Metaphysics and modern AI: What is causality?

In this episode of our series about Metaphysics and modern AI, we break causality down to first principles and explain how to tell factual mechanisms from convincing correlations. From gold-standard Randomized Control Trials (RCT) to natural experiments and counterfactuals, we map the tools that build trustworthy models and safer AI.

• Defining causes, effects, and common causal structures
• Gestalt theory: Why correlation misleads and how pattern-seeking tricks us
• Statistical association vs causal explanation
• RCTs and why randomization matters
• Natural experiments as ethical, scalable alternatives
• Judea Pearl’s do-calculus, counterfactuals, and first-principles models
• Limits of causality, sample size, and inference
• Building resilient AI with causal grounding and governance
  
  

This is the fourth episode in our metaphysics series. Each topic in the series is leading to the fundamental question, "Should AI try to think?"

Check out previous episodes:

• Series Intro
• What is reality?
• What is space and time?

If conversations like this sharpen your curiosity and help you think more clearly about complex systems, then step away from your keyboard and enjoy this journey with us.

Do you have a question or a discussion topic for the AI Fundamentalists? Connect with them to comment on your favorite topics:

• LinkedIn - Episode summaries, shares of cited articles, and more.
• YouTube - Was it something that we said? Good. Share your favorite quotes.
• Visit our page - see past episodes and submit your feedback! It continues to inspire future episodes.

Chapters

• 0:56 What makes a cause?
• 3:26 Causal chains, loops, and splits
• 5:21 Correlation: Limits and temptations
• 8:12 Statistical vs causal analysis
• 11:43 Randomized Controlled Trials (RCT)
• 14:31 Beyond RCTs: Judea Pearl and counterfactuals
• 15:40 Natural experiments in real life
• 19:05 Why not make everything causal?
• 23:08 The statistics problem behind causality
• 27:14 Truth, marketing, and claims
• 31:10 First principles, governance, and robust systems
• 34:15 Causes, objects, and world models

Transcript

SPEAKER_01: 0:02

The AI Fundamentalists. A podcast about the fundamentals of safe and resilient modeling systems behind the AI that impacts our lives and our businesses. Here are your hosts, Andrew Clark and Sid Mongalik. Welcome to the AI Fundamentalists. Today we'll be talking about causality. For those who are jumping into the series here, a quick catch-up. All episodes in this series are exploring different themes with the ultimate goal of determining whether AI should be able to think. To date, we've explored reality, space, and time. And Sid, keep me honest here, where are we with the series? Iggo as we go into causality.

SPEAKER_00: 0:48

Yes, yes, yes, yes. This is episode four out of five. So here we are at our penultimate episode in our series. And we'll be talking today a little bit about causality. And hopefully this will tee us up for our nice final episode, wrapping all these thoughts we've had up. And so today we're going to be focused on a really important question, which science has been at the heart of trying to answer, which is what are causal relationships? What are relationships where we have very clear causes? I had the pleasure of giving a version of this lecture to students at Stonebrook University. So I think we're going to be seeing a little bit of a recap for anyone that's learned about statistics in 2018.

SPEAKER_02: 1:28

Yeah, uh pre-LLM days, uh causality was it was a big thing. And like very much the adage from statistics, you know, correlation causality, like it's something that's so critically important and underpinning uh a lot of different things. Really excited to dig into it today. And it's uh I know it's an area both Sid and I are very passionate about. So it very much sets up as we go into that final episode of what is reasoning, what is thinking. So a very we've had we've had some some good intermission episodes. Very excited to get back into the meat of of the metaphysics series.

SPEAKER_00: 2:03

So we'll start by examining causality through a different field of study, which is called etiology. And I think this is really popular with people who study medicine, who are interested in studying the cause of certain illness, the cause of certain psychopathologies, and the cause of certain events in one person's life. Right. So if we want to study the causes for an event, we should understand what makes a cause. On a philosophical level, if you have cause A and effect B, we get the following three claims about how those two things must be related. A cause and an effect must be contiguous in space and time, meaning they must happen in the same place and happen next to each other in time. A must precede B, so the cause has to be before the effect. And in the constant conjunction principle, A and B must be observed to interact. Right? If A and B happens behind a closed door, we could infer that something has happened, but we could not say that it definitely happened. So really just trying to make rigorous what we kind of all know is that causes and effects are before and afters and that they're applied to each other. But let's think about that like very crisply as we like go through this episode. Now that we have this idea of a cause and its ensuing effect, we should consider the relationships that A and B can have. We could consider a causal chain where A is a cause of B, but then B is a cause of C, right? So we would then be able to make this claim that the cause A indirectly affects affect C, but only through B. We might consider a homeostat a homeostatic relationship where once B has caused C, C then causes A, and this creates a consistent repeating loop. So this is another kind of causal homeostasis. We also, I'm sure, can consider that things don't have simple causes, right? It's often the case that one effect is caused by more than one cause, right? This is called common cause. So A1 and A2 come together and they cause B, or vice versa. You have a single cause A, and it causes maybe intended impact B1 and unintended impact B2. These are considered the general class of relationships that we consider and that we study in causal relationships: chains, loops, shared effects, and shared causes. So with all of that in mind, we get into the question that statisticians and data scientists ask, which is well, I ran a couple of studies that were correlational. Can I now say that I have found one of these causal relationships? Can I now say that we have some sense of the etiology of how uh classically like wine consumption influences and uh influences income? Have we learned that connection? I've done the study and I've collected the data and I find that people who drink more wine make more money. Is there can we now make a causal claim out of this? And this brings us to the very classic statement in statistics, which is that correlation does not apply causation. That while we have this human tendency to interpret the patterns that we find in the world as demonstrating meaningful connections, this does not mean that we've actually learned something fundamental about the world. Uh we might said fine that there are deeper causal mechanisms to explain phenomena outside of just correlation.

SPEAKER_02: 6:06

And I think this is this is really the hits at the root of it's it's kind of like how we talk about with LLMs of people wanting to turn their brains off and things like that at times. It's like this is the causation is the holy grail of, I mean, science, research, and modeling. It's like that's what you want every model. It's like all models are wrong, some are useful, but it's when can we try and undercover uh uncover that causal relationship is really the goal for uh for as we're modeling. And there's all of the it even there's it's not even as black and white here. There gets to be, and we'll talk about a little bit, like some of the grayer areas of like uh in my econometrics background example, there's like ranger causality test. Well, it's predictive causality, it's not real causality test of like can one time series predict another one reliably, but it's even caused causality. So even within like the statistics realms, it gets watered down a little bit. But um it's the the nuance really much does matter. And that's where we've had a lot of back in the big data type era and stuff where people got a little sloppy of, you know, um, I think it's like shark attacks and ice cream, like those sorts of like relationships you you've heard are like um when a hurricane's gonna happen in Florida, more more strawberry pop-tarts sell out. Well, those are the correlations. It a storm does not cause Pop Tarts to be sold, right? Like though this is where it's a very slippery slope methodologic methodologically, and something to stay really tight on. But it's um just to help with that some of that framing of this is the constant thing. We've gotten a lot closer with it in methodologies, but it's the constant area of as we try and move past, like we keep doing that statistics is really focusing on on causality and sciences, and then you start having like the computer science world comes over, went back in the big data world, back with now the LLM world of like, well, we can just predict something well enough that it this nuance doesn't matter. It's part of the us uh the focus of this podcast, is is that these nuances do matter, and that that when we can try and build like well-built systems that have causal elements is when we really start having those useful systems.

SPEAKER_00: 8:04

That's right. And we'll use a practical example here. The big data world would get you to say a statement like As arcade revenue changes, we see a very strong correlation. In fact, a 98.5% correlation in how many computer science PhDs are worded. Does this mean that people going to the arcade are now inextricably drawn to being doctorates in computer science? Almost certainly not. But the data is there to show it. And we might be quick to come up with stories and narratives that make this basically plausible as a causal relationship. And in the field of psychology, we have a word for this, or like a theory for this, which is gestalt theory, which talks about this idea that we observe two patterns moving. And the human brain has a really strong circuit that says, there must be meaning here. I must be able to interpret this, and I must be able to say that there's some relationship between you know pictures moving back and forth. Like you know, you see like a circle and you see another circle following, and you're like, oh, the circles are following each other, but all you're seeing are shapes moving on the screen. And so this inherent circuit in the brain, which allows us to understand the world, understand motion, understand observable causality, gets into trouble when we try and study real causality and real mechanisms that connect objects and events together. So let's break this down then down into the two different types of analyses that we can do. And I'll make a bold claim that what humans do and what machines do is effectively statistical analysis, and that rigorous scientific methods get us something closer to causal analysis. Statistical analysis can help you make claims to say that A is associated with B, meaning that as you change or wiggle X, it's highly probable that this will change and wiggle Y. Right? As I increase the amount of lunch that we give to kids, how much improvement do we see in their school grades? Right? We assume that if we give the if we give everyone lunch, we see better performance, and if we take away the lunch, we see less performance. And that and then we could make a statistical claim about that, about that happening, that we see that generally seemed to be happened, we see an association between X and Y, and vice versa, Y and X. And you could use it to generate a prediction. You could generate a prediction about if we do intervention A, we would expect that intervention B would happen, but it would just be a guess. On the flip side of this, we would have a rigorous causal analysis. And that would let us say a sentence like A causes B, meaning that changing A will change B. That there's a fundamental connection between those two things. Now, you'll also find that those things are going to be correlated. A will be correlated with B. But that doesn't mean that the reverse is going to be true. That just because you do B does not mean that A will also always move. And instead of generating predictions, in this model, we can generate explanations, explanatory models of how the world works, rather than just predictive models. So I think the natural follow-up question is okay, you've told me that there are statistical and causal analyses, and I want to do the causal one. How do I do that? What does that look like? A really natural way to understand causal inference studies, as we start at the top of the episode, is medical research. In medical research, we conduct experiments with quantifiable outcomes that can be verified, can be checked for stability, and give us a sense of how medicines work. The holy grail of this is called the randomized control trial. This is probably one of the strongest methods we have for establishing causal analyses. And everything is in the name, but I'll break it down. In the randomized control trial, let's say that we're testing out a drug, and we want to see if this drug improves headaches, right? We say that patients who take this drug will receive X amount of inflammation decrease, which can be quantified. And from there we would then say, and then we have testimony saying like their headaches have gotten less severe. We would take a study group of, say, 100 participants and randomly choose them to either receive the drug or to receive either no drug or a fake version of the drug or some baseline drug, right? Just some sugar. We would then allow them to take the drug over a period of time, and then we would measure their differences. And then we would say that any differences between the groups is caused by being applied to the treatment group. Because we've removed a lot of these confounding variables, a lot of these other possible causal mechanisms, and really the only thing that should distinguish the two groups being studied is the use of the drug. And this is a very robust method. Um, you'll see papers out there that try 5,000 trials of a drug, giving 2,500 on either side of medication. You'll see ones where there's only 17. Um, as you scale up the amount of users in your study, you'll generally get more power and more ability to say strongly that there's a causal claim and not just a chance uh occurrence. But I think if you if you read any paper nowadays that's coming out about a drug, you'll find this exact same design because this is generally accepted as uh one of the only acceptable ways of making a causal claim.

SPEAKER_02: 13:52

Well, let's talk a little bit about uh uh as well. There's all like that's definitely the holy grail uh way of doing from science, but there's also some some interesting work by uh Judea Pearl on like uh his due calculus and and things like that, of like how to do other causal modelings, and also something that I know near and dear to both you and I uh is counterfactual modeling and things, which when we're start working in the modeling world, now of course it's like a uh the blind medical experiments are definitely going to be the best type of uh of doing the causal analysis, but in areas where we also have uh models of representations of realities, we can still be representing those counterfactual uh worlds of like with the same inputs tweaking certain outputs, how does that change? That's not it's not causal, but it helps us to try and like get a little bit closer to how does this again, but we have to be confident that that system's representing reality well. So, like can we build in some of those physics? And then we've um uh we've talked a lot about those and complex systems. So there's a lot of other things. You can't just like, hey, I'm gonna change an input and I'm now gonna have a causal, but there's other ways, and we don't have to have this just like the you know, p-value mining, like correlation, ice cream cause a shark attack type work. And then on the full other side, you have the the blind uh academic trial placebo studies. There is a little bit of a middle ground uh where we can be working, at least getting a little bit more confidence and trying to with um uh different types of uh, and there's a whole bunch of work there, and that's where Judea Pearl is one of the people that's really been focusing on on some of the other areas we can be uh trying to just get inspiration, causing some of the or pulling some of these causal methods into uh our more mundane workflows, specifically when we can't do those massive sample size blind studies.

SPEAKER_00: 15:39

Absolutely. And I think this tees up really nicely into the next idea, which is of the natural experiment, which is kind of this, we'll say mathematically degraded, but largely more scalable version of the randomized control trial. You could also imagine that there's some trials that you could never do. You could never take one community and say, like, let's cut down half their food supply. You can say, like, let's take one community and this one gets a hurricane and this one gets COVID and this one doesn't. And then let's see what happens. There's no way you could, you know, feasibly do this, ethically do this, or ever even hope to do some of some of these types of trials. And so, how then do we understand the effects of hurricanes? How do we understand the effects of food shortages if we can't do our randomized control trial? That's where we have the natural experiment, which is where we look at two communities in time or two people in time, and naturally we find that some event happens to one community that doesn't happen to the other. And if these communities are very similar in almost every other way, demographically, locationally, geographically, we would say that these are a strong match for each other. And we would say if the impact of this event caused this amount of depression, this amount of instability, this amount of health outcome in the community, we could then at least inferentially attribute this to the event. And we would have an almost causal explanation of what happened. We could repeat this over multiple communities, we could compute this over multiple times. Um, and in fact, this was part of the work that I did in my PhD, is that we would study different communities that have had, uh, let's say, for example, their first case of COVID. And we would see when the first case of COVID happened in every single community, because it didn't happen at the same time, and we would see how it affected mental health. And we found that there were strong associations that people's anxieties were going up in a quantifiable way. And the only and the only thing that was uh happening to these communities at that time was that they were getting their exact first case of COVID when we removed effects of, let's say, national news. And so in this way, we can learn stuff about communities without having to then apply the randomized control trial to them. We don't have to go to the community and say, like, okay, it's time for you to get COVID. We can actually understand just based on the natural world and the timing of events. Uh a weakness of this design though is that you can only study events retroactively, right? You can really only go into the past and you can only study what has happened. And scientists are clever people and they're smart and they know how to read data and they know how to extrapolate from events and write these kinds of natural experiments, but it means that you're either bounded by sample sizes that you can realistically do in a controlled research setting or large-scale events that have already happened. And economists are no stranger to this, and they have the power to go into into uh like poorer nations and then do these kinds of economics experiments where they, you know, inject wealth, inject factories, and try and observe what happens on scales that we can't always do in other in other countries and other spaces.

SPEAKER_02: 19:02

Yeah, and this is where like I think the you kind of put the gold standard, the second gold standard. Now we're getting into the little bit where kind of I talked about the counterfactuals. You get into like the the the middle ground, slightly not as methodologically correct, but gets us closer to those areas. And I think it's I've been reading Euclid's Elements, and that it's really about like the breaking down into the first principles and axioms and and and things like that, and the postulates and like the just the really the building blocks that we can be having for those first principles. So I think this is where and some some of the past modeling I did uh as well uh around the economic spaces, Tid is mentioning, is you can find what are those relationships we do have those first principle relationships with so that we are confident that two plus two equals four. There's some of these uh these facts that we just kind of take for granted. So, what are some of those as we're building up? Um uh and there's the more you dig in, there's there's less in like economics and things, but we do know certain relationships. So, can you build like on top of those factors? You definitely Do you know are accurate, like I mentioned earlier, when we're building um uh like if you have a physically correct or like a mathematically correct model where there's certain areas you do know as fact, we can get a lot closer to uh to what are those real relationships when we're building systems, which is what the goals are is how do we make systems that are more uh resilient, robust, accurate, and ideally have these elements of causal analysis if we can't always go to the level of like an academic study. And as Sid was alluding to as well, there's sometimes it's unethical for us to do certain of those academic tests. So, how can we try and walk into that? So, this is a very active area of research, still is like this causal inference and figuring these components out, but it's something that I think is I'd love to see as back to uh as like last couple episodes we talked about, kind of like what's some of the newer things happening in AI and where we want to see come back to. Is I think we had a lot of great progress on on causal inference and and causal analysis and things. And now with the LLM craze, we kind of like have parked all of that. But I'd love to see the trending back towards there's such a fertile area of research here is if we can be building on those first principles and we uh and things we do know and then uh and then working on robust analyses and counterfactuals and and simulations around the components we don't know as well, we can get a richer, uh, more accurate view and and more efficient systems versus predict the next word in in some of these areas.

SPEAKER_00: 21:28

That's exactly right. And it almost begs the question of well, you've told me about this wonderful causal method, and you've described sure that there's a few problems with it and a few difficulties with it, but why don't we just do everything in this way? Why isn't everything just a causal study? Then our models would be correct, they would be grounded in physicality, they'd be grounded in mechanics. Uh what's so difficult about basically taking these methods and just making everything causal? And we can go through a couple here. And I I think, you know, namely, it's still a form of inference, right? We can get close to understanding causation. We can definitely say things like uh, you know, if you're rolling a ball down a hill and it hits another ball, the other ball will will roll further down. Um, but for most things, it's not so cut and dry, and we're often still in a in a world of inference and saying that the causal mechanism is available to us, right? Even in the case of giving a medication, we might say, like, oh, the medication is is reducing inflammation, but what if the medication just gives you a little bit more energy, gets you out of your chair more, they're moving a little bit more, and now from that their inflammation has gone down. We don't know that. We can only infer the outcome on the other side, right? Even in effect, even if an effect is better explained by a different cause, the traditional causal method studies only let us infer connections between events, but they don't tell us implicitly or explicitly the entire causal chain, as we mentioned before, that's occurring. We also have the same trouble with statistical studies, which is that if you don't have enough findings, you might just have a chance finding. Right? So we wouldn't publish a correlational study with five uh subjects of participants. You definitely can't do that for a causal study.

SPEAKER_02: 23:20

And this is, I think, a key thing of uh we said a lot on here, bringing stats back is kind of like our unofficial, one of our unofficial slogans on the podcast. And statistics, I don't it has a marketing problem. Let's just be honest. They have a marketing problem, but statistics is so incredibly important to pretty much any part of life. Like policymakers as well, if they rely on data. I know with the government shut down, there was some lack of some data there and stuff, and there's this constant question of like how do you trust the statistical, like uh the uh the result, I mean the data that's coming, but like think about it, investments uh for anything in the stock market, anything with GDP analysis, any like even uh you know, Gallup polls or anything, everything doesn't, this is not an AIable thing longer term. Like you have to have a how do you trust some sort of data to know what the signal is. Even like usage of AI, people are throwing statistics around there, but numbers can say anything you want them to say, and it's the integrity of having quality statistics cannot be understated because everything is pinned on it. For instance, you could have the best causal study as Sid was describing for like uh let's test the the is this does this medicine actually reduce cancer or what have you? But if you as Sid is now bringing up, if you do a bad job with the statistics around that, or you have poor sample sizes, or you you you you don't have the segregation you'd have, like there's so many ways you can make that a statistically invalid. So like you the results are essentially meaningless. And every part of life relies on having, not every part, but like majority of uh of like major policy type decisions rely on having accurate data. So that can't be undersold. And I remember a couple of years ago, the the American Academy of Statisticians released a kind of a statement piece saying, let's no longer like publish p-values and things like that in papers, because that's very true. As it kind of had traditionally been that if you can find a statistically uh significant study, you can publish it. If not, it doesn't matter. But it's like sometimes the best publishing could be you you do a full study and you found out there's no statistical relationship, still publish it. But that's the whole academic conversation of people only want to find findings. But that that causes what there's the sample size, there's the statistical power, there's the white, like how are you interpreting? There's so many different elements of like statistics is really complex. And it's uh it's not something that can just be automated away with an AI test as an example. But all of this causal and the root of all of these things will comes down to statistics, data, and integrity. So that's a whole nother podcast for another day. But I think that very much underscores of that's a weak link in any of these causal now, like the best causal study has a weak link if you didn't do the statistics correctly.

SPEAKER_01: 25:53

As the marketer in the room, you say statistics has a marketing problem. I say marketing has a statistics problem because this is exactly what happens. Everything seems to have need to have a cause. You see, data has a cause, but they forget is that there's environment, like there were decisions and environment around that, and that trips people up so much when they're trying to put reporting together. I don't know. My short answer is I'm making this required listening for the every marketing operations and rev ops person that I know to get some knowledge and some and some good backup for like what just because there's this trend, you need all of the data around that trend and decisions made around it to actually make your story plausible or prove that it wasn't a real relationship at all, like you were saying.

SPEAKER_02: 26:40

Well, and this is where we're really, I mean, we are really bifurcating conversation going super tangential here, but it it's true. We're really living in a post-truth world for a lot of stuff. Like, pretty much whatever your AI bot wants to say, whatever your marketer wants to say to get you to click on something or buy something, that's what people are taking. Like there, it's just the integrity of what's true or not. Like someone will say, This causes this. No, like a marketer saying that means nothing to me. That's why like I don't read marketing materials uh often because it's like it doesn't mean anything, right? Because it if everybody's just saying whatever they think they can get the clickbait on. So uh, but that's where thankfully within uh we're trying that's where academia is still trying to have that bastion of the of the truth on it at times. But it it's uh it's a big, I think this could be another series for us is proper statistical controls.

SPEAKER_01: 27:28

Yeah, controls and reporting. Awesome.

SPEAKER_00: 27:31

Yeah, absolutely. And I think that we're all kind of highlighting the same problem, which is that like while everyone is capable of just going on a forum or report or on the internet or on LinkedIn and saying, hey guys, I found a causal claim. Only people who go through these rigorous scientific methods and really think fundamentally about how we understand and establish and debate causality are able to make these strong claims. And often people making the claims correctly are are drowned out writing these like really dry research papers. But I think that I want to emphasize that there is a burden on this explorer and the studier and the investigator to really think deeply about how they came to the claim they came to. Is it an ad hoc explanation that confirms our biases? Does it help us sell some product we're trying to sell? Does it simply feel good because the data looks really nice? Uh and so we should really think deeply like scientists if we want to make these kinds of claims.

SPEAKER_02: 28:30

And I I think that's the uphill battle we're we're facing that we want to like that's one of the lofty goals of this this podcast. Uh is that with the whole you know, LLM scrapie internet, finding whatever chat uh rooms say, and then like aggregating that up and calling that's that's what the right answer is. That's not. And so it is a thing for the the statistics organization in general or or researchers is how do we get better at marketing the truth and having uh like being able to show like when this is a real causal study, how do you have uh and there's lots of people working on this problem as well, but like that's the thing that you should be focusing on, like that because that drives as we get into like tying it back to the metaphysics, and I uh said you have some great closing stuff on this as well, of um as we're looking into like reasoning and thinking, and I think the whole basis of this series is like getting back to the basics, and uh what like I've said before, like the world changes really fast, but then it but it doesn't. So, like we're all on this feels like a uh you know a rat race treadmill thing here when an AI is changing, you can't keep up with all the AI updates every every day there's a new paper type thing. But then like nothing actually is fundamentally changing either. It's like we're in two different dimensions uh on speed of like the most successful people at anything are Michael Jordan as an example. He nailed the fundamentals better than everybody else. Like there's nothing so magical about like how are you the best basketball player? You're better at the basics than everybody else, right? Like there's that kind of thing. It's the same with like causality and understanding what like you'll read like a great book like the this one I was saying with like Euclid, and like when we're looking at Aristotle and some of these things, like it just it's like the comp or I think Richard Feynman uh from physics does an amazing job of this. Like you the most complex topics and you can explain them super easily and it makes sense of the logical this plus this equals that, and we're only on the first principles, and then the research side is trying to be showing how do you make those next postulates and things and proving those next little areas, and then you can incrementally adding to knowledge to adding that causal aspect. And like it's like we're trying to skip too many steps these days with like LLMs, or we're trying to make somehow we're gonna growth hack productivity in a company by like using agents and things, and there is some some there's some things to explore there for sure, but it's like we can't do that by skipping steps. I think there are things we can be doing there, but we it would behoove us all to come back to like the causal relationship and understanding and using like in the incremental of how do we have these things, and this comes back to like as we talk about AI trust and AI governance and what what what we're really focused on for AI risk risk uh management and and risk reduction is how do you know that that aspect is performing expected, and you want to be limiting the number of like stochastic, random we don't understand correlation things. There could be some alpha and things like that, and like the when you're building an AI system, there could be some areas of like we know this part is a little bit fuzzy, but we're doing correlation analysis. But we have a lot of rooted causal first principles, two plus two equals four things we're all comfortable with, and we limit the number the amount that we're not that is being stochastic versus the whole thing being like we're we it's an LLM, it's fantastic, we're predicting off of this, and it's everything's gonna be great, but we don't understand how any of it works or how how we're building up to it.

SPEAKER_00: 31:43

And and just to tap us even further back into metaphysics, we can go back to Rene Descartes, who makes the claim that causes are just as real as objects. And I think that in a very critical sense, this is true because the properties of an object arise from the causes that produce that object. If you create an LLM just based on statistical inference, just based on X sort of prediction, whatever you produce at the end can only be statistical. It cannot then step outside of that and produce or cause some object, which is like superior to the underlying fundamentals that are baked into it. Uh so you you run into this problem where if we don't consider the causes and we don't consider this this this causal method, we quickly lose track of what we're actually talking about. We lose track of the object because we haven't considered what caused or what created the object.

SPEAKER_02: 32:44

And this is where I think uh tying into a little bit of some of the current events we're having, uh world models. I know we're getting a lot of press right now, um, uh as kind of like the next frontier after LLMs. In theory, they sound great. They are trying to have some of these causal relationships and things. In practice, I'm not I'm not sold yet that it's not gonna be a next one like one year, two-year Silicon Valley hype train of like we're gonna capture the whole physical world with its causal relationships in a model and sell it, and it's gonna be and it's gonna capture all this. I'm not I'm not sold yet that that's gonna happen, but I think the direction is definitely a positive development.

SPEAKER_00: 33:22

I think to take all this together and kind of like put a put a bow in our thoughts here, as we're looking at this landscape of causal modeling and we contr and how we contrast with statistical and causal learning, we should remember that correlational findings aren't bad. People aren't gonna stop publishing them because they're really powerful and useful for creating predictive models, right? Trying to predict outcomes and futures based on data to infer about the future. These are things we can't do with causal studies. Causal studies really help us understand the world around us, to build these physical models, to build these world models that help enrich and deepen our understanding of the world around us, but don't always let us do things like predict stock prices the next day, because that would require a level of physical modeling, which is just not possible yet. So correlational studies absolutely have their place, and it's why we use them, and it's why data scientists have jobs, is because they help us understand and answer questions that we can't always study through causal modeling.

SPEAKER_02: 34:26

Totally agreed. And that's and thanks for time that in. It's like it's not that that correlation stuff analysis is bad, it's specifically like stock markets and things. I mean, is there really and some real properties? It's it's all human behavior analysis, right? And like follow the herd FOMO and predicting that more than he thinks there is fundamental things. This is like how Warren Buff and Charlie Munger got to be billionaires. There are actual processes, but most of the stock market is going to be that correlation type stuff. So there's definitely there's a need for it, but it's like the how do we ground ourselves in what's the, you know, we know, but what spade spade, right? We know what's correlation, we know what's causation, and we st uh and we embrace that. And we that's the part of like we we know what the correlate of the causation parts and models are and what's the correlation, and we're all aware of that. We put more controls around those stochastic correlation type things, but that's fine. It's when we start kind of merging and and blurring the lines that it gets a little bit iffy.

SPEAKER_01: 35:25

Well, Sid and Andrew, this has been a fascinating discussion. Um, I know that I'm as I work as I work in data daily, um, I learned a lot just from your discussion. And I think before we conclude, next episode is our last episode. We're getting to the final moment on what is thinking. Is that right?

SPEAKER_00: 35:48

I don't know if it's going to be the next episode in sequence, but the next episode of the metaphysics series will be how we close out the series. And I'm very excited to address our first question, which is what is thinking and can AIs do it?

SPEAKER_01: 36:00

Awesome. Well, for our listeners, if you have any questions about this episode or any of the episodes in our metaphysics series, please reach out to us at ai fundamentalists at monetar.ai. Until next time.