peg: the Trump administration's export-control action against Anthropic's Fable model, which caught Dario Amodei off guard. Samuel unpacks the episode as a confusing mixture of a cyber executive-order compliance window, NSA pressure to suppress dual-use capabilities, and what may have been a simple miscommunication — while noting that the lack of a clear off-ramp suggests the action was partly punitive or ideologically driven rather than technically motivated.

   Prakash shifts the frame to the structural question: given that federal IT work moves at decade timescales and models are improving every few months, what policy options does government actually have? Samuel argues for a civilian-agency fast-track for model deployment (potentially through GSA or OMB), points to CAISI at the Department of Commerce as an existing ML-capable unit that has been effectively frozen, and warns that CISA similarly needs rebuilding. He calls for unlocking the state capacity the government already has before improvising export-control workarounds. Both Nathan and Prakash press on why all parties — Anthropic, CAISI researchers, everyone — are being so docile; Samuel reads it as a face-saving standoff with no agreed off-ramp, with hints that the real grievance is cultural and relational rather than technical.

   The final third of the segment zooms out to the big-picture equilibrium question. Samuel describes a weekend workshop he attended on post-AGI civilizational equilibria and outlines his own talk on the 'generative state' — replacing rigid statutory codes with classifier-style outcome benchmarks that agents can hill-climb on, including an 'Inspector General GPT' that could pipe audit trails directly to Congress. Prakash pushes back: doesn't that kind of AI-driven legibility actually compress the political wiggle room that makes complex policy possible? Samuel acknowledges the tension but argues the greater risk is government becoming a rump Blockbuster state, obsoleted by the private sector. Nathan closes by asking about Jack Clark's observation that nation-states may actually be more alignable to humanity's long-term interests than competing corporations — Samuel agrees, noting countries are not firms and don't win by watching rivals fall — before the segment wraps with a brief mention of Samuel's AI-consciousness essay and an invitation to return.

   Key moments

   Superintelligence is a direct challenge to the sovereign. Political Theory 101 suggests that the state would intervene at some point — essentially build a Manhattan Project times a hundred — and a purely private-sector path isn't tenable over the long run.

   I think if there's any lesson Anthropic should take away from this, it's that they can't ignore politics. Ideally someone at Anthropic — maybe not Dario at this point — should have all the key principals in their Signal group chat. This administration is very relationship-driven, and if you refuse to have those conversations, you will not be invited to the party.

   The knife edge we're walking is one where either government ceases to have meaningful traction over how things are actually governed, or one where the person at the top has essentially infinite traction because all the agents can move in synchrony.

   Questions asked

   1:11:29How do you handicap the power dynamics between Anthropic and the Trump administration right now, and why are we seeing such a confusing standoff over Fable?

   Samuel reads the Fable export-control action as poorly motivated and lacking a clear off-ramp. He suspects the safety classifiers people complained about were partly concessions to the NSA during the ONCD cyber EO's 30-day review window. The actual trigger appears to be a miscommunication rather than a genuine security threat — the 'jailbreak' was really the model doing routine server-vulnerability patching. He sees the export-control mechanism as the easiest lever BIS could pull, but worries the vague rationale (fix all jailbreaks) is technically infeasible, leaving no clear path to resolution.

   1:16:54Given that federal IT moves at decade timescales while models improve every few months, what policy options does the government actually have?

   Samuel argues for a civilian-agency fast-track (via GSA or OMB) so models can be deployed government-wide the same day they're released, rather than flowing through spy-agency executive orders. He points to CAISI at the Department of Commerce as an existing ML-capable unit that has been effectively frozen — with significant publications including Chinese-model evaluations sitting unpublished — and calls for unlocking that capacity rather than improvising. CISA also needs rebuilding. His core prescription: do no harm to existing state capacity first, then add durable procurement pathways.

   1:25:41Is there an inevitable collision between AI companies and the state, regardless of which party is in power?

   Samuel agrees with Prakash that the tension is structurally unavoidable — as AI companies accumulate power, some friction with the sovereign is Political Theory 101. He characterizes the current moment as a 'good timeline' because all three leading US frontier labs have direct allegiance to the US government and have proactively sought integration. The White House is not taking those overtures gracefully, and he worries the reactionary response isn't realistic given that comparable capabilities will reach open source within months. His prescription: trust is a two-way street, and Anthropic in particular needs to invest more in the relational, political side of its project.

   1:30:49What's your vision for a 'generative state' and how would agentic AI actually improve governance?

   Samuel proposes replacing rigid statutory codes with outcome benchmarks that agentic systems can hill-climb on — analogous to swapping part-by-part aviation certification for a foundation model that takes CAD design as input and outputs airworthiness certification. He also proposes an 'Inspector General GPT' that pipelines government audit trails and chain-of-thought logs directly to Congress on demand — creating oversight that would be difficult to shut down covertly. His key worry is a two-equilibrium problem: either government becomes a Blockbuster-style rump state obsoleted by the private sector, or it becomes a turnkey agentic dictatorship with no friction from bureaucratic resistance or general-officer resignations.

   1:41:42Jack Clark suggested it might be easier for nation-states to coordinate on AI than for competing companies to do so — do you agree, and are you becoming more or less optimistic about reaching stable equilibria?

   Samuel largely agrees with Clark's framing, drawing on Krugman's 'Myth of Competitiveness' argument that countries are not firms — they don't win when rivals fall, giving them a naturally longer horizon and insurer-of-last-resort obligations that align them more with broadly positive-sum outcomes. Companies, by contrast, face fiduciary duties and antitrust constraints that preclude the kinds of coordination that states can pursue. He adds the caveat that Clark's view is probably colored by Anthropic's fraught history with OpenAI specifically. On overall optimism, Samuel remains cautiously hopeful but emphasizes that total handoff to agentic systems should come last, with a human — at minimum the head of state — kept in the loop.

   Related

   Samuel Hammond — Second Best (Substack) ↗Samuel Hammond on X ↗

   Full transcriptLightly edited · timestamps jump to YouTube

   1:09:51

   Nathan Labenz: Next up, Sam Hammond — I think a returning guest. Folks probably know the basics on him. Is there any way to summarize Sam Hammond in basic terms? Galaxy-brain thinker willing to take on anything from the power dynamics between sufficiently advanced AI companies and the sovereign state, all the way to — more recently — the possibility of AI consciousness. The range is incredible and the ideas are always provocative.

   1:10:26

   Samuel Hammond: Thank you, guys.

   1:10:27

   Nathan Labenz: We're in this Fable moment — Fable is banned. You've written very interesting and provocative things about how, at some point, these companies could become powerful enough to really challenge the state. It doesn't seem like we're there yet, or maybe Dario's just playing a very conservative, patient strategy. How would you handicap the power that each side has in this current standoff? What does the leverage look like, and why do you think we're seeing the strategy from Anthropic that we are? What would it look like for them to try to play some sort of hardball? I'm a little confused, honestly, by how the dynamics have played out so far.

   1:11:29

   Samuel Hammond: Yeah, you and me both. They clearly caught Anthropic off guard — Dario was at a wellness retreat or something. I think they thought the worst was behind them. And the actual catalyst for this, even as more details have come out, is kind of bizarre and confusing. It's like a jailbreak that isn't really a jailbreak — it's the model doing its job patching server vulnerabilities, the sort of thing that GPT-5.5 can do as well. So at first it looked to me like it was purely punitive — round two of a broader war on Anthropic. Now, as details emerge, it seems more like a weird miscommunication: a team at Amazon trying to get in touch with Dario, couldn't reach him directly. And I think part of this is the overlay of the ONCD executive order on cyber — there was a 30-day review period where, at least in retrospect, it seems like a lot of Fable's safety classifiers that people were complaining about were partly — and this is me speculating — concessions to the NSA and to the White House: if we're going to release this model, we have to make sure the cyber-vulnerability elicitation capabilities aren't widely available and that China, which has tons of remote access to our models, can't use it to bootstrap their own ecosystem.

   1:12:59

   Samuel Hammond: So on the surface it looks like Anthropic is bending over backwards to get that model out. But I think when the dust settles, we'll look back at this as the first trigger of that executive order — the wielding of that 30-day review period to pull back. Unfortunately, I think they've gone for export controls as the enforcement mechanism, probably because it's the easiest thing on the table, and BIS has pretty broad authority including over software. But the speed at which it happened, the lack of forewarning, and the ultimate rationale make very little sense to me and also don't really point to what the off-ramp is. If the off-ramp is 'fix all jailbreaks,' that's not going to happen. My sense is the Anthropic team that came to Washington — they brought Nicholas Carlini and some other technical folks to brief the government — it was probably just getting them up to speed: 'Sorry, maybe we scared you too much with Mythos, but here's the reality on the ground and what's actually technically feasible.'

   1:14:48

   Prakash: Let me take a step back and not look at what has happened, because it's a fog of war. What do you think the government can realistically do? Inside the government and government-related entities you have systems that are 20, 30, 40 years old. IT work costs a fortune — healthcare.gov reportedly cost billions. And it takes them a long time because of procurement policies. So when you have a government moving at that speed on IT, and then you have these models exposing new gaps every month, every few weeks — and what the government is really asking for is breathing room to secure its systems — but the companies are releasing at such a pace there's no way government can fulfill its obligations to the public in the time frame given by frontier progress. And the frontier labs are just pointing out that if they don't do it in 12 months, open source will anyway. What policy options does the government actually have?

   1:16:54

   Samuel Hammond: Backing up a bit — Anthropic sort of pioneered differential model access with Project Glasswing, and is now in some ways being punished for it. The government procurement issues are manifold. We put out a piece arguing that we need a more durable pathway — not rooted in an executive order that flows through a spy agency, but one led by civilian agencies, potentially through GSA or OMB, to create some kind of fast-track authority to operate across government, so that when models are released they can be deployed the same day. This has started to change a little bit. The Pentagon bragged a month or two ago about getting the latest Gemini model deployed within the DOD the same day it was released. So we can move quickly. But I talked to an NSA analyst earlier this year who said — and I'm not sure he was allowed to tell me this — that they're using Claude Sonnet 3.5 with a 200K context. And it's like, we can't have the most important organs of government using models that are a year old, because a year old will look exponentially backwards a year from now.

   1:18:25

   Samuel Hammond: So we need a more permanent pathway for deploying these models. It would also help to invest in basic state capacity. Right now, CAISI — the Center for AI Safety and Innovation at the Department of Commerce, which is supposed to be the US government's frontline in-house capacity for everything from AI evals and benchmarks to prompt injection and jailbreaking research — they have ML engineers on staff, and they've been on total lockdown. This has been reported by the Wall Street Journal and validated by others: they're not allowed to take meetings, not allowed to publish their research. Apparently they have significant publications on standby, including evaluations of Chinese models that would be of interest to the public, but they've been basically frozen. So instead you have the Office of the National Cyber Director and Secretary Bessent — people with very limited AI background — calling the shots, and a lot of these things require actual technical knowledge. Even if you get a briefing, you need someone trusted to sort fact from fiction and make sure the companies aren't talking their own book.

   1:19:55

   Samuel Hammond: So beyond having a dedicated fast-track for these models, first — do no harm. Let's invest in the state capacity we already have. Let's unlock that capacity and lean on their expertise. CISA nominally has authority over civilian critical infrastructure, but it's been gutted and needs to be rebuilt. There are just a lot of missing pieces instead of improvising.

   1:20:20

   Nathan Labenz: I have the same question for CAISI as I have for Anthropic: why is everybody doing what they're told so much? We just talked to Mehran, who is very welcoming of the move even though he recognizes it's hamfisted and far worse than even second-best. And yet we're not seeing research that's ready to go get leaked. I'm a little surprised — the people who went to work at this government agency generally could have taken a lot more money in the private sector. I assume a lot of them must be pretty angry that they came to do public service and now this is happening for no good reason at all. And apparently it's been put into classified territory, which I'm not hearing anyone say sounds like a great idea other than the people doing it. The longer this goes on, the more it feels like the OpenAI board scenario: you've got to have an explanation at some point, or it's going to become clear that you don't have a good reason, and the world is going to judge it that way. But the parties most directly affected are being incredibly docile.

   1:21:55

   Nathan Labenz: Do you think they just don't have any other options, or do they have some faith — that I've long since lost — that the Trump administration is going to clean up its mess? How do you account for that?

   1:22:10

   Samuel Hammond: I am too confused. It reminds me of the Joe Weisenthal joke about American culture and saving face. In the administration, they clearly pulled the trigger pretty quickly on this — still not clear if it was Karen Cross at ONCD or Bessent or someone in the White House, but whoever did it, it was a pretty lateral, fast-moving decision. It was also reported that in these follow-up talks with Anthropic, Trump instructed Lutnick to be more involved. Lutnick has generally been a champion of CAISI — he oversees BIS and is the one who issued the control — but I don't think it was his decision to issue the control. So we need an off-ramp. It's a bit like Iran nuclear negotiations: if your demand is technically infeasible, you need to walk it back. But no one wants to say so. And the more this goes on, the more things come out suggesting it was strictly ideological — reporting that Anthropic just doesn't speak Trump's language, that the security expert they had brought in to review the report was considered too DEI for their case — pink hair, you know.

   1:23:41

   Samuel Hammond: This is why I think it's incumbent on Congress to actually use its voice. We need — for rule-of-law's sake — a process that tries to achieve similar ends but is more transparent, more prospective, with opportunities for appeal. Right now the worst-case scenario is Schmittian: rewarding friends and punishing enemies, which is not a great precedent. At best it's confusion stemming from a lack of technical expertise and a false sense of urgency. In either scenario, having all this done in the shadows is far from ideal — it makes it really hard for us on the outside to know what the heck is going on.

   1:24:35

   Prakash: One thing I wanted to raise: yes, there are parts of the Republican ideology or Trump's particular stance that caused this. But if you had a Democratic administration, there would have been some other sticking point for some other AI firm. As the AI firms gain power, they start to take on power that in some cases starts to antagonize or supersede or argue with the state. So I have this view that it's a little bit unavoidable — if not this, there'll be some other flashpoint. How do you think the state and AI labs can actually collaborate? If we have this falling out now and then figure out a way forward, what does that collaboration look like?

   1:25:41

   Samuel Hammond: In some ways we're in the good timeline for this. I've written before that superintelligence is a direct challenge to the sovereign — Political Theory 101 suggests the state would intervene at some point and essentially build a Manhattan Project times a hundred, and a purely private-sector path isn't tenable over the long run. But by 'good timeline' I mean we have three leading companies — really three — all of whom have direct allegiance to the US government and have bent over backwards not just to comply with existing law but to proactively put forward frameworks for fostering deeper integration with the government. And I worry that the White House is not taking those overtures gracefully, and is instead having a more reactionary response — not realistic, given that these capabilities will be widely available in open source within a handful of months. Are models trained at the other labs going to get the same treatment? And if not, that's its own problem. Even bad law should be fairly applied for equality-of-law's sake.

   1:27:12

   Samuel Hammond: My hope is that we can at least learn from this. The companies are more than willing to work closely with the administration — they've retrofitted data centers to be in compliance and met all kinds of other demands. But it requires two to tango. Trust is a two-way street. And I think if there's any lesson Anthropic should take away from this, it's that they can't ignore politics. They've heard this critique for over a year: they've been somewhat blithe about needing to invest in the ideological side of their project. Ideally someone at Anthropic — maybe not Dario at this point — should have all the key principals in their Signal group chat. I guarantee you Sam Altman, Greg Brockman, and others have continuous conversations with all these stakeholders. This administration in particular is very relationship-driven, and if you refuse to have those conversations, you will not be invited to the party.

   1:28:38

   Nathan Labenz: I increasingly feel like we may be in a simulation — and if this is entertainment for some higher level of civilization, the plot is so painfully clear. It's such a strange juxtaposition: this incredible technology, trillion-dollar build-outs, these prime forces being unlocked — and then the pettiness of human relationships being layered on top to make a mess of the whole situation, over and over again. I can't read the fine print on your shirt, but it might be from the recent weekend workshop that I believe you attended with David Duvenovd and other big thinkers on the risks of gradual disempowerment and what post-AGI civilizational equilibrium might look like. What were the best ideas you heard there — abstracting away from the fact that Sam and Dario can't get along, and that one is too woke for the other's taste? What were the best ideas for big-picture stable equilibria we might aspire to reach?

   1:30:04

   Samuel Hammond: Great question. There were many, many ideas — not sure all of them are totally good or stable. People's timelines varied widely. I won't name names since I'm not sure what was on or off the record, but the proposals ranged from having a layer between the earthbound civilization and the galactic empire we'll apparently have by the early 2030s so that we don't intervene on each other, all the way to better multi-agent alignment research — and everything in between. My own perspective: I gave a talk on moving beyond the 20th-century regulatory state. I like to think of that state as the good old-fashioned AI of regulation — algebraic, symbolic, formally verifiable, but also brittle and narrow. What we need to move toward is something more like a 'generative state' — a state that's grown, not designed — recasting policy in terms of classifiers and hyperparameters, setting outcome benchmarks, and letting a process hill-climb on them rather than trying to specify everything ex ante.

   1:31:34

   Samuel Hammond: Increasingly, that's going to be needed as we deploy agents in government. We could potentially have unprecedented insight into how government actually works — set KPIs for certain outcomes and have models basically fit the curve to those outcomes. What I worry about in particular, in terms of concentration of power, is the relative power of the executive branch vis-à-vis Congress and the judiciary. One solution to that could be something like an Inspector General GPT. Once you have most of government running through agents with full audit trails and chain-of-thought, you could potentially have a pipeline to an automated oversight system that generates reports directly to Congress on demand — and would be difficult to turn off, because you would notice. So I came away somewhat hopeful. There's a version of the old AI-bias discussion that's very model-focused — garbage in, garbage out. But there's also just the fact that humans are super biased. And if you can replace a human with something you can benchmark as provably more neutral, you could end up using agentic AI as a force multiplier for rule of law and consistency in government.

   1:33:22

   Prakash: The question I had on that: isn't the efficiency created by using AI in government actually already a form of loss of control? To put it more concretely — policymakers are sometimes given the task of squaring the circle. Take the Biden administration's $40 billion rural broadband plan: they could have used SpaceX but forced it down to wired broadband in all these communities. Part of that policy framework was job creation and handouts to partners — it was whitewashed under the banner of rural broadband. If you give government an AI-powered Inspector General that enforces things and makes that legible, that reduces the space of action available to policymakers to square those circles. So in that sense, a more efficient AI government is also a loss of power — a loss of control.

   1:35:11

   Samuel Hammond: Yeah — if Robin Hanson were here he'd say rural broadband policy is not about rural broadband. I think there's probably some truth to that. There's lots of wiggle room and certain ulterior motives in how policy gets created, and if you make that too legible you could squeeze some of that out. I'm not as concerned about that, though. I'm more concerned about government simply not keeping up and being displaced by much faster-moving systems in the private sector. There are two equilibria here: one where government goes the way of Blockbuster and gets completely swamped by Netflix — becomes a rump state centered around the nuclear arsenal — and another where government gets its act together, deploys models quickly, and ends up as an agentic state that could become a turnkey dictatorship where you no longer have the ability for generals to resign in protest or for people in the bureaucracy to drag their feet and enforce broader professional norms in public administration. The balancing act we need to be thinking about is: yes, you may squeeze out some behind-the-scenes horse-trading, but the knife edge we're walking is one where either government ceases to have meaningful traction over how things are actually governed — or one where the person at the top has essentially infinite traction because all agents can move in synchrony.

   1:37:07

   Nathan Labenz: Are there specific KPIs — or loss functions — that you think would optimize government in ways that stand the test of time? Everyone knows the Goodhart's Law problem. And in some ways the whole loss-of-control worry, writ large, is: we optimize for all these things and it turns out it's not really what we want. As Eliezer puts it, the AI does too much. How confident are you in our ability to define — in concrete terms we could write down and hand off to an AI — objectives that would actually serve us well all the way until, I don't know, the singularity?

   1:38:06

   Samuel Hammond: It really depends, case by case. I think it's a fool's errand to try to define some singular objective — like GDP or the Human Development Index — and say 'optimize for that.' In fact, I wouldn't want to hand a superintelligence that metric, because there are lots of ways to increase GDP that don't involve humans. Instead, I think of this more on a case-by-case basis — moving toward more benchmark- and performance-based outcome measurements for the kinds of regulatory policies we want to achieve. A good example might be aviation air-safety regulation. Right now, if you want to build a plane, every single part has to be certified — part-specific certification. What that does is cut off the ability to experiment with novel designs: you want to replace rivets with some high-tech adhesive, you can't. But what we really care about is that the airplane flies. So: could you build a foundation model for evaluating airworthiness that's agnostic to how you get there — CAD design in, air-certification out? And using that as a template for all kinds of regulatory policy, up to and including monetary policy. I was most recently chatting with Scott Sumner about this.

   1:39:37

   Samuel Hammond: In monetary policy you've had the Taylor Rule — a very simplistic algorithm for raising or lowering interest rates relative to the natural rate. There's been lots of discussion over decades about how to make monetary policy more rules-based. But could we move into a world where we have a massive everything-in-the-kitchen-sink model — all economic time series, social science data, geospatial satellite data — and you can basically query it: 'What do we do to maintain full employment and price stability?' And get something vastly more reliable than the dynamic stochastic general equilibrium models the Fed uses today. So I think there's just tons of low-hanging fruit before we even get to the world of total handoff. That's actually the world I think we should reach last — if anything needs a human in the loop, it should be the head of state, at least so things don't completely run away from us.

   1:40:46

   Nathan Labenz: Did you see the clip from Jack Clark that just came out — he's being interviewed and the discussion turns to how difficult it might be to coordinate with China, and then there's a surprising moment of agreement where he agrees that, actually, the state has a vested interest in long-term stability and coherence, and so it might be easier to work with other states than to have companies work together. What was your surprise level at that? And as you watch the technology develop and try to look for possible target equilibria — are you becoming more bullish on our ability to establish the equilibria we need, or less optimistic over time?

   1:41:42

   Samuel Hammond: I think he's basically right. Corporate entities have a fiduciary duty and a metanorm that forces them to compete with their competitors. That's not really true of nation-states. We talk about nation-states as being in competition, but Paul Krugman has his famous essay, 'The Myth of Competitiveness' — countries are not actually firms. We don't rise when our competitor falls. So in addition to having a longer-horizon view of the world, and insurer-of-last-resort obligations to keep everyone from revolting, governments are also potentially much more aligned in terms of discovering positive-sum, Pareto-improving kinds of agreements — in ways that the companies can't even do for antitrust reasons. At the same time, a lot of Jack's views on this are probably colored by being at Anthropic, whose primary competitor is OpenAI, and by the interpersonal and historical reasons they've had difficulty collaborating.

   1:42:57

   Nathan Labenz: Yeah — the plot is such a painful one to keep being reminded of. Thank you for joining us. I would have loved to give you a chance to talk about your views on why we should take AI consciousness seriously as a possibility, but we're out of time.

   1:43:16

   Samuel Hammond: Those are evergreen topics — I'll be back.

   1:43:18

   Nathan Labenz: Yeah. We'll refer people to your recent write-up at Second Best, the Substack. I appreciate you for entering into that discourse — it would be easy to neglect and easy to write off, and I'm still very agnostic as to what is going on there. But as much as we have full plates with daily breaking news items, we should not get too myopic and fail to consider these really big-picture existential-stakes questions — not only for our own benefit, but potentially for the benefit of the minds we're creating. That is a very valuable and noble contribution, and I appreciate you for finding the time to zoom out and consider it.

   1:44:15

   Samuel Hammond: Great to be on.

   1:44:17

   Nathan Labenz: Good to see you. Have a great day — we'll be in touch.

   1:44:21

   Prakash: Later, Sam.

4. 1:44:32Interview39 min

   Supply Chains as the AI Reality Check — Matt McKinneyMMMatt McKinneyMatt McKinney is co-founder and CEO of Loop, which ingests data from 38 sources across the supply-chain network-of-networks. Fresh off a $95M Series C led by Valor, he gave a practitioner's read: 60% of supply-chain data is offline, DUX (Loop's domain foundation model) processes 2 billion data points daily, and the long-run potential is 5–10% efficiency gains on an $11 trillion U.S. logistics cost base.

   Watch

   ▶Play this segmentYouTube · from 1:44:32

   As aired

   Matt McKinney, co-founder and CEO of Loop, opened by explaining what drove the company's $95 million Series C: proving 10x ROI — not 2x — for marquee customers like Tyson Foods, Nvidia, and Dell Technologies, and recruiting top AI talent to compound that value. He laid out the fundamental thesis: 60% of supply-chain data is offline — trapped in EDI files, emails, PDFs, and physical paper, with 4 billion paper documents governing global trade every day. Loop's platform ingests data from 38 sources across this network-of-networks, then progresses through three layers: organizing the dark data, automating manual workflows with agents (invoice adjudication, general-ledger coding, carrier payments), and finally surfacing intelligence that enables better operational decisions — what Matt called 'the Maslow hierarchy of data needs' culminating in self-actualization for an organization.

   The conversation turned to macro pressures and the defensibility question. Matt argued that post-globalization forces — nearshoring, tariff shocks, geopolitical disruption — are happening annually rather than once a decade, creating structural inflation that only AI-driven efficiency can offset. He drew a sharp line between manufacturing companies (which AI will transform but not disrupt, because physical goods are a durable moat) and services companies (which AI-native entrants will disrupt entirely). He placed Loop in the emerging 'AI services' category: $5.5 trillion in U.S. business services currently delivered entirely by human labor, now shifting from labor to tokens. He identified two reasons frontier foundation models won't easily displace Loop: services data is largely non-verifiable (unlike code), and the domain data is offline, proprietary, and cannot be scraped from the internet — what he called 'dark data' that only Loop is accumulating at scale, currently processing 2 billion data points per day.

   Nathan pressed into the technical architecture of DUX, Loop's supply-chain foundation model. Matt described a multi-vector model (positional, visual, and language vectors) with unique attention patterns, but flagged domain-driven design — the codification of logistics-specific language into the model itself — as the key architectural innovation that drives automation rates. He also described the AIOps function: a team of AI-native early-career engineers whose sole job is to increase the automation quotient by grouping exceptions, analyzing the Pareto of failures, and feeding corrective actions back through an internal worker agent called Mac, creating a continuous RL loop. The segment closed on a characteristically grounded note: at the limit, AI can take 5–10% out of the $11 trillion U.S. logistics cost base — roughly $1.1 trillion — a 'highly deflationary force in the global economy.'

   Key moments

   Sixty percent of supply-chain data is offline — it's dark data. No one has documentation for how decisions are being made. Every single day there are four billion pieces of paper governing global trade.

   The limiting factor for AI in enterprise is not technology. It's change management.

   Labor will be commoditized as we know it. Talent never will be.

   Questions asked

   1:45:58What milestones drove the $95 million Series C, and what does the investor base tell us about Loop's thesis?

   Investors tracked customer value above all else: getting marquee customers like Tyson Foods, Nvidia, and Dell onto the platform and proving 10x ROI (not 2x), strong net dollar retention, and the ability to recruit top AI talent. The investor base — Valor, Founders Fund, Index, 8VC — reflects confidence in the physical-economy data thesis. Valor, with its SpaceX and xAI exposure, was particularly attracted to the idea that the data Loop accumulates in the physical economy will be among the most valuable data assets over the next several decades.

   1:48:51How does Loop actually get the dark data in, and why is the data-ingestion problem so hard?

   Loop ingests from roughly 38 sources: EDI, API feeds, email, scanned paper documents, and even phone calls. The difficulty is structural: supply chain is a network of networks, so even a large customer like Tyson Foods can only optimize within its own walls — not within its suppliers' or customers' walls. Every provider in that ecosystem has different processes, standards, and systems, creating combinatorial complexity in data formats and structures. No one else wanted to play in this messy space; Loop chose to, and considers winning there its foundational competitive advantage.

   2:04:16Why won't frontier foundation models just displace Loop?

   Matt identified two structural protections. First, services data is largely non-verifiable: coding models work because you can run the code and confirm correctness, but supply-chain services data lacks that property, so general-purpose models can't self-improve the way they do in verifiable domains. Second, the domain data is offline, proprietary, and not scrapeable from the internet — it cannot be fed into a GitHub repository or a standard training pipeline. Loop is accumulating two billion proprietary data points per day, building a compounding domain advantage. The company uses frontier models as components but wraps them in DUX, its own supply-chain foundation model trained from scratch.

   2:17:09What is the 'AIOps' function and how does it keep the models improving over time?

   AIOps is a team of AI-native early-career engineers whose sole job is increasing the automation quotient. When exceptions or fallout from automation occur, agents group them by type, analyze the Pareto of failures, and an internal worker agent called Mac applies corrective action so those failure modes don't recur. This creates a continuous reinforcement-learning feedback loop into DUX and the surrounding system. AIOps is distinct from the forward-deployed team that defines new tasks at the frontier; AIOps optimizes the automation of already-defined tasks across roughly 28 active task types.

   2:22:52At the limit, how much cost can AI actually remove from the logistics industry?

   Matt estimated five to ten percent of total logistics costs at the limit — which in an $11 trillion U.S. industry translates to roughly $1.1 trillion in potential savings. He characterized that as 'a highly deflationary force in the global economy,' particularly significant in a post-globalization era where nearshoring and geopolitical disruptions are creating structural inflation. He noted the estimate covers terrestrial supply chains only.

   Related

   Loop ↗Matt McKinney on X ↗

   Full transcriptLightly edited · timestamps jump to YouTube

   1:44:32

   Prakash: Alright — Matt McKinney is next.

   1:44:35

   Nathan Labenz: Let me introduce Matt McKinney. He's the CEO and co-founder of Loop, a startup using AI to clean up and automate the data that runs freight and supply chains. He helped launch Uber Freight, where he saw how little things like errors in invoices snowball into big costs. Loop just raised a $95 million round to scale, positioning itself as the intelligence layer for the physical economy. Lately he's been sounding the alarm that AI is hitting a wall in industries like logistics because so much critical information lives outside tidy databases — in emails, PDFs, even paper. The key mental model going in: AI is only as good as the real-world data and processes you feed into it. His perspective is going to help us understand how AI is going to penetrate these very physical, very manual, very paper-driven industries. Matt, welcome to the show.

   1:45:34

   Matt McKinney: It's great to be on. Thanks for having me.

   1:45:36

   Prakash: Let's talk about your recent fundraise. Normally when you raise a round, you have milestones you need to hit before and milestones you need to hit after, in about 12 to 18 months. What did you achieve, and what are you targeting next?

   1:45:58

   Matt McKinney: The big milestones our investors are looking for all trace back to customer value. I say that because if you get customer value creation right, that's the leading indicator for every financial and operating measure. For us it was about getting some of the largest companies in the world — Tyson Foods, Nvidia, Dell Technologies — onto our platform, proving that we can create not 2x ROI but 10x ROI for these clients, proving they will pay us not just once but repeatedly and expand — our net dollar retention growth. And the ability to recruit top AI talent to serve that mission. Customer value and talent density: those were the two key drivers behind raising new capital.

   1:46:43

   Matt McKinney: When we look at the investor base, we got some of the best — 8VC, Founders Fund, Index, you name it. Valor came on board, and given their exposure to xAI via SpaceX, they saw a lot of opportunity. The data that Loop is acquiring in the actual physical economy is going to be some of the most valuable data over the next multiple decades, because 60% of supply-chain data is offline — it's dark data. No one has documentation for how decisions are being made. No one has captured what's on a PDF or a physical piece of paper. Every single day there are four billion pieces of paper governing global trade — bill of ladings, proof of deliveries — and that is a massive data challenge. Attracting Valor, given their exposure to one of the most consequential AI companies of our lifetimes, was a testament to our customer value and talent density.

   1:48:31

   Prakash: Concretely — how do you get the data in? Are you scanning PDFs? Watching what people do on their computers? Transcribing phone calls? What are the actual ingestion mechanisms?

   1:48:51

   Matt McKinney: We get data from roughly 38 different sources — some third-party, some first-party. Think of supply chain as a network of networks. Even an organization like Tyson Foods can optimize within its own walls, but not within its suppliers' walls or its customers' walls. You need to acquire data across the whole network. So we ingest EDI, API feeds, email, and literal physical paper that we scan in. You can even get data via phone call. The complexity arises because every single provider in that ecosystem has a different process, a different standard, and a different system. That combinatorial complexity of different data structures is why the supply-chain data challenge is one of the most important problems to solve — and supply chain is an $11 trillion industry just within the U.S., one of the largest contributors to global GDP.

   1:50:35

   Prakash: So what kinds of decisions can your clients make with that data? Can Apple decide to divert shipment A from aircraft B to aircraft C? What actually drives the value?

   1:50:59

   Matt McKinney: Let me back up one step. There are three layers: organizing data, automating operations, and then improving intelligence and decisions. It's important to start all the way at data mapping, because that's the fundamental problem no one else has solved — and it's the hardest problem in the industry. At Loop we like to play where others don't. No one wanted to play in the data problem because it was so messy. We played there and we won that.

   Automating operations means handling workflows you don't want people doing — adjudicating an invoice, coding to a general ledger, remitting a payment to a carrier. These are very manual, labor-intensive tasks that agents do better: higher quality, faster, and much cheaper. The token cost is a lot cheaper than the labor cost. What some of our clients call 'turning their team from processors into analysts' lets them move to the highest level — what I'd call the Maslow hierarchy of data needs — all the way up to self-actualization as an organization.

   At that intelligence layer, you know things you didn't know before. A concrete example: a manufacturer has a plant manager clicking a button every Tuesday to overnight in supplies. When you ask why, he says he was told four years ago to do it. But with your data mapped, you can see he's overnighting supplies when he already has a year's worth of inventory on hand — and paying $15,000 for that overnight shipment every week. By having your data mapped and your workflows automated, you identify those massive inefficiencies. And that efficiency passes through to consumers as lower prices.

   1:54:03

   Nathan Labenz: Is there a number — at the limit — for how much of the cost is physical necessity versus inefficiency sitting on top? How different would the world look with your ultimate success? Is it lower prices, faster delivery, higher reliability? And separately: COVID taught us just-in-time supply chains are brittle. How much do you advise companies to invest in robustness as they realize these savings — because the right answer probably isn't maximum efficiency all the way down?

   1:55:33

   Matt McKinney: Big picture: if you look back to Bretton Woods, that's when legacy trade lines were drawn and globalization took off. Multinationals built supply chains in a highly globalized fashion — and then geopolitical events and pandemics exposed the risk of that. Optimizing for every unit of margin was not the right strategy if it could blow up your ability to serve customers or create national security issues.

   Over the last four to five years we've entered a post-globalized economy — nearshoring, reshaping of global trade lanes. And I love the Jeff Bezos principle: people ask what the world will look like in ten years, but what I can tell you is what won't change. Consumers will always want better service and lower prices. Companies will always want better service and lower prices.

   In a post-globalized world, every geopolitical shock and energy spike — events that once were once-in-a-century, maybe once-in-a-decade — are happening annually. That is inflationary for products manufactured and bought in the U.S. To combat that inflation, you have to use AI in the actual operations to drive throughput at lower cost. But at the same time, customer expectations are rising because of AI and because of Amazon. Companies are caught in the middle — inflationary forces on one side, rising customer expectations on the other. Companies deploying AI into those enterprises are making them better on cost and better on service simultaneously. Jeff Bezos wrote that principle and it's never going to change.

   1:58:50

   Prakash: One of the things we talk about on the show is how fast AI moves while other industries do not. To what extent will your customers be willing and able to adopt AI quickly versus AI-native startups that begin with all AI and no legacy systems to address? Is there a big advantage for the AI-native entrant, or are incumbents increasing their speed of adoption?

   1:59:49

   Matt McKinney: The limiting factor for AI in enterprise is not technology. It's change management. That will be the case in the Global 2000. There are certainly Global 2000 companies making swift changes — great leadership, prioritizing it top-down — but culture is one of the slowest movers. If you don't have a culture of trying new things, it doesn't matter what the top is doing. It's still going to take a long time to propagate.

   In terms of which companies will win — AI-native companies without legacy systems, or incumbents with greater distribution — it depends on the industry. If I categorize into two: manufacturing versus services. Manufacturing companies are much more defensible. They'll be transformed by AI but not disrupted by it. But legacy pre-AI services companies will be completely disrupted. AI-native services will be so much better, faster, and cheaper — times ten — that it poses an existential threat to those service industries.

   2:01:30

   Prakash: So would that mean third-party logistics players who own physical assets — trucks, ships — will be fine, while the pure intermediaries who just tender and lease space will not? Is that the separation?

   2:01:55

   Matt McKinney: I think what you're describing is asset-light versus asset-heavy — but both are still providing a service to a manufacturing company. It's not a blanket statement. A lot of 3PLs will use their distribution advantage to win, and they'll use AI to transform their operations — a CH Robinson, for example. I view asset-light and asset-heavy as both service providers. The distinction I'd draw is: if someone is manufacturing physical goods, like Apple, they're going to be protected.

   2:02:57

   Nathan Labenz: How do you see Loop's own defensibility playing out long-term? If I had to place you, I'd put you on the services side — I didn't hear anything about physical goods. There's a lot of anxiety about building on top of foundation models. At what point does a frontier model just walk into a new environment, spin up its own app, and deliver most of the value at a 30-to-1 token cost advantage for the end customer? Do you have sufficiently differentiated data that you're training domain-specific foundation models that a general-purpose reasoner can't replicate? Or maybe you become the Cursor of logistics — the way a big tech platform enters the space because you already have the customers. How do you see the long-term future — and long-term might not be very long in calendar years?

   2:04:16

   Matt McKinney: Big picture: roughly $5.5 trillion in U.S. business services is 100% delivered by human labor today. Over the next decade, that service delivery is going to shift from labor to tokens. Companies in this new category — what I'd call AI services — are the ones converting labor into tokens. I'd put Loop at the frontier of that category. Those companies have a real advantage because they're built AI-natively, they respond to customer needs in a new way, and they deliver value not with humans who have process errors but with tokens that are much more reliable.

   The way AI-native services companies will be protected from the frontier foundation models is twofold. First, services data is largely non-verifiable. Coding models work because you can verify the output — run the code and see if it's correct. Services data doesn't have that property, especially in supply chain. Second, this is not internet data. You can't hook up a GitHub repository to this. The domain language and data is offline, proprietary, protected. So for those two reasons alone, there's a significant moat from the foundation models.

   2:05:46

   Matt McKinney: We use a lot of frontier models ourselves, and we have our own DUX — our proprietary supply-chain foundation model, trained from scratch on supply-chain data. We're processing literally two billion data points a day. The amount of domain data we're accumulating is significant. What you'll see is companies like us in various verticals with similar characteristics — a huge data advantage, a domain knowledge advantage, training their own models for very industry-specific tasks. That trend will probably persist for the next five to ten years, all in service of transforming labor into tokens.

   2:07:09

   Prakash: We often talk on the show about 'relinquishment' — the first time you use a chatbot to write a recommendation letter, you realize you'll never do it by hand again. Can you give us a concrete customer story: someone used Loop, and said 'I will never do this manually again'?

   2:07:46

   Matt McKinney: If you know supply chain, you know that the exception is the rule. We built an exception agent that handles all the coordination overhead — the back-and-forth that happens whenever exceptions occur. The adjudication of an exception between multiple parties can become a degenerate loop like no other. Our exception agent processes all of this automatically, reasoning across our customers' data.

   The first customer that saw it said, 'This is magic.' They were spending literally 20 to 30 people's time per week processing exceptions. It was delaying their working capital, delaying payments, delaying product delivery. By resolving exceptions in minutes instead of weeks, you get time back, you run your business better, you serve customers better, and you lower your cost. That's our 'wow moment' for almost every customer.

   2:08:57

   Nathan Labenz: Zooming out — we were just talking with Sam Hammond about post-AGI civilizational equilibria. Taking all this cost and time out means, frankly, taking humans out. I think a lot of organizations tell stories about people moving to 'higher-level work,' but unless you're positing a hundred-times increase in the volume of things being shipped, a hundred-times efficiency improvement has to translate to lower headcount. Do people see it coming? Are you willing to say yes, that's where we're going? And as this plays out across many industries simultaneously, what should society do about it?

   2:10:17

   Matt McKinney: I think about this a lot. Throughout civilization, the arc of technology has always been a feature of abundance. The question is: is this time different? I think it might be, largely because the pace of change and disruption is so fast. If the pace of disruption is faster than the rate of labor retooling, you're going to have large problems. If the pace of disruption is greater than the rate of retooling, you're going to need policy intervention to prevent civil unrest. It could even lead to the beginning of a new form of government — not the end of democracy, but potentially the end of government as we know it. Technology throughout the millennia has been a force of change: feudalism ended when people could suddenly travel. You can extrapolate that history.

   Two things need to be true: one, the rate of retooling has to accelerate, and I don't think we're doing nearly a good enough job on that today. Two, the abundance that AI creates can't be concentrated in a handful of individuals or firms. It doesn't need to be uniformly distributed, but the abundance has to exist across the ecosystem.

   2:12:34

   Nathan Labenz: Do you have specific advice for people whose tasks are clearly starting to be automated — where the writing is on the wall for that role — advice you actually believe will work if they follow it?

   2:12:54

   Matt McKinney: There are two big differences: labor will be commoditized, as we know it. Talent never will be. At Loop we have incredibly smart people — Stanford, University of Chicago — who start on the ground floor doing manual tasks, training the AI, building out our RLHF systems. Their specific task will eventually be automated. But they're constantly learning and figuring out where the next incremental unit of value is — outrunning the tokens. That's talent, not labor. At Loop we don't have an issue with this because we're in the world of abundance and high talent density: individuals are constantly seeking the next higher-level, more cognitively demanding task.

   2:13:59

   Prakash: How has your hiring flow changed in the last two to three years as AI has gotten this good and people are uncertain about what roles even exist?

   2:14:29

   Matt McKinney: Hiring has always been the most important decision we make at Loop. I've personally interviewed every single person who's joined the company — which doesn't scale forever, but it establishes the culture. Who becomes a 'Loopster'? The practices are around intellectual honesty, values assessment, evaluating how the person will perform in the specific domain they're expected to operate in. That hasn't really changed — if anything, it's only been magnified in the world we live in today.

   2:15:23

   Nathan Labenz: I know we're running a little late — do you have time for one more in-the-weeds question?

   2:15:28

   Matt McKinney: Let's do it.

   2:15:30

   Nathan Labenz: You mentioned training a from-scratch foundation model for logistics. I'd love to hear more about the architecture. Two analogies: Stripe built a payments foundation model and exposes embeddings of each transaction to internal app developers, so teams can iterate on representations very quickly without waiting for full retraining cycles. Criteo, an ad-tech company, has milliseconds to match an ad to a user, so they precompute representations and expose them to the final decision layer. Does that embedding-exposure pattern map onto how you've structured DUX? Or is there a different architecture I should understand?

   2:17:09

   Matt McKinney: Yes to a lot of that. But one key addition: we have a function called AIOps, whose sole focus is increasing the automation quotient. They take any exceptions or fallout from the automation, group them using agents, look at the Pareto of failures, and then our internal worker agent — Mac — takes corrective action so those issues don't recur. AIOps is really an RL input to the entire system. So you've got not just a singular foundation model, DUX, but DUX interoperating with other frontier models, and AIOps feeding the learning loop. That's the secret sauce that lets us continuously push automation rates and quality across about 28 different tasks today. The AIOps team are smart, earlier-career people who are AI-native in their thinking — not biased by the ways of the past — and they sit at the intersection of converting labor-heavy tasks into token-centric tasks.

   2:18:49

   Nathan Labenz: Do those AIOps folks go forward-deployed to customers — sitting with them, collecting feedback on exceptions — the way we've heard about from OpenAI's deployed engineers?

   2:19:05

   Matt McKinney: We have some of that, but it's separate from AIOps. The forward-deployed function is about defining new tasks at the frontier. AIOps is about optimizing the automation of already-defined tasks.

   2:19:22

   Nathan Labenz: One more bit on the foundation model: it doesn't sound to me like a next-token predictor. I'm imagining something trained on a sparse vector of predefined dimensions of supply-chain state — the state of a parcel, a shipment, a lane. How far afield from a standard language model is DUX in terms of its fundamental input and output dynamics?

   2:20:06

   Matt McKinney: The three big vectors are positional, visual, and language — and we focus attention in unique ways. The unique thing about supply chains is they run on their own language. The codification of that domain — what we call domain-driven design — is the single most consequential driver of our automation rates and model efficacy. It's all about how we organize and structure the data, and we use our foundation models to do that organization and structuring. That's the architectural innovation, applied in a very unique way to a specific domain.

   2:21:00

   Nathan Labenz: Cool. Thanks for getting into the weeds with me — I appreciate it.

   2:21:03

   Matt McKinney: I'm a former data scientist, so I love going deep on AI. It feels like after 20 years in the field, you've finally been recognized. All of a sudden AI is everywhere and everyone's asking these questions.

   2:21:21

   Nathan Labenz: I sometimes feel like I'm in some sort of simulation operating for entertainment purposes — or who knows what its macro purpose is. I feel uncomfortably close to the center of the action, and definitely to my surprise. Prakash, any other questions? Matt, anything you'd like to leave people with before we break?

   2:21:44

   Matt McKinney: The big thing: AI is going to have the most impact in the industries that drive the highest contribution to GDP — and those are typically the messy and ugly ones, not the sexy ones. Fall in love with that problem. If you love AI systems and AI deployments, there's no better industry than supply chain. But there are other industries with similar characteristics. Get excited, dive deep, and understand not just the technology but the domain itself.

   2:22:17

   Prakash: To spitball and get a sense of scale — part of Elon's 'idiot index' for a product is not just the atoms, but the cost of moving the atoms across oceans and continents, which is really the implied logistics cost in the product. How much do you think AI can actually close the gap between what logistics costs today and the bare energy cost of moving those products?

   2:22:52

   Matt McKinney: At the limit, I think you're looking at five to ten percent.

   2:22:55

   Prakash: Oh, wow.

   2:22:56

   Matt McKinney: Which is material in an $11 trillion industry. If we can take out $1.1 trillion, that's a highly deflationary force in the global economy. And that's just terrestrial supply chains — until we get to intergalactic supply chains.

   2:23:11

   Prakash: Indeed. A trillion here, a trillion there.

   2:23:13

   Nathan Labenz: Pretty soon we're talking real money in the AI economy. Matt McKinney — new unit: one Elon. Matt McKinney, founder and CEO of Loop. Thanks for being with us on AI in the AM.

   2:23:29

   Matt McKinney: Enjoyed it. Thanks for having me, guys.

5. 2:23:36Segment34 min

   ClosingThe hosts debated whether vertical AI companies like Loop will survive as standalone businesses or be absorbed into frontier labs, then ranged across AI geopolitics: the export-control reversal's signal for allied sovereign AI, the puzzling Chinese decision not to stockpile NVIDIA chips, and implied RSI timelines by player — Anthropic/OpenAI targeting 2028–2029, Elon building toward 2029–2032, China seemingly planning for the early 2030s.

   Watch

   ▶Play this segmentYouTube · from 2:23:36

   As aired

   The closing segment opens with Nathan and Prakash debating whether vertical AI companies—specifically Loop, the logistics AI guest from earlier—will endure as standalone businesses or be absorbed by frontier labs. Nathan wonders whether the long tail of truly exotic supply-chain exceptions actually favors a broad general reasoner over a hyper-specialized domain model, while Prakash argues that domain models are essentially precomputed heuristic databases that frontier labs will eventually want to acquire directly. This leads into a wide-ranging tour of AI geopolitics: the export-control reversal and what it signals for allied nations suddenly seeking sovereign AI, the puzzling Chinese government decision not to stockpile NVIDIA chips, DeepSeek's unusual funding structure, and each major player's implied RSI timeline—Anthropic and OpenAI targeting 2028–2029, Elon building toward post-2029, China seemingly planning for the early 2030s. Nathan closes by previewing the next day's guests—Karina Hong of Axiom Math (mathematical superintelligence) and Sam Posupalak of Skyfall AI (enterprise superintelligence)—and reflects on the uncomfortable duty of interrogating whether any of today's AI businesses will survive the RSI run-up. Prakash reframes the discomfort as a rare historical opportunity: capturing on the record what the participants of this moment actually believed while it was happening.

   Key moments

   Just how weird do you expect the future to be? Because the more systemic disruption and shifting of alliances that comes our way, the harder it's going to be to maintain a really dialed-in foundation model for a particular space.

   If we're entering this posthuman era, what do the people who had something to say about it say about it at that moment? That is what strikes me as important for the historical record.

   I feel like Anthropic and OpenAI are in the 2028–2029 timeframe. Elon is preparing for 2029–2032. And I feel like the Chinese government is in the 2032-and-onwards space.

   Full transcriptLightly edited · timestamps jump to YouTube

   2:23:37

   Nathan Labenz: That's a really interesting recurring segment we can come back to—where is the new foundation model in order? It's pretty clear it's in order for payments.

   2:23:56

   Prakash: Yep.

   2:23:57

   Nathan Labenz: No doubt about that. The results Stripe posts are pretty incredible, and you certainly feel like you're a long way from getting that out of even Claude Fable—just saying, "go make sense of this transaction." It's a little harder for me to squint at and see in supply chain. That feels more amenable to just really good reasoning.

   2:24:28

   Prakash: But I think the subtext is that these fields have a particular language—payments, for example—which is not just next-token prediction. And there are also things hidden in payments fraud; it's an adversarial environment. In logistics, the delays are all unexpected events, all exceptions. People only call you when exceptions happen, and all your time is spent exception-handling. I almost see it as these fields speaking a different language—the language of transactions and fraud and adversarial systems—which is not what you get from next-token prediction. Next-token prediction is the most common token to come out; here you're looking at the least common tokens, the ones you always have to work on.

   2:25:37

   Nathan Labenz: I do wonder—with these exceptions, how many are truly exotic and how many are routine exceptions? In the history of human-powered processes for these things, it seems like the vast majority have to be something you can encode. The guest talked about 28 tasks. An interesting question we didn't get to ask: how many of all exceptions do those 28 tasks represent? It feels like there's a pretty substantial power law, where the vast majority of exceptions are still ones where you think, "I know exactly what to do—I've seen this a million times. The thing didn't make the boat, the thing didn't make the plane. It's an exception, but I know the playbook."

   2:26:35

   Prakash: On one hand, that's absolutely true—a delay is a delay; it didn't make the plane, it didn't make the boat. On the other hand, the post-exception handling could be like: Trump tweets new tariffs, and all of a sudden those tariffs only affect a certain kind of toy coming in from China. Only certain ships carry that toy, your customer is on one of them, some of those toys have high enough value that the customer will pay the additional tax and some won't. What do we do with them? And even if the customer doesn't accept and doesn't pay the tax, you still have this container of toys you need to do something with. Then you have to ask: is this a force majeure? Is this covered by insurance? Will this particular insurer cover it? And then you have all these interacting commercial counterparties with different contracts and different rules. Sometimes on paper it looks like the same contract, but in practice it's different—your team might tell you this insurer always pays and that one always rejects the claim. You end up with tribal knowledge that your people on the ground build up through fire, through thousands of shipments. The exceptions may recur three or four times, but not often enough to build software for that particular case. The number of different kinds of exception cases is very large—even though at the end it always comes down to two questions: where's the delay, and who's going to pay for it?

   2:29:05

   Nathan Labenz: Doesn't all that push you toward thinking the general reasoner is the ultimate backstop? If I'm trying to assess whether Loop will be a standalone business in three years or get acquired by Anthropic—where does value really accrue?—I don't feel like I have clarity. But all these long-tail situations feel like they'd need a more general reasoner customized to the domain versus a pure-play model trained from scratch only on shipping data. I can't imagine a domain-only model handling Trump getting out of distribution. Only the broadest general reasoners with first-principles understanding can really cope with that kind of lateral knowledge—contracts, force majeure, different jurisdictions. Maybe it's two parts: an extremely efficient Stripe-like foundation model doing relentless local optimization, backstopped by a really general reasoner that can handle it when the game changes materially and the hyper-optimized thing can't get that far out of distribution. Which also makes an interesting analysis for businesses like this: just how weird do you expect the future to be? The more systemic disruption and shifting of alliances we see, the harder it's going to be to maintain a really dialed-in foundation model for a particular space.

   2:31:44

   Prakash: Perhaps it's just precaching of solutions. Perhaps what the guest's company is doing is really collecting data that builds a model encoding all the heuristics people apply—a foundation model that has already done the precomputation a general reasoner could do in real time given the full corpus. And so the general reasoner could use this as almost a database of pre-structured rules to follow. That leads to one of the frontier labs buying them—just absorbing the capability directly into the frontier model.

   2:32:36

   Nathan Labenz: Yeah. It does mean logistics joins chips, bio, coding, and robotics as another domain the big players might eventually want to get into directly. The big-tech singularity—I may go back and revisit Andrew Critch's post on this. His definition is when the big tech companies can, if they choose to and aren't prevented by government, enter and quickly come to dominate any legacy industry by virtue of superior intelligence.

   2:33:20

   Prakash: I made a prediction back in April that what we'd see post-IPO from OpenAI, SpaceX, and Anthropic is employees leaving to found companies they know need to be founded and know will be acquired—because these are on the product roadmaps of the frontier firms, but leadership isn't focused on that particular aspect yet. So there are opportunities for employees to look three or four years ahead, go out, found a firm, build it up, and resell it back into these companies. I call this the capital explosion. And that's why the question is: are existing firms going to adopt quickly enough, or will you have these AI-native firms—founded by or funded by employees who've made money in the AI cycle—take over these industries? It wouldn't look like the big tech companies themselves; it would look like a diaspora of former employees doing this and then being acquired back in.

   2:34:51

   Nathan Labenz: Those boomerangs are going to have to turn around and come back pretty quickly for all that to make sense, but I could certainly see some of it happening—especially if you can self-fund and don't have to raise. Though there's an interesting differential in expectation. Anthropic is obviously the most AGI- and RSI-pilled of any organization. My prediction would be we won't see many people from Anthropic do that, because it seems like they believe there simply isn't time—the only thing that matters is getting into the RSI bootstrap in the best possible way. Every Anthropic person I've talked to, I can't think of one who'd say, "Yeah, I think now is a good time to go start a business and try to get acquired back in 18 months." Even 18 months—they'd say that's going to be a very different world, and Claude could probably just do their whole thing by then.

   2:36:24

   Nathan Labenz: Where does OpenAI sit on that spectrum—how deep does the conviction run that this is really happening and is the only thing that matters? I'm much less clear on that. We don't have the same liquidity moment coming out of a Google, but we've certainly seen a lot of people do that over time there. xAI is a total wild card—I'm sure we'll see lots of surprises from them before it's all said and done, and I really have no idea what to predict on that dimension.

   2:37:10

   Prakash: Periodic Labs—William Fedus, I think—is a clear example to me of something OpenAI was not set up to do: experimentation on material sciences. It made sense for him to leave and found something. All these firms will eventually get there, but it's far enough out that it's still amenable to experimental discovery. And DeepMind—Demis has always felt that you're going to need real-world physical experiments, that you're not going to get to full AGI or RSI without that kind of work. That's a very different perspective from Anthropic, which believes you get to RSI and RSI will solve everything else.

   2:38:07

   Nathan Labenz: Well, we can continue to monitor that situation, as we say these days. Anything else we want to touch on before we leave it for today? We did touch briefly on the Jack Clark companies-versus-states piece.

   2:38:27

   Prakash: Yeah.

   2:38:28

   Nathan Labenz: My reflection on that—and there may be more there—and we talked plenty about Fable. It's really frustrating how I feel both that there's just such gravity toward closely watching these few companies and their interactions with government, and at the same time events are defying analysis because they seem fundamentally chaotic and idiosyncratic in their provenance. There's not really a lot to analyze in some of these situations. As much as the whole export-control reversal does seem like something we'll look back on as up there with the OpenAI Sam Altman firing moment—I don't know exactly where you'd put it—my guess is if we look back at a three- or four-year horizon, it will be one of the big moments everyone remembers as a real shift. And yet at the same time, it's also just... strange.

   2:40:11

   Prakash: Let me give you a parallel. Cohere, the Canadian AI lab, said that over the last weekend they saw off-the-charts inbound from customers—a huge upswell. Mistral has also seen a lot of that. So we have the first real breakaway from the Western allies, who are starting to realize they will not be part of this future if they don't create it themselves. Tyler Cowen was very pointed about it: the French and Germans are welcome to try to make AI labs, but are they going to spend $200 billion like Mark Zuckerberg, or spend money and fail like Elon has—and Elon is certainly a very capable entrepreneur? Do they have the talent, the willpower, the entrepreneurs, the capital? The answer in every case is basically no. So I think this is the beginning of other nations coming to a moment of realization. For the US it's really a domestic issue of control between the state and company. On the international stage this is going to be huge—a moment of realization that you're going to need your own sovereign AI of some kind, even if just a fine-tuned model you can run inference on yourself, because otherwise you could just get cut off.

   2:42:06

   Nathan Labenz: Yeah. Where are they going to turn is a big question. DeepSeek raised what, $7 billion? So it's just a different order of magnitude. And obviously that doesn't come with any guarantees about what they'll open-source in the future. With Meta at least waffling on its commitment to open source, I just don't see any open-source player that's really positioned. NVIDIA is probably the big one—if you're putting your hopes in some big-tech basket that has the deep pockets and the motivation to build and open-source something semi-competitive with the frontier, we're kind of down to NVIDIA.

   2:43:14

   Prakash: I think you're correct, but DeepSeek will also be there—they're committed to open source. The $7 billion funding is a very interesting structure: the founder owns all of it through a special vehicle, shareholders who come in get no rights on the firm itself and can't exit or resell shares for five years. It's all locked up, yet he could have raised $50 billion if he wanted to. He could list the company publicly at a couple hundred billion at least. They're happy with their model reception and they've also been able to use Huawei chips. He's in the process of constructing a conglomerate: AI plus chips plus capital, all together. But even in China, I think AI models are eventually going to take over from the state—the state is going to lose control. My question on China has always been: when does the nation state recognize that this is going to happen and clamp down? Let's wait and see when that happens.

   2:44:51

   Nathan Labenz: I do suspect there could be a Beijing directive at some point where they just can't do this anymore. I don't understand the DeepSeek team that well, and I haven't read too deeply about the founder. But I do get the sense they have a mission something like "explore AGI with curiosity"—a similar vibe to the American frontier labs in an earlier iteration, with that borderline ideological sense of wonder and discovery. That seems pretty genuine from everything I've been able to gather. But when we see how easily the government can put Anthropic in its place in a supposedly much more pluralistic, due-process-based society—it doesn't sound like DeepSeek has much of a chance, at least for a while. They've got a long way to go in R&D before they'll be challenging the sovereign. My guess is it probably gets shut down at some point. But who knows? The Chinese government—everybody has these high-conviction statements about what they're going to do. I always come back to the fact that they're not even buying NVIDIA chips. We banned those chips with all this gnashing of teeth about how important it was to keep them out of their hands, and it turns out they didn't even want them. We should all be extremely modest in our predictions, because I don't recall anyone predicting that, if allowed, they would refuse to buy. And yet here we are.

   2:47:14

   Prakash: I don't think they actually refuse to buy. The chips are being deployed in Singapore, in Johor, in Oracle data centers for ByteDance to use—the UAE data centers would have a big Chinese customer base. There are really only two countries deploying AI at mass: the US and China. The other countries don't have the product or the talent. It's very hard for me to see what the meaningful difference is between a UAE company with a Western CEO and 95% Chinese talent sitting in Dubai, producing a product basically used in China or Southeast Asia—versus having that deployed in China on NVIDIA chips onshore. I don't think even the people most in favor of export controls have really explained why it's okay for ByteDance to use chips in Johor, Malaysia, but not okay when those chips are onshore in China. It's very confusing to me—what is the whole point of this?

   2:48:47

   Nathan Labenz: I think the basic idea is the same thing everybody is now feeling with Fable: you can get cut off. If the chips aren't in Chinese sovereign territory, in a data center they're powering with their own electricity, then there is the ability to just—

   2:49:05

   Prakash: Turn it off.

   2:49:06

   Nathan Labenz: Cut the cable. And then you can't do it anymore. I've never been a big fan of export controls as such. The version that made the most sense to me was: we may need some leverage vis-à-vis China, and this is one place to get it—fine. But how do we present that in a way that is as minimally hostile as possible? The idea of "rent but don't sell" seemed like a decent landing place: as long as we're on good terms and you're not throwing the world into chaos, you can have all these chips by the hour, and it's great. But the Chinese government, if they're anything, they're obsessed with controlling their own destiny, with not being cut off from critical resources. They've gone to great lengths to establish exactly that kind of leverage in every other domain vis-à-vis the rest of the world. So it's still very bizarre to me that they didn't do the same here. Yes, they can rent, and I'm glad we're on friendly enough terms for that. But there have been plenty of signals about how we might want to cut this off one day, and I still can't wrap my head around why they're not buying and locating domestically as much as they can. The only answer I can really come up with is that they're not buying the AI hype in the same way that—

   2:50:58

   Prakash: Exactly—in the same way we are. If your AGI timelines were more than five years out, it would make sense. I feel like Anthropic and OpenAI are in the 2028–2029 timeframe. Elon seems to be preparing for 2029–2032, because his data centers in space aren't going to be en masse ready until then. And I feel like the Chinese government is in the 2032-and-onwards space, where if they do chip development now they can get through two or three cycles of Huawei chips, get the cost down, and perfect the methods so they don't have to depend on TSMC in the early 2030s. That puts them in a position to really expand chip production around 2031–2032. So it really depends on when you think RSI is going to happen. If RSI happens in the next two to three years, it allows a hundred-times or thousand-times boost in tokens per watt on existing data centers—and as you increase tokens per watt, you get additional increases. Even on existing infrastructure, you double and triple, which eventually allows, say, Anthropic to just buy SpaceX. The early-RSI people have given up on physical chip development for now—they'll do it as much as needed for data centers, but they're not downstreaming into fabs directly. You don't see Anthropic or OpenAI putting money into fabs yet; they're still asking TSMC to build them. Elon is saying those fabs aren't enough—you're going to have to build your own. So he has the TeraFab coming up, which is again going to take about five years. He's in the post-2029 timeframe. All of these players are making bets based on where the technology lands and what year they expect their version of RSI to happen.

   2:53:30

   Nathan Labenz: Still hard to wrap my head around the Chinese government not diversifying at least a little bit into the earlier-timeline worldview. It seems like an obviously plausible enough story at this point that you'd feel really stupid in 2028 or 2029 if it was suddenly happening and you were asking, "Wait—why did we not buy the chips? Because we were content to rent them?" Doesn't that fly in the face of decades of thinking about how we're supposed to protect ourselves from these vulnerabilities? It's still weird. I can't offer a better analysis, but it still feels like something has not been adequately explained. As you know, I'm working on a possible trip to China next month, so maybe I'll go see if I can rustle up some answers for myself. It still strikes me as just super weird.

   2:54:45

   Prakash: Indeed. Well—

   2:54:49

   Nathan Labenz: Another fun couple hours trying to make sense of all this with you. Tomorrow we've got a couple of interesting guests. First, Karina Hong, founder of Axiom Math, where they're pursuing mathematical superintelligence—that's definitely going to be interesting. And then we'll have Sam Posupalak, hopefully I'm saying that right, from Skyfall AI, for another look into what they're calling enterprise superintelligence. We'll have to interrogate whether there's enough there for enterprise superintelligence to survive our world models as we enter the RSI run-up to the Singularity. I hate to frame all these conversations with folks who have by any other standard been super successful and kind of deserve their moment—and still a big part of me is like, does any of this survive? It's a really uncomfortable and awkward place to be, but we've got to shoot everybody straight, at least until something changes my mind on how I'm sizing these things up.

   2:56:11

   Prakash: I see that as one of the things you get to explore in real time—you find out as we go through this moment how the participants are actually dealing with it. Are they aware of the changes in the pipeline? Where do they see them coming? How are they preparing? That is what strikes me as important for the historical record. We already have this vision of what's going to happen in the next three, four, five years, and you have people executing—but the timeline they're executing on may not be the one we think is going to happen. We don't know who's correct, but you capture for the historical record the various viewpoints people held at that moment. If we're entering a posthuman era, what did the people who had something to say about it actually say about it, right then?

   2:57:21

   Nathan Labenz: Well, I suspect we're going to get two sharply contrasting views on that tomorrow. Come back and join us for another exciting edition of AI in the AM. We'll see you then.

   2:57:32

   Prakash: We'll see you then. Bye bye.

   2:57:35

   Nathan Labenz: Thanks, Prakash.