Language, Brains, And The AI Mirage with Eli Sennesh

What if today’s most powerful AI systems are closer to a free-floating hippocampus than to a thinking mind? We dive into the messy borderlands between neuroscience, semiotics, and political economy to ask what LLMs really do, why they feel authoritative, and where their limits begin. Along the way, we explore how humans negotiate meaning in real time while models operate in a frozen field of correlations, why that matters for education and writing, and how the surveillance stack turns our lives into tidy sequences for machines to memorize.

Together with our guest, we unpack grid cells, place cells, and the hippocampus as a vivid analogy for sequence modeling. Then we press on the big claims: can a next-token engine think, or does it merely interpolate? Why do these models stumble on math unless we bolt on tools? And how did the training corpus—heavy with ad copy, business speak, and now model-made text—nudge outputs toward a bland, consensus voice that can be tuned to institutional aims?

None of this unfolds in a vacuum. We follow the money to examine power costs, chip monopolies, and a rush to constant capital that favors server farms over genuine productivity gains. The result looks like a bubble stitched to state-capital priorities and fragile cloud infrastructure, not an inevitable march toward “superintelligence.” If planning is back on the table, we argue it needs new objectives: replace the one-size value function with interpretable quotas for health, learning, resilience, and ecological limits, and design cybernetic feedback that respects agency instead of erasing it.

Curious about a future where meaning stays alive and tools stay honest? Listen, share with a friend who’s wrestling with AI’s promises and pitfalls, and leave a review to tell us where you stand.

Musis by Bitterlake, Used with Permission, all rights to Bitterlake

Chapters

• 0:00 Setting The Terms: Minds And AI
• 2:47 Credentials And Cross-Training
• 6:07 Structuralism, Language, And Mind
• 11:32 Theory Of Mind, Recursion, And Language
• 16:42 Chomsky, Tomasello, And Continuity
• 23:17 Sensory-Motor Roots Of Cognition
• 29:07 Categories, Autism, And Meaning
• 33:47 What LLMs Are: The Hippocampus Analogy
• 42:17 Grid Cells, Place Cells, And Memory
• 49:32 Can LLMs Think Or Just Memorize?
• 55:07 Do LLMs Learn And When?
• 1:01:27 Load, Learning, And Conditions
• 1:08:27 Feedback Loops And Social Systems
• 1:16:17 AI’s Market Capture And Misuse
• 1:24:17 Consensus Machines: LLMs As Ideology
• 1:30:57 Surveillance, Memory, And Control
• 1:38:47 Fragile Infrastructures And Risk
• 1:46:07 Budgets, Bubbles, And Power
• 1:55:17 Profitability Crisis And Rentier Turns

Transcript

SPEAKER_01: 0:48

Hello and welcome to Varmblog. And today we're here with Eli uhinesh. And we are talking about some of the structural and informational ideas um around AI and cognition. In Nicholas Viral's Soul of a New Type, which people should watch the show on probably before watching this one. Because we're gonna be uh you will often be responding specifically to that show. Um so hi Yale, what you I wanted to ask you since you're actually you actually study a lot of this stuff, what are your uh qualifications to information theory and sociology?

SPEAKER_00: 1:34

Let's see, sociology I would not say I have formal qualifications, and I would hope I haven't claimed to do so. My apologies if anyone was deceived by some means.

SPEAKER_01: 1:45

Um regarding that may have been my misunderstanding, so right.

SPEAKER_00: 1:50

Um, so regarding both information theory and psychology and neuroscience, I basically did a PhD where I worked in two labs at the same time. One was a probabilistic machine learning lab, so information theory was part of the coursework there, and the other was a cognitive neuroscience and affective psychology lab. So that's really where my training in those two subjects comes from. Then I did my postdoc in systems neuroscience.

SPEAKER_01: 2:24

Okay. Um, and my background is nowhere near as qualified. Uh I am uh married to a bioinformaticist PhD candidate, our student. I don't think she's actually at the candidacy yet. Um, and um I know educational psychology, the elephant's graveyard of psychology, except for left theory, which is the even elephant's graveyard of the elephant's graveyard of psychology, since everything's stuck in late 1920s. Freud. Um okay, so uh I I wanted your general response because I uh I consider myself influenced by functional structuralism, but I'm pretty pretty skeptical of classical structuralism because whenever someone tells me the unconscious is is structured like a language and I go, Oh yeah, prove it. Um I don't think I've ever gotten a proof.

SPEAKER_00: 3:31

So um okay, so I think it would probably be a good idea to define what you mean by classical versus functional structuralism.

SPEAKER_01: 3:41

Functional structuralism to me would be like not my field. Okay. Functional structuralism would be like we talk about the structures in an uh almost empirical way, like these are the relations, um, this is how they model, this is how the networks model. Um, and it does bear uh a it has a historical legacy to structuralism. Structuralism as a science doesn't even come from sociology, it comes from linguistics. Um but there's a bunch of early 20th century structuralists, Altousseur Lacan and a bunch of people in France who argue um that uh the mind is basically structured like a language, and that's how we have to understand that. And I have said, like, okay, that made sense to me as a 19th-century understanding of the way language works, the way like Boolean theory is a 19th-century understanding of cognition. Um, but from my understanding in educational psychology and cognitivism, those kind of assumptions have not borne out when trying to map the structures of the mind. There are structures of the mind, they're real, so there's a there is something that semiotics is getting to, but the assumption that it mirrors language or are semiotics in language is a bad or false assumption. That's my uh general understanding, and uh that goes into educational psych uh my understanding from educational psychology, but you've worked in uh you know psycho um neurology and and uh um various other forms of psychology. So, what's your understanding there?

SPEAKER_00: 5:34

So, my understanding would really be that language is sort of this thing that we evolved relatively late. And that's not because there was you know no precedent for it in evolution, that's because we needed to develop fairly advanced theory of mind before we could acquire sophisticated language. So if you have you know an organism that's earlier in the lineage, then effectively you still have you know social signaling to conspecifics, and I don't mean this in the economists' way of signaling, I literally mean you know, dogs can well not dogs, we bred dogs, but you know, wolves and coyotes, for instance, can use gestures and sounds, as can apes, to convey information to other animals. But in order for that to get elaborated enough that in order for it to get elaborated enough that you can make language out of it, firstly you need more theory of mind. And as part of what might have evolved as part of theory of mind, you do need, you know, that famous R-word recursion, so that then you can build up the individual signals into grammatical structures that you know recurse as Chomsky described.

SPEAKER_01: 6:58

Now you you have probably listened to me critique Chomsky in linguistics. Um and uh I was gonna ask you how much in the great language wars, because I tend to think Tomasello's critique of Chomsky is actually pretty hard to refute. That not that Chomsky is necessarily wrong, but that Chomsky's explanation could also be explained perfectly well by by external mechanisms or uh coherences with with forms of which the mind sees, which means that you might not have anything like universal grammar. Um and Chomsky himself has reduced like the amount of universal grammar he thinks would exist over the you know between the 1950s and when he retired. Um, but but how much do you think that that theory still holds in in like cognitive psychology?

SPEAKER_00: 7:59

In cognitive psychology, I'd say it's basically a flame war starter at this point. Like it's it's really the kind of thing. So I am thinking of a conference I was at in 2022. There was a workshop after the conference, and of course, there's you know, a party, and you know, tongues get loosened enough for people to start saying what they really think about things. And you know, we basically ended up in sort of this fight where my side is saying, you know, well, I think, you know, sensory motor processing can pretty much explain most of the major forms of cognition, you know, and that therefore bottom line for it, well, what you could call bottom line for like these more sociological views, that humans are very, very continuous with animals. And our cognition is very, very continuous with our slightly earlier ancestors in the lineage. Gradual change, no disjunctive break. And this other guy, um, I'm gonna call him John, that's his first name. I will not reveal his last name because of course, you know, I remember this, but he didn't say, oh, you can go and you know, tell people this scientific gossip with my name on it. Anyway, you know, he's quite senior, very tenured, etc., etc. And he basically insisted, no, I don't think that poetry, literature, you know, um, classical music, I don't think any of that can be explained as continuous with sensory motor processing. I think you really did need to develop this additional sort of neurocomputational building block of grammar and recursion in the very late evolving frontal areas of cortex. Which is to say that it's hard to actually give definitive evidence one way or the other, partly because it's hard to make cross-species comparisons, partly because it's hard to do, you know, invasive recordings in humans that would allow us to test the hypotheses. And partly because when you get down to the detailed theoretical level nowadays, that is the level of really trying to, you know, formally model what is going on inside you know the brain, then you know, the different uh theories or the different camps can become a little more mixed in with one another, and you need more nuanced category divisions.

SPEAKER_01: 10:35

Yeah, I mean, I I think about um this in reg I've been thinking a lot about this in regards to like Ian hacking and types of people, believe it or not, not so much in language, but that uh a lot of the debates we have around like whether or not there was autism at the rates that we have now a hundred years ago is actually as much an epistemic problem as it as it is a scientific one, because do we look for a category before the category existed? And if the category itself has a fluid you know definition over time, which autism actually does, is that because we are understanding like autism in the universe as a condition that exists outside of the set of conditions that that uh we have called it, or are we actually um changing our definition of what we're looking for? And this would have been just labeled other things in the past. And to me, the the second answer is transparently obvious, but I realize that a lot of people are secret platiness and don't realize they're secret platiness.

SPEAKER_00: 11:49

Um like when they tell me like language means things, and I'm like, they mean it language means what we agree that it means, like so these categories like like that's why you need the theory of mind in order to get language, that you need to have at least enough of an understanding of what's going on in the other person's head to think, how do I make a noise that will convey what I intend it to convey rather than conveying something completely different and throwing them off?

SPEAKER_01: 12:22

So um as uh so I have gone from a skeptic in the LLM wars to a skeptic and counter-skeptic simultaneously. So I am skeptical of the business claims of LLMs, I am skeptical of the um of the economic utility of some of the LLMs for reasons of hidden cost and video trips, this, that, and the other. But the actual scientific utility of which I used to think there was very little, I have changed somewhat. Um, so has uh Nico, uh so has Nico, actually. That's part of why he wrote that book. Um, and I I wanted to get your experience with this because uh I admit that while I do think a lot of what LLMs produce is slop, I also know that I can personally train them to do very complex tasks pretty well with sufficient feedback. Um now, I guess the question that I have for you is how is that like people?

SPEAKER_00: 13:28

Okay, at some point I'm just gonna have to compose like a formal talk on this or something because people are gonna keep asking. I would say that LLMs are sort of a lot like if you tore the hippocampus out of your head and just ran it alone.

SPEAKER_01: 13:42

Okay. So for those explain what that means for layman, but I have an idea. Yeah.

SPEAKER_00: 13:48

So, you know, for laymen, first of all, hippocampus is the Latin word for seahorse. It's this sort of seahorse-shaped structure in your brain that is involved in sort of taking your short-term episodic memories that your brain accumulates over the course of a day, and slowly, often during sleep, actually, transforming them into your long-term semantic memories. So a semantic memory might be, you know, you know what a computer is, you know what a tree is, you don't have to remember a specific time you saw a tree in order to know what a tree is. Instead, you remember something more like uh this fuzzy collection of every time you ever saw a tree. Now, what this comes down to is that the hippocampus is very, very good, or really, you know, its specialized job is to take sequences of inputs from the rest of the neocortex, you could say. Now, I can hear my advisor yelling at me, neocortex is the wrong term, it's isocortex, whatever, fine. Point being it takes these inputs from the rest of the neocortex, especially through enterhinal cortex, which has what are called grid cells. Okay, so what are grid cells? Why do I bring this up? A grid cell, if you sort of plot the firing of grid cells over a physical space, like a table or a maze that a rat's running in, or just an open field that an animal is moving in. Whenever the animal moves, say essentially whenever it hits sort of the center of a sort of little square, you could draw on the field that is on the space of movement, the grid cell fires. Actually, whenever it hits the corner of a hexagon, you could draw on the space, the grid cell fires. So what this helps to tell us is that there's some machinery in your head that is taking sort of the raw sensory motor stuff of your experience and turning it into, you know, a neat, we call it a cognitive map, like a map of physical space that is in your head. And then the sequence of locations that you move through on that map becomes part of the input and output of the hippocampus. So in the hippocampus, we'll find these nice things called place cells. They fire when the animal is moving through a specific place, a specific landmark. Now, of course, you know, that's the classical finding going back to the 1940s. You know, that's from uh Tolman, I think. So these are very old, you know, well-established, well-replicated findings. Nowadays, there's of course a lot of subtlety and arguments about, you know, are place cells just one kind of sort of, you could call it semiotic location cell? Like, you know, are there uh essentially there's all these other kinds of cells that will fire for other things in the hippocampus as well. So are they do all doing the same job in a more abstract representation, or are they all doing slightly different jobs in a concrete representation of what the scientist can see? This is all you know subtleties and debates, but one thing we can get right to is that uh a hippocampus helps to memorize sequences of events. So when you memorize a poem or a song, you're using the hippocampus. When you memorize the route to work or to your friend's house, you're also using the hippocampus. Remove the hippocampus and your ability to memorize new sequences goes away, even while you retain more or less the entire rest of your cognitive faculties. So the point about LLMs being what do they do? They memorize sequences. You know, they memorize all the sequences of text tokens in their training set. And then they can reproduce them when prompted auto-regressively. So to start criticizing a little bit, because I have to say, you know, I personally don't use LLMs because I think they're an overly fancy party trick, you know, like I don't always trust having this thing looking over my shoulder while I work, like all kinds of reasons. But you know, broadly based on how they are trained, which is you know, memorize the input corpus and interpolate on it, you know, do I think they can think? No. Do I think that they can memorize and generalize among what they've memorized? Yes. Does that do something interesting that um for instance Markov chains 20 years ago for generating spam emails couldn't do? Absolutely.

SPEAKER_01: 19:25

But it is actually still on the spectrum of a Markov chain.

SPEAKER_00: 19:31

Um it's auto-regressive, so the like the mathematical definition of a Markov assumption doesn't hold.

SPEAKER_01: 19:42

Okay, got it.

SPEAKER_00: 19:44

Like a Markov assumption would be you sum up the entire state of the world in one token of fixed dimensionality.

SPEAKER_01: 19:56

So I guess one of the questions that this leads me to is what does it mean when people keep on asking? I have also basically said I think LLMs don't think that they uh that they are memory processes that you can train on an instance. I've heard a lot of people say they don't learn. I that to me is a a more complicated question because I'm like, well, what do you mean by learning? Do you mean the idiomotor effect, in which case, kinda? Do you mean the um do you mean feed you mean sustained feedback sensitivity and memory? And I'm like, well, uh, they won't learn from instance to instance, but they'll learn in the instance.

SPEAKER_00: 20:44

I mean, like, I think, yeah, exactly. Like they'll, you know, is it learning if you only learn when the training software is running? Right.

SPEAKER_01: 20:57

Um, and you know, I will also say some humans do that, but uh on certain things.

SPEAKER_00: 21:06

Um, but I So I would actually I'm actually, you know what, I'm gonna object on behalf of humans and just sort of say learning is metabolically expensive. The brain spends 20% of the calories you burn every day. And so creating an environment in which the system can the system meaning your body and brain can make the metabolic investment of actually learning rather than disregarding novelty. You know, like that's something you have to produce the conditions for. If you bring in a child who's, you know, hungry, thirsty, and halfway traumatized by being around adults all the time, they're not gonna learn shit.

SPEAKER_01: 21:52

No, absolutely not. Because they're I mean, we we refer to this as cognitive and emotional load in the educational world, but basically it means like you don't have the processing capacity to handle the situation, you would in a more ideal situation, and that was why we did things like feed kids in the morning back when we had those sorts of programs.

SPEAKER_00: 22:14

Yeah, um, we called it allostatic load in our literature.

SPEAKER_01: 22:22

Uh remember, as I like to explain people, when something moves from psycho real psychology to educational psychology, it often gets stated in more basic ways. Uh I was about to say dumber ways, but sometimes you actually do need to state it in more basic ways.

SPEAKER_00: 22:37

Um, like allostatic is a little bit of a allostatic is one of those words where like 10 people know what it really means and properly understand it, and those people are trained physiologists, and everyone else using the word is kind of bullshitting. Like there are people who won't read a paper with the word allostasis in the title, and despite my having written one of those papers, you know, I kind of get it why you wouldn't want to do that.

SPEAKER_01: 23:08

Yeah, it's like me when I mention feedback loops, and I always forget because I get I fall into common language and I refer to negative feedback loops, but I'm like, no, actually, a negative a bad feedback loop is the positive feedback. Never mind, guys.

SPEAKER_00: 23:20

It's just well, like bad and yeah, I mean positive feedback loops can either I remember like the mental sensation of actually getting what positive feedback does. Because I feel like the pop impression of positive feedback is that you just draw an exponential growth curve and line goes up.

SPEAKER_01: 23:41

Right.

SPEAKER_00: 23:42

Therefore positive feedback is bad. But actually what'll happen is that you know the line going up will hit an upper limit and just flatten out. So positive feedback can do one of two things. It can either switch a system between two more or less discrete states, like two, you know, we'd call them attracting states. And if you put it in the middle, it'll tend to gravitate to one or the other depending on what's closer. Positive feedback can do that. It can also cause oscillations where a system bounces back and forth cyclically. Yeah.

SPEAKER_01: 24:18

And in the literature I've read in uh in social ecology, they tend to they tend to focus on the oscillating feedback loop. The other is mentioned, but it's not sexy, so it doesn't get the literature in that field.

SPEAKER_00: 24:34

Um I mean as a labor person. Like one of the problems we might have in Marxism and in socialist sort of scientific socialism generally, is distinguishing the state, you know, the state switching or the phase transitions from the oscillations. You know, we've been predicting the collapse of capitalism and a phase transition to something else for a long, long time. And unfortunately, it often seems to oscillate at something like a scale of you know 50 to 70 years, depending on the thing you're measuring.

SPEAKER_01: 25:18

So I mean it actually does kind of look like a Kondrav raid plus whatever Schupenter was talking about. Um uh creative destruction, blah, blah, blah, blah, blah.

SPEAKER_00: 25:29

Um like it's been about a decade since I I'd have to go hunt down the book I read that had, you know, Kondrave waves in it. But like, yeah, like it can be unfortunate when we you know see something that looks like positive feedback and go, ah, this system is about to destroy itself and phase transition into something else entirely. And then it reaches the top and turns back down.

SPEAKER_01: 25:57

Well, I I would say that I mean I would say one of the things I like about Nico's work is that actually does engage with information theory and whatnot, which most Marxists today uh included not including me in practice, but including me in training, are humanists. We don't really um, you know, I don't think we're uh Elaine Sokel needs to write a book about us bad because we're just using all the metaphors from science wrong. But I do think sometimes I I I even go, man, some of these Marxists really do that shit where we have a theory for everything, so we don't have to actually fucking know anything, and like we just talk about whatever. Um for me, the big one is like the assumption that the forms of capitalism that we're given uh are always about the crash, and that's actually why I started getting into looking at other forms of society because I'm like, okay, well, how long did it take feudalism to go down? What was you know, was it even a mode of production?

SPEAKER_00: 27:01

And the my argument is not really kinda, um, so um and I mean as as long as we're admitting to these things, like I am not a true believer in the neo-feudalism thesis, but I would say there are tendencies towards modes of extraction versus modes of accumulation that you can sort of see, you know, overlapping and trading off against one another. Yeah, I actually sympathetic to saying, you know, we've entered a for the moment, you know, we're in a heavily rentierized, like structuralized moment in capitalism that like you know, if you think that there's literally someone sitting there collecting rents, empirically you're not wrong, but do we understand this well enough to predict that we've phase changed into a whole different system entirely? Yikes, I really don't think we understand it that well.

SPEAKER_01: 28:12

Yeah, I I would I would agree. And I would even go so far to say, if and if we did, we might actually have to argue that prior epochs of capitalism weren't capitalism. Um, so uh which I mean, uh I've said this the other day. I mean, that could indicate that your theoretical framework is now incoherent and you have to rethink it.

SPEAKER_00: 28:33

And I don't want, you know, uh uh it could, but like the problem is you need to accumulate a lot of converging evidence in order to actually, you know, motivate a paradigm shift. Like you can't point and say there's a lot of inconsistencies, therefore we should paradigm shift.

SPEAKER_01: 28:55

No, I will you you pretty much can only paradigm shift way at well after the fact on this one. Like you'd have to, like it would have had to have been obvious that like this rentier mode functions fundamentally differently. Uh, that markets don't exist anymore, blah, blah, blah. And I just, I mean, I'd say the labor market very much still exists, um, and and stuff like that. Uh, to get it back to this AI though, and the soul of the new type. Definitely. Um, although it's interesting because I do think we have to like one of the things I liked about Nico's initial engagement with with talking about LMs and AI is like they're for lack of a better term, teleological orientation towards uh market capture is a problem. Because one of the things that that I've said about them is they're marketed for things they're not good at, and because the things they're good at would increase uh labor productivity and efficiency, maybe if you could get the power cost down, right? Um at least for certain white anything that's heavily throughput and you and you don't need a lot of quality control, or you don't care if there's a lot of quality control, see modern translation software. Um you can use it for, but but uh it I have to check it like crazy. Like we, you know, I use it now because you know, I'm actually kind of being mandated to do by school boards and stuff. Um and I have to check it constantly, and students do not know how to use it in an effective way, and teachers don't know how to teach them to use it in an effective way, and Despite the fact that, like, there's entire companies that have come to be to try to offer LLM-based products to education, they can't teach people how to use it in an effective way either.

SPEAKER_00: 30:58

Um, I think part of the problem is that when you're talking to a person, a real person, you know, both of you are adjusting the language you use to each other and to the circumstance in which you are speaking to one another. Whereas LLMs are sort of there's this popular meme, you know, showing an LLM as like a gigantic HP Lovecraft, you know, many-faced HP Lovecraft monster wearing a little smiley face mask. And like it's actually not quite that. It's more like they're the statistical average of every author in the training corpus, weighted by how much text they contributed to the training corpus. So LLMs are sort of really good at giving, you know, acting like sort of the ghost of the imagined average normie, but no one in specific.

SPEAKER_01: 31:59

Right. Well, I mean, it reads like ad copy a lot of the time for that reason. Yeah. Right?

SPEAKER_00: 32:05

Like again, you know, weighted by how much they contributed to the training corpus, you know, now consider how much of the text in the training corpus is effectively ad copy or business communication.

SPEAKER_01: 32:21

Yes. And then the the hyper recursion that the datasets now have LLM generated text in the data sets.

SPEAKER_00: 32:31

So what yes, which people have actually proven does cause model collapse, actually. So someone's gonna have to put more and more labor into actually filtering the LLM generated text out of the training corpuses for future LLMs.

SPEAKER_01: 32:55

Now, one of the things that you pointed out cognitively, this I think we should think about is learning and thinking are heavily caloric processes. Like they they take up a ton of resources. We don't think about them because no, you can't think your way into losing weight. Uh, but that's because thinking has like a steady state cost, it's really high. Um whereas other physical activities don't, they're very they have a variable energy input cost depending on what you're doing, and thus they're more manipulative for stuff like weight loss. But if you actually looked at where most of your calories went in your body, your brain eats up a ton.

SPEAKER_00: 33:37

Um 20% is the number.

SPEAKER_01: 33:40

Yeah, quite a bit, considering it's not 20% of your body. Um I I I have been thinking about this a lot because my issue with LM, particularly when people start talking about mystical super gods like uh AGI, which by the way, um, not to speculate, but I do wonder if allegedly this does not come from just doing too many amphetamines. But um I see you just your head just hurts when I mention AGI.

SPEAKER_00: 34:16

Um, so I actually I mean, here's the thing it's sort of like I know a lot of people who are very into this kind of thing and have done for you know more than a decade now. Since you know, my like initial initial training was in computing. So around the time these ideas started to become cool, they were cool with you know computer nerds who all knew each other. And the thing is, you know, of course, you can draw very dramatic conclusions from thought experiments that have these seemingly intuitive um premises. I'm trying to look up, you know, who was Nick Bostrom?

SPEAKER_01: 34:56

Nick Bostrom. Yes, yeah. Like, you know, I there's so much brain damage that Nick Bostrom has wrought. But go ahead.

SPEAKER_00: 35:06

Oh, well, what I was going to say is, you know, the problem is trying to actually prove that those assumptions hold regarding any natural or engineered system, right? Because their assumptions are assumptions. Nick Bostrom is not a neuroscientist. Neuroscience and cognitive psychology are by no means finished fields. You know, this is not Newtonian mechanics where we can retrospectively look back and say, ah, yeah, we've got the correct theory and it works on pretty much every experiment we give it a shot at. You know, this is set and done. This is settled science. No, it's not settled science. And so how the hell do you know that those assumptions hold? Could you try to build a system deliberately such that the assumptions would hold? Yeah, you totally could, but it's going to take decades of your life and you know, potentially unbounded amounts of money. And by the way, no tech VC is going to fund you for basically saying, you know, give me the equivalent of an entire university department to myself for 30 years. Which is where the overblown claims come from, right? You know, I act I'm not again, I'm not going to reveal who all my friends are, but I have a friend who started a company and was told to put the word superintelligence in the pitch to raise money. So I am not, you know, I'm not trying to say AGI is impossible. I don't think it is, because scientifically we could understand the brain. I do think that it is if it ever arrives, you know, because there could be economic barriers to it, there could be infrastructural or regulatory barriers to it, you know, there could be all kinds of reasons why producing sci-fi type AGI is just not practical, even if it's possible in theory or in physical principle. So if it ever arrives, it's going to take a lot more collective effort than anyone is claiming it will right now. There will be no one neat breakthrough that gets you there.

SPEAKER_01: 37:28

I think I think my my thought about it is if we assume the caloric inputs of the human brain and then we like map a brain that can do all of it at once, the power ratios for me. I was reading uh uh more of everything all the time, I think is the name of the book. It's by an astrophysicist critiquing uh this industry. But like we're talking about the power of a sun um for some of these for some of these transactions, and and yeah, I think maybe you could you could theoretically come up with maybe there's a way to have the power of a sun on earth without depleting every single resource that we have very quickly, but we don't have that. Um and so you know, I just I find you know, even even in my more utopian, like let's think about what this would be like under socialism things. I'm just like I still don't think we we would easily do that. That the energy costs would be they seem to me like they'd be enormous. Um and the energy costs to a AI are quite large, um, and they don't seem to be making a headway and making them more efficient. I mean, Deep Seek did some, but but I don't see a whole lot of evidence that we are getting more energy efficient on these uh on these things, and I also don't see there's many incentives for them since the profit model is actually almost solely based on NVIDIA chips being put into server farms, not on any other out-end product.

SPEAKER_00: 39:02

So, you know, um I mean the funny thing is, you know, the funny thing is speaking as a socialist, one of the reasons the industry acts this way is because, you know, basically they're trying to spend on constant capital rather than variable capital, meaning labor. Right. You know, not that labor in this field is cheap. It actually, you know, due to sheer volume of demand, it is not cheap. You can make a good living as a you know generative AI research engineer, thingy thingy. But if you thought about, well, how much would you actually need to just look at LLMs and say, well, this would never be scalable enough to run at AGI scale without consuming a son's worth of power, therefore we will invest in finding a whole new paradigm that will be you know much, much, much cheaper and closer in power efficiency to the human brain. And by the way, this is the kind of pitch that you know people put in their grant proposals and in their papers on the academic science end. You know, people will say the brain takes about like two watts of power to run, your data center takes gigawatts of power to run. Actually, I don't think it's two watts. I that's an embarrassment, I should unsay that. Point being it is on uh, I believe like two digits number of watts. Less than a hundred. You know, data center takes gigawatts, therefore, let's try and understand the brain so we could build better engineered systems. And broadly the preference by investors is no, let's just build more of the same thing we already know how to build, rather than take the risk of basically paying people to produce research papers that have you know no guaranteed success. Right.

SPEAKER_01: 41:05

Although it both early classical economists and Marxists would tell you that the preference for uh fixed capital versus variable capital is a disastrous preference in the long run because eventually market will drive everything down to cost, and you have no labor market to fluctuate around. Um which will be very bad for you.

SPEAKER_00: 41:31

Like, yeah, I mean, this is why I always try and tell people, you know, you don't have to be like a true believer, quote unquote, Marxist to just observe that everyone you know in business and industry are basically following trends that they know have to be altered. It's you know, they're following straight linear line, you know, straight lines, trajectories that they know are gonna have to be altered at some point, and they're effectively waiting until the people in charge get together in a room over whiskey and cigars and decide that it's finally time to allow something different to happen and decide to you know deploy their authority in suitable ways. So, like everyone knows they're coasting from you know, sort of fad to fad or crash to crash.

SPEAKER_01: 42:27

Yeah, I mean and and and for people who don't understand this, I think this is also why we've seen this shift, uh, not just in this area, but in general towards rents, in it's been the weirdest thing for me is arguing with like non-Marxist socialists and heterogos economists that like that the profit rates have been increasing since since 28. And I'm like, no, revenues have been increasing, but profits have not. If you look at her item, like like if you looked at the item in aggregate and looked at how much money you make off selling one item of thing, those profits are down, and it shows up in GDP. Like, no, GDP doesn't go negative very often, but like we would be as static to hit the relatively low uh GDP growth of China as a relatively low compared to its historical norms, which uh between the 70s and you know the the mid-18s was 10-20 percent. Now it's you know five, and it's a mature economy, it makes sense that it's that. But I mean, my point, my point is on this, it's just like this all affects this stuff. They you can't separate this technology from the economic basis of which it is emerging. And the other thing that we have to be honest uh about, and you know, one of the things that actually does make me think about why this set of billionaires seems so clueless, both socially and even in other things, compared to other sets of billionaires, is they did, I mean, uh not to sound like a classical economist, but there's a giant free rider problem on all these tech industries because they free-rowed off the government and then ventured capital until they were big enough to have monopsony power. They did not actually get this from uh standard production, and the one that came closest is still not, you know, as like AWS services and Amazon's back end. Um, and even that is limited on you know on what it would have Democrat government backend.

SPEAKER_00: 44:37

And the thing about Jeff Bezos and Amazon is that he's always made, you know, first of all, Jeff Bezos himself has always retained enough of a share in his firm, despite it being publicly held, that he can effectively cast a one-man controlling vote.

SPEAKER_02: 44:53

Right.

SPEAKER_00: 44:54

He always makes sure of that. And two, he has always chosen to reinvest in capturing market share over actually distributing out the profits as dividends to shares, to shareholders. Like Amazon in a lot of years does not make a nominal profit because they are reinvesting so much, and this is a deliberate choice. Obviously, Amazon.freak and com is not a charity case that is, you know, taking on it, they're not a zombie firm in what I think would be the Japanese sense or the 2008 sense. No, he's just prioritizing you know being the winner in the markets he competes in over actually returning a profit to shareholders, at least temporarily.

SPEAKER_01: 45:46

Well, this is a sense in which I think why people might might find the yeah, I I recently had Alex Focili on a talk about the neo feudal thesis, and I've had uh Nico on in a show that will come out before this one will, but has not come out yet. Um uh talking about profit rates. But one thing that is like feudal uh uh development isn't just the rants, um it's that you there's so much extraction that there's a low investment in in the future. Now, the difference between us and the feudal period is there wasn't a way to really invest uh in the feudal period. Um but there is for us, and by investment, you're like, oh, but there's all this stuff in the stock market. I'm like, no, a lot of that is wealth transfers, people. By invest, I mean investing in something that's going to produce a physical tangible commodity uh or enable physical tangible commodities to be produced downscale. I mean, you could argue with futures, it's not a direct, but it's an indirect, blah, blah, blah, blah, blah. But there's something going on there. That's not what we see right now. We're seeing basically stock capture, and yet the profitability is in the physical tangible commodities, the the data centers. And it's it's you know, it's it's overvalued AI firms to the fact that they don't even admit that it's overvalued now. Um, you know, like you got Bezos and Zuckerberg talking about a bubble, and it's weird because you have people like Derek Thompson over there at Abundance who should know better trying to argue that no, they're wrong about their own businesses. And I'm like, like the people doing this are telling you it's happening at this point. Like, what is wrong with you? Um, which I bring people go, okay, well, what does this have to do with the soul of the new type? Well, under what conditions would the kinds of things we're assuming about structural cognition emerge if this is the economic context for it? Like, I just don't know what kind of investment would lead to a more functional AI. Um, if if part of why it seems profitable is revenue cap uh revenue capture by dysfunction. Um, and and the other thing we have to admit is so far there is no evidence of increased enough productivity to justify the layoffs that have been tied to AI, if they're actually really tied to AI, which I'm actually somewhat skeptical of myself. Right? Like, I'm not sure that all these layoffs are actually about that, but but that is what we're being told that they are about. Um, and you're like, well, but the job market's still strong. And I've also been like, okay, how many people retiring? How many people are over 67? Um, and they're like, what do you mean? I'm like, the the the largest generation in the history of the United States, actually in the history of the world, is dying now. Um that's gonna make a lot of assumptions about how job growth does and does not go uh and where unemployment's gonna show up a lot harder to track. Um because you just have a lot of people who are leaving the labor force for other reasons and not necessarily being replaced.

SPEAKER_00: 49:06

Um today's market conditions in terms of inflation, unemployment, like basic goods, is that you know, rather than having no idea what might be causing any given phenomenon, we have five to seven ideas what might be causing any given phenomenon. And the trouble is working out what combination of all the known factors is actually doing something.

SPEAKER_01: 49:35

That's a very valid point. So I guess to go back to some of the things that that uh we were talking about, um uh so so uh to tie this back into the initial discussion. Um a lot of the discussion about cognition emerging from uh Nico is about issues around um what we might call information through classical semiotics. And by classical semiotics, I'm not so much mean Cesare and that sort of thing, or are Charles Sanders uh uh purse, I mean Umberto Echo and maybe Rowan Bart. Um uh what do you what do you make of those semiotic assumptions? I mean, I I think I do think semiotics is kind of uh it was a sexy field in the 80s, 70s and 80s, and it's kind of not so much anymore.

SPEAKER_00: 50:38

But uh what do you make about what do you make about some of those assumptions and and how it might affect the assumptions he's making about AI So I think that at least as regards LLMs, Nico is mostly you know pretty spot on actually. The sort of division between LLMs and humans that I would think is the interesting distinction to make.

unknown: 51:02

Okay.

SPEAKER_00: 51:02

Because the thing about LLMs is that because they have you know a training phase, and then you just run them, since they're effectively grinding out little bits of dead cognitive labor over and over and over again, and since they aren't properly thinking, you know, they could, so to speak, be the first things to produce speech while to produce speech in let's call it a fixed frozen semiotic field. Like for an LLM, the relationship between a word and what that word actually means, which of course for an LLM is you know like a point in a high-dimensional vector space, math, math, math. But the point being for an LLM, that's actually fixed. For human beings, it never has been. And that's sort of you know, the let's call that the rational kernel of postmodernism or post-structuralism that does have to be rescued at some point. And you know, by knowing about like information theory, statistics, and a bunch of these other fields that you know go into making modern AI work, like we get a more formal scientific language to talk about you know the relationship between the sign and the signified that sort of goes beyond the false dichotomy of it's either fixed platonistically or it's completely mutable and you know, always being reinvented ex nihilo. Like that's a false dichotomy, you know, it's sort of a more statistical relationship of I say a word, what do you know, and then your there's your background of what most people you've known have meant by that word. There's also your knowledge of me as an individual, what do I usually mean by that word? And then maybe there's even your own prescriptivist tendencies. But then there's also the question of like, where are you and I each trying actively to move a conversation by using a word? So all of these are sort of influences on you know the statistics of words in living conversations by living people, whereas it's all frozen in LLMs.

SPEAKER_01: 53:25

That seems to be a pretty significant difference, and uh not just in the way that signs work, but in the way that like uh social recursion works. And I this is one of the things that actually worries me about kids who offload thinking to uh to LLMs. I've been told to tell students it's like, oh, you're you're you now how you you know you you can now have the social effects without being social, basically is what I'm told to sell this to them. Like you can get feedback on your writing without being in a in a writing group that causes you massive amounts of anxiety. Not gonna get into the anxiety argument. Right. I mean you're getting you're getting statistical feedback from a variety of tokens, most of which is ad copy. Like I actually what monitor like I I have monitor writing feedback from from uh from these things. Teachers like it because it does reduce their time, but I when I actually read what it suggests, I'm like, this is wildly inconsistent and sometimes self-contradictory evidence, you know, writing feedback that you have no social context for why it's giving it. This is the point. Like this approximates the social, but since its inputs are static effectively, it doesn't actually have the damp the dynamic impacts of social intelligence, which I actually do worry will have an effect on human intelligence. Not like not that we lose it, but like we are just uh training people to misunderstand how they how they think and what they can do these things for, and thus uh having them underdevelop their own cognitive capacities. And I, you know, I know I sound like a doomer when I talk about young people, but I have reasons for thinking that. Um what do you make of that?

SPEAKER_00: 55:18

Um, I would say that eventually we're probably going to start as a society trying to pull ourselves out of the hole and developing a more real understanding. And if I can say something to hopefully preempt a little bit of the train crash and help people understand what's going on before they experience it live, I would say that in both a positive and in a negative sense, LLMs are ideology machines. And that doesn't mean ideology as in like lies, but as in they are almost literal embodiments of the sort of textual consensus reality that you might have once you know read about in say your local hometown newspaper or seen on you know the three national television channels. Right? Like this sort of 20th century consensus reality. And I think that might actually be one of the hidden political angles to you know why investment in LLMs is sort of like the opportunity to produce a consensus reality and then be able to fine-tune it to you know your liking as a corporate or individual social actor.

SPEAKER_01: 56:45

Yeah, and it being more integrated in your life than television was, and thus more total. Like that. I I I will say that the area where I'm convinced where LLMs are both effective and damaging is large data sets and um and surveillance. I actually do really think that. I mean, like, you know, and if I think about what the hippocampus does, it makes sense to me. Um to go back to what we were talking about earlier.

SPEAKER_00: 57:12

Um right, like why would you want to feed, you know, every photo you ever take of yourself into a giant hippocampus that's going to memorize them and make all the symbolic and semiotic connections between them? Like, you know, that that's not Big Brother is watching anymore. That's you are performing for Big Brother, whether or not Big Brother is even there. Like you're now designing your life as this series of you know selfieable moments.

SPEAKER_01: 57:44

And if you think about it from this point of point of view, in a real sense, to me, it does make sense that that the social media people got in on this. Um, because as I I commented, I think it was maybe 2013 or 2014. I'm like, man, the liberals I know not care about privacy anymore, and also are encouraging people just to give it all up for monetary reasons in a way that would like I mean, what would the Stasi do if they like the TikTok and and and and Facebook algorithms? And uh and I and I pointed this out to people and they're like, Oh, I don't know what you're concerned about, and now they're freaked out because, like, oh, where the government's involved in this and they're using this. I'm like, this this was always possible. I mean, this total information awareness network stuff going all the way back to the Bush days, which they've done but had no means of correlating or doing anything with, and that was back before every household had broadband internet, right?

SPEAKER_00: 58:47

Like that was when broadband internet was a luxury for upper middle class people in western countries, right?

SPEAKER_01: 58:53

And when I tell people what your broadband internet stream can tell them about you, like they know what if they want to, they can figure out what room in the house you're in off that, like and people think, oh, you're being paranoid. I'm like, No, I'm not, I know how it works, like um, you know, um uh it makes movies funny. I was I watched uh one better after another a couple weeks ago, and I was like, Oh, they have to pretend there's a bunch of cars from like 1992 laying around so that they can do any of the stuff they need to do to make this movie work. Um because good luck with hot wiring andor not being surveilled in a modern car.

SPEAKER_00: 59:33

Um so I mean, funny thing, isn't it? Um Hyundai and Kia, I'm told, you shouldn't buy nowadays because it's so easy to crack into their ignitions. They're like they have digital fobs that you carry with you, and then you know, I think the RFID tag or the Bluetooth pairs up with the car and allows you to start the car, allows you as the legitimate owner, quote, legitimate owner. As in, it's not you know a physical key with lock and tumbler.

SPEAKER_01: 1:00:06

Oh, there's those are rare now. My Subaru doesn't have one anymore either. Like, but but some of them have really weak.

SPEAKER_00: 1:00:12

I own the model of car that I own because when I went to the dealer to get a car for the first time, I said, don't give me any of those digital fobs, I want a physical key. And that narrowed it down to like two models in the entire dealership.

SPEAKER_01: 1:00:25

Yeah, like an Elantra has one. Um, and I I have I do think people should be slightly alarmed about all this. I mean, I really do.

SPEAKER_00: 1:00:36

Like, um uh I mean, much more likely, I will say, you know, much more likely than a like super. super totalitarian ultra surveillance state that you know would have that would make you know mid-20th century regimes giddy is uh something more like what we're getting with these Hyundai's and Kiyas where effectively if someone cares to put in enough effort they can break the system and effectively you know violate the social relations that are supposed to be the legitimate function of these systems. And then they can steal your car. But they can also break the system spy on you. So you're going to get like a lot of criminal stalking charges that make use of new digital means. And then yes sometimes when you have like you know a stupid ass small town sheriff who doesn't care what's legal or what's not then he might just you know decide to watch where you're fucking going all the time via your cell phone or something like that.

SPEAKER_01: 1:01:52

Like there's a lot of things about these where they're supposed to be very locked down and effectively only available to powerful governments and they're just not actually locked down that well yeah well I mean the other thing I think we're learning is powerful government systems aren't as locked down as we thought either. I mean right exactly like like I know that like most governments have their own uh their own uh internal uh network systems but their their public facing ones are are so dependent on non-governmental systems like AWS that when AWS goes down a whole lot of government functions and educational functions go down canvas went down Zoom went down like if I had been in I I was off that week but if I had been in person I couldn't have done my job like even though I have internal systems for FERPRA reasons so much of this is still interfacing with non public systems that if those non-public systems go down or they're compromised or whatever I can't do my job or I don't know I mean I've been thinking about this I'm like what does FERPRA mean in the age of AI? Because uh we we have been trying to figure out what to do about that because on one hand they want us to use AI more on the other hand they in my state this is not all states there are some states that are basically scraping kids' data like crazy and but which I consider a violation of federal law we actually do have lists of programs we can't use including Chat GPT because it scrapes data so thoroughly um that's almost definitely illegal if you could just get a judge to actually basically if you could get standing for a suit and then get a judge to understand what the fuck is going on that's almost definitely illegal. Right. And I was even hearing uh I went to uh ISTE a year ago for those of you know what ISTI is is the uh the educational uh the institute for educational technology uh institute for the study of uh of technology and education I think um but anyway uh it's basically a giant ad for these corporations particularly with AI just gonna point that out there but there was a lot of discussion about like with these data concerns that schools would need to develop their own AIs and their own internal AI systems and then we're like with what money and with what power like if you think it's expensive to have teachers like if we have to do our internal AI systems like we have to do something that I don't know that the most valued uh revenue producing um company in the world right now can't make profitable because it's so expensive like so um so here's where I'm I would like to actually ask you a sociological question which is where the hell is the where the hell are the actual budget constraints right now because and I don't mean that in any fine grained sense I mean really if these things cannot be made profitable then why don't they stop it seems like for lack of other viable profitable investments that everyone is taking a venture that these bubble this bubble will eventually be profitable even if it's not as profit as it is right now sometime in the future but there's lots there's lots of evidence that that is breaking that like that companies are beginning to doubt they're gonna get that payoff the problem is why I'm not willing to predict that the bubble's gonna burst like you know we're talking in late October of 2025 but like in the next one or two months is we've actually seen a period where this almost happened a year and a half ago and it didn't.

SPEAKER_00: 1:05:41

Oh you mean um the Silicon Valley bank thing in 2023?

SPEAKER_01: 1:05:45

Yes yeah I do um and and we there's also been several reports about the profitability not being there that have come up from major business institutions in MIT and it spooked investors but but but they recovered or even the deep seek thing which spooked investors but they recovered and I am wondering if um if people are so afraid of rev of these revenues driving up and having to figure out profitability again that they're gonna try to keep this floating for as long as possible which also means that the that the repercussions of when it when those budget constraints do come are going to be massive um and and let's not also forget sorry to roll my eyes but you know to say something very stereotypical I am a millennial and I just heard that there's going to be a once in a lifetime economic crash for the third time in my life for the third time in my adult life and I'm only 36. Right I am uh I am the Amazinial I am the youngest of the Gen Xers and um and this is like my fourth so I mean but here's the thing here it's that point where you find yourself reading you know American affairs going gosh it would be nice if the bourgeoisie would exercise some bourgeois discipline over its own goddamn behavior yeah to subject the rest of us to the discipline of bourgeois social relations all the time I mean well I but you know we in Marxism we often talk about the decay of uh proletarian relations and the K of the proletariat I talk about it too too but I've actually been really like thinking about why the hell is our bourgeoisie this bad I mean it's bad enough that a lot of people could have believed until Trump that the actually the bourgeoisie wasn't running anything that it was just a bunch of PMC people running the intermediary Trump has removed that ability to argue that but it's just like it's like uh guys there's like our masters are are are bad at doing the thing that they're supposed to do like yeah like their their whole job is to make the fucking numbers add up and they don't like you know go read Seth Ackerman in Jacobin you know 2012 and his references to you know Kornai the Hungarian economist like the hard budget constraint and the discipline of running an operating surplus like that's the entire fucking point of capitalism you know that is its only and chief defining virtue is that every enterprise that can't make the number add up to something more than zero goes out of business. The creative destructive part it's why it has been productive historically but right now we don't have any evidence that it's able to do that. Um I mean it does lead to for a Marxist at least to this weird kind of state of affairs is can you have capitalism that's not profitable and thus all the benefits that we put into capitalism not actually happen and it just decay not necessarily into feudalism but but the reason why to go into some historiography for a second the reason why I say is feudalism a mode of production kinda it's actually it's about four different modes of production some of which overlapped um and uh what you really see is like uh what Chris wick uh wickham calls the feudal revolution which is actually a decayed re a decayed order that tries to re-establish itself in several different ways and in that sense I'm like we're not there yet but I actually could see this going that way uh the other thing that we have to deal with is like the state is going to be despite all the neoliberal rhetoric we see out of the Trump administration the state is more and more and more and more directly involved in capital production. Now it all it actually was a neoliberalism too but go ahead like oh I mean of course it was but like everyone who's remotely observant has been calling for a few years now oh yeah it's oscillating back towards state capitalism right let's yeah the the war the the the the war economies of the early 20th century which was our answer to the last time there was it wasn't a lifetime financial crisis every so many years um uh because i i do like to point out that like in the long deray of history the middle 20th century is a weird development it's it there there's not a whole lot of time periods like it like um and i think go ahead like even from the normiest point of view you know it had the coinciding of let's see military Keynesianism and the war economies in the West the height of the social democratic movements in the West and a serious challenger in the form of the Eastern bloc which had not yet decayed right so like that was back when you would go you know read the Wall Street Journal and they'd say we're terrified of communism because their GDP growth rate is 12 and ours is four numbers pulled out of my ass of course but like but not far from the truth I remember even as late as the 1970s people looking at the the the South Korea the Republic of Korea's economic numbers versus um uh the DPRK's economic numbers and the DPRK was kicking its ass like um now that obviously the DPRK's economy has recovered somewhat in the last 10 years although people who don't follow it I do but a lot of people don't follow it don't know how much it's liberalized it just can't liberalize with the West because of sanctions it's liberalized with Russia and China. But and by liberalization I only mean economic liberalization. I do not mean uh uh any other kind of liberalization I just mean a go away future go away fukurayama yeah keeps trying to say futurama i've been on too many today um well fukuyama is interesting to me because i I think uh I've been listening to off uh to bunker cast a lot more lately and I think they're right that we're at the end of the end of history but just barely like because we still we still don't have real competing ideologies like like my how is when I talk about like you know China they may internally be socialist uh in their in their political orientations they in the 1982 uh constitution declare themselves to have achieved socialism however no chinese premier has said that in like uh I don't know the last three uh and they keep on talking about you know socialist reforms necessary to fully achieve full socialism and that date keep getting pushed back but my my point has always been but even if you think they're socialists they're totally integrated into the capitalist economy in a way that makes us code co-relational even with bricks and stuff even if a the IMF falls down like um so was the eastern bloc in the first place that's also true you know this is maybe my own hot take not yours but you know it's always jumped out to me that sort of the decay and decline of the Soviet Union began roughly around the same time that the West had its own profitability crisis um I think it's also interesting that that's also when the Soviet Union was taking out IMF loans which people don't ever even know um uh you know I mean and to be fair I'm not sure the Soviet Union would even have like claimed at the time that they were some entirely separate planet from you know maybe they would have claimed it from the West they certainly wouldn't have claimed it claimed it about the non-aligned states that they traded with correct i'm not I'm not sure they would have even claimed it from from the west they probably would just thought they were getting leverage to eventually pull the rest into communism uh depending on who you were talking to like like they were you know for at least up until the very end they wanted to win the Cold War. Right right I mean this is the paradox of uh as I was like the paradox of socialism in one country is like they never really acted like they believed it either like sure that they were like not trying not to freak out the Allied powers and and and encourage revolution everywhere but they never acted like they thought they could develop their economy totally autarchically they never they never did um I mean you start you start seeing that ending even when Stalin's alive so it's which is you know to bring things back to the end of the end of history like brings us to a sort of paradox of neo-Salinism where like the people looking back nostalgically at you know Mao's China or Stalin's Soviet Union are effectively like more Stalinist than the actual Stalinists ever were oh yeah like people are making things up that these things never claimed in life well for example that like there was no revolutionary bourgeois in in in the settler colonies which I'm like okay that seems like a small histographic in the United States and Canada right including the United States and Canada. MLs will claim this today and I'm like you realize that Stalin would have literally had you shot for saying that like oh yes the United States it really just has no forward momentum like it has no driving force from its own capitalist classes profit motive like no no it was all just uh settler colonialism uh a category which did not exist in in Soviet historiography and came out of dissident Maoist in the 70s in the in the Americas actually predominantly in the United States um like it's you know um and was picked up by a lot of by a lot of a lot of this stuff and I'm not even saying that there's no truth to settler colonialism as a historiographic understanding of what happened in the in the Americas.

SPEAKER_00: 1:16:38

No I mean like look let's be frank we're living in a country that still has the Homestead Act on the books.

SPEAKER_01: 1:16:43

Right.

SPEAKER_00: 1:16:45

There's no point pretending that's not true.

SPEAKER_01: 1:16:47

Right um but the idea that it somehow like undoes any form of history uh I say this as a I'm a critic of of of like Marxists who believe in linear progress and I want to be very clear about that. But like um I find that just to be like kind of nutty to me particularly because they still keep France in and I'm like how are you going to argue that the French Revolution they also did settler colonialism they also had slavery.

SPEAKER_00: 1:17:25

So to be a little more positive rather than just complaining like I think the problem with some of these categories that have arisen on the American left and I'm especially thinking of like what tends to be sort of this blurring boundary between Marxists, Marxists, Leninists and what we could call social justice radical liberals often found like in academia and nonprofits. Like the problem with these things is that they don't help us to see how our own time is different from other times. You know they have what I've heard called a very kernelist view of history where there was like one moment that happened or one event that happened well into the past and that set the course of pretty much you know all of society up to today and therefore we have no you know our responsibility is to understand that founding moment say in for the New York Times it was 1619 when the first slave ship landed. Although rather how 2025 is different from 1975.

SPEAKER_01: 1:18:33

Right. Although uh for uh yeah the the 1619 project or if we're if we're completely honest it should probably be the 1492 project like um I mean I'd love to see the anglophone left start to be shall I say fairly critical to some of these people in the way that they left in certain other countries has.

SPEAKER_00: 1:18:57

Like you know I I'm thinking of you know the again the translated review of Huriya Butelja's book where you know the translator basically notes that like the francophone left thought this book was trash and excoriated it. And then the Anglophone left uh via Verso books was just like oh we love this and like that's about like a very historically contingent and arbitrary difference of opinion to me. Maybe the Anglos could listen to the francos a little bit more.

SPEAKER_01: 1:19:31

But anyway point being this author we only like our francophone left from the post-Marxist part of the 70s even when we're Marxists we don't like our actual existing francophone left that exists now.

SPEAKER_00: 1:19:41

Well okay I guess point being for the listener like this author has this book she says they say 78 1789 will answer 1492 which I will just point out is very silly because she's Tunisian like Columbus didn't actually sail to your country your country was part of a large well known empire that both traded and fought with the Europeans the entire time for hundreds of years.

SPEAKER_01: 1:20:18

Well but they bring on the the internal presentism I I I have really seen this I've had people on my show who would argue I mean you see this in certain MMT Turquals too that there is no distinction like oh you should think of guilds in the medieval in the medieval sense as like uh advocacy groups today and I'm like but advocacy groups don't keep trade secrets well they don't keep trade secrets and they also don't have the ability to enforce via independent violence or via the state any kind of licensure.

SPEAKER_00: 1:20:53

Right. I mean it's just like there is such a thing as you know licensure issues in the modern day you know there was the license raj in India and then there's like your neoliberal your capital and neoliberals complaining about hairdressing licensure in like the present day United States.

SPEAKER_01: 1:21:13

But this seems to be a very different thing from medieval guilds right I mean and I I've seen this claimed a lot and uh I've also talked about like uh I have to go to other modes of of discussion now because historiography is a is afraid of any historiography theory that would have any explanatory power at all. And I and I mean that I'm I I know that sounds like me complaining but you to tie it back to the original point of this this show when we talk about Nico here which we haven't talked about a lot in a way um to me Nico's hopes about ai are limited by the fact that I think he's he treats the technology as not having certain trends that must that you would have to undo in development not just not just expropriate but actually change if you were going to use it for socialist means like it's it's not just a matter of uh I used to argue a little bit with Lee Phyllis about this it's not just a matter of you taking over the Walmart supply line you have to kind of change what the Walmart supply line is for and what it's doing in a way that might require you to do different engineering.

SPEAKER_00: 1:22:26

I'm gonna send you the link to this but you know as part of a journal club that I run with like a couple of colleagues you know I gave this short like technical talk and this is technical so it's for AI and neuroscience people but you know I titled it Abolish the value function and that you know partly this was an actual criticism of you know the mathematical construct in the field of reinforcement learning called the value function where I was saying these systems are inflexible because you know they have a non-interpretable um reward function which then gets aggregated across time into this thing called the value function requiring them to be very inefficient and require these massive amounts of training data and GPU time blah blah blah blah blah blah blah but also I'm sure if you sit through the lecture you know what'll probably jump out to you is that you know basically I'm referencing um you know Murovsky it's all a hidden reference to like Murovsky's more heat than light where he was talking about how you know value theory has often been about trying to find a conserved quantity associated with a variational principle that then drives the gradient flow of political economy. You know it is 19th century physics math that was ripped off and relabeled so that classical economists could then say here's how an economy works here's how we show that it reaches equilibrium and what Murowsky goes through and shows in that book is that pretty systematically not only are these like borrowed metaphors and borrowed math but in order for that to hold as I mentioned you would need a conservation law which would mean a law of conservation of value saying that value can be neither created nor destroyed in the economy which you know from both a like a non-Marxist socialist and a Marxist sense and an MMT sense and even a bourgeois neoliberal marginalist sense is complete nonsense creating value is the whole point of the capitalist economy.

SPEAKER_01: 1:24:48

Right no I mean this was the problem with uh trying to map technocratic assumptions onto capitalism in the early 20th century for those you don't know the technocrats actually tried to come up with a theory of value based on energy not even labor energy expenditure and and the value would not test that I've heard people try to do that again. Right and I'm like I would love for that to work I have no idea even more than labor how you would measure that um um and they didn't figure it out either but the other thing is they tried to say that what they were doing was not like socialism because you could do a capitalist economy on it and I'm like how how in the world could you do a capitalist economy with a steady state theory of value based off of energy expenditure?

SPEAKER_00: 1:25:36

Like um yeah so to get back to Lee Phillips actually because you know I know him I like him I would consider him a you know quite friendly acquaintance and you know a good political colleague I greatly enjoyed his book you know People's Republic of Walmart but yes I was making you know little margin notes in my Kindle saying you know all of these assumptions about planning assume that you have a value function that you're going to optimize. And for a capitalist firm that value function is their operating surplus. And now the problem is that the classical socialists like you know up through maybe Lenin and then the USSR like their vision of socialism was to basically say you take the entire economy under vertical and horizontal integration as you know the absolutely massive org chart of a single firm.

SPEAKER_01: 1:26:31

Mm-hmm and teleological development of everything that was the that was the approach.

SPEAKER_00: 1:26:38

Yeah yeah and I mean besides you know the philosophical assumptions of teleology if you had you know before your firm encompassed the whole economy simply when there ceased to be you know meaningful budget constraints and market competition uh the operating surplus would no longer like reflect the thing that it's supposed to reflect within capitalism. Right. So you would have to like if you wanted to take over the economy and run it this way as a socialist, you would have to come in and replace the value function that goes into these planning and optimization algorithms. You would have to wholesale replace it with something new that could then indicate some kind of you know sociological metric or econometric construct you know that you had invented and could assure yourself reflects something meaningful even when the conditions of capitalism no longer hold.

SPEAKER_01: 1:27:37

Right. Interestingly this is consistent with I remember getting people not putting together Lee's position on growth on bright green growth eco-socialism versus uh degrowth now I have said that I think the whole debate's wrong like that I actually agree with you you'd need a lot of you'd need a lot of investment but that if you're running a socialist economy I don't know what what surplus growth means in the same way I just don't and nor are you necessarily investing for having to reinvest you you actually want the investment to be like we have planned obsolescence everywhere earlier economies really didn't do that and um because they didn't need to and we all by earlier economies I mean early early capitalist and pre-capitalist economies um and uh I kind of my my pushback only was the same thing like if we're not doing if like there's no GDP growth because we don't give a shit about gross domestic product in that way what are we even talking about like why would you be pro-growth versus the anti-growth we're talking about how you uh you how you allocate resources and the resource allocation would have a different function than surplus uh accumulation um even if you were doing surplus you some surplus accumulation they take care of people can't work to to develop technology etc you that's not your primary modus operandi anymore like and it just seemed like then you like then you have to actually come up with social relations that make all of the incentives and the discipline mechanisms actually work out.

SPEAKER_00: 1:29:23

Like you know I've due to my professional history I have some engagement with you know planning algorithms and optimization constructs and things where you aren't using what I would call you know a value function. Like you have something that's mathematically similar but the point is it has measurable interpretable units that you could refer think about as being much more like output quotas. So like you know someone gives you an output quota and you try to turn your crank and come up with a plan that satisfies the quotas you know while leaving margin for error and margin for surprising things to happen, et cetera, et cetera and of course you know the math works out under the assumption of either A, you're running this on a computer so it's not human, it doesn't have a will of its own, or B, you know, you're working in the context of a vertically integrated firm or even command economy where you can tell people to do things and know that they will carry out your orders. So like it's a very you know it runs into the limitation like this kind of stuff that it is a bit Gauss planes esque.

SPEAKER_02: 1:30:36

Right.

SPEAKER_00: 1:30:37

There is an assumption that you can give orders and they will be followed and therefore all you need to do is give like socially beneficent orders. Somehow But of course, in the real world, like that's not true. The Soviet Union was never a perfect, you know, hive-minded command economy where everyone did what they were told.

SPEAKER_01: 1:31:00

Absolutely not. Particularly over time.

SPEAKER_00: 1:31:04

Right, like particularly over time. And like the need to apply repression in order to conform to that model is part of why people got cynical and disenchanted with the project, even when they were born and raised as like ardent communists. Right. So then there's the question of how do you arrange social relations so that people want to contribute to you know a democratically decided upon plan.

SPEAKER_02: 1:31:31

Right.

SPEAKER_00: 1:31:31

Or and you know, some for democratic substitute some word that means good if you happen to not like democracy. Right.

SPEAKER_01: 1:31:39

Um I mean for me there is some hope in some cybernetic inputs of making that making those inputs more participatory of the people who make them. Um so uh but even then, I mean, like we talk about this in Cyberson, as I pointed out before when I had a specialist on uh Alyende's uh experiments in Chile. Uh it actually wasn't CyberSyn that made that feasible, it was uh it was some of the cybernetic um social dynamics being applied to uh Cordonas, the Cardona system, um, which Cybersyn would have made coordinatable, but was not actually the thing. So it's like I'm like I've I've pushed people on their misunderstanding of what cybersyn even was. I'm like cybersyn barely got off the ground. We don't actually know if it would have worked.

SPEAKER_00: 1:32:31

Um okay, so I'm I'm gonna say something really counterintuitive, but I almost think in a certain way more leftists should go to grad school or something. As in I wish people no, no, because really, like a lot of this stuff actually does give me a sense of persistent optimism for what's possible. Right? Because I don't think that these are insurmountable problems and will never get anywhere. They're the kinds of problems where you can put you know a team of people to work for like five, ten, fifteen years and they will come up with something.

SPEAKER_01: 1:33:07

I I think that they should go to grad school, but in the sciences, but not but but not in engineering for reasons that I do not entirely understand, but engineers always end up being the most this is a you know the you know the you know the psychological research I'm referring to. Like engineers tend to be dogmatists in the extreme of whatever ideology they pick up, and like and like I will say like I'm path dependent.

SPEAKER_00: 1:33:37

You know, I got lucky that my PhD advisor was like, well, I had multiple advisors, but one of them, you know, Jan Willem, was like this very you know open-minded and even killed person, and then Lisa, Lisa Feldman-Barrett, author of How Emotions Are Made, she's you know, managed to become scientist famous. But point being, you know, she was someone who actually had her first experiment fail in grad school, famously so, because the theory behind it was just wrong. And she has spent most of her career disproving the theory that she entered grad school in order to prove correct. And so, like everything about her advising style and how she teaches science and what she passes down to her students in a very path-dependent way is completely dependent upon this experience of like going to do confirmatory normal science in the Kunian sense and having it blow up in your face.

SPEAKER_01: 1:34:35

Right. She's actually one of the people who do the who does the thing that Pop Arians say science is supposed to do, but actually studying the sociology of science only happens rarely.

SPEAKER_00: 1:34:46

It does happen, but it rarely happens when like a very senior famous person dies or retires. Like it's another reason why you know lifting the age cap on holding tenured professorships in the 90s was possibly actually a bad idea.

SPEAKER_01: 1:35:04

Oh god, I didn't even think about that. Um, but we will go on. I actually have like when I talk about redesign, what would Varn make a university? I'd be like, there's job protections for everyone, and there's no tenure. Um and decisions are made in committees like tenure, but this hierarchy system that's quasi-medieval that tenure is based off, that you guys tell me is something like good job protection. Y'all have not worked as academics. Are you lying about your experience of doing so? Because that was a nightmare field. Um, and not just in the humanities.

SPEAKER_00: 1:35:44

Like this is another one of those things where we could do to listen to people from other countries, because you know, the Anglo University system has this thing where we're all looking, especially in the United States, especially, especially in the United States, softer in other Anglo countries. We're looking over our shoulders and going, oh my god, if I leave the academic track, I'll have to deal with at-will employment. Meanwhile, someone like France, Germany, Italy, like China, um, Japan, Singapore, like, you know, pick any country developed enough to have universities that are like seriously taken seriously on the world stage, and they'll tell you what? No, I have an employment contract. Yeah, like it says I get seniority after you know n years, and the degree of seniority rises over time and protects me from being arbitrarily fired for no reason. As a foreign professor without ever making me a little bit, without ever making me into a mini a miniature medieval lord. Right. Exactly.

SPEAKER_01: 1:36:46

Um the as what I worked in.

SPEAKER_00: 1:36:49

In some places, even auto workers on an assembly line can have that kind of job security.

SPEAKER_01: 1:36:54

In France. I mean, like in uh in the Nordic countries, and like look, I get I'm a big proponent of hey America, don't don't try to do Nordic socialism because they're not even really trying to do it anymore. But Nordic job protections and Nordic contract, like like, oh, you lose some productivity. I'm like, oh yeah, we've been so great at productivity lately. Uh and also the Germans are still pretty damn productive last time I checked. I mean, not not since they can't rely on Russian gas anymore, but but uh it's not their workforce's problem why they're not. Um so it it's I find this kind of uh uh fascinating to me. But yeah, when I worked in um I worked at in so in South Korea for two years as a lecturer, which is a non-tenured professor. But if I had stayed at that job for five years, it would be hard to fire me. You could do it, but I'd actually have to violate my professional ethics uh group. So I affectedly had tenure, even though I could never get tenure as a as a lecturer, I could get like promoted to reader and some other things, but like uh research uh research teacher.

SPEAKER_00: 1:38:02

Um but but even in that system where there was tenure, I there wasn't the the the the nightmare system I had saw in the United States, where like full professors are basically lords, like um uh you know so my very even keeled and like open-minded advisor, Jan Willem, you know, from the sound of his name, he's Dutch, right? And he basically explained that like he had done academia in a number of different countries, and he said, you know, in the United States, your PI is the head of your mafia, and your different mafias are facing off against one another, you know, sort of competing in the pseudo-market for government grants. That sounds correct. In the United Kingdom, where the system was more that like the principal investigator was like the king and dispensed favors accordingly, you know, according to personal views of favor and disfavor among lower level scientists, where lower level means like these are actually people you know in their 40s with families and PhDs who've been working for 20 years, like they've been in the workforce for 20 years and have effectively no independence because you don't have independence until the king decides you're ready to become a fellow king. Yeah, it's uh it can vary a lot.

SPEAKER_02: 1:39:34

Yep.

SPEAKER_01: 1:39:36

I I I've I've actually a I as I uh I made this joke out so chili. What are the one of the ironies about right now is um the one actual semi-feudal institution in American life is the one that's being destroyed. Um so like Yep. Yep. Not to depress you as a as a as a are you still would you still consider yourself an academic?

SPEAKER_00: 1:40:01

Like um uh let's say for the moment, no.

SPEAKER_01: 1:40:08

Okay.

SPEAKER_00: 1:40:08

I'm not gonna go into more detail. Um I don't like to announce what job I'm in until it has officially started and I have, you know, like yeah.

SPEAKER_01: 1:40:22

Yeah, no, no, that's fine. Um for me, I I I uh when people are like, oh, you're an academic, I'm like, I used to be, but I am not anymore. I don't I don't have, I mean, unless you consider high school teachers academics, which no one does, um uh, you know, I think it's not part of your job to like review papers in journals, publish in journals, and attend conferences you're not an academic. Like, nope. I attend conferences for fun. Like I have and I have occasionally written papers for fun. Like I have gone to I've gotten into conferences and fields I don't have high degrees in, which I find kind of funny.

SPEAKER_00: 1:41:04

But you know, um uh I mean that is the positive and open side of peer review, right? Which is that like if you can pass peer review in a field, then you can publish in that field whether or not you have a formal qualification for it. The negative side of this is that of course no one can stop physicists from barging into other people's fields. That's not such girl.

SPEAKER_01: 1:41:28

Um I did sort of offend uh some some uh some Ivy League uh intellectual historians when they're asking me, Well, what's it like writing an intellectual history versus literature? And I'm like, Well, if you do literature well, there's no difference. If you do literature review well and not and not theory gallblogy gook, there's no difference at all. Which I think actually made every side of that equation kind of mad. Um, so but I was just like, I go into archives, I look up stuff, I historically contextualize it. That's what I was trained to do. Yeah, honestly, it's kind of weird.

SPEAKER_00: 1:42:05

It's weird to me coming from STEM that this would piss people off because, from my point of view, the literature review is the part of the paper or the monograph that you write, precisely in order to contextualize your novel contribution in the recent history of the field. It's not meant to be an intellectual history, it's meant to like it's meant to provide context for the other part of the work.

SPEAKER_01: 1:42:35

Right. Whereas I find it like uh I won't go on the theoret theoretical when I talk about the theoretical term of the humanities, people are always surprised because I know so much theory, but I'm like, I think it, I think in in in literature it was a disaster because it didn't lead to better theory, it led to misapplying old theory to papers where the theoretical constraints were social, but now we're talking about literature and art, and then somehow that being taken as radical, thank you, Verso, for publishing all that shit as if it was somehow political commentary. Um, and um and I've always pointed out that one of the first things that you notice in a lot of these fields is like you know, the the lit people, as soon as they get tenure, you know what they never do, publish on literature anymore. They start writing weird history books or communization theory or whatever, like um, you know, not trying to convince you that there's something meaningful about post-structuralism when applied to Faulkner. Um, but anyway, uh we are topic.

SPEAKER_00: 1:43:43

I don't know what you did your degrees in, but I did not do mine in literature, so I am not going to comment on air. I had enough bad tweets without saying something with my real face and name. That's totally fair.

SPEAKER_01: 1:43:56

I'm interested, I actually did work in I worked in literature and in uh philosophy separately. Um it's two different degree, two different undergrad degree paths. Uh, because I thought I was gonna be a lawyer, which is that I've always thought it was analytic, by the way. Uh uh historical continental and analytic modern. So I did Hegel and backwards, and then the what I did modernly was all analytic philosophy. So actually, I studied a lot of Derek Parfitt. Um so which I don't consider myself an analytic philosopher. In fact, I actually uh uh I'm one day I'm gonna debate Ben Burgers on on the usefulness of analytic philosophy after a certain time period, but I I did train myself in like I read Roger Penrose, I read Derek Parfitt, I read Brandom, I read that's what I was reading. Um and what I got interested in uh in grad school, but at this point I was in an MFA, so who the fuck cares what I thought? Um um uh before I got into educational, was actually trying to bridge the difference between continental and analytic philosophy. I wanted like I wanted a dialogue, and there were some hints that like with Ray Braz with Ray Brazier, uh Brazier Brazier, I'm never quite sure how I said that name, and and stuff and brandom going on in Europe that that was possible to happen, but it didn't seem to actually take fruition after all. Um and I just moved on to other stuff like educational psychology and educational anthropology. Um uh so that you know I am I am in all those human those humanities that like butt right up to real science, but like don't go there. Um so I mean technically education is social science, but I don't think I believe that. Um but anyway, uh still teaching as a bit of a side remark.

SPEAKER_02: 1:45:57

Go ahead.

SPEAKER_00: 1:45:58

Um I mean you have this within the sciences as well, where there's like engineering, physics and mathematics, the theoretical side of any discipline, so like biology, chemistry, physics, neuroscience, all of them uh are a lot more mathy and theory-oriented. And then there's an experimental side that often is uh very close to being like uh no real skills except wet lab skills. And it turns out, you know, the reason for this is of course that like there's just that much wet lab stuff you have to learn in order to be minimally competent to run an experiment yourself.

SPEAKER_02: 1:46:41

Oh yeah.

SPEAKER_00: 1:46:42

It's just hard to find people who are basically going to do twice the work in order to learn theory and experiment at the same time. But it does mean that you can have a lot of experimental work that like butts right up against the theory but completely misuses the terminology and concepts. And then theorists who are out of touch with experiments, and it has occasionally had me think like, you know, if I was God King of the Earth for a day, like one of the things I might try to do is like maybe emisserate myself a little bit by like trying to sort of raise the waterline of what we teach people in K through 12.

SPEAKER_01: 1:47:19

Uh this this is a this is a whole different show. Maybe I'll have you come back so we can rant about this together. But like I am, I am uh I am I have been terrified at both people talk about the lack of arts, but I'm like, but also STEM is taught so poorly, it's taught so decontextualized. It's and then math math in high school is so I am good at math. I did not know that until until well into university because I did not see the point of symbols per and I am I do have uh dyslexia for numbers just like I do for letters, but like I I compensated for that. Um but I did not know that I was actually good at math until well into college, um, because I did not see the point of symbols pushing. I was like, well, I don't know.

SPEAKER_00: 1:48:09

I'm pretty sure I was actually bad at math until grad school. Like, I I mean to like really contextualize, there's this thing about learning programming early that just makes you bad at math. Because high school math is sort of taught in this very platonistic way. They like have you learn trigonometry by getting out a unit circle and like pointing to you know the 45 degree angle, and then what the trigonometry, you know, trigonometric functions are for the 45 degree angle, and you ask, what about the 46 degree angle? And they say, get your calculator. And you ask, how did the calculator get that? And they can't tell you. And it's like, hey, so how did God decide the value of these functions if you can't give me a formula I could actually compute with? This is programmer brain. You know, then once you actually get up to like college-level math, they will show you that you know you can use calculus to derive, you know, an infinite, a convergent infinite series that allows you to approximately compute these numbers to you know however many terms you want. And like you'll learn eventually some topology and like what infinitesimals really are, what limits really are, like open intervals and open neighborhoods. And that's all very nice. And you'll finally learn and like enjoy enlightenment of like so. This is the difference between how I can push symbols around on a piece of paper and like geometry, and here's how they're actually connected. But until I got that, I would really say like I was not good at math, I was a computer geek. Yeah, that's uh and at no point did the pedagogy actually try to make the connection actively. Instead, it was just like, look, you need to have faith that you know the value of sign exists for 46 degrees, it exists for 47 degrees, because the calculator gives you like the calculator or what is it called, a slide rule, like you know, you used to have log tables printed in the back of the textbook.

SPEAKER_01: 1:50:18

I used to use log tables, and I actually am I I we had calculators when I was in school, but I actually was trained on a slide rule in elementary school.

SPEAKER_00: 1:50:26

So yeah, like you just had to have faith that these things give you like close enough approximations. And if you ask close enough approximations for doing what, by the way, the answer is for doing engineering, like really for doing like mechanical engineering and civil engineering, because that is what the STEM curriculum is actually structured around, is you know, producing enough like engineers for the 20th century.

SPEAKER_01: 1:50:52

Which we don't do anymore.

SPEAKER_00: 1:50:55

Um even though we have the same STEM curriculum, uh so it's it is all actually structured by political economy underneath, and you just don't find out what you went through until you're well into adulthood, right?

SPEAKER_01: 1:51:08

Or if you're the average person, you never found out what you went through. Like, you're just like, what was that weird training even all about? Because I remember they would tell me, like, oh, we're teaching you uh trigonometry so you know how logic works, and I was like, Really? Because lot the only thing that I've ever studied in math that actually works like a logic function is a geometric proof.

SPEAKER_00: 1:51:30

Um, so which is not actually definitely the only proofs we had to do in high school.

SPEAKER_01: 1:51:36

Absolutely. I didn't know about other proofs until I was like in grad school studying um psych stats. Like it's just um that's what I actually also that's when calculus actually made sense why I needed to know it. Because when I first learned, I'm like, this is just weird shit. I don't know why I need to know what a cosine does. Like, like um and uh my my students today in high school, like I will explain to them like why you actually need to know the Pythagorean theorem and what you can actually do with it, the number of things you can actually do with it. And I'm the English teacher, and they're like, Why are you teaching us this? And I'm like, Well, because someone's gotta teach you that this is actually useful, because it was a revelation to me when I learned that it was actually useful. Um, but you know, I I often tell people like the thing about math is like it doesn't actually work like the other sciences, but you you can't do the other sciences without it. Um and uh uh and logic, logic, uh predicate logic looks like math even in the analytic sense, but it isn't. Um, you know, it's related, it you know, even neurologically, but it's not the exact same thing. Um although, okay, to to end on something about actually what we started this topic about, and this would be our last question. Um why is it that LLMs can't do math that well? Like, because that does not actually make sense to me.

SPEAKER_00: 1:53:06

Um I'd have to look at the guts of an individual LLM to tell you that because I'm sure someone has actually come up with the idea of if you've get if you're given a prompt that looks like a math question, then just try to find the math question and turn it into symbols and feed it to a like you know, a calculator prompt inside the computer and answer with that.

SPEAKER_01: 1:53:36

Right. And you can train them to do good math, but like if you like if you just throw math, and they'll do good coding. Well, they'll do inefficient coding, but they will code. Um but but the math functions, it it's wild. What I math teachers have told me what they get from the AI, and it's like okay, like that seems bizarre.

SPEAKER_00: 1:53:59

Um so if I had to like if I you know, I don't have access to the training corpuses, so I'm really venturing a guess here based on my knowledge of how these things are trained. I would guess it's that most of the time if you're just trying to predict like the next word in a sentence, then you're kind of going to do vibe math. Like, I I know this sounds very fuzzy and vague, but like, you know, if you remember Stephen Colbert coined this word truthiness, then there's like sort of or I guess the better word might be sort of authoritativeness. And it's spitting out the thing that sort of sounds sounds right, even though it has even though without being architected with like a coding prompt inside to actually run something that would calculate the answer, it doesn't have a proper ability to like check. I guess think of you know, think of someone think of someone who doesn't actually know math, but they still want to answer the teacher's question because they're on the spot. And so they try to come up with something that's that's gonna get by the teacher.

SPEAKER_01: 1:55:29

Yeah, that makes sense. That does make sense.

SPEAKER_00: 1:55:32

Yeah. Like if a real child does that, it's because they didn't do their homework and they're bullshitting, but you know. If an LLM does that, it's because they physically cannot like they don't have the working memory, you know, they can't visualize a piece of paper on which they could then, you know, sketch like the arithmetic procedures they were taught. Yeah. Which is more or less what actually I don't know about you, but that is how I actually did mental math until eventually I had done it so many times that I started memorizing the answers to the simpler problems. Or like I would visualize, you know, a number line. Like my fourth grade teacher third fourth, third, fifth, you know, all of those like taught us all of these kinds of tricks. But all of which involved sort of like our sensory motor imagination rather than just memorizing text.

SPEAKER_01: 1:56:34

That makes a little sound. But I actually I remember doing that a lot myself was actually it it it like before I had more internalized more math, actually going like I think I had like this vision of a number line a lot of the time, like literally thought out the number line in my head. Um and I I guess that's also part of my the mind is structured like a language uh critique is like, well, I know there's people who don't make mental images, supposedly, uh, but uh I thought in pictures way more than I thought in letters or words exactly. I mean, there's still signs and they still have semiotic content, but like it's not quite the same as a language. Um I think maybe even Nico would would admit that like sign signs encompass language but are larger than that. But I I I would like to go maybe push him on like what that exactly means in a dynamic social system. Um yeah, you know. Um well, thank you, Eli, for coming on. We've been all over the place. Um, I'm gonna have to retitle the the the uh the title to this episode, but uh I'll figure it out. And uh this was just Nico response to Nicholas V, I mean not Nico, Eli response to Nico via Real, and really that's only about a fourth of what we did today. But um uh uh we'll have you back on. We might actually have like I might actually have a symposium on education because I actually think so, like I know you would agree with me on this, but most people would be horrified. I think socialists should oppose progressive education. Define progressive education, uh education as understood by Dewey and assumptions and by also whatever Alfie Cohn is talking about this week, um, which are like we did which we need to educate for social justice. By that we mean we need to change standards, push against standard uh standardization, we need to uh focus on soft skills and social skills, we need to make school more about reparative justice, although by that we mean uh uh administrators not having to deal with students, really.

SPEAKER_00: 1:58:53

Um I mean, I would instinctively dislike the vibes of this sort of thing, but really like a substantive critique would really be look, we live in a world where kids are going to have to grow up and labor, and if we don't give them any skills to do that, aside from navigating social relations, then we're spitting out kids from the education system who are only qualified to basically be hot bourgeois bullshit artists. Right. And there aren't enough jobs available doing that.

SPEAKER_01: 1:59:23

Right. But these are from the same people who tell me that teaching code switching is impressive, that students shouldn't have to do it.

SPEAKER_00: 1:59:30

Um again, like just to be purely substantive and not vibes-based because I know myself, I will say too much. Uh you know, we have to admit that the education system is in a social context that is not ideal, uh that is not ideal according to the educators, and that we are arming and equipping kids to deal with a world that they don't control and which is in many ways going to be hostile to their interests.

SPEAKER_01: 2:00:01

Maybe that sounds a little too aggressive, but like you know, you're I would be more aggressive than you, which is actually but but I I think about this a lot.

SPEAKER_00: 2:00:11

I mean, I've spent decades toning myself down and still have the instinct to think that I'm being harsh, which is which is a funny thing, Eli.

SPEAKER_01: 2:00:21

So have I. I go back and listen to if you ever find anything for me from like 2012 or 2013, like I will, or even 2019, I'll be I'll be quite frank. Like what I would say, stuff like checking your privilege has never changed anything in history at all ever. Um like which I I now would nuance as it has rarely changed anything in history. At all ever.

SPEAKER_00: 2:00:48

But um now you're talking like a scientist. That's one of those lovely weasel words where you leave open the possibility that someone can come at you with an exceptional example you haven't heard of. But everyone knows that what you mean is never in my experience.

SPEAKER_01: 2:01:06

Exactly. Which is like which is still, I mean, fair. Like you come up with something not in my experience. That's fair. But um yeah.

SPEAKER_00: 2:01:15

But if you've got a fairly broad experience, then it might be a fairly strong statement that someone should listen to. Exactly.

SPEAKER_01: 2:01:22

All right. Well, thank you so much. Um do you want to plug any of your work?

SPEAKER_00: 2:01:27

Um just that abolish the value function less uh lecture.

SPEAKER_01: 2:01:32

Okay, send me that and I'll put it in the show notes. Thank you so much, Eli.

SPEAKER_00: 2:01:35

Sure. And I'll send my Google Scholar link. Cool, cool, cool.

SPEAKER_01: 2:01:38

All right, take care.