Season 1 Episode 6: Meet the AI Resistance

Nick: Greetings all. Welcome to this episode of The Last Enclosure.
Mike: I’m Mike.
Nick: I’m Nick.
Mike: Last Enclosure is our way of helping you navigate the confusing world we’re in and understanding the political and economic stories that are enclosing our minds in order to give you the tools you need to dismantle those enclosures.
Nick: And feel better and less anxious, or at least less bad about yourself in the context of hypercapitalism. That’s the goal. You’ll let us know if we’ve achieved that goal.
Mike: So what are we talking about this week?
Nick: I believe that this episode is Meet the Resistance.
Mike: I think episode three was us doing the people who were boosting AI. Right. So we had buckets of contributors. Buckets of miscreants. Miscreants, yeah. Right. Yeah. So this is our bucket of books.
Nick: We’re going to get canceled like Hillary Clinton up here.
Mike: I didn’t call them deplorable, so it’s fine. That’s right. The only D word shall not be used. That’s right.
Nick: That’s right.
Mike: But yeah, these are buckets of joy, or at the very least, buckets of your everyday people who are pushing back against excessive AI developments, giving you a sense of what the landscape is. I think by comparing this episode to episode three, you get a lay of the landscape about how AI is evolving and what kind of fights we’re having so that you can be better informed and make your own decisions about what AI is and how it is acting in your life.
Nick: I think as a jumping-off point to the previous episode, which will be available in that order, we examined the odd arc of Bernie Sanders in becoming aware of AI as a social and technological phenomenon, and then, dare I say, fumbling the bag—in the sense that he became a useful entity for many of these AI boosters while he was trying to resist it. So that’s an example of “don’t be like Bernie,” but maybe be more like the people we’re going to examine today.
Mike: Right. Yeah.
Nick: So whose Top of the list here? Are we going to do this ascending or descending?
Mike: Yeah, I don’t know. I mean, what kind of—let’s leave with your experiences of AI resistance. Do you know anybody in your life? Can you think of anyone or anything or any stories, anecdotes?
Nick: I mean, the bias in my existence—Michael and I are both proud residents of the Bay Area—is most of my friends are survivors, if not actively in favor of this technology.
Mike: I know who you’re talking about. Yeah.
Nick: Yeah, yeah. Many are our beloved friends who shall remain unnamed. I am actually the most pleased by a lot of the political figures in California, obviously, as they have to deal with it because Silicon Valley is here, the companies are in San Francisco, and so there is kind of a cohort who have not been incidentally co-opted the way Bernie got. And I want to applaud them. I guess that’s my starting point.
Mike: Yeah, so I mean for me, this is kind of a difficult topic because there’s a lot of nuance that I think even this episode will not provide. One thing that makes it so hard to talk about is that as we established throughout this entire series, AI is a bunch of different technologies, and so what does it mean to resist AI? What does it mean to be kind of pro-AI, right? Like there’s a cacophony of different pieces here that inform what it is people are reacting to. And so what you end up getting is sort of like this inconsistent mix of people who are doing different things. In some ways the notion of there being a resistance is sort of wrong, but in some ways it’s not exactly wrong, but in some ways it is wrong, right? So all that is to say is I think in this episode, kind of to set the table, we’re mostly focused on an American context around American consumers and their reactions to AI. But just know very broadly, there is actually a whole global ecosystem around AI resistance. And in places where the AI supply chain is dirtier—you and I watched, I think a few weeks ago, we watched Ghost in the Machine, a documentary about the AI supply chain. A lot of people in that documentary were talking about the dirty supply chain of AI, right? People who are in the global south who are being conscripted into very challenging labor, often reviewing or annotating sensitive materials that are psychologically harmful, right? Or you got people who are physically in the supply chain, digging up minerals and stuff. And those people are in much worse conditions than we are.
Nick: Right. The emergence of AI has drastically impacted and immiserated their lives in a real material way.
Mike: Yeah. I mean, in some ways it’s sort of the, you know, what is old is new again, right?
Nick: Who always pays the price? The global south.
Mike: The global south. And a lot of those supply chains exist because they were already being used for other technologies, right? So in some ways, AI is not necessarily this novel agitator. It literally is just an acceleration of what has been happening, right? And the questions of who gets harmed, and who benefits—the answers are the same ones that we’ve been giving for all the other externalities, right?
Nick: Yeah, and it is true so far in this series we’ve been pretty focused on the figures and technologies coming out of the United States and less of the other people caught up in this chain, right? And so I think it’s great that you highlight our friends in the global south; they need our help more than anybody. As a point of comparison, I think we can also, like you’re mentioning, kind of restrain ourselves to talking about American figures because-

Mike: Well that’s our context, right? I think it’s fine to be open to the audience saying this is our lane, and we’re trying to help, you know, in many cases, Americans, Westerners kind of make sense of their lives, and hopefully, indirectly by affecting them, they can help us steer this ship in a better direction, right? And that’s sort of what is optimal, what is the best to do, right? So I’m fine being open to that fact, and I’m fine talking openly about that.
Nick: Oh, totally. And I’ll say as a point of comparison, there are many, many policymakers and leaders in Europe who are both skeptical and resistant to the emergence and subsidizing of AI in the European sphere. Europe as a policy establishment has always been skeptical of American tech, the companies that thrive off of it, and they’ve used their regulatory power to keep that from invading Europe as much as they can.
Mike: Right.
Nick: So I do want to put them to the side and say good things are happening there, but for very different reasons. And I don’t think we have to touch on that in this particular session.
Mike: It’ll come up a little bit. So Europe, interestingly enough, with kind of last digression before we dive in.
Nick: No, for sure.
Mike: Yeah. Europe has its own AI scene, and I don’t want to praise their AI scene, but their sourcing is a little bit more ethical in some cases, right? Like they have language models—not a ton of them—but they have language models over there where the data is sourced from public domain resources, as opposed to gobbling up tons and tons of copyright. In a lot of cases, American companies are better positioned to do that. Like Google basically has access to all the world’s information, right? So they were always going to train on everyone’s data.
Nick: It’s almost inevitable.
Mike: Right, if you’re a European startup, you don’t really have access to a Google’s worth of data, right? So, I’m not trying to say Europe is doing this responsibly and they’re always better, but I’m saying Europe is providing examples of maybe more “responsible” ways of developing this technology. There are also companies in Europe that are not responsible. So again, I’m not trying to paint a picture that Europe’s better, but I am saying
Nick: Right. It’s not uniformly.
Mike: Right. And in these kind of strands, we can see glimpses of the possibility of something that is more responsible, more thoughtful, less harmful, right? But anyway, diving into our focus in the discussion in the American context, especially around American consumers and American citizens, right? I’ve got my own list of buckets that I’ve seen online around interacting with people in real life, etc. And so our first cohort of people is probably our least organized, our least structured. I’d call them abstainers or refusers. So abstainers are people who, to varying degrees, do not want to use AI. They think it’s wrong morally, either the technology is ontologically evil in the sense that it is-
Nick: And because of the groups who are being exploited, as we kind of touched on. Right.
Mike: Right, right. Like there is this obviously intertwinement with AI sourcing and with the usage of AI, which I do agree with. And also, kind of, at a high level, when we talk about AI—and in this episode, we’re talking about generative AI, and a lot of people are responding to generative AI—AI historically is a very boring technology. If you play video games, for example, AI was a term used to describe the non-player characters, right? AI as a term is just a pretty terrible term because these technologies are not actually connected in terms of actual technical features, but they are connected in lineage, right? There is a field called artificial intelligence where people design architectures and infrastructure and code that is designed to automate processes, right? And that’s what these things have in common. But the thing that is causing our currently politically charged environment today is predominantly generative AI, which is things like your ChatGPTs, Claude from Anthropic, right? Grok, the naughty, naughty large language model, right? And there are non-language model-based generative AI technologies. I’ve talked before on this podcast about EVO from Stanford, which is basically used by biology research departments to sequence RNA to study evolution, right? Which is a very boring use case, has nothing to do with affecting people. I mean, in some ways, actually, it’s almost a novelty, right? It will help biologists do evolutionary research, right? But it’s not going to motivate the development of data centers, and it’s not necessarily even going to translate immediately into biomedical research, right?
Nick: And certainly the peer science, like you’re talking about, is always the hardest thing to monetize and should be. I mean, peer research is called out for a reason. There’s more integrity to it and there’s more of a general humane mission, educational mission.
Mike: Yeah, but there’s also just the point that this technology, even within generative AI—if we’re narrowing ourselves to that part of AI—there’s so much divergence between what is possible with this technology that our fights are hard to pin down, right? So there are people in the abstain bucket, as I’m calling it, who are like all generative AI is evil. And if generative AI was just Claude and Grok and ChatGPT, I would almost agree with them, right? But then I see things like EVO, which are literally being used by biology professors in some narrow context, does not motivate or require mountains of data centers to be built and could actually help biologists. And it’s like, okay, I don’t think the technology is inherently evil. What is happening is that there is a cohort of people who are promoting bad or inefficient use cases,
Nick: or overpromising in a deceptive way, those use cases.
Mike: No, some of these cases are bad. I mean, I really hate talking about generative AI. I really do, because it’s just like—I’m not just trying to be deferential—I think literally all of these sides have a point. You know, not that I’m necessarily above it all, but I said this online actually to Taylor Lorenz, who is this new media tech reporter, kind of whatever, but generative AI—at least the large language model thing that was developed—is a little bit inherently problematic because it’s tied up in this notion of—we were watching the documentary Ghost in the Machine—this notion of intelligence or what have you. For me, that is partially the problem, but the real problem is in the pursuit of this idea of creating a “mind” an “artificial intelligence” they’ve built a machine whose inherent use case is unscoped, right? So if you’re talking to a chatbot, there is no way for you to actually verify its output without any external scaffolding, right? And so by design, this is a technology that is structured or organized to convince you that there is someone else on the other side.
Nick: There’s a black box element to it, a Turing trick element to it.
Mike: It’s not even a black box. I mean, the parts that are black boxy don’t actually matter. The fact is that when it talks to you, all the UX and UI is oriented around you treating it like a person, as opposed to a program. Right now, there are cases where this doesn’t matter. Maybe you’re playing a video game, you want to have a character talk to you. For example, there are mods for Bethesda games like Fallout 4 to plug LLMs into an NPC, right? And so who cares if you know the masked man on the corner in Fallout 4 is using a large language model? We could talk obviously about the downstream consequences of if you’re using ChatGPT, their supply chain is dirty, they’ve got people in the global south. But just the technology itself, the Transformer and the large language model powering it, that’s a use case where, okay, fine. It’s not scoped, but you know you’re talking to a character in a game and that’s fine. But if this is being used in an ed tech platform or this is a tutor, you are now relying on that and its inherent expertise to guide the process, and that is problematic. Now the transformer architecture itself, which is just a machine learning architecture that can be used for any purpose, right? For example, text-to-speech, or in the case of EVO, the Stanford transformer model, like RNA sequences, right? That’s a scoped domain. You can verify the outputs, right? If the RNA sequences look wrong, you can say, “Oh, that’s not the right RNA sequences.” For text-to-speech, if it got the translation wrong or the audio transcribing wrong, you can go, “That’s not what I said.” Right. But if you’re using an AI as a tutor and you yourself don’t know the domain, then you know you will be led astray. And there’s really nothing you could do about it. There’s no way to fix that, there’s no way to rectify that. And so it is bad that we were given this technology in this way, right? There is something a little bit problematic about the technology. Now, my view has always been that because the technology exists, and because many people who don’t know better are going to find the technology, rather than shaming people for using it, we should inform them about this and empower them to make better choices, and in that way we begin to scope the technology, right? By educating people and
Nick: We as the user base, therefore, are better utilized ourselves.
Mike: And then obviously, as citizens, we should regulate it so that we can actively ban the use cases that are probably not good for kids, not good for vulnerable people, right? So I’m not saying there is no political action—this is not a political issue. It inherently is a political issue, I completely agree with that. Right. But given the world we find ourselves in,
Nick: It’s a deeply political process, right?
[00:17:10.960] Mike: Right. But given the world we find ourselves in, you know, a lot of the abstainers—I don’t really want to. This is the group that I feel kind of the least comfortable talking about because it’s a really, really broad bucket, right? There are people who are abstainers who themselves just say, “I don’t want to use the technology, it’s bad.” There are people who are actively sneering at anyone. I heard someone say, “If you use generative AI, you’re evil, you’re a fascist,” right? And their logic is it’s
Nick: Inescapably. Right. Well, it’s painting with too broad a brush, even.
Mike: Yeah, but their logic is obviously given the source of where this technology came from and in its use cases, right? Sam Altman is pushing this on children, whatever, right? Yes, it is inherently bad. And I agree that the origins of technology rooted in this race to AGI does have these problems, right? But I’m not an abstainer because my pragmatic stance is like abstaining and sneering at people for using it will not stop the harms. Let’s find a way to scope this, let’s find a way to, you know, I wish there was a world where we could unilaterally, as a hive mind, reject this and ban it. Right? If that were possible, I would be an abstainer. But you know, it’s not possible. The other thing is, incidentally in the chaos of this deployment of AI, there are people who are finding use cases that are actually beneficial to them. And who am I to take that away from them? Right? So, I’m not saying again, like people when you try and talk this way about AI, go like you’re providing justifications for its existence. I’m like, no, it already exists. These are the way people are using it. What I want to do is, as long as it exists, educate people about the “better” use cases so that we kind of scope and constrain this into a mundane technology, right?
Nick: This also sounds very much like the free trade, fair trade debate around coffee beans. Obviously it is semi-exploitative in certain parts of the world, like the beans from Sumatra versus the beans in West Africa. But also there are communities there who rely on this trade simply to sell their beans and subsist. And the critics you’re talking about, the abstainers, are like, “Well, anybody who even walks into a Starbucks is a terrible person.” That feels like…
Mike: Yeah. And I want to be clear, there are abstainers who are like, “Hey, look, I personally abstain. Sure. I’m not judging people.” There are abstainers who are judging people; there are abstainers who are kind of in the middle, right? And their reasons vary too. I mean, some people are mostly centered on the harms to them and their industry, right? There are people who are more centered on the environmental harms, right? There are people who are centered on the global supply chain harms, right? There is kind of no one type of standard. This also is not a left-right issue either, although people will say, “Oh, well, you know, it’s the leftists who are the ones who are kind of against the technology.” Like that article that I mentioned in the last episode, the Bernie Sanders episode, said, “You know, the left’s missing out on AI,” right? Like there are abstainers who are not on the left-right spectrum, but who are anxious about job loss and they’re saying, “Hey, you’re accelerating my job losses.” I’ll give you, I guess, like I’ll throw you a bone and say, “Hey, maybe if you were to put a blindfold on me and I had to guess people saying that word left or right, I’d say left,” but it’s not always true. And in this case we have an increasingly polarized environment where the left-right distinction on AI does not matter anymore.
Nick: The only time—this is partially joking but a little bit serious—and I think the only element of this that is definitely right-wing is the fact that Grok has been programmed to favor essentially right-wing talking points by its overlord Elon Musk. I mean, that’s been made pretty clear. So in as much as personal bias can be said to be right-wing, that’s maybe the one piece that is clearly on that side of the spectrum.
Mike: Yeah, I mean that’s with the AI bot itself, but with the people I think I’m saying it’s like
Nick: the actual users of—the actual users or
Mike: the actual abstainers, right? Yeah. There the left-right distinction may not be as valuable, right? If you give me a description of what they’re concerned about, maybe I could go, “Oh, the environmentalist might be,” but in general where lots of abstainers—because AI, the way it’s been rolled out, I think I’ve said this multiple times. It’s actually a very good quote: AI is like the guy that showed up and started pissing in everyone’s beer at the party, right? You know, the beer is yellow, it’s actually not beer, it’s piss now. AI literally was promising us the world and it just started pissing in everyone’s beer, right? So once you’ve pissed off enough people, it doesn’t really matter what your political allegiance is, you’re just gonna get a mad populace, right? It’s just that abstainers have sort of this more kind of supply-chainy, more broader notion of what it is that they are concerned with.
Nick: And their objections could be as wide-ranging as environmental or social justice and all these other various topic areas. Is there anybody who you necessarily want to sort of not necessarily cite but throw out their name in an article or something, or should we sort of leave that to the generality of the bucket of the abstainers?
Mike: Yeah, we’ll keep it general. Keep it—there definitely are
Nick: and you know them when you see them, I guess, is the point, because they’ll identify themselves in a way.
Mike: Yeah, yeah, yeah, you’ll know them when you see it. And some of the people we’ll be talking about have been engaging with abstainers, right? So they’ll kind of pop up throughout the story. The other thing I’m going to point out—and maybe this is more of a general question—is I do wonder with some of these buckets you can kind of have flexibility, people going in and out of the bucket, but for the abstainers, or at least some of them, I do wonder if we got our perfect world where, imagine you have a transformer model, you have an LLM built on a transformer model that the data was sourced ethically, right? So there’s no copyright issues. You’ve got it running on a home lab module instead of a giant ass server, right? You’ve got kind of your
Nick: which is how you came into this whole story, was by developing your own server stack in your place.
Mike: Yeah I mean some abstainers would say because my models are coming from Google or whoever, they’re still not ethical.
Nick: It’s the fruit of the poison tree; can’t be helped, yeah.
Mike: But there is a world—I mean because there are models again, like I was mentioning, that are not trained on everyone’s stolen data that can be run locally. I mean, there’re not a lot of them and they’re kind of “less good,” right? They know “less things” because there’s just less data given in the model, right?
Nick: And some of them are even outside of Europe, so you know.
Mike: Right. But like, is it that the technology is evil, or is it that the way it’s been rolled out, the way it’s been developed, the way it’s being pushed, and the way that it’s being sourced is evil, right? So my kind of open question is—and I don’t know—are there abstainers for whom you could tick off enough boxes that the technology goes from being evil or dangerous or problematic to—I don’t wanna say good, but neutral-
Nick: Maybe less bad.
Mike: Right.
Nick: In a tolerable way.
Mike: Yeah. I mean, and look, this is sort of my catch-all bucket, so I don’t know the answer to that. I’ve kind of designed the bucket in a way where it’s like maybe I can’t know because it’s too broad, right?
Nick: Yeah. It’s casting a wide net, but just to be inclusive.
Mike: Right. Yeah. But you know, I do actually know abstainers in my own life who are kind of like, “it’s mostly personal,” they’re not judging people, or ones who are kind of focused on specific issues that can be mitigated. And I do imagine there is some type of abstainer who would be like, “hey, you know, now that you’ve checked off all these boxes, it’s fine.” I guess maybe I still don’t want to use AI, and maybe for good reason, right? A lot of the use cases of today’s generative AI, the large language models especially, they’re not scoped right. So like, they’re not wrong in the sense that like, “hey, maybe we shouldn’t be using AIs as oracles to answer questions that they inherently do not know the answers to.”
Nick: If you say “not wrong,” do we have to take a shot?
Mike: Sure.
Nick: Viewers at home, listeners at home
Mike: Yeah, I want Nick drunk at the end of the episode.
Nick: Yeah, so okay, that sounds like the one-issue voter kind of phenomenon. So we have our abstainers here in this very, very broad bucket in that sense. Is there a slightly smaller bucket that we can elucidate?
Mike: No, the next bucket’s even bigger.
Nick: Oh, didn’t know we were going that direction.
Mike: Yeah, we’re just going to bigger buckets. It gets smaller, later.
Nick: Okay, okay.
Mike: The next bucket, I’m just calling plain Jane. So plain Jane is basically, you know, going back to AI pissing in everyone’s beer. Right. Well, now you’ve pissed off again, like you have these red districts, you know, they love business projects, they hate data centers.
Nick: Right. They hate data centers and they object to them in the same way that they hate windmills, wind power projects that are mostly—and here’s the irony—being set up and built in red voting districts. Right, right. So because there’s less regulatory means of the voters pushing them out.
Mike: Yeah, I mean, AI is basically a story of people who are being promised the moon, right? And so they are kind of frustrated by the demands of AI. And then in return, in return for higher electricity bills in certain regions—not everywhere, but in certain regions—they’re buying up land where it’s cheap, and some of those places actually do have water shortages. People have said that the water usage claims of AI are overblown, and that’s true actually per query, per input, the water usage is not that bad, but you have to consider that in some of these regions where they are developing data centers, they’re already having water shortages, right? So it is an additive problem, not necessarily that AI is cumulatively creating a whole new class of problem, right? Kind of going back to our point about the supply chains in the global south. A lot of those supply chains existed before AI existed; AI is just sort of piggybacking, exacerbating the harms in those supply chains, right? So water use per query might be small, but it is additive; it is an additive risk to somebody who has a data center in a water-poor region, right?
Nick: This, just to ground this example, you’re reminding me of the massive, truly massive data center complex that Elon is putting into Missouri. Am I thinking correctly here?
Mike: There’s already one in Memphis.
Nick: Sorry, Memphis. I’m-
Mike: That already exists,
Nick: My apologies.
Mike: It’s called Colossus.
Nick: It’s very big.
Mike: Yeah, it’s very big, and I mean that area was chosen because their environmental regulations are so poor, and so I think that is one of the densest regions for asthma in the country. And so the citizens there, predominantly black citizens who are poorer, right, who already have this high asthma risk, are now being exposed to—because they couldn’t get the permitting for the power, they’re using gas generators to power it. So they’re getting even more pollution.
Nick: And the asthma rates are increasing, and this is also even more problematically a very, very dense population of African Americans that have all these other compounding problems already being put on them, and then the data center comes in. So, yeah-
Mike: That’s the poorly permitted data center that uses gas power to, you know, exhaust more pollution. Right? Because there’s a world where maybe Elon did it correctly, didn’t break the red tape, and not that the data center would be inherently good. It’s being used to germinate CSAM, right? Because Grok is—but you know, it wouldn’t be adding to the problems in this way where it’s overwhelmingly exacerbating the worst aspects of these environmental injustices, right? And this is, of course, Elon being Elon—doesn’t give a shit about that. I mean, right? No. He probably is actually gleeful it’s happening, right? But what were you gonna say? Sorry.
Nick: Elon Musk does not care about black people—to quote George Bush. Um,
Mike: No, Kanye West.
Nick: Kanye West, sorry. That’s my second joke of the pod. Yeah, so this—it’s interesting that you’ve set up this even larger bucket because the complaints of the people in this township in Memphis are—this is not a highly red district. These are pretty reliable Democratic voters in a relatively blue to purple district, having the same complaints as some coastal areas where you’re really talking about just NIMBYism.
Mike: We—I mean, we literally had it in Pittsburgh here in California, right? Right. Pittsburgh is up the street from us and recording in the Bay Area.
Nick: So this is Pittsburgh, California, for those outside California. That’s, you know, plenty of development is occurring there, but there is an element of resistance of data centers.
Mike: Resistance to data centers. Yeah, data centers. Yeah, and these homeowners there not being happy with that, right? So I mean that’s why I said plain Jane literally is the bucket because it’s like you’ve got normal people who are literally kind of like against this. I mean, but I can expand it even further, right? So when I was on Blue Sky talking about this, the left is missing out on AI article. The whole point was that the EAs were setting up a schismogenesis campaign, right? They were trying to paint the left as—or sorry, they’re trying to paint basically the only people who are against AI are people on the left. And if you’re gonna be against AI, at least be like Bernie Sanders, who you know is deferential to its capabilities, right? Please, right? Don’t call it a sarcastic parrot. Don’t say it’s just a text generator, right? Please be afraid it’ll take all your jobs, right? Like Bernie Sanders.
Nick: Be fearful and in awe of this emerging supertechnology.

Mike: Either respect it or be fearful of it. Right. But when that article came out, I was screaming at the top of my lungs at Blue Sky because I had people who were on the left going, “Oh, I agree with this point. We need more nuance on the left on AI.” And I agree with you, but do not let people who are not on your side set the terms, right? They are using this as a wedge issue to hide the fact that AI antagonism is high and it’s nonpartisan. So the day before that article came out, Jeff Gearing—he’s a YouTuber, he’s a home labber like me. I mean, not like me, he has a better home lab. My home lab is terrible. But Jeff is roughly apolitical. Literally, you go to his channel to learn about the newest device, learn about how to set up something on your home lab, right? And he had a video, I think titled something like, “Getting Sick of AI,” because he was talking about—in this video, earlier in the year and a little bit of last year—so there’s something called the open source community in software. Open source basically is people pitching in, developing software collectively, right? People can contribute at will, freely, right? And so the open source community is like the commons of software. And, very importantly, actually, a lot of proprietary software is built on open source software, kind of undergirding it, right? Like there’s this meme from XKCD. Yeah, so the graphic is like you’ve got a bunch of Jenga towers, and there’s really one tiny one that’s holding up the whole block, right? And that’s open source, right? So-
Nick: Oh, as a specific example of this, the truly fantastic groundbreaking Sony game, Shadow of the Colossus, was built very much on a few elements of open source code to the point that they had to stop selling the game outright and claim ownership of it. And that’s a singular example of how this can happen.
Mike: Yeah, I mean most projects are like that. It’s just that normally what you do if you’re building something that expensive is you actually look at the licenses before you acquire something.
Nick: Yeah. You’re more careful than Sony, who is never careful.
Mike: Yeah. So the reason I bring up open source—and Jeff’s video talks about this—is because open source maintainers are being spammed with AI submissions to fix their projects, right? So open source requires people in a community pitching in and saying, “Oh, we found this bug, here’s a patch for it, here’s a patch for it,” blah, blah, blah. And there are people who are trying to earn their wings, as it were, by finding bugs in code, submitting to these repos, using AI to identify bugs and problems. The AI is obviously in some cases hallucinating bugs that don’t exist. So these people who are submitting to these open source projects with AI, many of them don’t know what they’re doing. They’re relying on the AI to say something that’s plausible and submit it to the maintainers, hoping that the maintainers go, “Oh, thank you for helping us,” right? And so maintainers are being inundated with just slop nonsense pull requests to fix their projects, and they can’t keep up. And in some cases, they’re just like, “Well, fuck, I’m out. I’m not gonna maintain this shit if I’m gonna be torrented with nonsense,” right? So, like, this is bad. These are the commons of our technology society, and-
Nick: It’s foundational to everything else in terms of how we use technology in the modern day, and yet now it’s being eroded by the slopulation of this from AI technology.
Mike: Oh sorry, Jeff’s video, which is also a blog post, is called ” AI Is Destroying Open Source and It’s Not Even Good Yet.” But in the video he does say, “I’m kind of getting sick of AI,” because the hype basically is obscuring the fact that—well, not so quietly, very openly actually—AI in its current form, the distributed AI, is eating everything that we rely on, and if we don’t fix that, it’s going to be a big problem. Right?
Nick: And so these kinds of critics who you’re talking about, are we placing them in the same bucket as West Coast NIMBYs, black voters in this memory?
Mike: It’s plain Jane. It’s literally everyone. But they—just like the abstainers—they have their own reasons for being here, right? So Jeff—
Nick: But this is across the political spectrum.
Mike: Jeff uses AI. Jeff’s not saying AI is—I mean, Jeff has, as far as I know, used tiny local models in his home lab. Yeah. But he’s not saying, “Hey, AI…”
Nick: So he’s not abstaining.
Mike: He’s saying this is not responsible. Right?
Nick: Right.
Mike: Like, “I’m frustrated with this because look, this thing, if you scope it, has some uses.” I’ll admit that. Right. I don’t know if Jeff would agree to the fact that maybe we spent too much to get what we got, right? That’s sort of my position.
Nick: We have been saying that pretty loudly. Right. This being the “a billion dollar industry masquerading as a trillion dollar industry” critique.
Mike: Right. So, again, this is a plain Jane bucket because we all have different reasons for being in this bucket, but we’re all here. Some of us are people who are—I wouldn’t say anti-AI, I wouldn’t say pro-AI either, but we’re kind of AI-accepting. I don’t really know if I’m in this bucket; I might be in a different bucket we’re going to talk about later. But the point is that you’ve just gotten the normal person invested in a way they didn’t care. Also, aside from Jeff, there are people who are definitely not political, so gamers. If anything, gamers lean more regressive. I’m not saying that’s a truism, in
Nick: their viewpoints,
Mike: yeah, yeah. I mean, as a community, right? It can be kind of insular, whatever. But gamers are frustrated because guess what? You can’t buy computer parts. Because Sam Altman—Samuel Horde—all the RAM Altman has taken. It’s actually not just his fault. I mean, literally a lot of these parts manufacturers are cartels, and so they are deliberately reserving a supply for the data center build-out. Sam Altman sort of is like a canary in a coal mine, in that his claims that we need seven trillion dollars of investment or whatever was a signal that there would be huge data center demand. But, you know-
Nick: Which is also a policy standpoint he has been pushing, thus making Sam Altman the Grinch who stole Black Friday, I guess, or will be very soon. Nobody’s getting their new Xbox or-
Mike: Yeah.
Nick: The device for their car.
Mike: The downstream consequence of that is not just for enthusiasts too. So less
Nick: Common consumers will be impacted.
Mike: Yeah. So earlier this month, or maybe it was actually late June, Apple and Microsoft were saying, “Hey, next year’s electronics are gonna be more expensive because the parts are more expensive, guys. Don’t know what to tell you, right?” And RAM goes in not just PCs and home labs like mine, it goes into everyday consumer electronics. So this is again another way in which the average person, the plain Jane, is going to be harmed by AI, and what are they getting out of it? Well, they’re being told that it will cure cancer, maybe, it will teach your kids, maybe, but it doesn’t actually do those things.
Nick: And the theme of it is “don’t worry, it’ll all be worth it.” That is the theme of the promise being given by figures like Sam Altman. And this
Mike: technology, specifically the large language models, cannot do that, right? Right. There might be some offshoot like EVO or whatever that is going to be focused on some niche problem. Sure. Right. But if you can spin out these modalities, these functionalities, out of the large language model and into a scoped technology, then we don’t have to give Sam Altman all our money to cure cancer. We would build a transformer for that purpose, right? And it would not require every data center on Earth to power it, right?
Nick: We don’t have to build the God machine, which is a definition and a topic we’ll get into in a different episode, just to have one little singular track of research solved. It’s not even overkill, it’s hyperkill. I don’t know how we can describe it, but yeah.
Mike: So I think we’ve kind of outlined this bucket pretty well. I mean, it’s everybody who is basically annoyed that they’re
Nick: You might just be in this bucket, dear listeners.
Mike: I think the average person is in this bucket. Probably the next bucket I would call AI accountability. These are people who have basically been doing AI ethics for years. A lot of these people were featured in that documentary you watched, Ghost in the Machine. And so these are people like, you know, DAIR, Distributed AI Research Institute, people like Alex Hanna and Emily Bender, and these are people who are academics who are concerned with the ways in which AI has been marketed, deployed, etc. They’re kind of squarely focused on these supply chain issues and the ways in which AI is actively harming society more broadly, right? You know, I guess you could call them abstainers, but from when I have interacted with these people on social media and when I consume their media, I don’t get the sense that they hate AI, but they definitely hate the current version of AI and the current rollout of AI.
Nick: And it’s important to make that kind of distinction to have the critique land more readily.
Mike: I mean, I do think that they encourage not engaging with this iteration of AI, but it’s not like—so in theory, my hypothetical distinction is something like an abstainer would kind of say maybe there’s no constraint that would ever satisfy making this technology good, and maybe the accountability people do have some constraints in terms of—so
Nick: There’s in their sort of thesis and their premise, there is a version that could be acceptable to them down the line.
Mike: Maybe. But maybe not with generative AI. The thing is that they also, as practitioners of AI, know enough of the differences to appreciate these distinctions, right? Look, they’ve definitely rallied against large language models, and very clearly for good reason. I, as a user of large language models in my home lab, have said they’re not suited for purpose; they are not scoped, they are not designed to do what we’ve been told they do. But given their knowledge of the field and its history, they presumably would know, for example, that EVO exists and that there is a usage for transformers that is not this and that could be salvageable, right? So maybe generative AI in the current form is not salvageable, but there’s some version of generative AI or just future AI that’s not based on a transformer that is salvageable, right? But also they’re distinguished in the sense that they have an epistemic position that is not the average person’s, right? They are in the field. They have been in the field. They’ve been critiquing the field for years. So their critique, even if there are abstainers who are in this group, lands differently because they’re coming from a place of somewhat of authority, right? Because they work with AI. It’s not just the average person who’s like “AI is threatening my job, therefore it’s immoral;” it’s someone who has a technical background.to say these are the reasons why this is a bad technology. And then maybe some of them do actually think there’s some way to salvage it. Maybe some people in the group don’t believe that. I mentioned DAIR as one example; there are other people who are not part of DAIR who are in this group of AI accountability, right? But my notion of what I’m pointing to is somebody—probably an academic, but not necessarily—who has the requisite background to understand technology at a technical level and is critiquing it from that vantage point.
Nick: And has a historical grounding to understand how this emerged in the first place.
Mike: Right, right.
Nick: Because I will say—and maybe this is a good moment to note this—the broadest feeling that I’ve garnered from friends and colleagues in the internet more broadly is that the current era of this AI debate kind of came out of nowhere. It was a little niche and then it exploded into the zeitgeist pretty broadly and pretty quickly. And the figures you’re talking about now are the ones who could have told others that you know better. They’re coming from a place of expertise. Mike: Right. I mean these people have been having this debate inside the field forever, right? So like, maybe decades. Yeah, so like the Ghost in the Machine documentary we watched—a lot of people in that documentary are exactly the same people I’m talking about and thinking about. They’re one-to-one the same people. And within the field they’ve been saying, look, AI as it’s been rolled out and as we’ve been thinking about it for the last decade or two has just been not suited for purpose, even before the transformer thing took off, right?
Nick: Yeah
Mike: I think off-air I told you about the story of Joy Buolamwini. And computer scientists—and you know she was working on an image system. Her colleagues did not give the image system enough data to recognize black people, right? And it’s like those types because it was not trained on enough skin tones to actually detect that skin color, right?
Nick: Are you saying that the AI literally didn’t see color?
Mike: It saw color, it just didn’t see black, right? So it ignored the—
Nick: These are all sounding like DSA talking points.
Mike: brownie points for the obvious pun, but yes, I mean, the fact that society itself is now having these debates more openly is sort of just downstream of the fact that the field itself was having these debates five, six, seven, eight years ago before Sam Altman decided to monetize technology that’s not suited for purpose, right?
Nick: We are just now all having to catch up at this moment, yeah.
Mike: Right. So another bucket I would say is what I would call “scopers” — which is a terrible name, I know — but it’s sort of what I would consider myself. Maybe someone like Cory Doctorow with his reverse centuar book
Nick: I was gonna ask about Doctorow eventually, so yeah, that’s good.
Mike: There’s also a Bay Area creator, Dr. Fatima is her name. She’s actually a physicist by training, but she had two videos on her YouTube channel about AI. One of them is how to anti-AI better, because she’s actually having the exact same discussion that we’ve been having in our series. She’s talking to the “abstainers,” especially the ones who are using shame and rage and vitriol to police the usage of AI, and talking about more effective ways to nudge people towards the least harmful alternatives. A scoper, in my reckoning, is somebody who is focused on the fact that we’re not all uniformly aware of what a technology is. You cannot update the hive mind at once and get everyone to understand the environmental impact, the misuse impact, and all these impacts that exist. Even if you could, there are probably scenarios and situations where people are forced to use AI. So how can you nudge society towards a better equilibrium given that world? I hate the phrase that “AI is here to stay.” Obviously, if we had a functional Congress and experts who could get up on stage and tell us, “Here’s the technology, here’s the pros and the cons,” and we had a rational consensus about those harms—a rational cost-benefit analysis as a collective—then we could decide how to handle AI. But in the absence of that, we’re going to have this residue of local models. I think even at the larger scale, a company like Google or Amazon is probably just going to buy OpenAI or Anthropic if they go out of business. So we’re going to have this residual of AI. The sooner we have people illustrating less harmful use cases or usage of better-sculpted models, the sooner we have the escape valve we need to normalize AI in a way where it becomes boring business automation or boring normal use-case technology. In the absence of that, what you have is this name-and-shame game, which will probably affect some portion of the population, but not nearly enough numbers to actually steer away from the most harmful use cases. If consumption was a cure-all—if consumer choices were so powerful that they unilaterally could prevent harm—we would have solved climate change already. You have people like Karen Hao, who I would put in the AI accountability bucket with her book Empires of AI. She was saying that this is not an individual selection problem. Just like in the fast fashion industry, naming and shaming consumers is not how you fix the problem. You have to go upstream, regulate it properly, and provide better choices.
Nick: And it requires collective action. Right. There seem to be previous examples of these kind of things like-
Mike: This is the double movement we’ve talked about in episode two. Yeah.
Nick:This is the double movement.
Mike: This is sort of like factories were terror—this is my go-to example because I mean, we don’t really think much about factories today. I mean, you’ve mentioned multiple times, and I appreciate the fact that there are parts of the world where factories are still dangerous and that people are getting away with murder, right? But relatively speaking, you or I don’t think about factories, and so far as we do, we think of it as a policy issue, right? And that is hopefully the world where AI goes to. And by AI I mean this general generative AI, large language model thing, right?
Nick: And the infrastructure that supports it will go in that direction. Right.
Mike: Yeah. And that includes, you know, having a populace that knows enough, even if they hate the technology, to understand okay, there are tasks that it can be scoped to that are mundane and safe. I don’t keep talking about this, and I’m working on a blog post where I go into detail about different use cases.
Nick: Maybe Michael will get it out of his system at some point.
Mike: No, I mean I’ve gotten it out of my system, but the reason I’m bringing this up is because are you two? Are you also tired of AI?
Mike: I am. I am tired of AI, yes.
Nick: Guys, educate yourselves. We need to save Michael the effort of being a crusader about this.
Mike: The reason I’m reacting this way is because I hear it. There’s someone in the audience going, “What are the good use cases, idiot? What are the good use cases?”
Nick: There must be, right? And there are a couple, and we’ve cited them, and there’s probably others that will emerge.
Mike: I want to be specific, and I’m saying this could be scoped. And like, what does that mean, right? So like scoping in most cases—this is my view—I have basically developed and have not published a political compass of AI use cases, okay? And basically the axes for these use cases are: you could think of one axis as basically cognition, right? So are you offloading cognition to the AI? Are you doing a task where you are kind of giving into the AI’s innate capabilities? You’re trusting it, right? Which you should not trust it because it doesn’t know anything. It is a social, associative engine.
Nick: It is barely an it.
Mike: Right.
Nick: Actually.
Mike: Yeah. Well, it is just a shape of language. Literally it’s just a shape of language. And like you’re asking a mirror that resembles a shape of language, you know, “mirror mirror on the wall, who’s the tallest of all,” right? Like you know that you can do that, and you know, like you can even do it multiple times and not be harmed by it materially, but like that’s not really what it’s for or what it should be used for. The other axis is like measurability, right? So is it a use case that can be measured? If you can do something where you can actually have an objective criterion that you yourself can apply, and
Nick: we know it just means not just APM. Like actions per minute is not objective enough-
Mike: I don’t know if the listeners will know. So that was my metaphor in episode three, four. People feel very productive with AI because it makes them do more things per second.
[00:51:25.199] Nick: And the feeling of productivity is not the same as productivity. Right.
[00:51:29.199] Mike: So large langauge models by design, and I’ve said this elsewhere many times, are what I would call resemblance of an affective design. So affective, A-F-F-E-C-T-I-V-E. I don’t know how to spell.
Nick: That is correct.
Mike: But-
Nick: That’s my main role, is just to affirm spell check, yeah. Yeah. So AI is gonna take my spell checking job away from me.
Mike: That’s what Cory Doctorow was using AI for, spell checking.
Nick: Oh wow.
Mike: And I arguably I don’t actually think that’s a good use case, but it is measurable. Oh, okay. But because AIs use tokens, they don’t actually know how to spell.
Nick: Right.
Mike: They can spell. Wait, they do spell correctly.
Nick: A spell checker actually knows in your Word document program, does know words. LLMs do not know words.
Mike: Yeah, and it’s deterministic, right? So the LLMs actually can spell, but because it’s a non-deterministic process.
Nick: It’s an emergent property of the program. It is not something baked into it.
Mike: Oh no, sure, right. But the thing I’m pointing out is that the real problem is that it may not be consistent. Now there might be scaffold that you could deploy to actually make it more consistent, but at that point you should just use a spell checker. I mean, they’ve gotten worse.
Nick: Just call me.
Mike: I mean, they’ve gotten worse, and so I am sympathetic to needing because I mean God knows, I spell checking I’ve gotten so used to it that I’m pretty bad at spelling. I mean-
Nick: it’s a good idea. That’s ruined us all as a writer, but you know.
Mike: But you know, my point is, you know, because LLMs are effective design machines in this psychological effect sense, you will feel good when using them because that is the immediate feedback. If you do not design external feedback, if you do not design external scaffolding. So like one good example of external scaffolding is there are use cases I call fetch. So fetching is basically information retrieval, right? A lot of people just ask LLMs questions, but you could do something called RAG to actually connect—retrieval augmented generation—to connect an LLM to a source of data. And that by itself actually is not good enough. You probably need to develop systems of measurement. So, like in data science, there’s a term called precision and recall. These are terms actually that determine how accurate given a piece of data is the servicing of that retrieval, right? If you do that and you have a RAG system, you could actually benchmark how reliable the LLM is at fetching the data you needed to fetch in that context, right?
Nick: There’s a bunch of mathematicians whose heads just exploded in rage because you said benchmarks, and that’s a whole other rabbit hole for us to
Mike: These aren’t benchmarks. Precision and recall are not really benchmarking. They’re literally just—people were doing this before at one of my earlier jobs. We were doing data detection. We literally had to define content types. Like, how reliably can you build a pattern that detects a name? Right. And so precision recall are ways of benchmarking. I’m not even talking about LLMs; we’re talking about literally regex and dumber pattern matching systems. Right. How accurate is that system at retrieving the content type you’re looking for? Right. You can do that with LLMs too, right? So, you know, if you have a use case where you give the LLM a data store and you need it to reliably retrieve that specific piece of information, right, you know, you can do this. Now, which types of retrieval tasks benefit from this are probably small. I’m not trying to say that this actually saves LLMs and they’re gonna be a trillion dollar industry and that this is all worth it—and I hope all the kids lose water in the future, right? But I am saying, you know, you don’t need data centers for this. You don’t need ungodly amounts of compute for this. And the sooner we tell companies this and they find scoped use cases like this, the sooner Sam Altman loses his power. Right. And we all plug in LLMs to the most boring mundane tasks we find and stop talking about the God machine. And I could stop yelling.
Nick: And then Michael can have a good Sunday. All right, so without getting into the very large use case rabbit hole, scopers as you’re describing are the ones who are trying to literally narrow the scope and expectation of the use of this technology to a responsible, lower level than we currently have.
Mike: I think there are scopers who are more optimistic than I am. And look, even I have quote-unquote use cases that are good for me that I would maybe say don’t translate to general utility. Sure, sure. Right? Like so Doctorow in his Reverse Centaur book, you know, he identifies use cases that he thinks are valuable to workers. And you know, this is a story you’ll recall I’ve told you many times: Doctorow and Ed Zitron were at Clarington West, I think last year.
Nick: Oh, they’re great debate.
Mike: Yeah, it wasn’t—it was not a debate. They were literally—so this is the context. The story basically is, you know, Doctorow and Zitron are talking about the AI bubble, they’re in agreement, and then the moderator asked a general question of how is this going to end up? What is the AI bubble? What does it look like, right? And then “can it be a good bubble?” she added at the end. And Dr. Doctorow tepidly is like, “Well, let me give you some case for that,” and you know, that’s when Zitron kind of his ears perked up. And so Zitron was like, “Give me, you know, give me specific good use cases for generative AI,” because generative AI is the bad AI, right?
Nick: Therefore what’s the good one?
Mike: Yeah. Yeah, like, “I want you specifically because the bubble’s built on generative AI, and for this to even have any residual value, right, we need to find—you need to articulate a use case that specifically the generative AI that matters, right?” And so I think Doctorow ‘s first use case, which was kind of like pixel mapping, right? It basically helps video editors instead of having to one by one map a region of pixels if you needed to blur an entire audience’s face
Nick: For anonymity.
Mike: Right, yeah, you could do that in one fell swoop as opposed to doing it one by one, having to actually identify the regions in which this is the face, this is not the face, right? You know, and you know, arguably that use case, right, it kind of predates this particular AI craze.
Nick: There are different programs dedicated to doing that in different ways already. But it could do this.
Mike: Yeah, I mean he was basically talking about deepfakes. And before deepfakes were about swapping someone’s face over with someone else’s face, there was mundane video editing usage for AI that was this sort of efficiency for mapping use cases—or there were use cases basically that did not involve revenge porn and all these nastier things, right? But Zitron didn’t like that use case. So Doctorow gave another use case which is actually I think more interesting because it actually was a large model. So the Innocence Project, which is justice—I mean how would you describe it? It basically is a way to investigate wrongful incarcerations
Nick: And reinterpret evidence that could not be used at the time to exonerate people-
Mike: filling a genuine gap in our justice system, right? And so I think very early on—this is like a GPT-3 level model—they used to actually identify terms in cases where inaccuracies were more common. Like a misapplication of law even? Yeah, there’s certain linguistic indicators that they were able to use to actually process cases faster. Right? Again, not relying on the model to actually do the work of “is this person innocent or guilty?” Right? But using their knowledge of legal language, they were able to use heuristics they can give to a language model to help them identify cases in a way that they help them process things faster. This sounds like a really good scoping of this technology.
Mike: Yeah, I mean it’s well-scoped. The people doing it have the requisite knowledge to actually know what they’re doing, right? They’re not asking it for legal advice. They’re not asking it to be a lawyer.
Nick: It’s a functional process, not an interpretive one.
Nick: Yeah. And so that’s again, so that’s the good scope. That’s the example that Dooctorow then gave
Mike: that’s Doctorow gave—that’s the second example I think he gave. I forget the third one, but that one stuck out to me because even I at that time, because it was 2025, I was kind of skeptical of whether large language models specifically have a use case that is broadly
Nick: any humanistic value?
Mike: Not even that. There are use cases I have on my home network that are beneficial to me that I feel comfortable doing. I use it, for example, as an orchestrator. So when I have new tasks coming in my pipeline, it adds stuff to my calendar. Now, you can use scripts to do that, right? I didn’t know programming, so this is one way where I added an efficiency.
Nick: Most people don’t know code, so of course that’s a broad use. Right. Yeah.
Mike: But you know, this is a case where presumably they’re using the technology as is, right? They’re not necessarily—maybe they did build scaffolding around it—but they are. I mean, this is GPT-3 era, so this is pretty early on. This is before we had Claude Code and all these harnesses, right? So again, scope matters, right?
Nick: And that was Doctorow’s point though.
Mike: And he doesn’t use this word “scoping” as it’s really invention of me.
Nick: You’re kind of-
Mike: Applying it retrospectively, drawing the circle around myself and saying these are my friends Dr. Fatima and Corey Doctorow. I know them very well.
Nick: We’re all in the resistance!
Mike: But yeah. I mean, long story short is I think—and having located myself in this bucket—teaching people about use cases and giving them the option—not mandating it, giving them the option to have an informed approach to AI—I think takes away Sam Altman’s power. You know, now there are abstainers who argue that if you provide good use cases, you’re kind of saving the technology from itself, right? You’re actually saying instead of making it fall and implode and say, “Oh, it’s completely useless, it can’t do anything.”
Nick: So maybe don’t throw the baby out with the bathwater. There’s something here. This sounds like the most nuanced of our buckets so far.
Mike: I’m not saying don’t throw the baby out with the bathwater. I’m saying my position—I don’t know what Doctor or Fatima believe for them, other scopers—but my position is something like this: my fear is the residuals of this bubble. There won’t be many, but I think large language models actually survive the bubble.
Nick: After the AI bubble has burst, what’s left? What do we work with, right?
Mike: Not what do we work with? What will remain? I think among the wreckage are these large language models—both the ones that can exist in the home lab like mine, right? I’m running it at home, only increasing my electricity bill. And ones that will run in servers because Google and Amazon are probably going to survive the bubble, and they will buy up the remains of all these large language model companies. So my fear is if we do not come up with a narrative to scope the technology now, we give these people the time to scramble and come up with a new narrative that they can push on the public. An educated public will be able to resist those narratives and go, “Oh, I know this is a dumb machine for sequencing tasks,” as opposed to whatever inanne lie they’ll come up with next after the bubble and when the economy recovers. So I’m saying my view is: let’s not throw out the baby with the bathwater. I mean, I guess that is a function of my view—that is what will end up happening. My view is sort of like the residuals probably will include large language models from this bubble popping, and rather than us pretending that the bubble will take the technology away and letting these guys set the narrative for the next go-round, we collectively educate the public about not just large language models and transformers and the current generative AI crop, but the history of AI and this obsession with automating intelligence. That way, next time the tech industry tries to come up with the next big AI thing, the public knows: “Okay, I know what the purpose of this technology is. I know that the next iterations of the technology may not be any closer to what they’re saying. I know that the field’s been pursuing this for years. I know that there’s actually been dozens of AI bubbles before where these promises have been made, and I know enough now that this will not have an effect on me.” And maybe some of the people will find usage of large language models that actually are, quote-unquote, not throwing the baby out with the bathwater. Sure. And maybe that’s good. I don’t actually care—I feel fine, I feel not immoral for using AI in the way I’m using it. But I don’t really care about that. What I care about again is educating the public in a way where they at least understand this goes into the round hole. And when we put it away, no one else can take it out and use it in the wrong way.
Nick: We just dump it to crump it. Yeah. I am confident too that information is naturally diffusive, right? There will be a more educated user base and citizenry in the future. I also feel confident predicting there will be a new crop of different tech bros trying to package the same thing or a similar thing again and overpromising the moon. But maybe it won’t be, “Hey, we’re gonna achieve superintelligence with this,” but it’ll be “Hey, we’re gonna make your wallet into a different wallet” or whatever. You know, there’ll be a thing next. There’s always a next trend.
Mike: It’s probably quantum.
Nick: Probably.
Mike: It’ll probably be quantum, and they’ll probably say with quantum we can actually make a real intelligence, a real boy, right? “Oh yeah, you’re a quantum system and your intelligence is”
Nick: Oh god, I could do a whole quantum rant. Let’s save that for its own episode. I’m excited for that, for sure. Wait, we’re getting off topic from your very good point about Doctorow and this particular bucket. Does that satisfy this definition, like he’s kind of in that moment?
Mike: Yeah, I mean, for him, I think there are actually kind of positive use cases. I don’t really know. I mean, he feels confident using Ollama, for example, for spell check, which again, I don’t think that’s a good use case.
Nick: But then Zitron jumped on him at this discussion.
Mike: Oh, but so him telling us he uses his for spell check was way after the Zitron discussion. Oh, okay. During the Zitron discussion, the only thing that came up was just, “Hey, I’ll make a case for the bubble having some residual,” right? Not even because I’m happy about it. I mean, Doctorow and very open that this is a complete waste of everyone’s time and resources. Yeah.
Nick: Right. But correct me if I’m wrong, he also said something to the effect of there’s always a good bubble and the bad bubble. Like there’s always something that gets passed on and is the result of it.
Mike: It’s not him, but like it’s a very common common form of economic analysis.
Nick: Yeah. Right.
Nick: It’s kind of Schumpeterian.
Mike: It’s Schumpeterian, but it is there—actually this is actually a field of economics and business.
Nick: Disruption economics, essentially.
Mike: It’s basically like technology business cycles. Carla Perez is like one big name I can think of. I think she actually is one of the people that coined this notion of a GPT, a general purpose technology. And that’s like a technology that diffuses because it fulfills so many use cases.
Nick: Yeah.
Mike: You know, whether or not—I think we talked about this in episode three or whatever—whether or not large language models are that. I mean, probably they aren’t, but transformers might be.
Nick: Assuming we don’t know yet, and we need to be prepared, yeah.
Mike: No, I mean like we kind of know in the sense of like transformers can be trained on any type of data, but what matters is sort of the scope and the scale, right? Like, is this a use case? Again, like going back to my scoping criteria, is this a use case that’s measurable, etc., right? You know, because large language models themselves kind of enclose the transformers, it’s the point I made before. It’s the big labs enclosing transformers and then presenting us with this technology saying, “this is the way you must use transformers, right? You must go through us to use this technology,” right? Whereas if you do what Stanford did, right, you can build your own model for RNA, right? You can build your own model for talking to whales if you want to. You can build your own model for, you know.
Nick: What you’re describing is a technology that is specifically designed to defy scoping in a way. Like it is made to obscure the attempts to scope it at its sort of fundamental building blocks. And that’s the abstraction of language, and it gets very, you know, sort of large and unwieldy to measure and to deal with.
Mike: Yeah, it’s not designed to be scoped because we’re being sold a machine that could do everything, right?
Nick: Right.
Mike: That is the whole point, right?
Nick: This is the nature of the fraud. I always call it a fraud.
Mike: “We’ve solved intelligence, ask it to do anything.”
Nick:Right.
Mike: Right? It is like a human person in the room with us, ask it to do anything.
Nick: This ain’t your grandpa’s Ask Jeeves. This is the whole new-
Mike: The actual thing that language models are kind of in—and in IT and software, we have these terms called best of breed and best of class, or best in breed and best of class. So best in breed is a thing that does one thing really well. It’s the best of its kind, right? It’s one tool that does this one thing really well, is the best of doing that one thing. Best in class is like we’ve stapled together a bunch of different softwares together, and we’re offering you this package that does a little bit of everything kind of okay, but it covers so many ranges that you’ll be good, I think, right?
Nick: So it’s breadth, not depth, kind of
Mike: Yeah. And I mean it’s kind of a joke because I used to be in marketing and in security software. The joke is sort of best in class—you got a lot of these larger software companies, they just buy up their competitors and they staple their products onto their core product and they sell it.
Nick: Things that weren’t even designed necessarily to go together nicely or modularly are then being sold as a bundle.
Mike: Right. Or even to single package. That’s what large language models are. They kind of have all these different modalities of prior machine learning technologies, right? If you’ve got these multimodal models like GPT-4.0 and onwards, they can read images, which is OCR—optical character recognition. They can do translation because they’ve been trained on different languages. They can do all these different machine learning tasks. Before, you would actually have to have different models from scratch to do all these things. They can kind of do all these things. They do them okay, but because they’re probabilistic, they’re not reliable. These things try to give you an everything machine that does a little bit of everything, but very mediocrely. Unless you scope it, you’re not going to get much value out of it. And even if you do scope it, if you’re not measuring the task efficacy, then you don’t even know. The effective design of the model user feedback loop is just, “Oh, I got an output that resembles what I want, I feel good,” or “I think I’m doing things that resemble what I want, this is good.” The APM—the actions per minute—of “Oh wow, I’m doing a lot of stuff,” but did you check if you’re doing them correctly? Did you check if you’re doing them accurately? Is this actually what you want? Going back to APM coming from video games: okay, you’re able to fire your gun in the first-person shooter faster than the other guy. If your aim is dog shit, then it doesn’t matter; you could shoot more clips than the other guy.
Nick: You will still lose Call of Duty.
Mike: Right. So, you know, this is a technology, as you said, that resists scoping by design. But we can, as users, now that we know this—now that I’ve educated the world single-handedly—we can begin scoping it and hopefully the actual use cases are mundane, less harmful, and less stupid.
Nick: And boring. Let’s get to the boring point. Yeah, totally. So are there any—there any buckets that are actually smaller than the ones we talked about?
Mike: Yeah, these are kind of like I’m kind of playing with these. I really should have blog posted first. Honestly, my ideas that I write are more thought out than the ones I’ve talked about verbally, but loosely there’s a pro-labor contingent of people who are concerned with AI. Maybe Sanders goes here, but Sanders obviously is discussing labor on terms that the industry is comfortable with, right? So—
[01:12:13.039] Nick: He’s also, you know, to jump in here, he’s also presupposing, “Oh, AI will already disrupt and harm all these people. No, no. Therefore, we have to help them later.”
Mike: No, no, I know. Right.
Nick: So he’s—that’s—he’s a useful idiot. Right.
ike: Literally, that’s my point. Like he’s discussing AI in terms that benefit the industry such that I would make an argument that he goes in bucket one of episode three, you know, of the boosters, right? People who are true believers in the capability of AI, right? Because functionally, that is what the role he is playing in the debate, right?
Nick: He himself has failed the Turing test and now he’s on the other side of it and has issues.
Mike: Right, without knowing, is hecaptured and is actually—while he’s trying to do pro-labor stuff, you know—is captured and actually is in bucket one of the boosters, right?
Nick: So there are real pro-labor figures and resisters.
Mike: Doctorow goes here too. He’s saying, “Hey, the reason why AI sucks partly is because you are a reverse centaur; you’re being made to do things for a machine,” right? Remember, let’s go back to Grandpa Shelby, right? Large language models don’t inherently know anything, and you have people going behind the scenes cleaning up the mess after the model, right? If your job is to clean up after AI—here’s a less sympathetic example—the people at Meta, as of I think May or June of this year, have called the place a gulag because they’re randomly conscripting engineers to build puzzles to train AI, right? And they’re being given keyloggers to actually monitor their inputs into the computer to feed AI.
Nick: Given keyloggers, have keyloggers imposed on them to put them into this work gulag to soften all of the edges for this machine.
Mike: Right. So if you can’t control where and how you work and what you’re doing, and you’re being made to keep up with the speed of a machine as your job and to the demands of this, right? Going back to our discussion of people in the global south too, I mean, they are being fed into the machine and in doing so they are, as Doctorow puts it, becoming reverse centaurs. And that’s horrible because you’re being grafted with this-
Nick: Onto this inhuman thing. Right. Yeah. So exploitation exists at all levels of the machine, not just at the lowest point, also at the highest point. Meta certainly being the pinnacle of this thing. Yeah.
Mike: Right. And so there is a vision of this technology—well, besides it not existing—there’s a vision of technology that Docker points to and reverse centaur, where a person is in charge of how and where they use this. And if it actually saves them time, they get to actually benefit from that. Not just that the machine is not being imposed on you and you’re not being forced to work in ways to serve the machine. So it’s not your machine; the machine is serving you, and you get to reap the reward from that, right? Because there is a world where the bosses actually know the best use cases for AI, they make you do that, but all of the efficiency gains, all the time given back to you, all the productivity given back to you, it’s taken away from you, right?
Nick: Yeah, it’s sucked up to the top level, leaving you the exploited one.
Mike: Right. Yeah, exactly. So again, the technology by itself, even if it’s scoped, is not inherently good. And I don’t even think that Doctorow was arguing that it’s inherently good, but he’s again arguing at what I believe is the pragmatist view that this technology can, if we scope it and we identify use cases that actually benefit people, be less bad.
Nick: Right. And certainly this is a good moment to cite some real resistors in the Meta example. The Meta engineers who’ve been shunted into this AI gulag are certainly existing under the threat of firing. There have been massive layoffs at all the big tech giants, most harmfully, I would say, at Meta. For context, they hired a bunch of new people during COVID. They’ve been slowly trying to—they would say—rightsize the workforce. I think they’re just trying to save costs and do these mass layoffs. So the recent heroes that I want to cite here are a cohort of these fired employees from Meta who have decided that they’re going to sue and go to court for wrongful termination—that their jobs were terminated because AI chose them incorrectly. AI was used to measure their productivity and their working hours, and factors weren’t taken into account, things like sick leave, legitimate time off, maternity leave, paternity leave—all these things were ignored by the AI. In the same sort of spirit of how the LLM can’t see color, it couldn’t see that these were good workers and fired them. Now they’re suing in a class action. Those are some of the pro-labor resistors that I would say is a great example. They were brave in standing up, maybe not trying to get their jobs back, but trying to say, “I was harmed and I should be compensated.” So there are brave people in any equation, and those are among them.
Mike: Yeah, I wonder if it’s AI that they use to tag these people. I mean, who knows? But-
[01:17:39.920] Nick: Meta claims that they didn’t do that. I think the evidence will show that they didn’t.
[01:17:43.279] Mike: Right. Well, that’s yeah.
[01:17:44.880] Nick: That’s the usual line. Yeah.
[01:17:46.479] Mike: Yeah. But besides the pro-labor people—of whom, you know, I want to be mindful of this too—they’re pro-labor people not just in the US, right? So I mentioned Doctorow, who is technically a Canadian author in the US, but he’s kind of Western-centric. But there are non-Western-centric versions of this. You know, there is actually, I believe, an AI resistance kind of list that one of the Q&A hosts at the documentary we saw was talking about; Karen Hao mentioned it too. I’ll link it in the show notes, and it talks about AI labor resistance across the world, right? In circumstances that are much more dire than I would argue the Meta employees here in America. Right. They get to sue. They have that sue and the working conditions—you know, while I would never downplay mistreatment at work in terms of an HR environment.
Nick: But it’s very much by Western standards.
Mike: Right, exactly. The people in the global south are literally fighting for their lives in terms of recognition of the resources and recognition of material harms like psychological harm, et cetera. The Meta employees, if they were screwed, could probably get another job at another tech company and they would be-
Nick: certainly in demand, for sure.
Mike: So not that no one’s suffering is ever warranted, but I do want to highlight this is a global issue, and we have kind of limited ourselves partly out of knowledge from talking about the broader context and the severity actually and the dire nature of this conflict outside of the Western world, right? Because we are kind of in the global core, as it were, isolated from some of the worst effects, even though we are starting to experience some of that, right?
Nick: Yes, we live in the metropole, as our professors would say.
Nick: But before we have to beat the clock, are there any other bigger and or smaller buckets that we can touch on?
Mike: Yeah, one more bucket, I guess I call counterbuilders. Maybe this goes under a subset of scoper. So my thesis has always been—and this is why I’m a home labber, partly for other reasons, I’ve always had a home lab—but some of the logic here is we’ve been given this one technological cycle where we’ve actually been given the tools that can actually deflate the bubble, right? So it is not a coincidence that now, this late in the bubble, a lot of companies, as we said in the last episode and even the one before that, are turning to Chinese models to offset the cost of AI, right? And this is not to praise Chinese models or open source models. I’m not saying that. But what I am saying is it’s extremely interesting that the Googles of the world and the OpenAIs of the world have created open source versions of language models. Now, obviously, they don’t compete with the scale and the scope of the ones that they’ve created running on their data centers, right? But what people are learning, going back to the scoping story, is “Hey, I don’t need a Ferrari to drive down the street,”
Nick: If you’re just going one block.
Mike: And by the way, the promise that I could drive a Ferrari from New York to LA with my feet up was a lie anyway. So I shouldn’t even want to have a frontier model doing my kids’ math homework because it actually is not good at that. And there’s no way to make it good at that, right? So instead, what I’ll do is I’ll run a model on my home lab that does not interact with my kid in any way, because that’s not a good use case for a lineage model, right? So the local tools kind of inherently scope themselves because, one, they’re not as powerful as the large models, right? They can’t pretend to do the things that these larger frontier models are pretending to do and doing poorly, right? And so you already have—this is almost the bubble’s way of telling you, “Hey idiot, this is the way to use the technology.” The labs have given you the right way to use technology: deploy it locally and scope it to very small use cases, because that’s all you’re going to get out of this technology, right? You could play pretend with the big boys who are running these giant models in their massive data centers, but it’s not actually doing anything, right? So what makes this bubble so interesting is—again, I want to make it clear, I’m not praising open source in any particular way—but I am saying this is a lesser evil, and it’s very interesting that the companies themselves actually provided this of their own accord. Right.
Nick: It’s ironic for sure.
Mike: Yeah. Like this is a tool that will teach everyone: if you want to use language models for whatever reason, this is the proper scope. Now you need to find use cases that actually match this ambition, this scope, right? As opposed to paying Sam Altman billions and billions of dollars a quarter to pretend to have a god that doesn’t do anything, right?
Nick: Oh man, yeah.
Mike: So counterbuilders are people who are taking these local language models and finding places to use them, right? I’m not so self-aggrandizing to recognize myself as part of them because I’m kind of just playing around. I don’t even know that my use cases scale to anything meaningful, right? But there are people who are more technically inclined than I am, and they probably have found use cases that could translate to business opportunities that make sense, where the AI is scoped and it’s doing things that don’t harm people. And they are helping pave the way—actually materializing the use cases, not just pointing to them like I am, but actually materializing the use cases that will make this a mundane technology. And in some ways—and I’m not giving them credit either—the Chinese labs are kind of doing this too by simply providing the local models and businesses around the local models, but it’s sort of a separate kind of discussion. I don’t really want to include them as part of this cohort. I’m saying the act of trying to build infrastructure—sustainable infrastructure, mind you, not all of it is sustainable—but the act of trying to build sustainable infrastructure around local models is the thing that will deflate the bubble and help scope the technology in ways that are less harmful.
Nick: And maybe even land the plane of this business cycle without it.
Mike: No, it’s too late for that. Too late for that.
Nick: Okay. Yeah
Mike: We talked about it in episode four, Nick. Do you remember? Four bubbles.
Nick: You know I’m the optimist.
Mike: In this infrastructure bubble, it’s too late for that. Too late to save it. Right. The plane is going down.
Nick: Right. Plane’s going down.
Nick: I had a sign of hope, just before we have to sign off here, that the mayor in conjunction in Manhattan, in New York, has put a one-year moratorium on approving and building data centers. While that sounds like a very small legalistic change, it is buying that time that I think these local polities need, because obviously the current federal government is pro-bubble in a sense. Trump, the current Congress, they’ve all been captured in their own way. They’re all, you know,
Mike: I think people have bought into the race dynamics, right? So it’s like there’s this national security issue. Right. Even if we’re not making God, we’re making something that’s going to be an economic supercharger.
Nick: And certainly the language out of the president’s mouth and various allies in Congress is, “Oh, we have to allow this unregulated, unscoped use of this technology so that we can beat the Chinese at it.” I was like, okay, you’ve become captured, you’re a useful idiot. So critiquing those figures on one hand, I want to raise up the political leadership in New York for drawing a line in the sand and saying, “Let’s educate the voter, let’s let them have the time to have this debate before we approve of and go into debt deeply to pay for all these data centers that we don’t know we need yet.” And even if it’s just kind of a stalling action, a year, or maybe more, of letting the voters get comfortable with these ideas, I think is enough to really move the needle. So that’s my hope, and that’s one example of a local hero. And there can be heroes in any of these polities, whether it is in Memphis or the Bay Area or San Francisco specifically. There are these figures and they are trying to push back in real ways.
Mike: Yeah, I do have one more bucket. I do think there is a genuine kind of recognition for this to our listeners probably, right? So I have many of my listeners so far who are friends and family who are not as chronically online as I am, and so they’re not aware of the nature of the debate to this degree.
Nick: And they’re much happier than us. So yeah, they’re much happier, right? You know, they don’t know about people trying to nudge Sanders, you know, to become a
Nick: Third-party candidate for the moment,
Mike: Yeah. So you know, if you’re undecided, right, but you are curious and you’re exposing yourself to resources like hopefully this and other resources, you know, don’t feel like you have to have a strong opinion on this because we are all still learning. You know, I understand that as the abstainers have pointed out, this technology has many grave harms, and I don’t want to whitewash it by saying,
Nick: Well, we’re not minimizing any of that. We’re saying be informed, be skeptical, and therefore be informed.
Mike: Right. But that there is space to recognize there is a type of person who is still learning, but they are maybe skeptical, curious, right? At very least they know something is off, right? They’ve been told all these promises about what something could do, but something’s off and they don’t know exactly what it is, right?
Nick: The vibes are off for sure.
Mike: Right. So sit with that discomfort, find resources. I’m not saying it has to be me, right? That help you learn what you need to know in order to make an informed decision. And, you know, please continue to self-educate yourself; that’s the best you could do.
Nick: In a democracy, an educated electorate is a requirement. And there’s nothing wrong with the low information voter, there’s nothing wrong with the low information consumer, as long as they are open to learning, because in every field when you start absorbing it, you start at the bottom. You start at no information and you go up from there.
Mike: And I didn’t have these opinions a year ago, right? Like I learned, right? So I think it’s possible for anyone to take your time, don’t feel rushed that you have to find yourself in one of these buckets. And not that these buckets are absolute, right? I’m just giving you my lay of the land. You know, so as long as you have that inclination to learn, I believe that you’ll find the proper scope, the proper kind of perspective on moving forward with this technology.
Nick: And anybody trying to gatekeep you is trying to sell you something, let that be my summative thoughts. And thank you again, everybody, for joining us on this episode of The Last Enclosure.
Mike: I’ve been Mike.
Nick: I’ve been Nick, and please join us next time and free your mind.