A Collectivist, Economic Perspective on AI

0:00If you've been following the conversation around artificial intelligence lately, it often feels a bit untethered from reality, doesn't it? One minute we're hearing about these utopian futures where AI solves everything, and the next it's a dystopian nightmare about to change humanity forever. It's just, it's a lot of noise. A huge amount of noise. And that's precisely the challenge, right? Is this popular AI framing truly grounded, or are we maybe missing a crucial piece of the puzzle? Our aim in this deep dive is really to reveal a fresh, maybe more grounded perspective. perspective, viewing AI not just as some like sentient thinking machine, but through what we're calling a collectivist economic lens.

0:42Our mission today is to explore how blending these economic and social concepts with computational and inferential ideas might pave the way for a brand new, truly human-centric field of information technology, moving beyond that traditional, often quite individualistic view of AI. Yeah, the old view. So, okay, let's start by unpacking the current AI narrative, because the terms themselves, They have a history. AI was actually coined way back, like in the 1950s. A long time ago. Yeah. And then in 1959, an AI researcher, Arthur Samuel, he coined machine learning or ML. And ML really became this sort of powerful intellectual bridge, didn't it?

1:19Connecting data intensive research across fields with computer science. That's a really crucial distinction. And it's striking how the emergence of large language models, LLMs, you know, really triggered the return of that older AI phrase to prominence. Right. The underlying ideas for LLMs were still, while firmly in the ML tradition, they use sophisticated optimization, these gradient-based methods with just massive amounts of human language data. Huge amount. But their outputs, they were so strikingly fluent, it gave the appearance of a single sort of human-like entity speaking. You know, it is striking how much the conversation around AI still seems to hark back to that 1950s vision.

1:57You know, thinking machines or even artificial general intelligence, AGI, this idea that we just want machines to outcompete human cognition. But when you look closer, that goal starts to show some cracks, doesn't it? What's fundamentally missing there? Well, the core problem, I think, with the thinking machine goal is that it fundamentally misses a huge piece of human intelligence. We are social animals. A vast amount of our intelligence isn't just what happens inside one skull. It's profoundly social. It's cultural. And beyond that, this sort of narrow view often treats the social consequences of the tech as almost an afterthought.

2:34Which seems crazy. Which, given the massive societal impact these things are poised to have, is just, well, it's unacceptable, really. So if that individualistic AI goal is problematic, what's the alternative? What's the path forward? You mentioned this collectivist perspective. Tell us more about how that fundamentally shifts the approach. In essence, it involves revisiting an older idea, actually, that complex systems through pretty simple local interactions can give rise to emergent, intelligent, system-level behavior. And here's where it gets, I think, really interesting. In these future systems, data will be increasingly partial, noisy, biased, local, contextual, even strategic and incoherent.

3:16Wow. OK. That sounds like a mess. It sounds like a bug. But surprisingly, this isn't a bug. It's actually a feature. A feature. How so? Well, to unpack that, we need to consider two crucial interconnected themes, social and uncertainty. OK. True intelligence involves not just knowing stuff, but knowing how to act when you only have partial knowledge. And crucially, knowing how to interact with others whose knowledge is also partial. Right, like in the real world. Exactly. So if we take another look at an LLM, while it seems like this single thing, it can be better understood as a collectivist artifact.

3:50You're implicitly interacting with a vast number of humans who contributed data, opinions, language. So in that sense, you could even say an LLM is analogous to a culture. Cultures are these rich repositories of narratives, abstractions, collected wisdom, you know. That's a really interesting way to frame it. This idea of uncertainty then, it naturally leads us to think about how current LLMs, well, they kind of fall short in managing real world uncertainty. They do, massively. They can be systematically overconfident and they often lack the ability to truly understand where their data came from or how fresh it is.

4:26Which makes them inherently untrustworthy for anything high stakes. You wouldn't bet the farm on it. Definitely not. But a key takeaway here is how human intelligence, both individually and especially in groups, excels at navigating uncertainty. Think about it. An individual human foraging for fruit, maybe they find some, maybe they don't. It's uncertain. Yeah. But once a collective creates a market for produce, the uncertainty about finding fruit just drops dramatically. And that certainty then allows for totally new ventures, right? Someone can open a pastry shop because they can actually depend on fruit being available reliably.

5:02The market manages the uncertainty. That's a great example. And this leads us to another really counterintuitive insight you mentioned about how intelligence works in groups. This probability matching idea. Ah, yes. You'd expect like optimal individual behavior, right? But in experiments, say a mouse maze with more food on the left branch, maybe two to one ratio. The standard setup. An optimal mouse rationally should always go left, always. But real mice, and actually humans in similar situations, they often probability match. They'll visit the left branch, say, twice as often as the right, matching the ratio.

5:39Which seems inefficient for the individual. Exactly. Why not always go left? This really highlights an aha moment, I think. While probability matching might seem suboptimal if you just look at one mouse in oscillation, It can actually lead to high overall social welfare in a collective setting. Think about it. If every mouse went left every time, all the food on the right would just sit there. An unexploited resource. Ah, right. Wasted. Probability matching prevents that waste. It spreads the exploration. It suggests it might be an evolutionary equilibrium, something that's good for the group's survival, even if not perfectly optimal for one individual at one moment.

6:15Okay, that's fascinating. And this connects right back to the idea that if our technology, our AI, is going to truly help rather than hinder adaptive decision making on a large scale, it needs to be designed with these kinds of collectivist principles in mind. So applying this lens, we can start to see how the Internet, you know, despite being this amazing hub for collectivist activity like Wikipedia. Yeah, prime example. It often features markets that are, in a sense, kind of defective. They fail to properly reward creators or value data effectively. or even foster trust. And this often stems from, well, a foundational lack of thinking rooted in microeconomics and what's called mechanism design, thinking about incentives.

6:57So it's not just about efficiency, like recommendation systems. Well, that's the surprising thing. Traditional recommendation systems, yeah, they might make existing markets for, say, physical goods more efficient by suggesting purchases you might like. Right. But often they don't involve money changing hands directly within the recommendation system itself, or a deep consideration of the incentives of the system. They just make an existing market work a bit smoother without fundamentally rethinking its structure. Okay, let's make this concrete. The recorded music market, that's a good example, right?

7:32In the classic online model, musicians upload music, often for free. Yeah, hoping for exposure. And the platforms profit through subscriptions or ads. There's often very little direct incentive baked in to actually pay the artists fairly. In fact, you could argue there's even an incentive for the platform to use generative AI to replace the musicians, potentially cutting costs. Ouch. That's bleak. It can be. But if we connect this to the bigger picture, a collectivist alternative like the one seen in companies such as United Masters where the author of the paper we're drawing on is involved, they designed a three-way market.

8:08It involves musicians, listeners, and brands. Adding the brands is key. Absolutely. This design fundamentally incorporates direct incentives. When a brand uses an artist's song in an ad, for instance, the musician gets paid right then and there. Instant feedback. And audience reaction to that ad, that song usage, is measured. Positive reactions incentivize other brands to partner with that artist. It creates a positive feedback loop. So it's creating real value and job. Exactly. This collectivist AI system, this market design has genuinely created opportunities. United Masters has signed, I think, over 1.5 million musicians now, and their music gets used by major brands like the NBA, Bose.

8:51It's a different model. That's a powerful example. Now let's shift to another complex area. Data markets. You can picture this maybe as a three-layer thing. A user gives data to a platform for a service. Like a social media app or something. Yeah. And then that platform turns around and sells aggregate insights from that data, or maybe even the data itself, to third-party buyers like advertisers. Right. This raises a really important question for the user. Are they truly incentivized to participate, fully or honestly? The critical problem is that the user often loses control over their privacy.

9:26Yeah, you click agree, and who knows where it goes. With potentially no new service in exchange for that specific data sale? The privacy laws can become, well, unbounded. You don't know the downstream effects. However, a collectivist design, thinking economically, can create a different incentive structure. Imagine platforms formally guaranteeing privacy by adding a specific amount of statistical noise to the data they share. Okay, like fuzzing the data. Exactly, a level of noise they contractually specify. Users could then literally shop among platforms for a privacy level they're comfortable with alongside the service quality.

10:01So competition based on privacy guarantees. Precisely. Platforms that attract more users, maybe because they offer better privacy, get more data overall. This improves their service, potentially, even as the data buyers might pay a bit less because the data is noisier. It sounds like a complicated balancing act. It creates a complex game, absolutely. And understanding the outcomes, the equilibria of that game, needs both economic thinking and machine learning thinking working together. Okay, so this brings us really nicely to these three complementary thinking styles you argue are crucial for designing systems in these social environments, systems that take uncertainty seriously.

10:40A trifecta. Computational thinking, which is about algorithms, data structures, the mechanics. Yeah, the CS basics. The inferential thinking about learning from data, making predictions, handling uncertainty. Statistics and ML territory. And finally, economic thinking, which focuses on incentives, strategic behavior, markets, equilibrium. Right. How agents interact. Let's dig into inferential thinking a bit more. It's not just running calculations, is it? No, not at all. It's more than just descriptive statistics. Imagine, say, a hospital database. You can compute averages for the patients you have.

11:15That's descriptive. But inferential thinking is about designing algorithms to extract value and predict outcomes for new patients, future patients from that same population. And crucially, it includes asking what if questions. Causal inference. Exactly. Causal inference. What if this patient had received a different treatment? Can we estimate the population level effects of different treatments? It's about drawing conclusions, often causal ones, about unseen entities or future scenarios, and critically providing some measure of uncertainty around those conclusions. Got it. Then we shift to economic thinking.

11:52This is about designing incentives to shape behavior, especially when the people or systems supplying the data are strategic agents with their own goals. Precisely. They're not just passive data points. They might act in ways to maximize their own benefit. So it's not just about finding the single best solution like optimization often aims for. Often, no. It's frequently about finding an equilibrium, a state where different agent strategies are in balance, where no single agent has an incentive to unilaterally change their behavior. And it's compelling to see how the field of mechanism design kind of flips game theory on its head.

12:26It's sometimes called inverse game theory. Inverse. Yeah. Instead of analyzing a given game, you start with a desired outcome like fairness or efficiency or maximizing social welfare. And you ask what game, what set of rules and incentives would lead to that outcome as an equilibrium. Designing the rules to get the behavior you want. Exactly. And for these large scale collective systems where one agent's actions can really influence others. Think big platforms concepts like sequential play where one player moves first become important. This leads to things like stackable equilibria, where a leader makes a move and followers react optimally.

13:04Information asymmetry is a huge challenge here, too. Agents know different things and might strategically hide knowledge. Mechanism design addresses this, often through designing clever contracts. There's even an emerging field called statistical contract theory that explicitly integrates inference and uncertainty into contract design. That sounds powerful. Can you give us a real-world application? Sure. think about regulating self-driving vehicles it's a huge challenge definitely you can frame it as on one hand an inferential challenge it's a bit like the FDA testing drugs you need to control the rate of false positives saying a car is safe when it isn't and false negatives saying it's unsafe when it is statistical testing okay makes sense but it's also fundamentally a microeconomic issue the AI companies building these cars are strategic agents they want to design their vehicles and their internal testing protocols specifically to pass the regulators tests because that leads to profit.

14:00Ah, so they might gain the test. They have an incentive to. So statistical contract theory offers tools here. Regulators, acting as the leader in the Stackelberg game, can design testing protocols and requirements essentially. Contracts that allow them to control those overall false positive and false negative rates for the public, even when they know the companies, the followers, are strategically choosing how to design their systems and present their data to pass the test. So the contract anticipates the strategic behavior. Exactly. It uses advanced statistical concepts, sometimes called e-values, which quantify evidence strength, to create menus of options or payoff structures.

14:40Companies pick what's best for them from the menu, but the regulator still maintains the overall statistical guarantees. It's pretty sophisticated. Wow. Okay. Let's also touch on the really critical issue of bias in AI systems. How does this collectivist perspective help illuminate that problem? Well, it highlights that future AI systems won't be monolithic. They'll involve many diverse participants, humans, other AI organizations interacting in these dynamic networks. Okay. And maybe surprisingly, information asymmetries will likely be common. Participants will often strategically withhold specific knowledge or data to gain value or influence.

15:17this will probably lead to the emergence of hubs of specialized expertise within the network. But not everyone knows everything. Definitely not. And the network shouldn't just be seen as data flowing freely. It's better viewed as a web of contractual interactions, ideally based on principles from mechanism design, governing who shares what with whom, under what conditions, and with what incentives. And the data itself, as you said earlier, will often be very local, contextual, maybe even fleeting. Absolutely. Knowledge relevant in one part of the network might not apply elsewhere or might expire quickly.

15:51This creates a real need for ways to locally vet or validate information coming from outside your immediate context. How do you manage that? Well, there are emerging techniques, like something called prediction-powered inference. This allows you to take a global models assessment, maybe about bias or performance, and adjust it using local ground truth measurements. It lets you mitigate unwanted biases coming from the global model or even intentionally impose desirable local biases if appropriate for that context. Interesting. So this complexity, these networks, asymmetries, local knowledge, it sounds like it creates new roles.

16:27Oh, inevitably. Just like in previous technological shifts, we'll likely see the emergence of many new specialized roles. AI auditors, data brokers, knowledge aggregators, privacy guarantors, maybe even bias mitigators. And these roles become points for oversight. Exactly. They create important touch points for regulatory control, for society to step in and mitigate biases or enforce standards that align with legal or ethical considerations. So bringing this towards education, how should academia adapt to foster this collectivist perspective? The source material talks about a missing middle kingdom in AI education, right?

17:08Referring to a diagram. Yes, figure four in the paper. It shows how academia has developed these pairwise blends of the thinking styles we discussed. You've got machine learning, blending computation and inference. Right, the dominant blend currently. Econometrics, blending economics and inference. Common in social sciences. And algorithmic game theory, blending computation and economics. Mostly theoretical computer science. But the argument is these pairs aren't enough. Exactly. If you connect this to the bigger picture, these existing pairwise blends are often insufficient for the real world systems we're now building.

17:39Algorithmic game theory, for instance, often makes little use of real data or inference. Econometrics, while strong on inference and economics, often doesn't leverage large scale computational techniques from modern machine learning. Right. And machine learning historically has used very little from incentive theoretic economics or mechanism design. It treats data as passively given, not strategically produced. So you get these gaps. You get gaps, you miss crucial connections, and frankly, it leads to an impoverished education for the emerging AI workforce who need to build and manage these complex sociotechnical systems.

18:14What's really essential, the argument goes, is a true tripartite blend. All three together. Computational, economic, and inferential thinking integrated from the ground up. This is needed to bridge those gaps, to make academic research more relevant to industry's real problems, and to provide a common language and conceptual toolkit. It's truly a necessary middle kingdom that connects the technical side, the engineering, with the social sciences and humanities. Okay, so to summarize our deep dive today, the core argument is that integrating these computational, economic, and inferential concepts is absolutely vital.

18:50It's vital for envisioning the future of information technology in a way that's grounded and human-centric. And this tripartite blend, it allows for a much more subtle, more nuanced discussion of really complex issues like fairness, privacy, data ownership, transparency, moving way beyond simplistic kind of black and white distinctions we offer here. Absolutely. And just to build on that, it's crucial to remember that other disciplines are also incredibly essential for shaping genuinely human-centric technology. Cognitive science, social psychology, law, ethics, the humanities. Right. It's not just these three.

19:24No way. Think about behavioral economics, for example. It explicitly links cognitive and social psychology with economics to help us understand how real humans actually behave, often deviating from purely rational assumptions. We need all these perspectives. So thinking about the listener, what does this all mean for you? The takeaway seems to be that AI needs to mature. It needs to become a true engineering discipline, maybe like chemical or electrical engineering did centuries ago. Yeah, that's a good analogy. Those fields managed to bring incredibly complex physical phenomena under control.

20:00They developed modular, transparent design concepts. And importantly, they had foundational theories. Foundational equations like Schrodinger's equation in quantum mechanics or Maxwell's equations in electromagnetism to guide them. A solid theoretical medrock. Right. But AI currently often feels like it's kind of winging it when faced with these incredibly complex cognitive, social, commercial and scientific phenomena. It doesn't quite have those universally accepted foundational guiding principles or equations yet. That's a really important point. And it raises the final question, the provocative thought perhaps.

20:33The path forward for AI, for information technology more broadly, it demands more than just piling up more data and throwing more compute power at problems. It needs something deeper. It requires, and I'm quoting loosely from the source here, are overarching hard-won general scientific and humanistic principles, including rationality, experimentation, dialogue, openness, cooperation, skepticism, creative freedom, empathy, and humility as daily companions on the journey ahead. Wow. That's a call to action. It really is. So maybe consider how can you in your own learning your work or just how you think about these technologies apply these broader principles when you're engaging with AI.

21:15It's not just about the code. It's about the context and the values we embed within it.