Season 1 Episode 4: The AI bubble is actually four bubbles
“If you’re looking for a sign companies back down and go: ‘Got you, it’s a bubble!’ They’re never going to do that. They’ll keep pretending everything’s fine until the last death throes.”
Mike and Nick describe the AI bubble not as a single failure, but as a stack of overlapping collapses. From the physical reality of energy constraints in the Strait of Hormuz to the shadow banking fueling data centers, they map out how the bubble isn’t just an economic issue but increasingly a political one, too.
- Overlapping bubbles: The economy is experiencing 5 bubbles, with at least 4 of them concentrating around AI spending.
- The energy ceiling: Silicon Valley ignores the material world at its own peril. Between energy costs and supply chain chokepoints like the Strait of Hormuz, “scale is all you need” is about to hit a wall no amount of VC can overcome.
- Capital as labor (The Lie): The promise that AI replaces human workers is an economic fantasy used to justify mass layoffs and degrade remaining jobs.
- Enclosure of ML: For decades, machine learning was an open academic field. Big tech effectively enclosed this commons, rebranding a general-purpose architecture (the Transformer) as a proprietary product.
- The “Siri Plus” residual: When the bubble pops we’ll be left with scoped, boring automation that works for specific tasks but never achieves the “general intelligence” used to justify trillion-dollar valuations.
Timestamps
0:00 What is the Last Enclosure?
2:38 AI bubble in context of Hormuz
8:03 Bailouts?
11:10 Bubble 1: Frontier AI companies
15:33 Bubble 2: “Wrapper” AI companies
17:13 Bubble 3: GPUs, infrastructure, and data centers
19:13 Bubble 4: AI financing (Banks and private credit)
22:33 What drives AI “value?”
30:18 The bubble in the real economy
33:37 Harms of the capital-as-labor myth
38:56 Residuals from the bubble
42:49 AI not causing layoffs
44:10 Why did this go on for so long?
49:08 AI is kayfabe and paper tigers
52:34 Types of bubbles
55:22 LLMs as the last enclosure
57:17 Accountability
1:02:56 Recommendations
Deeper learning
Mike on the AI bubble: misaligned.markets/what-to-expect-ai-bubble
Zitron – OpenAI Losses Increased Nearly 8X in 2025, Spending Hit $34 Billion: wheresyoured.at/exclusive-openai-financials
Harvey AI wrapping frontier models like Claude: help.harvey.ai/articles/what-ai-models-does-harvey-use
America’s Data-Center Build-Out Is Falling Way Behind Schedule: wsj.com/tech/ai/americas-data-center-build-out-is-falling-way-behind-schedule-e408a9a8
AI Boom Runs on Debt: aol.com/articles/ai-boom-runs-debt-global-144019676.html
Reach out
Site: rss.com/podcasts/the-last-enclosure
Bluesky: bsky.app/profile/lastenclosure.bsky.social
Transcript
00:00:00.160] Nick: Welcome everyone to this episode of The Last Enclosure. We’re
[00:00:04.080] Mike: your hosts, Mike and Nick. We’ve got a special episode for you today—a follow-up to our last conversation about AI. I think this is a different through line, talking through a very common talking point that is emerging even as the AI bubble is deflating or popping.
[00:00:28.239] Guest 2: We
[00:00:28.399] Nick: live in strange times where the bubble popping occurs quite so slowly. But yeah, the important emphasis today is a concept called the permanent underclass. This is probably a phrase that’ll be bandied about with increasing frequency. It is being used in tech circles, in oligarchic circles, and in the policy monkish side of things. So hopefully we can distill this into a conversational package and examine it in a
[00:01:01.200] Mike: in a timely way.
[00:01:03.200] Nick: Yeah.
[00:01:04.480] Guest 2: I
[00:01:04.640] Mike: Mean, obviously there’s a lot to talk about with this particular phrasing—its origins, its implications, exactly why people are freaking out about it. Although very interestingly, I just saw a podcast episode from Derek Thompson earlier today. Not that I watched him, but he was calling the claims that AI is going to take jobs—the “job apocalypse,” as they’ve been saying—overblown. Of course, Thompson was one of the earliest voices saying AI was taking jobs, so he’s jumping on the bandwagon as usual. But anyway, there’s this tension where, if you’re living where we are living, people are still talking about the permanent underclass as a serious kind of
[00:01:54.000] Nick: As a note, we are speaking to you from the Bay Area, California. So sort of the heart of the beast, if you will.
[00:02:00.159] Mike: My bad. Yeah, we mentioned that in the first episode. Yeah, yeah, yeah. We’ll
[00:02:03.519] Nick: Try to reemphasize it. Yeah. We are speaking to you from a place at least of witnessing it,
[00:02:08.240] Mike: working near it, and inside of it. And the lion’s den. Indeed, yes. So, yeah, there’s some sense that there’s still this fear about the permanent underclass in tech, right? So there is this bubble—not the economic bubble, there is this social bubble, there is this cultural bubble of people working in Silicon Valley and SF, and they are still feeling as if AI, if it won’t take jobs, will allow them to create opportunities that wouldn’t otherwise exist. There’s this sense that for all the flaws of LLMs, and we talked about this a bit in the previous episodes, you know, the one thing that it does to the user, even in cases where it is useful, is it creates this effect of efficacy, of forward motion, right? You basically—I liken it to this idea of APM or actions per minute in video games, right? So there are people who are using LLMs to automate workflows. Now, whether or not they are judging the efficiency of these workflows, I don’t know. But the fact that you can automate all this work, be it redundant work or work that’s not even useful, whatever, but you—it just feels like you’re doing more things per minute, right? Your APM is going up. And so even if you’re
[00:03:27.840] Nick: In the video game metaphor, this is sort of like, “Oh, I’m crafting, and I’m also training my pet, and I’m also gaining these skills or learning this magic—look at all the things I’m doing to get better and upgraded.” In real life, seemingly only the LLM is offering that kind of high-intensity, high-pace skill improvement all the time, every day.
[00:03:48.080] Mike: Yeah, it’s not even skills; it’s just literally I can offer so in Silicon Valley—task handling, yeah. Task handling. So people in Silicon Valley are not using chatbot interfaces; they’re using agents. An agent, for the audience, is basically a large language model that can prompt itself continuously until a task is done. So the interface is different, but this idea of agency workflows being different from LLMs is sort of a misnomer because it’s the same technology—it’s the exact same technology. It’s only that you’ve gotten the LLM to prompt itself to finish a task over time, right? And so if you’ve got, say, dozens of agents doing loops or whatever, your visible APM has gone up. Now, whether you’ve asked yourself whether these tasks should be automated, or if even the workflows are being automated correctly, that’s another question, right? But the reason I think—sorry for the long tangent—the reason why I think this permanent underclass idea is still relevant, even this late into the bubble, is because for true believers, their APM is up. They must be doing something more efficient. And if their APM is higher than your APM, well, they will join the permanent overclass, and you will remain in the permanent underclass, right? And so even as the bubble is deflating, for those of us who are in the cultural bubble of SF, there is this sort of self-affirmed notion of: well, even if AI has no ROI, my APM is high, this will not go away. Someone will find a way to make this work because everyone’s APM is going too high for this to just go away.
[00:05:30.560] Nick: Right. And of course, the responsible thing at this juncture is to point out that there’s a lot of the fallacy of metrics, right? I’m able to measure more output, therefore I am improved. I’m offering more, I’m increasing value. When in fact it can be a misleading metric. In fact, often is. But we have to measure and we have to kind of sell ourselves and pitch ourselves using something, and that thing should be as objective as possible. So this is kind of what we’ve settled on in this time period right now—this APM sort of abstraction.
[00:06:10.319] Guest 2: Yeah,
[00:06:10.480] Mike: I mean, going back to metrics, that’s a good point. Like, there’s a Peter Drucker business—you can call it business school or business idiot if you’re trippy to Exitron—the biggest
[00:06:22.079] Nick: Bubble deflator. Shout out to our boy Exitron.
[00:06:25.279] Mike: Drucker says whatever isn’t measured doesn’t get managed, right? That’s actually an inversion of an earlier quote by—I want to say RF Ridgeway, I’ll put it in the show notes—but Ridgeway says something to the effect of: whatever gets measured gets managed, even when it’s useless to do so.
There’s this sense that we live in a society where people look for metrics that are easy to correlate to measures of success, efficacy, or progress. There are abstractions that map crudely to those things, but sometimes when you go too far in measuring, you end up measuring things that don’t matter. Because they look like they matter, and because those things are going up and up, you just follow the trend line—basically maximizing something that doesn’t really matter.
So the question for LLM agents is: of the tasks they’re being automated for—keep in mind with agents, because the LLM is prompting itself over and over, and because modern agentic AI workflows involve reasoning models, these are models that actually have to emit tokens to themselves before they start a task. It’s as if we’re going back to our Leonard Shelby metaphor.
[00:09:06.879] Nick: Signing the tokens?
Mike: Yeah, yeah. If you remember in our last episode, we talked about Leonard Shelby, the metaphor of Grandpa Shelby.
Nick: Grandpa Shelby, yeah.
Mike: Right. Again, because the model has amnesia, this is as if Shelby is writing before he’s given a task—writing down all the things he has to do, then looking at that screenboard he just wrote on and saying, “Okay, right.” You or I might think to ourselves if we’re doing a task, but an LLM literally, verbatim, has to emit tokens in order to imitate that process of thinking: “Okay, my first step is to do this and do that, do this, that, that.”
Thinking wastes a lot of tokens before the task has even begun. So you’ve got an LLM that is “thinking,” writing on the draft board before starting a task, which costs money. Then, in order to do the task as an agent, it has to keep prompting itself—not with reasoning tokens, but with actual tokens—to keep continuously finishing the task until it’s done. This is a very, at best, inefficient process.
Now for the user who is able to automate these workflows, especially if they’re subsidized or they’re not paying the cost directly, of…
[00:09:08.399] Mike: You know, it feels like they’re doing a lot and they’re doing a lot for cheap because they’re not paying the full price of the tokens. And obviously, we’re in a bubble because a lot of these workflows may not be sustainable if you actually were paying for them on a per-token basis. And that’s sort of like the big lesson that a lot of people are learning as we’re recording this: even if it feels like our APM is up, it may not be worth it to do these tasks in these ways. The workflows could still be valuable; it’s just that the cost of doing them this way is way too high. So there’s this core mismatch at the heart of the AI bubble. There’s a question of what happens when the token reality hits everybody and what they’re going to do. Most likely, they’re going to cut back on tokens. And my thesis has always been—going back to my writing at Misaligned Markets when I talked about the bubble—that we have this technology for better or for worse now. I believe it’s better for harmful technology to not have been invented, but these things exist. What will likely happen is people will begin scoping them to tasks where the token cost actually matches the ROI of the task, right? And maybe that means agents survive, but it probably won’t be the agents we have today. Very likely, though, you will have bespoke workflows that are kind of very basic. To give you an example, I have a home lab, so I’ve played around with the technology just to understand it. And I was reading subreddits like LM Dev and a few others, and the people who were actually at the very beginning of the AI takeoff around 2024, 2025 were doing kind of consultant work where they were basically pitching themselves to companies automating low-hanging fruit, like basic classification tasks using on-premise large language models. So you feed a model a bunch of images that need to be approved—say contracts or something like that. And this is not actually a real workflow, so don’t worry anyone who’s like
[00:11:18.639] Nick: saying our application resembles this. Yeah. Yeah.
[00:11:22.000] Mike: But you have some type of document that has a certain type of schema, so you can actually codify—you know, this document has a name, it has these features, right? So when you see a serial number or something in a document, this score. When you see these other features, you give a document this other score, right?
[00:11:42.000] Nick: This is relatively objective, right?
[00:11:43.759] Mike: Yeah. So that’s the point, right? There is value in having workflows that revolve around tasks that are measurable, not just in a “my APM is up” kind of subjective way. Going back to what I call the effective design of LLMs, which you can only judge if you’re using them as the tech pros—the oligarchy has intended—is just “wow, these outputs really match my feeling or whatever, I’m gonna run with them, I like this.” Right? And now you’re putting it into a workflow where it’s like, “no, now we’ve got some external validation of did it actually meet this criteria.” Now, this does not make the technology valuable. I think it was incredibly wasteful that we had this bubble. But again, the local models aren’t going away. And very likely, even when or if OpenAI and Anthropic implode, their models probably aren’t going away. Some tech companies like Google or Amazon or whoever will scoop up their models, right? It’s just that they will be put towards boring and hopefully less harmful uses and less
[00:12:52.480] Nick: ambitious uses, probably.
[00:12:54.480] Mike: I think there could still on the margins be ambitious uses, but they will probably be bespoke and mundane.
[00:13:00.559] Nick: The kind of over-radical and drastic overpromising that we are currently witnessing has to subside at some point.
[00:13:08.320] Mike: That overpromising—I mean, yeah, the mismatch again of the bubble is, and we can get into this because it goes back to us exploring this question of the permanent underclass. You know, the big misalignment of the bubble is these guys, many of them, thought that they were making AGI, right, artificial general intelligence. And so they were promising God or a machine that could do anything because they believed it was imminently possible. Now
[00:13:39.919] Nick: in their early stages they saw line goes up, its capabilities are increasing, therefore at some point it has to be AGI.
[00:13:46.480] Mike: Yeah, so it was a kind of a presumption. If you even look at the academic literature—at least on this side of the machine learning divide—I’m thinking of people like Richard Sutton, who is a prominent researcher. Sutton has this thing called “The Bitter Lesson,” which revolves around the idea that you can have a very simple algorithm, and in some ways, scaling that up is actually more efficient than trying to create bespoke rules that anticipate certain contingencies and so on.
For example, in AI’s earliest days—and I use “AI” in this sense deliberately; dear listener, if you remember in the last episode, I was talking extensively about separating all these technologies from one another—the history of AI begins with what are called rules-based AI. They call this “good old-fashioned AI” (GOFAI): basically logic and symbolic AI built on rules, code, and schema that tells a machine what to do in specific cases.
To give you an analog, this sounds very early on to my ears. Good old-fashioned AI actually never went away. In fact, some of the irony is that one of the reasons why LLMs scaled is because, after the models stopped getting better, the language learning labs started building bespoke scaffolding and tooling that made the models perform better on tasks, even if the models themselves were not getting more capable. You get what I’m saying?
They began implementing rules; they began adding back some of these rules and toolsets from the old days of AI. Not literally, but the premise is that in the machine learning era, there’s this divide: do we even need good old-fashioned AI anymore? We’ve got these algorithms that can learn and self-improve on their own, right?
Well, it turns out the reason why people were going ape shit starting around late 2024, 2025, was because the labs had begun building workflows. They began building workflows that would apply to specific tasks, and the model began to get better at those specific tasks because it had the right tooling to do them, as opposed to just the model getting “smarter.”
Now they were telling us, “Hey, our models are getting too good too quickly,” and that was kind of true around 2023, 2024. The leap between, say, GPT-2 to GPT-3 was a big deal. Before, GPT-2 was literally autocomplete. I know people say AI is just “spicy autocomplete.” That is true. We can go back to the Grand Paul Shelby example and say that language understanding gets you certain things for free, and that’s why LLMs are a little bit more than just autocomplete. But back in the earlier days, they were literally just autocomplete.
Then at some point, when they began training on more data and making the models bigger, they began to exhibit other modalities. Their ability to do these things was limited, but they were able to do lots of little bits of lots of things, which gave the illusion of there being something here.
Even for me—I’m so sorry, I’m dominating—but even for me, I was an AI skeptic. I’m very close to somebody who was an AGI doomer, and people in our community are, so it’s kind of all around us. SF is so weird though, because you’ve got this hybrid of people who are equally optimistic and equally pessimistic. In my circles, they were doomers, but they are simultaneously—the important element of that is they
[00:17:45.039] Nick: are enamored of the growth and the capabilities. Oh, sure. They are seeing the line go up. Their emotional reactors
[00:17:50.160] Mike: can be either. They are flip sides of the same coin. Right. But I knew doomers, and you know, maybe you may know a booster or whatever. And, you know, in 2023-ish, when I was predicting—because early on, and I say in 2021, you’ve got people trying to create policy-based models, which I don’t want to go into technicality, but there’s this idea that started appearing in the early 2020s that you know you could have basically a model trained in one area would have some transfer learning in other areas. And so I made a bet with a friend who was a doomer, and you know, I was like, “I don’t understand transfer learning, you know, probably is useful in the margins, right? I could see a model that’s good at math, maybe some of that logic translates to basic programming, right? Because some programming code involves numbers. I’m not saying all of it does, but you know,” but transfer learning appeared to be much broader than it otherwise would have. And so as the models started scaling, to jump from GPT-2 to 3 to 3.5 to 4, it actually freaked me out. I was like, “Oh my God, am I gonna become a Utakowskian doomer? Am I gonna,” you know, but then I got in my home lab, I started playing with these models. It’s like, “Okay, that’s not what’s happening here,”
[00:19:15.599] Nick: right? Um, but somewhat illusory. Right. As you’re witnessing it, yeah.
[00:19:21.599] Mike: And I think for me personally, the reason why this sort of all played out this way is I think language itself encodes a lot of these modalities. And I think I mentioned this discussion, this talk in the last episode. There is an old veteran AI researcher, Michael Woodruff, I believe is his name, who gave a talk for the Royal Society: “This is not the AI we were promised.” He says the same thing. So you know, it’s not just me using my grandpa Shelby metaphor. There are people in the field who are alluding to the same thing. That language packs in a lot of these features that give a little bit of affordance towards doing some types of tasks for a machine. Right. Now that does not mean we get what we were promised. You know, these AI bros have oversold and overpromised us something that is not
[00:20:17.279] Nick: anything. It means that language itself already has baked in so many of these sort of convergences that it’s handing off—that it’s allowing the capabilities of these LLMs later on to thrive.
[00:20:32.240] Mike: Yes, yes. No, I mean yeah, the modality comes from language use.
[00:20:36.640] Nick: Yeah. Props to the people, not to the programs. Yeah. Yeah.
[00:20:41.519] Mike: Yeah. But in this nugget of doomers and boosters, right, there is a question of: where does this permanent underclass narrative come from? And you and I were talking about this a little bit before. And obviously it’s not that we don’t see eye to eye, but I think there’s a bit of unpacking of
[00:21:04.240] Nick: two different AI skeptics don’t agree. My God.
[00:21:07.839] Mike: Every AI skeptic has to be on the same page. If they have the same beliefs and it’s a—I can’t be AI skeptic and use AI at the same time.
[00:21:13.920] Nick: We need to assemble the brigade today.
[00:21:17.680] Mike: But yeah, there’s a sense of where this narrative came from and who’s kind of pushing the narrative per se, right? And for you, I mean I don’t know your firsthand or secondhand experiences—what is that sense for you?
[00:21:32.720] Nick: Well, it comes from the first time I was exposed even to the phrase in this era. The permanent underclass was a very top-down agenda—not conspiracy—where the sort of figures on top, the founders, the designers, the Elon Musks, et cetera, were the ones enamored of this: “Oh, the coming permanent underclass is right ahead. I, with all my wealth, need to be prepared for it; need to be on top as this radical shift occurs.” So that’s again, that’s a very top-down thing, and no phenomena is ever just top-down, right?
[00:22:13.680] Mike: Yeah, for my lens unpacking the AI bubble, there’s really a lot of narrative threads here because there are different parties. I just told a long, roundabout story about the people who were saying “scale is all you need,” then hitting a wall, and then turning back to good old-fashioned AI—to rules, scaffolding, and tooling to have machines that did stuff.
So you have true believers who were part of this camp. By “boosters,” I don’t just mean people who are AI bros; I mean people who are actually true believers in the sense that they’re techno-optimists. On one side of the coin, there’s a guy named Eliezer Yudkowsky—you may have heard of him, but for those of you who don’t know, he’s like a neckbeard who, for some reason, the media has turned to to talk about AI doom. He’s one of the key voices leading us into discussing the possibility of AI killing everyone on Earth—or literally everything in the light cone of this planet. If you can see it in space, the AI will colonize it. It’s like a 4X game; it will colonize every planet, every species, every asteroid it encounters to maximize whatever the fuck its objective function is.
[00:23:48.480] Nick: And he’s in the Bay Area and also promoting himself as this kind of expert. Right.
[00:23:53.839] Mike: And on the flip side of that coin is a 10-year optimist who’s not just a booster today who’s like, “Oh yeah, my APM is up,” but not only is my APM up, but like in 10 years imagine how far the AI species will travel, right? And some of these people—you know, I say optimist kind of lightly because some of them generally are utopian. Some of them are optimists in the sense that they think they agree with Yudkowsky: okay, AI will kill everybody, but that’s actually good because the species that supplants us—this superseding species of AI—will be better than us in every way, so we should want to become mulch for these creatures because that’s how you maximize utility in the universe, right? So yeah, this is kind
[00:24:41.039] Nick: of a bad, you know, script-doctoring of like Skynet in a Terminator scenario, but that is essentially what they’re talking about, expanded into the long run.
[00:24:51.200] Mike: Well, it’s like Terminator if I guess like if Sarah
[00:24:54.000] Nick: Connor wanted to help the machines replace humans, conquer the Earth, and then leave the Earth and then continue being Terminator. I
[00:25:00.640] Mike: mean this is closer to maybe some Japanese anime where it’s like, you know, these stories where the villain wants to kind of replace the world with another world, right? Final Fantasy—was it six like that? Which one had that plot?
[00:25:17.200] Nick: Uh, five did. Okay, so I’m sorry, yeah. Five.
[00:25:20.000] Mike: Right. It’s just a very common plot in Japanese media where the villain literally wants to replace the world with another world—a better world, a purer world that is pure by some metric of their own imagining.
[00:25:32.720] Guest 2: And
[00:25:32.880] Nick: inevitably,
[00:25:33.599] Mike: notably, one in their direct control.
[00:25:36.400] Nick: And that is the obsession of, again, from my point of view, these oligarchic figures. Their inspiration—like a Sam Altman—is trying to sell this idea again. We cannot peer into his mind, but he’s trying to sell the idea of this permanent shift in the capabilities and economic outcomes of the entire world is imminent due to the emergence of AI, and you have to get on the right side now. So the ownership class, shifting all of their resources into these companies and these projects and data centers, etc., are trying to position themselves on top to not survive it, but thrive in this new era. And that is the sort of top-down view of the permanent underclass. Can we speak to the experiences of the kind of men and women in the middle? The worker—the engineer who says, “Well, I want to survive. I don’t want to be pushed all the way back down to the bottom of the food chain in this new labor economy.”
[00:26:43.680] Mike: So there’s something to what you’re saying. I think what the commonality of all these people is that there is a shift, right? But their views are different, right? So I’m laying the table. It’s
[00:26:56.559] Nick: It’s difficult to paint them with one brush. Yeah. But the “they want to be on top when the change occurs” motif is itself relatively consistent.
[00:27:06.480] Mike: Not completely. So for some of these people—I’m laying out the rogues gallery of characters here just to show you the variance actually in all these views—like the people who are talking about being mulch, they literally think they are going to die and they are happy with that because they want a progenitor that is better than them in every way. Right?
[00:27:24.960] Nick: To have given birth to the next thing. Right. Yeah.
[00:27:26.799] Mike: Which is obviously edgy teenager shit.
[00:27:30.319] Nick: It’s edgelord logic, but again, they have enough power to get it into the minds of as many people as possible right now, and to scare and intimidate and influence the career aspirations of people who we know and are working on. I mean, this
[00:27:46.640] Mike: is definitely a subset, although supposedly I believe Musk and somebody argued about this. So Musk purportedly is not on the sublimate everyone’s side. I mean, not to his credit, but you know, because he wants to—closer to your point—he wants to rule over the future, right? Not be mulch. But yes, our God Emperor of Mars, for sure. Right. But you have even within the people who are, let’s say, naive—or I guess for lack of a more nuanced word, emphatic about AI capabilities, right? In either direction, the direction where it kills us or the direction where it helps us, even within that group, you have this kind of kerfuffle of people who are like, “Okay, I want to be on the team that is not part of the permanent energy class,” or “I want to be on the team that is going to usher in the future of intelligence in the universe,” which is the side that—you know, they want extinction, but they just want to bring about the extinction themselves.” For
[00:28:52.400] Nick: anybody at home is like
[00:28:53.759] Mike: when did this come to
[00:28:54.960] Nick: influence, predominate sort of the debate? The answer is very quickly in the last couple of years, and you are sharing the sentiment of my personal hero, Han Solo. I’m out of it for a little while, and everybody gets delusions of grandeur.
[00:29:11.279] Mike: Like that’s where we’re at. This is the delusions of grandeur. The insane thing is, okay, I mean, I know I live here, right? So of course I’m familiar with this stuff. I never thought this knowledge would be useful to understand the cultural milieu, let alone the economic machinations or the product. Right. But like these were the voices of Silicon Valley, they didn’t come from nowhere. Elias has been here. They’ve been here
[00:29:32.319] Nick: the whole time. Yes. The call was coming from inside the
[00:29:34.720] Mike: building. You guys now are learning what I had to learn, you know, I would say seven, eight, nine years ago.
[00:29:40.640] Nick: Welcome to the jungle.
[00:29:41.680] Mike: Right. I mean, when I was learning this, I didn’t think it was relevant. Like, okay, I know about some eccentric people. Cool. I mean, hopefully this knowledge is never useful. And now it’s like, oh my God. Like I think—but then Sam Altman wasn’t even a billionaire. So right. Yeah. So speaking of Sam Altman, I don’t think he believed
[00:30:03.279] Mike: What bucket should we put him in? Yeah, so let me try and lay out the buckets. I know I’ve been kind of roundabout—I’m sorry, dear listener—it’s just that within these buckets, you’ve got these weird subgroups, but let’s just set up the broad table here.
Bucket one are people who believe in capabilities in either direction. They are AI hypists, right? They believe in AGI, either as destructive AGI or productive. All the true believers go here. They don’t agree, right? There are people who want to use this to create a permanent owner class, people who think it’s going to kill everybody, people who want to bring about the extinction of everybody because that’s actually good for some reason. So, this bucket is a true believer bucket. They believe in AGI. I don’t know what they’re doing now that it’s becoming more and more apparent that we’re not getting AGI yet.
Bucket two, I would say, are managers. This is probably more to the listener’s day-to-day experience. Managers are people who are in positions to deploy AI or to influence AI deployment. Managers can be people who are either bought into the delusion—a lot of them are going through their paces right now because as the AI agents burn through their tokens, they’re being asked, “Hey, why are we spending thousands of dollars per day and nothing’s changed about our company fundamentally?” So these
[00:31:30.960] Nick: are like mid-level white-collar. Mid-level, mid-upper, mid-level,
[00:31:34.240] Mike: yeah, white-collar managers, right? So some of the managers are actually maybe
[00:31:38.240] Nick: they used to be technicians, but they come up into the phase of how I’m getting to choose how to deploy this within a company.
[00:31:44.160] Mike: So some of the managers are people who fell for it, and they actually, in some sense, are similar to the guys in the first bucket. They are true believers—maybe not in the AGI vision, but in the fact that LLMs today are delivering way more value than they are—and so they are emphatically pushing it on everybody because they believe it to be true. Now, I don’t have a sense of whether the current deflating of the bubble is giving these people room for pause, but as late as this month, there are people demoing tens of thousands of dollars worth of AI demos, just plowing through tokens and not caring. So maybe they still walk among us.
And then the other half of this bucket: there are managers who don’t believe in the capabilities of current LLMs, but they want to seem like someone who does because it’s in fashion. For their own job security, if they’re the ones setting the pace of AI deployment in their companies, they can keep their jobs safe and secure, right? If you’re the one telling everyone, “This is where AI should go—not towards my job, but towards those jobs over there,” and you seem like you know what you’re talking about, everyone’s going to defer to you as the voice of AI at your company. So there’s a bit of a
[00:33:03.519] Nick: combination of staying on trend, but also deploying these systems almost offensively. Trevor Burrus, Jr.
[00:33:10.079] Mike: Right. It’s almost like a bit of a witch hunt. You’re the witch, you’re redirecting the hunters.
[00:33:15.119] Nick: And you’re masquerading as the witch hunter general. Right. You’re like,
[00:33:17.839] Mike: “Over there. I’m not one of them. Yeah, we love AI here. And you guys should be put over there, right? Yeah. And I know what I’m talking about. You guys better make sure AI is all in those places, right?”
So bucket three is I would say tech leaders, right? And so there’s some sense in which there are people who actually are closer to technology and they probably knew all along that this was not sustainable or viable or valuable.
[00:33:52.480] Nick: That the use cases were not as numerous or as essential as were being advertised.
[00:33:56.960] Mike: Yeah. I mean I’m sure amongst some of these leaders, some of the people probably fall in bucket one. So I don’t want to pretend that every tech leader is cynically deploying AI knowing it didn’t work. Regardless, they were responsible for—but they’re culpable enough because
[00:34:11.039] Nick: They do have a background in the technical side and they’re not just misguided business heads.
[00:34:15.360] Mike: Oh, absolutely. No, there’s no doubt about that. And there is some sense in which the incentives for these guys is just, “Hey, look, the world went ape shit when Sam Altman released ChatGPT 3.5 and everyone’s stocks went up.” And so if you are going to be a good leader for your shareholders, you’re going to make that stock grow by leaning into the thing everyone is screaming about, right? So of course you’re going to say “AI, AI, AI” over and over again, even if you know that it’s not as big a deal as they’re saying it is. Right? So let me give you an
[00:34:55.440] Nick: antidote. Be an innovator, at least be—I don’t know. At least be an excavator, be a booster of it, right?
[00:35:02.239] Mike: Yeah, yeah. And I’m not trying to make these guys more sympathetic than they—I’m not trying to make them sympathetic at all, is what I’m trying to say. But just to understand their incentives: their companies had kind of structured set growth. AI provided a breakout opportunity. They are fiduciarily kind of obligated to grow the company stock. It’s just how
[00:35:26.800] Nick: they have to chase that dragon as they see their duties. And
[00:35:31.360] Mike: All their peers were doing it, right? And so, you know, why not do it? They gave me an example of this sort of approach—there’s one story where Jensen Huang was asked about the future of quantum computing. He said something that was kind of dismissive of it, because someone was hyping up that AI is going to lead to quantum computing very shortly, something like that. I don’t remember the exact story, but something like that. And you know, Jensen added some reality to the situation, which is like, those things aren’t related and quantum’s not a thing yet—this is not even applicable.
[00:36:14.880] Nick: Just if you’re being
[00:36:15.760] Mike: If you’re being honest, you can’t make that statement. And then he caused some stocks to tumble, right? And so, you know, if this is the reality and your incentives are to provide a return for asset holders, some of which include our retirement funds—again, I’m not saying we’re culpable personally, I’m just saying this is a system we are all shackled to—then these guys are just sort of playing their part. And obviously we could be saying, “Hey, why don’t you actually speak out? If you see something, say something,” right? But you know, if your peers aren’t doing it and you’re the idiot ignoring free money, your shareholders are going to rake you over some coals and their
[00:37:02.239] Nick: Ethos is if you see something, sell something.
[00:37:04.719] Mike: Yeah, I mean that’s the case definitely, partially because, well, one, even if they were responsive to this nuance, their peers probably weren’t. There’s a prisoner’s dilemma here, and that’s sort of inescapable and unavoidable unless you have a society where people can solve collective action problems and or there’s an incentive to, right? So they must defect
[00:37:38.880] Nick: in order to betray their peers simply in their mind simply to stay relevant. Right.
[00:37:45.360] Mike: Yeah. I think there might be at least two more buckets, but I think the one that matters and the one I think you’re kind of pointing to is like the guys at the absolute apex, the guys at the top—the Sam Altmans of the world, the Elon Musks of the world?
[00:38:00.239] Nick: This is the ownership class, right?
[00:38:02.239] Mike: What do they believe? Right. And I think once you get this high up, you’re kind of dealing with personalities as opposed to dealing with a particular class of interest. And so all these guys have idiosyncrasies that make you want to tackle them differently. I don’t know how much this matters individually, because they’re so—yeah, I mean, I don’t know how much this matters. This reminds me of a conversation that Karen Hao, who wrote Empires of AI, had. She was embedded at OpenAI early on, right as they were moonshotting. So Karen Hao had a conversation with Ed Zitron—you know, again, famous bubble popper. And Ed was pressuring her on, like, “How do these people believe this nonsense? Are there people who actually believe this?” Going over the buckets I’ve talked about—she kind of touched on them, not in this structure, but as far as I know, I’m one of the few people who have put this together in this way—being people who are true believers, and so
[00:39:18.320] Nick: come on the podcast.
[00:39:21.360] Mike: And Ed’s like, “You cannot be people who believe this.” He was saying there’s a class of person who says these things so they can get money and power, and I think that is true. I think even within the buckets we’ve discussed, there are people who are very clearly oriented that way. But there are also people in these buckets who genuinely believe because either the effective design of LLMs made them fall in love with it, or they were, based on their machine learning knowledge, enthralled by the idea of encoding intelligence in machine learning. To give you an example from the upper echelon class, Dario Amodei—the CEO of Anthropic—was an OpenAI researcher, and he and a cohort of people left OpenAI because they weren’t doing safety well enough. He left in protest, effectively.
[00:40:18.960] Nick: This is what led to the ouster of Sam Altman in its original—
[00:40:22.719] Mike: No, no, no. This topic is before that.
[00:40:25.199] Nick: The prototype is before, right?
[00:40:26.239] Mike: The ouster does involve people, and we could talk about that. We did a little bit in the last episode.
[00:40:30.400] Nick: The safety-minded people. The safety-minded people. Yeah, yeah, yeah.
[00:40:32.639] Mike: But the ousted is just literally kicking sand out of the company. Right. But separately from that, there are researchers who left OpenAI beforehand because they’re like, “you know, OpenAI is not safety-minded.” And so Dario studied—I want to say he studied neuroscience and then computer science, or sorry, cognitive science. I believe that’s what he studied. Deeply interested in this question of intelligence, deeply worried about AI alignment. And so for years, he spent his professional academic career on these questions. And so you know, when he goes over to Anthropic, like, yes, it’s a money-making machine, or that’s what they’re trying to build. I don’t think that disposition disappears. Now his incentives obviously change, right? But I do think Amodei is separate from Sam, not because he’s safer and he’s more secure—moral. In fact, in some ways, Amodei might be more dangerous because he actually believes something and might be emphatic about that, right? Whereas an Altman reads more like a kind of blank slate, if you will, right? Someone who will say anything.
[00:41:50.320] Nick: A Sam Altman is just a George Costanza principle of “it’s not a lie if you believe it’s true.” Right. I think that’s again, without getting to the depths of his mind, I’ve safely and firmly said that for a while.
[00:42:01.280] Mike: Right. So you have in this last bucket a group of people who belong together because of their class and because of their similarities, but each of them has different reasons for acting the way they do. Sometimes they act in concert, but in a lot of ways, it’s not necessarily that they’re acting together. It’s just that their incentive structures individually lead them to act this way, even if they are dispositionally different. Elon Musk is another example of this group. I mean, Elon is probably harder for me personally to read. Elon was talking about the Yudkowsky and AI doom view since the 2010s. He famously said AI is like summoning the demon—I think 2014 is when he said that. He was
[00:42:51.679] Nick: kind of a pupil of that
[00:42:53.039] Mike: school for a while. He read Nick Bostrom. I think Bostrom has connections to Yudkowsky, right? All this idea of superintelligence, right? Trevor Burrus—
[00:43:01.360] Nick: To the original Berkeley clique that is still there, essentially.
[00:43:04.960] Mike: Yeah. Bostrom is Oxford, but I think via the broader effective altruist slash rationalist project, they have intersected. These are
[00:43:14.800] Nick: the EA folks. I’m sure our audience has heard some little snippet of this at some point for sure. Right. Yeah.
[00:43:20.079] Mike: I use EA loosely because technically you could say Eliezer and Bostrom aren’t EAs, but there is this broader diaspora of people who are rationalists—I guess that’s what you would call the broader culture, Bay Area rationalist. And of that diaspora, you’d get all these groups like EA, long-termists, which is kind of like EA.
[00:43:41.679] Nick: This is a Sam Bankman-Fried, right?
[00:43:43.199] Mike: EA is Sam Bankman-Fried. Long-termist is more like—I think Jeff Bezos considers himself a long-termist.
[00:43:48.960] Nick: We’ll come back to Jeff Bezos in a second. But yeah, that’s who we’re kind of putting on this big dartboard.
[00:43:52.960] Mike: Right. So Elon has kind of waffled between Dario’s position and Sam’s position, but also Elon, in some ways, he’s more legible than them because he’s obviously a craven capitalist, right? To some extent, Elon will—just like Sam will—say and do whatever benefits them in the moment. I mean, what came out of the trial between him and Sam was Musk’s money that actually seeded OpenAI as a nonprofit because again, Eliezer—specifically Eliezer and Bostrom—had convinced Elon to put money towards saving us, quote-unquote “saving us from AGI.”
[00:44:29.679] Nick: From this demon that they were summoning, right?
[00:44:31.760] Mike: Yeah. Now back then OpenAI wasn’t making a product. They were doing research for profit. Right. So you know, but in this pursuit of
[00:44:45.039] Nick: Can I just say it’s still funny that in the open trial between Sam Altman, the known liar, and Elon Musk, the awfulest guy in any room, it’s Elon Musk who loses? I think that’s fascinating. On the merits, clearly, but it’s
[00:45:05.360] Mike: Most of his stuff is self-owning. That’s his specialty. When he wins, it’s just because the barrier to stopping him was too high. But we could talk about this later—my thoughts about it. We’ll come back around to it.
[00:45:28.159] Nick: No, the point is in the trial, a lot of this is revealed—the original sin of where these deals were made to fund it, to push it, and that’s where Sam Altman got his funding.
[00:45:37.840] Mike: Yeah, and so early on—I forget what year came out on the trial—Elon is egging OpenAI when they had a small breakthrough to become a for-profit. So it’s like, this is the guy that was publicly biting his nails about AGI, and then when it becomes possible to monetize it, he tries to monetize it. He wanted to roll OpenAI into Tesla. And ironically, if that had happened, we may not have the AI bubble as it exists today. Elon Musk could have possibly saved us from this entire nightmare. This entire nightmare.
[00:46:17.360] Nick: Well, because if he’s like, “oh, we’re summoning a demon,” and then he pays to fuel the summoning of that demon, that’s hypocritical at best.
[00:46:26.239] Mike: It’s hypocritical at best. One read is that he’s cynical. One read is that when he created Grok, he was saying something like, “Oh, well…” — he’s an idiot, obviously. But to the extent he understands these subcultures, and these subcultures are also full of idiots too, but very wealthy, educated idiots. So, you know. To the extent that he understands what these guys are saying, he’s like, “Oh, you know what? If we make an AI that just wants to seek the truth, it won’t kill us.” Right. So if he actually believes that — this is us mind reading, obviously, I’m not saying I know we’re
[00:47:04.800] Nick: having to read the tea leaves here.
[00:47:06.239] Mike: In Musk’s heart of hearts, right? If Musk actually believes this loophole — this idea that making a maximally truth-seeking AI will save everybody — allows him to be consistent. He was afraid of the demon back when we didn’t understand that maximally truth-seeking AIs could be made and that that’s how you avoid the problem. Right. Yeah. And then, concurrently as his business interests change, well, you can solve this problem with the maximally truth-seeking AI, of course, right?
[00:47:37.199] Nick: And his beliefs and his fears also change to align with the profit motive. Typically and inexorably. So I said a second ago, and I’m so grateful that you set up these different buckets to put these sort of deplorable figures in—either culpable or truly terrible. I want to ground this in a specific example just so we can examine it a little more, gristle off the bone here. Very recently it was revealed and announced and hyped that Amazon, behind the scenes, Jeff Bezos personally was leading an effort to implement large swaths of AI decision making and services within Amazon itself. It went horribly by all metrics.
[00:48:22.559] Mike: Is it Bezos or is it Jassy? Bezos is out of Amazon. I think it was Jassy, right? I mean Bezos
[00:48:28.960] Guest 2: was
[00:48:29.599] Nick: Pushing it, like saying, “Oh, this is interesting, in terms of who’s on the ground behind—you know, I don’t know whose head should roll for this—but it resulted in crashing of something that should never crash, which is the Amazon storefront.” A terrible amount of screw-ups on orders. Like, if we are inheriting the example of “Why is Facebook great? Because Facebook never goes down,” then that’s the only rule, that’s the only metric for the success of Facebook. And that’s fine, that works. Amazon is at least—your basket never gets emptied randomly, right? You know, it remembers the stuff you want to buy. Even if you’re like, “I shouldn’t be buying all this stuff,” at least Amazon remembers. Even that fundamental lowest level, lowest threshold for success was botched because they got so excited, they got so in their own heads about it—that these AI capabilities were true and that they were going to follow it up and it was going to make everybody all this money, and therefore they could fire or reposition all of their employees. It was going to make everything better. And now we’ve seen it at scale for the first time in a huge tech company like Amazon.
[00:49:44.239] Guest 2: Yeah,
[00:49:44.480] Mike: Yeah. This is a story playing out, you know.
[00:49:47.119] Nick: In real time, right now, as of the publishing of this episode.
[00:49:50.800] Mike: Oh no, no, not just that story, but that Amazon reporting is back from March. I talked about it in my “What to Expect When You’re Expecting the AI Bubble.”
[00:50:00.079] Nick: To Misaligned Markets, check it out now.
[00:50:02.400] Mike: So, you know, those stories were happening and companies like Amazon were at the vanguard of that comedy of errors, I guess you could say. And what surprised me is after that reporting, I have friends and I know of companies in the area that as late as April and May were accelerating their AI adoption after reading stories like Amazon having those major issues. So at this stage where we’re at, these are now second-order effects.
[00:50:40.159] Nick: Yeah. The willingness of the Amazon heads to adopt this regardless of the result—which was a terrible result, objectively—is now encouraging other firms, smaller competitors, bigger ones, to say, “Oh, well, it’s legitimate now because they’re doing it. We get to do it, or we ought to, we have to, to avoid becoming pushed down into the permanent underclass,” right? This is how we have to contend with this.
[00:51:06.960] Mike: And now the irony is everyone’s starting to hit the wall at the same time, right? So you’ve got the token spend is out of control. Uber went through their entire year’s budget—I think I mentioned this in the last episode—entire token budget for the year in four months.
[00:51:26.880] Mike: Verifying. And there is a company—this came out even more recently, since the last episode. This is newer. There’s an unnamed company; an unnamed source said because they did not meter their company’s token spend, they spent, I think, $500 million on AI cost this quarter. They didn’t mean to. They accidentally spent $500 million. I think so. I had
[00:51:55.039] Nick: to check the—yeah. Yeah, just check the episode notes for that one. That’s a good grief. That’s ridiculous.
[00:52:03.119] Mike: Yeah, so this is sort of like—when is there a
[00:52:08.000] Nick: stockholder revolt sort of against, right? Because second-order effects is one thing. That’s kind of the people, the business heads in one bucket, all kind of hyping each other up. But there are other ownership groups that have to say at some point, “You are literally burning through all of our cash. This is our money.” They do have rights. Like, I haven’t really witnessed that yet. Maybe because the bells and the red flags aren’t being raised high enough. Yeah, so this is interesting, right? Except on this podcast.
[00:52:34.800] Mike: Not this. I mean, you know. So I laid out the buckets of people who are kind of like the core actors of the bubble from the tech side. For the rest of society, we’ve obviously got investors and we’ve got the media. The media has just been saying this is the fourth industrial revolution, don’t get left behind, right? And in some sense, the media is actually where the permanent underclass idea—I don’t want to say it originates, but where it gets steam, right? I think within tech itself, you have obviously people who are upper middle class. There’s a good video on this by how many works where he’s talking about basically—and I agree with this—Silicon Valley was sort of the last egalitarian bastion. I put this in quotes because obviously last great meritocracy where it’s like if you come here and you work hard and you got equity in the right company, you could actually have a good life, right? You put in the work, you get the equity, you exit, and you did something for the world at the same time, right? How the world should work, right? That’s why there is such this mythos and this sort of air of even moral kind of finality of Silicon Valley. It’s like this—if the 90s was the end of history, Silicon Valley is sort of what that looks like, right? In theory, right? In quotes, right?
[00:54:01.440] Nick: Does that make Palo Alto the city on a hill? Like is that essentially what was being sold as an idea?
[00:54:06.559] Mike: I mean no one said this out loud, but I’m saying this is sort of the way people quietly said this is how it should work, right? People come in, they put in their hours, they get equity, they get the exit, they did something for the world. This is why the system—and you can fill in the blank for what the system is. It could be capitalism as a whole, could just be the technocracy, could be just tech companies as a whole, right? Whatever. All these sort of modern technocratic apparatuses, this is why they work, right? Look at the results. People who came in here who were poor, went to college and studied computer science. That’s why “everyone should code” became such a moniker, because this was proving the vision of what the end of history should look like, what it was expected to look like, right?
And so the people buying into the permanent underclass narrative now see that, well, that was kind of a lie. Now that all the VC money is locked up in AI and AI increases your APM, well, time to join the side that is pushing AI for whatever reason, right? Whether they’re Doomers, Boosters, Nihilists that want us all to die, Elon Musk, right, Dario Amodei who wants to save us while working on AI, right? Regardless of why they are boosting AI, join them so that you’re not in the permanent underclass when the AI inevitably increases your APM and everyone else’s APM is too low to keep up with you, right?
So media is an important part of obviously feeding that fear. Every time they were reporting job loss numbers like a freaking horse race, right? They are feeding the idea that it’s coming, it’s coming, everyone’s losing their job to AI. Now they’ve started to walk it back as the bubble has become harder and harder to deny, but the damage of the idea was done. I mean, there are people still even now worried about losing jobs to AI, even though I think it’s become apparent that, you know, if you lose your job to AI, it may not be because of AI, it might be because someone wanted to cut costs and they’re just pointing to an accountability sink to justify their decision, right?
And you asked a good question about where are the investors? They’re burning money. Are they sleeping at the wheel? Well, it’s their money is enabling OpenAI—has, I believe, either Information or Zitron reported negative 133% operating margins. If you or I had to run a company that poorly, right, we are losing 33 cents for every dollar.
[00:56:46.400] Guest 2: It’s
[00:56:46.639] Nick: only profitable if you take out all the unprofitable elements.
[00:56:49.760] Mike: Right. So the investors obviously—well, one, you know, they don’t understand what work is.
[00:56:57.679] Nick: Yes, and they’re also—let’s point out another element of this network, this web. And those are media figures and critics who are enamored of and have been tech journalists for a long time, starting as far back as the era of Steve Jobs, who were sort of hero worshippers. And that was a critique that was very much eventually validated within tech journalism: “Okay, we have to stop worshiping these heroes.” Back then, you know, when Elon Musk was Iron Man and when Steve Jobs was still alive and he was still this aspirational figure. But now, because they’re not permitted to become these fanboys and girls, they still have to sell something, and they’re selling—and therefore perpetuating—the myth around AI and therefore a lot of these associated myths like the permanent underclass. They’re selling this whole package, and investors seemingly are buying it. Eventually, you know, it comes down to basic taxpayers are buying it. And as much as there is now a better understanding of the basic elements of all these systems and how they’re not what we were told they were, even as that understanding grows year by year, month by month, quarter on quarter, it hasn’t reached a saturation point yet, seemingly. It hasn’t reached the revolt stage. And I’m saying that because I’ve been looking around, and we’ve been commenting to each other, “Well, when is that?”
[00:58:28.400] Mike: Right. Yeah, I mean there are a lot of things at play here. The real thing is investors don’t know how work works, right? So they were just like—the people who were being one-shotted by Claude who were saying “my APM’s up” without understanding if the APM being up is a good thing or a bad thing. Investors, I think.
[00:58:55.119] Mike: I mean, these people aren’t—I mean these are begging for me?
[00:58:58.639] Nick: No, I know what we’re talking about. No, no, I know you are.
[00:59:01.119] Mike: I mean these people are usually just like the workers who are obsessed with APM and they’ve been impressed or enthralled by AI. Trevor Burrus,
[00:59:10.639] Nick: Jr. Well, and the worker groups are also doing it because they’re terrified of being downsized. Some of them are
[00:59:15.519] Mike: rightfully so. Some of them are—some of them are what?
[00:59:18.159] Nick: Some of them are more craven, you’re saying?
[00:59:19.440] Mike: The ones I’m specifically thinking about—the ones who are talking about APM, or I’m using APM as my metaphor—are the ones obsessing over this: “I can do more, it’s just one more token, bro, one more token, bro.” They’re for the most part well-off tech workers. They are upper middle class. They are the people for whom the promise of meritocracy in Silicon Valley existed and has gone away, or has started to go away. But they’re still subscribed to it via the last remaining bastion in their mind, which is: get your APM up high enough so that you can reach escape velocity.
The features of large language models that made these workers susceptible to this idea—people who should know better—are the same features that made investors fall for the delusion too. And doubly so for investors, because they don’t do any work. They don’t do the actual grunt work of the companies they invest in.
[01:00:18.159] Nick: As of today, we haven’t reclassified investing as working. And thank God.
[01:00:23.519] Mike: Maybe they did at some point in the 90s or whenever they made their money and exited, right? But I’m saying they are far enough removed from these processes where they are looking over the shoulders of everyone else and saying, “Oh, that looks good, that looks good, it looks fine, it looks fine,” right? Because they don’t have any other means of doing that. And now that this has gone on so long, now that no one called foul earlier, even if they know that this is not working out, their incentive—sort of just like all the other people who are trapped—CEOs, the leaders who are kind of trapped in that they probably know the technology is not uh
[01:01:04.800] Nick: They’re locked in because not only did they make these promises, but they have burned through all this money pursuing these promises, and either it works to some degree that it launches them far enough that their growth overwhelms the losses—which is also not rational by any metric and based on the current losses—and they somehow come out the other side, as you’ve been describing.
[01:01:27.519] Mike: Or they know this is not gonna work and they’re looking for their exit liquidity, right? So there’s a story of Elon Musk and it’s unload the debt on somebody else. Right. Like the SpaceX IP. The SpaceX the
[01:01:39.360] Nick: Taxpayer, anybody, yeah, right, right.
[01:01:40.960] Mike: The SpaceX IPO—when Nasdaq was fast-tracking SpaceX, a lot of regular indexes with access to Nasdaq would get exposure. The S&P didn’t follow suit. But still, you see the plan here. I don’t know who knows what is happening, but obviously, among the people who are still boosters, I think some of them are cynically so. You’ve been doing this song and dance for so long, you can’t stop now. You have to keep going; there’s no other way out but through. Which makes some of these people more dangerous, right? I believe we still have a bunch of true believers—I think at some point they’ll wake up—but among the people pretending to be true believers, I think they are just cynically aligned. Their incentives make it such that they have to keep this song and dance up because they have to find a way out.
[01:02:41.679] Nick: They are entirely locked into it. Yeah, they have to, because there’s no other way. And I was validated recently when I heard Elon himself speaking in defense of this behind-the-scenes merger of his AI company and SpaceX. It’s natural because people were questioning, like, “Well, you can slam these two companies together. One is bleeding money, the other is very profitable, and I guess the books work out, but what are the efficiencies? How do they organically fit together?”
[01:03:21.039] Mike: We’re going to take all the matter in our light cone. What—all the matter in our light cone? What don’t you understand?
[01:03:26.559] Nick: Yeah. Well, “Why are you so dumb, you media critics?” And they’re like, “Okay, but what’s step one?” He’s like, “All matter in the light cone—that’s the step. Why aren’t you listening to me?” And no, he basically said—and we’re making fun of the idea of “they’re going to get everything,” the Borg, “resistance is futile”—he’s like, “Yeah, well, we’re going to have data centers more efficiently, we’ll launch them into space, and then they’ll cool themselves easier up there, and that’s what we’re going to do. That’s what SpaceX is going to do.”
It’s like, my God, he’s finally off his rocker. He’s saying these words in public. Data centers in space sounds like a rejected comedy bit from the Muppet Show to my ears. And again, I don’t know what our audience thinks, but we’ll find out if they continue to do it. I don’t think we’re going to see data centers in space. Mostly because I don’t think we’re going to see data centers in Tennessee. Like, that is also a bubble on its own that is being pushed. So—
[01:04:28.239] Mike: Yeah. No, very good call. Yeah, I mean, I look at the reporting of someone like Zitron who’s showing a lot of these data centers are not being built or they are behind schedule. Even as the money’s sort of—
[01:04:37.920] Nick: —on paper been dedicated to it, it’s just
[01:04:40.320] Mike: piling up debt. Trevor Burrus, the data center, yeah, exactly. The data center, the AI bubble—and I said this in the last episode—the AI bubble is closer to actually a data center bubble, right? So you’ve got like a tiny bubble in the form of the AI startups like Anthropic, OpenAI. And that would be bad, you know, by itself, but it wouldn’t be that bad because the losses would be kind of centered on Silicon Valley.
[01:05:02.960] Nick: But the Trevor Burrus, J couldn’t therefore be just unloaded onto all the rest of the world and the country and other people and classes and such.
[01:05:09.440] Mike: Yeah. But the data centers, we’ve got some of the world’s biggest banks have invested in lending money to these. And the Trevor Burrus,
[01:05:17.440] Nick: it now starts to look like ’08 again if we’re getting the banks involved.
[01:05:28.320] Nick: Terrible. So
[01:05:29.119] Mike: It literally is like somebody had the bright idea to replicate ’08 with data centers. On top of the banks doing it, we have these entities called private credit or non-depository financial institutions—shadow banks, some people call them—and private credit has lent money for the data centers too. In some cases, they’ve taken money from the bank to lend out to other people to build data centers, right? So there’s a very circle-jerky motion, not unlike the one in Silicon Valley where the AI companies are feeding each other money: OpenAI to NVIDIA, to CoreWeave, et cetera. But separately from that AI company bubble, there is a data center bubble that involves these entities—some of which we don’t even know who they are—lending money to build data centers that may not ever be built.
[01:06:19.519] Nick: Yeah, and then this self-licking ice cream cone kind of process. I’m thinking now back to the second Trump inauguration, and without diverging us right into purely political territory, you know, on stage with the incoming president were all of the great tech oligarchs. And even at that time there was the plan being pushed for Congress itself, on behalf of the American people, to contribute to and pay for the Stargate initiative. Right. And I remember the Stargate thing eventually going off the headlines and it didn’t get picked up by Congress because they hate the idea—it’s clearly a boondoggle. But the second that Elon said data centers in space, like he’s still trying to get some form of Stargate out there of this concept, like we need to get the American taxpayer on the hook, otherwise it’s only gonna hurt us.
[01:07:17.760] Mike: The tech holder. That’s how their companies work, though. I mean, that’s why Silicon Valley loves working with the government, right? So like it’s a natural extension of their existing business relationships, right? And for Elon specifically, you know, it’s a little hard to know if he is just that big of a—he is dumb, but it’s hard to know how serious he is. I used to think that it was a given time. Right.
[01:07:43.760] Nick: We don’t know, yeah. So
[01:07:45.039] Mike: I used to think that some of the stuff was tactical—not to his credit, but it’s become apparent that I think this is what happens when you have too many yes-men around you: your worst impulses are not curtailed. We’re seeing this with Trump, obviously, in every dimension of Trump’s life. And so Elon grew up as a sci-fi nerd, and of course he likes data centers in space. He’s now a billionaire toddler who gets to live out his fantasies. I’m sure data centers in space sounds futuristic; it sounds right.
[01:08:16.000] Nick: For a person suffering arrested development, as we have all sort of armchair diagnosed him as, this is too perfect for him. Right? For his mindset.
[01:08:26.399] Mike: Yeah. I mean, it obviously doesn’t matter if he believes it or not, right? Like, did he believe we’re going to go to Mars? And now he’s saying we’re going to go to the moon, which is—you know, he should have been promising that before Mars.
[01:08:38.479] Nick: But yeah, I mean it really only matters if he does get the taxpayer on the hook, which seems to be they’re constantly probing and poking at
[01:08:48.399] Mike: the smartwater, but other inventions. I mean, even though Tesla’s a meme stock, right? I mean, they’ve not really created a new car—and I mean I guess if you call the Cybertruck a car, fine—but you know, they haven’t created
[01:08:58.479] Nick: a new successful vehicle in over
[01:09:00.800] Mike: 30 years, yeah. So, you know, if these companies are meme stocks and Elon knows he could say whatever to rile up markets, then why not speak a little bit to your childhood fancy, right? Why not kind of placate your inner toddler and just say shit that the uncritical media will just parrot and amplify and not question you on, right?
[01:09:28.560] Nick: Yeah, in that sense, it’s a—if not a great tactic, it’s a great strategy. So in that, we don’t have a defense against it, purely speaking. Right. Yeah, so having said that, this all sounds like a very dire sort of assessment. Let me inject an idea here that is not—if not hopeful, I think it is revealing. And that is: at a certain point, I saw in the permanent underclass discourse a fear emerge in the eyes and in the sense of these oligarchs—and this was not Elon specifically, but there was consistent reference to when this new era dawns ahead of us, we all need to have enough wealth put away that we’re not the people relying on UBI. That was the dividing line. There’s going to be, you know, the underclass is fed and watered by UBI; everybody else has made it, their rocket ship took off, metaphorically speaking, and they have enough wealth they can sit on their Scrooge McDuck mountain. That’s the distinction. And as soon as I saw that, it’s like, okay, they’re now seeing that there’s enough of a consensus coming around UBI, for one thing, that it’s a fear for them, it’s a calculation for them. And this is a tacit admission, I guess, is kind of what I’m pointing at. And it to me it’s just not a reveal, it’s a bit of a tell in a larger sense. Because for all of its—you know, the debate around the implementation of the UBI, creating cultural and policy consensus was one of the goals of everybody before the era of the tech oligarchy. It was: how do we rebalance this equation for the worker? How do we build the economy for what it’s meant to be, which is to help, in a utilitarian sense, distribute wealth in the most frictionless way possible. And their fear is actually that wealth will be distributed this way, and they want to stand sort of against history and or thwart history and say, no. They want to stymie this progress. And I think that is—that’s them revealing themselves.
[01:12:03.760] Mike: Your diagnosis is right. I think though the timing is not particularly interesting. I mean, the real thing is that they’ve been doing this for over a decade, right? So you have Andrew Yang, for example, talking about UBI. Obviously Andrew Yang currently—
[01:12:25.199] Nick: —aspiring tech bro.
[01:12:26.560] Mike: Right. Yeah. I mean, he kind of was because he had a company, but he kind of always has been, I should say. Yeah. Right. But it’s not just Yang. I mean, actually, OpenAI—I forget the relationship to this other entity, but I want to say there’s some research arm in this EA-adjacent philanthropy universe. I forget who. But I believe Sam Altman was connected to the study. They did a UBI study, and this was a while ago, right? So I’m saying like—
[01:13:05.920] Nick: I think I remember what you’re saying. This was, if I’m not mistaken, Y Combinator taking pitches for “design your own UBI distribution system, bring it to us, and we’ll build around it.”
[01:13:17.439] Mike: Yeah, yeah. I’m sorry. For some reason I thought it was Open Philanthropy. But maybe Open Philanthropy did their own. I’ll double check that.
[01:13:24.479] Nick: Well, I think this was in the water, but yeah.
[01:13:27.520] Mike: That’s why I had the Sam Homeland connection in my mind. Why Y Combinator makes sense. Tech has explored some notion of this for the last 10 years because they were always afraid of the “revolt,” and they were looking for the bread and circuses even as early as the late Obama years. They have been trying to get ahead of this, but almost always through placation and words. Their preferred UBI strategy in many cases is just having a social safety net floor and saying, “Hey, we have a lump sum payment you get every month. That’s all you need.”
[01:14:16.319] Nick: Even as we are currently fighting against the longtime, misformed policy that is minimum wage, they’re like, “Okay, we’ll give you that at least—everybody minimum wage,” and then we just suppress that as much as we can over time. Which isn’t what UBI was originally conceived of and should not be. Trevor Burrus, Jr.
[01:14:35.359] Mike: There’s a whole universe of UBI options, and a genuine debate would look way more multifaceted. Arguably, though, the Y Combinator study did actually provide some good ammunition for a more expansive UBI program. I don’t remember the details, and I can’t go into super great detail, but I remember being surprised at the time that a tech company or a tech-adjacent entity was even involved in it.
[01:15:00.960] Nick: It didn’t feel like an organic fit. It felt like this was sort of inserting themselves. This was the
[01:15:09.359] Mike: Frontiers moving in again. They hired researchers. And you know, with some of these entities that are nonprofits in the space, they do care, right? So you have arms of these entities that actually care about the issues that are engaging. I mean, to go back to the AI story and what we were talking about in these buckets of people, you’ve got people who actually care about the thing that the state, right? So you brought up early on Sam Altman’s ouster from OpenAI, right? That happened because at OpenAI there were people who were doomers who genuinely believed that AI could kill everybody. Very depressed moralists. And they were like, “Sam Altman is antithetical to the mission of preventing that from happening,” right? And even amongst people at Anthropic, who are probably amongst that bucket one and bucket two, you know, there are people who believe AI actually could be a force for good and they’re trying to find good applications for AI. There are people who believe AI is a force for bad and they’re trying to do AI safety research, right? You have all these amalgamations of different types of people working together in the same entities, sometimes resulting in outcomes that are kind of contra to what you would expect the organization or entity to actually pursue, because they’ve got an arm within their institution or whatever that is aligned in this way.
[01:16:38.159] Nick: And I think I’ll make a guess, I’ll make a bet here. I feel safe making that bet. At some point, this UBI program, as it was being designed, would have paid people out in crypto. That’s just a guess, but it feels like the kind of solid top-down manipulation that was going to be attempted. I think there
[01:16:59.119] Mike: was a premise behind Sam Altman’s WorldCoin, which I barely understand. Yeah, yeah. I mean, WorldCoin is separate from the Y Combinator UBI study, so they’re not one and the same, but like, you know, Sam—a kind, though. Yeah, but with Sam, a lot of his stuff just reads like grifts. If you know about his first company, Loot, at Y Combinator, Loot
[01:17:22.159] Nick: basically the audacity of naming a company Loot, by
[01:17:25.039] Mike: The way he did it is not lost on me anymore. Basically, it lied about the subscriber numbers and they were inflated, so that’s how he kind of put himself on the map, right? So Sam has always been kind of like a bit of a schemer, a scammer, right? Even more so than—again, I don’t want to make this seem like I’m an Anthropic fanboy, I’m not—but you know, Dario Amodei actually is not a schemer. He literally was probably obsessed with this AI question, and just like the Yudkowskis of the world, who, again, as crazy as these people sound, to some extent actually believe something. Now, to what extent that has kind of corrupted their broader sense of understanding of the world, and to what extent are the incentives that reward them for having those beliefs? So Eliezer, for example, is somebody who now gets rewarded for repeating this message of doom, even as it’s become apparent that at least for now LLMs are not going to be the pathway to whatever it is he’s—as the public is
[01:18:35.279] Nick: increasingly aware and skeptical of these pitches and promises, he is still pushing the doom message with greater and greater intensity.
[01:18:43.680] Mike: Right. So there is this tension—not to focus on any one particular person—but these people genuinely came to this moment as they were with these beliefs, right? But someone like even Eliezer, who does not produce anything as far as I know, is making, I believe, $400,000 or more with his nonprofit, right? So now you could argue there is a question of whether or not it’s just
[01:19:09.840] Nick: his paycheck for working at the nonprofit.
[01:19:11.920] Mike: Yes. It might actually be $600,000. I don’t actually know. In the show notes, listeners, every correction I will issue, I will specifically say all the information I’m saying here so you have the correct version of
[01:19:24.239] Nick: the errata of
[01:19:25.439] Mike: my hazy memory of all this chaos of the AI bubble.
[01:19:31.680] Nick: This is the sea in which we swim.
[01:19:33.119] Mike: Yeah, but there’s a question for me. It’s like, okay, this guy clearly came to this vision—when you read his early writings, I think he genuinely believed this—and then now he’s getting paid $400,000 to $600,000 to say this, and it’s obvious that it’s not true anymore, and he’s not updating his point of view. Well, maybe the incentives now have captured him such that even if he doesn’t believe this anymore, he wants to keep saying it because he gets paid to say it, right? Same is true for Dario and you know
[01:20:01.600] Nick: He’s not as locked in, but he’s essentially locked in in his own path. Aaron Ross Powell
[01:20:05.439] Mike: They’re all locked in in some way, right? Trevor Burrus Because
[01:20:07.600] Nick: Their reputations are stinked on it.
[01:20:09.680] Mike: Right. Their money. They’re little like
[01:20:12.319] Nick: Legal culpability, all these things
[01:20:13.680] Mike: Are locked in. I mean they all dovetail. Trevor Burrus, Jr.: They’re locked in these buckets. Right. Yeah. Some of these we can’t actually know what they are because it requires mind reading, but like some of these you can probably assume, you know, like if you’re making that kind of money, you know, it’s probably hard to distance yourself from access to that amount of money, right?
[01:20:28.960] Nick: I also would hazard a guess, and again, this is just an assessment. Yeah. Based on their skill sets and their training and their, you know, their what school and where they got their degrees or not, I don’t think they have real separate skill sets. Like they’ve come into a place and a time where tech was sort of on the ascendancy and they went all in because it made sense and they had a certain expertise that worked. But they can’t exist in any other frame or in any other era. Or at least not in the same way as they would as they’d like to. Or they couldn’t be the expertise, they couldn’t be the gurus that they have become. I
[01:21:11.199] Mike: I think that’s right. I think this is an environment that is perfect fauna for this exact type of—call it a grift, call it a vision gone too far, right? Whatever it is. And they all came here with their particular idiosyncrasies, whether they were true believers or not, but the result is that they ended up pushing the same false narrative. As it diffused to the media, to the extent that it did, it kind of just became—look, we’re not going to do the work of picking apart these different buckets, these different people, but many of them are directionally saying, either cynically or not, that AI is capable of some great things. We’re just going to repeat this over and over again until you’re scared about losing your job and whatnot. If
[01:22:06.239] Nick: You flood the zone enough, and it creates a kind of inevitability in the minds of the working class and the middle class audience for this. And that inevitability is intimidating; it’s terrifying. Using targeted phrasing and rhetoric—like “the permanent underclass,” even those words coming out of my mouth, I’m a little bit scared. Even as we’ve been sitting here parsing it and tearing it down, it’s a potent set of ideas and phrases, especially when paired with the even more terrifying and potent rise of AI narrative. I think this is a healthy place to wrap up. Also, immediately after listening to this podcast, please check out Ed Zitron’s work. Ed, come on the show anytime.
[01:23:03.199] Mike: I’m gonna keep logging it until it works.
[01:23:07.279] Nick: A lot of the bubble popping and myth busting—I didn’t say myth busting because it’s probably copyrighted—has been done by Ed and is incredibly valuable to me and to Michael as well. Is there a particular one you can point them to? I’m trying to think on my own… just type them into YouTube if you don’t know his catalog. I’ll put
[01:23:33.199] Mike: something in the show notes that it’s a good starter. I can’t off the top of my head think of anything, but stuff’s moving so fast that he actually showed up on Bloomberg, which is mainstream. He’s now mainstream. We’re at the height of the bubble because Ed Zitron is on your dad’s Bloomberg terminal. Yeah, he’s on your dad’s TV trying to speak truth. I think people are waking up. I will put those in the show notes and all the other references I’ve been making, half-remembered or totally remembered. They’re all there for sure.
[01:24:09.520] Guest 2: uh
[01:24:09.920] Nick: So yes, thank you again for joining us on this episode of The Last Enclosure.
[01:24:14.720] Mike: I’ve been Michael.
[01:24:15.840] Nick: And I’ve been Nick, and we’ll see you on the next one. Thank you again. Take care.