Everyone loves a trilogy, so let’s wrap this up. I told you why agent-based modeling never broke through in economic research. I told you how it could eventually matter. Now I’m going to tell you how AI could, and maybe should, actually make it happen. And as someone who is not fully “AI-pilled”, I will confess that there aren’t that many avenues in which I am particularly excited about AI contributions, particularly in the arts, and even where I see opportunities for great things, I remain pessimistic that Gresham’s Law is going to hit us hard and fast while we are still learning to culturally and institutionally separate the wheat from the slop-riddled chaff.
But let’s put on our more optimistic glasses for the moment and think through what the opportunity for AI to guide agent-based economic modeling into a new (and first) golden age.
First, let’s define a few terms. When I refer to Agentic AI I am specifically considering an LLM model that is capable of autonomously pursuing goals, taking actions, and adapting their behavior in response to new information, all based on a set of pre-defined rules and reward functions and with little to no need for continuing human instruction. When I refer to an Agent-Based Model (ABM) I referring to a computational model of an economic or social system in which individual, heterogeneous agents (consumers, firms, workers, etc) make decisions according to specified behavioral rules, interact with one another and their environment, and collectively generate aggregate economic outcomes.
I don’t want to be wholly redundant with my previous posts, but it’s important to revisit a key theme that ran through both posts: the greatest strength of agent-based modeling is it’s ability to accomodate incredible depth of individual and institutional heterogeneity. Which is a nice way of saying that it is parameter rich. Which is itself a nice way of saying that the models can generate any result without the prospect of credibly investigating the mechanism behind those results and, more importantly, suffiently culling enough noise from them to learn anything about how the world works. For those less familiar with the curse of dimensionality, let me explain.
If a model has 2 parameters you can simply run a grid search of every possible combination of those parameters and observe what kind of outcomes are generated. If you had an agent-based model with 100,000 agents randomly interacting you could still investigate the parameter space by “running” the model 1000 times for each possible combination of parameters and look at the full distribution of outcomes, reporting the mean, variance, and individual quantiles of every outcome of interest for each combination in your grid of parameters. Your results might look like this:
This is a useful way to design and investigate an agent-based model. But you probably already see the problem. Sure, toy models with two parameters are cute, but how do I answer complex economic questions? Toy models can only take us so far. And that’s the rub.
Once you start really start trying to advance economic theory of large, complex economic systems with an agent-based model you’re going to have a lot of visible parameters and lot more less visible parameters. An ABM will be composed of:
Agents (people)
Collectives of agents (firms, clubs, mosques, families, states)
Rules for agent decision-making
Rules for collective decision-making
A context, network, and/or phystical environment with it’s own rules
Every single item on that list will in turn have attributes with both
Initial conditions
Distributional characteristics of those inititial conditions i.e. mean, variance, etc
Are they fixed, exogenously varied, or endogenously varied over time?
Take the first list and add up the total agents, collectives, rules, etc. Then multiply those by each other and by every decision that has to be made in the second list. That number will likley be in the thousands. Congrats, that’s how many true model parameters you have.
With modern macroeconomics, and I am speaking here as a non-macroeconomist, your goal is to reduce that parameter space using basic logic, empirical observation, and modeling decisions that collapse heterogeneity to a small number of representative agents or types. You produce equilibria predictions at each stage of your model and compare them to objective reality. Once the model is sufficiently micro-founded and validated, you can then generate predictions and investigate mechanisms.
The whole point of agent-based modeling, however, is not to collapse the model to a small numbe of representative agents. It’s to endogenously grow aggregate outcomes from the individual endogenous decisions and how, through interactions within the model, they aggregate into predictions for a social system. So what’s the problem? Let’s assume every agent type and collective is characterized by 5 attributes and that each attribute is characterized by 3 statistical properties. Each rule comes with 3 functional variations, characterized by 3 weights each. We’ll then say that there are 5 total agents ( 2 individual types and 3 collectives), for each of whom half of attributes are fixed and half are endogenous. They manner in which they are endogenous has 3 functional variants. If we want to do a true grid search, we’re looking at least at a 5000 combination parameter grid space.
And we are no where near done yet, because we haven’t considered how time works. How steps and sub-steps are ordered. What decisions are synchronous, asynchronous, hieracherially ordered or random. We haven’t considered the social or physical environment What’s you start digging in, your parameter space starts gaining orders of magnitude. And then you realize the real problem is that true grid search of the entire parameter space would take until the heat death of the universe.
And that’s what research expertise really is – competently making decisions about what can and should be varied, what can and should be assumed, etc, until we have a tractable question and well-executed research endeavor to answer it. The ABM space is too big for a researcher, however, because too much of the competent execution of research is unobservable, possibly even to the researcher themselves, as the model and the interactions of each decision become intractably complex. Too big for a human researcher. But what about an Agentic AI?
What if we created a base ABM, handed it to an Agentic AI, and told it to recursively evaluate and update the model until the model and the results it generated met a pre-defined set of necessary meta-conditions that the researchers defined as i) a falsified theoretic prediction, ii) generated via observable, definable mechanisms, iii) dependent on a maximal set of modeler decisions, and iv) and independent of set of unlimited, but fully observable modeler decisions?
This is going to take a long, but not necessarily intractable amount of time because we can also define a series of “golden section” adjacent rules that allow the AI to decide with a parmeter, rule, or decicsion is sufficiently irrelevant to ignore or sufficiently constrained to tractably investigate as key outcome determinant or source of outcome sensitivity.
As long as I am giving unsolicited advice to graduate students everywhere interested in agent based modeling, I would suggest forcing such an Agent AI modeler to exclusively use a constrained, pre-defined modeling language or library. Further, you probably want to force a literal, numerical maximum on the number of lines or characters of code. And, lastly, you’ll want to come up outcome rules by which the Agentic AI has to itemize and defend every single deviation of the final model from the base model handed to it. By constraining AI to a limited variety and quantity of coding objects, you are in turn ensuring that the resulting model and the internal mechanism will be transparent to the researchers managing the AI and the broader discipline evaluating them.
Agentic AI offers researchers the ability to iterate through the prospective solution space for modeling problems otherwise intractable via brute force. Further, it allows researchers transparently reveal the instructions that guided the agentic AI dimensions on which the generated model is sensitive to both varied parameters and modeling decisions. It allow for two hierachical levels of meta-research transparency.
That’s what I would be coding write now if I was a PhD student interested in contributing to economics and other social scientific disciplines via agent-based modeling.
There’s plenty of press about sticker shock on halloween candy prices this year, but the day having arrived and the candy now opened, I can tell you the biggest shock to me was how notably inferior it was to standard “full size” candy equivalents. The Kit Kats are thin and flavorless, the Twizzlers taste like bottom shelf generics. I had heard that chocalate products had reduced their cocao content, but Twizzlers? I tried a few other gummmy candies from other brands and they were inedible. As a livelong afficianodo of Halloween childhood sugar crashes and adult weight gain, this is a generational crime.
It’s a good reminder that shrinkflation often hits quality harder than quantity. Whether there is a direct tariff or not, the reality is that everything costs more to import than it used to. Cross product subsidies exist in every dimension when you’re trying to fill a cargo ship. Your kid’s candy haul this year is one of the purest ways to experience the thousands of sources of consumer surplus lost as we’ve crippled our own economy for no reason but spite and ignorance (all dusted with the unmissable taste of xenophobia and racism).
I used to worry about the first week in the office after Halloween and all the willpower that I was going to have to muster not to sample a Twix, Kit Kat, Snickers, or Twizzler every time I needed to get up from my desk.
“You Can See Everything” is one of the great film-making experiences of my lifetime. That’s not even a particilularly hot take these days. If you need a recommendation, yes, four stars, you should see it. The singular dominant theme, for my money, was the grand telephone game that “truth” is distilled through as it skips from source to source, medium to medium. By the end of the film you may find that, like me, you have updated your priors a half dozen times, left with your hands up in the air as the crushing weight of subjective reality threatens to abandon you in a fog of nihilism. Then you get home, rewatch the documentary (“The Inventor: Out for Blood”), scroll through a few bullet points of the actual case and remind yourself that no, this person is very much a crook.
But that’s not what I want to talk about. What I want to talk about is confidence.
We are told, repeatedly and with hardly any caveat, that confidence is a good thing. An admirable thing. A sexy thing. But confidence can very much be a trap, particularly if confidence becomes your comparative advantage.
The parable of the young genius has become rooted in the tale of frustrated and constrained greatness. They are precocious and unruly, possibly with a history of mild behavioral issues or sullen solitude. They follow the “normal” track for a few years, often to a prestigious college, where they catch a glimpse of great possibility and simply can’t be held back any longer. They drop out and become an entrepreneur. They pursue investment and promise unblinking greatness and unfathomable innovation with the kind of a committed sureness that only youth can provide.
They are not tradditional frauds, at least not yet, because they believe it everything they say. And why wouldn’t they? They’ve never been wrong. The first time they stepped into an environment where they might be challenged they left before any incontrovertible truths might be controverted. Let’s stop the story here because it rarely actually starts in such a way and, if it does, it usually ends quickly. Which is for the best. It’s not crazy to suggest that the best thing that can happen at this stage is failure. Whether it’s an expert in your field explaining the fallacy underlying your grand promises or venture capitalists simply sniffing out a bad investment, the train stops and you can return to building yourself. You had a good life experience, a new perspective on your own education and skills, and the a greater appreciation for what it means to have something to contribute. No harm, no foul.
But that’s not what happens to a lot of people. Not just undergraduate drop outs either. It’s what can happy to anyone that a market anoints as special. That hype can fuel you until you’ve been running on it for so long you realize, perhaps too late, that you’re no longer selling a product or even yourself. Your selling the story of you. You change your title on LinkedIn to “thought leader”, “visionary”, or “futurist”. You think you are selling yourself, but you’re not because there’s nothing actually there to sell.
And this where the trap is. What you are selling is your confidence in the words coming out of your mouth. The problem is that the words themselves don’t have any value. You don’t know how things work. You don’t know, actually know, biochemistry. Or stochastic modeling. Or tokenization and neural network modeling. You know of them and how to reference them in a manner that obscures your own lack of understanding. The only thing you know, to your creeping dread, is your continued existence in this market is dependent on continuing to dazzle others with the kind of confidence they can only dream of posessing. You’re a carnival barker without a carnival. And eventually someone’s going to walk into an empty tent and they’re going to be pissed.
And that is terrifying.
When you watch “You Can See Everything”, pay attention to how hard they are working to maintain the image that they know what they are talking about. The 3X speed videos and obsession with rapid vocabularly adoption. They know of chemical reagents. And mechanism design. And blood. And lasers. And whatever it is they desperately need to convince people they know about. And notice to whom they scrupulously avoid ever interacting with: anyone with actual scientific or technical expertise. They are always cramming for a exam they didn’t study for. I’d call them dilettantes, but dilettante implies non-trivial, if thin, knowledge about a lot of things. Genuine knowledge. They aren’t dilletantes, they’re frauds.
Imagine Elizabeth Holmes was never sent to prison. What would she be doing right now? It’s pretty obvious to me that she’d be doing exactly the same thing. Selling grand visions wrapped unwavering confidence because that is the only thing she knows how to do. The film trope of the villain confessing to the hero (and the audience) “I’ve come too far” comes to mind.
I’m not saying being confident is going to turn you into a fraud. What I’m saying is that you need to have the skills, knowledge, and other forms of human capital first so that you can offer those skills to the world confidently. Confidence is a useful social lubricant in a market full of imperfect information. It’s useful as a complement, often necessary but never sufficient. To sell naked, untethered confidence itself is to be a confidence person.
I was told when I was a graduate student that the modal published research paper was cited zero times. They then went on to say that the median was one or two, while the mean depended on how good the journal was. Well, someone went looking in political science:
“The modal number of citations to a political science article is zero is true, but most political science articles have more than zero citations is also true.” — @tompepinsky.com
So, at least in political science the median appears to be 26 and the mean 71. That’s a lot higher than I expected, but then poltiical science articles tend to cite far more papers.
My only real thought is this: what question is each number answering? The mode really just let’s you know that the distribution is truncated at zero. It’s a mechanical artifact more than anything. The median, or any of the quantiles, on the other hand, seems only informative if you have a benchmark in your head of at what point an article seems to to be moving the literature. Is 26 substantial, or is it just inframarginal noise for the 90% of articles that don’t really matter? It seems to me all these numbers only matter as fractions of the maximum. If 8123 is the ceiling, how does that change how we interpret a paper with 7 citations?
Most of this seems more like academic appointment accounting, which is almost by definition uninteresting, but it’s a great reminder that statistics have no meaning absent a context that goes beyond other statistics. That said, with just a minimal context, it’s amazing how far knowing nothing but the mean, median, and mode can take you.
You can read it as the narrative is designed. There are good guys resisting a group of marauding (in some cases literal) children. But I think it implies a vastly more important question that we’ve been learning the answer to every day for the last two years: are there rules? If someone does not care to follow the rules, what are the consequences? If someone is creating bespoke rules on the fly with the implied promise that you are taking on great personal risk if you do not follow their rules, are they actually rules or are they something else?
Funny thing about high level governance, it takes on a far greater “fog of war” than a more optimistic observer of institutions might expect. The shape of the current US federal government and the actions it takes appears to be determined far more by the actions of individuals willing and capable of making decisions in a fog of war. That favors people who know that rules aren’t real if they are unenforced. But it also implies the potential for far greater consequences if an authority ever decides to impose consquences for prior transgressions within the fog of war in the interest of resetting precedent for the future.
This, to me, is the principal lesson of the last 10 years from US governance. The story and data in this paper are based in Singapore, but the lessons are universal. Social norms are critical, but if you grow reliant on them absent any formal enforcement, corruption will eventually take root. There are just too many people who figure out that the concept of “being in trouble” isn’t real. If transgressions do not receive instrumental punishments beyond social shaming and reputational damage, then individuals will simply find institutional and social pockets to inhabit where that shame doesn’t reach and a negative reputation might even be viewed as an asset by other would-be corrupt agents.
Read this paper (I mean, at least skim it) and internalize its core observations. This is important.
Here’s the Polymarket…market… for bets on the outcome of the US midterms in November over the last 7 days. You’ll notice a 7 percentage point (13% relative to a 54pp baseline) increase in the expected probability of the leading outcome. That’s a very big change, most of which happened over 3 days. In a $12.8 million market where the largest shareholder has $73k on the line, that is a very, very big move.
I don’t have any insight into what is happening here, I would simply note that there was to my knowledge no singular news item or event that appears to have sparked this. It’s just a great example of what makes betting markets so interesting. There is some tacit knowledge held by a relatively small number of inviduals, but within and through this market this information has now become at least partially observable in the form of a price. Now, to be extra, super, you just won’t believe how clear about this, I don’t have any insight to lay on top of this, no personal model to share, and I would strongly advise very nearly everyone against trying to participate in this market for personal profit. But, nonetheless, this is useful information. If you made me choose between a world with sports gambling or everything else gambling, I might very well choose “everything else.” At least the world captures some value from the revealed information that is, at least in part, subsidized by enabling a very real and serious addiction. I don’t know that we can say anything similarly positive about sports gambling.
Update: longer timeline as of 10am this morning (7/11 to 9/21)
Inspired by Canada’s attempt to integrate into the European Union, here is a non-exhaustive (and highly correlated) list of advantages that Americans have benefited from, been subsidized by, or captured rents from over the last 100 years:
Immigration
The world’s largest free trade zone (50 states, 3.8 million square miles, 116m-350m people)
Friendly neighbors
Two coasts to transport goods from
Serving as the world’s reserve currency
The lowest perceived probability of defaulting on its national debt
The absence of historic or prospective hyperinflation
Possessing nuclear weapons that can both credibly be used and, more importantly, not used (NB: I’m not excited about this one, but I at least can’t rule out that it has deterred antagonistic military action)
The world’s most prestigious higher education institutions
Relatively high levels of interpersonal trust
I have no doubt that there is many, many more. I just wonder which of these we will still be intact in 10 years. Are they all still intact now?
Addendum: This is not meant to be a list of simply good policy choices. Those are, in actuality, far easier to achieve. Rather, this is a list of physical and societal attributes that either simply have to exist or take a significant amount of coordination across time and, more difficult still, generations, to achieve.
I try to avoid nakedly partisan output, especially in this forum. I give my self a little more leeway on social media, but still try to remain within reason. And over my entire liftetime, the intellectual capability of the president has always been a target for insult (sometimes horrifically so) by the opposition party. But we have arrived, at some point in the preceding months or years depending on your vantage, at the arrival of a sitting President who genuinely appears to be limited in his ability to self-regulate.
A sundowning old man, addicted to AI slop, surrounded by supplicants and apparatchiks, with control over the US nuclear arsenal.
If you let the post play, it shows 7 posts over 10 minutes. The rapidity of posts is, I believe, relevant because it signals that it is unlikely to be the product of a staffer in charge of his social media. PR staff would, regardless of content, prefer that posts each have their own time at the top of the feed to generate attention. By firing off so many posts in rapid sequence, each post is buried by the subsequent. This is not professional PR.
As for the posts, it is precisely the kind of content that would generate a lot of interest from juveniles and those in cognitive decline. It’s the kind of slop, AI or otherwise, that gets a lot of attention from individuals with zero opportunity cost of time on Facebook. Read in conjunction with the all-caps posts of threats to ban trade with whole swaths of the globe and other similarly illegal executive actions, it’s just part of the evermounting evidence that the President is not in full control of his impulses. Putting aside what Thomas Schelling would say about fragility and faillings of madman theories of conflict, it’s brings up one of the most important reforms we can and should be pushing for in the next administration.
Age limits on representation.
Reliable, unbiased cognitive tests would be first-best, of course, but the problem there is a) failure is endogenous to the control of the branch administrering them, and b) extrication in the wake of a failure is the kind of stress test that our system does not seem up to at the moment. For all of it’s coarseness, counting days since the record of your birth is straightforward and relatively easy to enforce before being allowed on the ballot in the first place.
One of the realities of modern medicine is that our ability to preserve the body is likely to outpace our ability to preserve the mind at a level adequeate to the task of high representative office. At the same time, our abilitty to extricate and replace someone unfit to their duties in the public sector will likely always be limited relative to similar processes in the private market, where executive services will always be at will in any context where the executive is not themselves the owner.
From a public choice/political economy perspective, age limits will not be an easy legislative or constitutional feat. Representatives will always favor the status quo rules that helped them get their job and subsequently keep it. This is only made more complicated by the basic human optimism regarding our own current and future capacities. And perhaps the simplest obstruction is also the biggest: the reward for winning an election is the expected incumbent advantage in the next election. To limit the prospective future terms for any sitting official, in their current position or the one they aspire to, is to reduce the value of what they already earned, and nobody likes that.
In the face of all that, however, remains a simple truth. The citizens of the United States and the broader world are all made less safe by a sitting President whose self-control and capacity to understand complex contexts has waned well-beyond the threshold of reliable decision-making. That’s a cost big enough and bad enough, wrapped in all the risk aversion and loss-aversion who can possibly imagine, that it should at least have a chance to motivate real legislative action. At least a chance.