Recorded on May 5, 2026, this video features a talk by Benjamin Recht, Professor of Electrical Engineering and Computer Sciences at UC Berkeley, focused on his book, The Irrational Decision: How We Gave Computers the Power to Choose for Us.
Professor Recht was joined in conversation by Marion Fourcade, Professor of Sociology and Director of Social Science Matrix. This event was co-sponsored by the D-Lab; the Algorithmic Fairness and Opacity Group (AFOG); the Center for Science, Technology, Medicine & Society (CSTMS); and the UC Berkeley Departments of Sociology and Electrical Engineering and Computer Sciences.
About the Book
Mathematicians and engineers of the 1940s set out to design machines that could act as ideal rational agents in the face of uncertainty. In this pursuit, a cluster of foundational mathematical technologies — including information theory, linear programming, game theory, and neural networks — emerged as a foundation for a mathematical formalization of rationality, reshaping how we think about human decision-making itself.
The Irrational Decision traces how a narrow mathematical framework for computing came to define rationality in economics, public policy, and popular culture. Recht’s talk will discuss how these seminal computational methods have evolved into a robust discipline and industry, with success stories in accelerating computers, regulating pharmaceuticals, and deploying electronic commerce. These examples will highlight how automated decision systems excel in specific sweet spots with clear rules, well-defined goals, and well-constrained contexts. They will also show how, outside these narrow contexts, the rational program tends to absurdity. Given these strengths and limitations, the discussion will explore how to best harness 80 years of unfathomable computational progress while preserving human agency and judgment.
About Benjamin Recht
Benjamin Recht is a Professor in the Department of Electrical Engineering and Computer Sciences at UC Berkeley. He is the recipient of a Presidential Early Career Award for Scientists and Engineers, an Alfred P. Sloan Research Fellowship, the 2012 SIAM/MOS Lagrange Prize in Continuous Optimization, the 2014 Jamon Prize, the 2015 William O. Baker Award for Initiatives in Research, and the 2017 and 2020 NeurIPS Test of Time Awards. He is the author of three books: Optimization for Modern Data Analysis (2022), with Stephen J. Wright; Patterns, Predictions, and Actions: Foundations of Machine Learning (2022), with Moritz Hardt; and The Irrational Decision: How We Gave Computers the Power to Choose for Us (2026).
Podcast and Transcript
[MARION FOURCADE]
Hello, everyone. Thank you for being here. It’s great to see a packed room.
My name is Marion Fourcade. I’m the director of Social Science Matrix. Today is actually a bittersweet moment for me.
It is our last talk of the academic year. But it is also the last talk I will host as director of Social Science Matrix. And specifically, I just want to begin by the fact that it’s my last chance to publicly thank our intrepid staff, Sarah Harrington, Chuck Kapelke, who’s our communications director, and Eva Seto, for the incredible work that they do to support our activities and our community.
It makes me especially happy that our last event this year perfectly encapsulates what Social Science Matrix aims to be, an interdisciplinary community of the best kind. And it also encapsulates my own view of interdisciplinarity, which reaches beyond the social sciences and seeks to engage more directly the harder sciences that are transforming the world today. And so our guest today Perfectly represents these aspirations.
Benjamin Recht is a professor in the Department of Electrical Engineering and Computer Sciences here at Berkeley, and probably the first from that side of campus that we have invited to give a lecture. But there’s a reason for this. He has a unique ability to communicate about difficult topics with broad audiences.
He has one of the best and most entertaining blogs, and I guess we should now say Substack. He’s an anti-rationalist rationalist of the best kind. And at a moment when machines seem poised to take over our lives, Ben turns our attention to what cannot be optimized away, what we humans are uniquely good at, and why we should hold the line as hard as we can against the computerization of every decision we must make.
Professor Recht is the recipient of a Presidential Early Career Award for Scientists and Engineers, an Alfred Sloan Research Fellowship, the 2012 SIAM/MOS Lagrange Prize in continuous optimization. So the credentials are very much there to talk about this. The 2014, uh, Hamman Prize, Jamieson Prize, uh, the 2015 William O. Baker Award for initiatives in research, and the 2017 and 2020 NeurIPS Test of Time Awards.
He’s the author of three books Optimization for Modern Data Analysis in 2022 with Stephen Wright. Patterns, Predictions, and Actions: Foundations of Machine Learning with Moritz Hardt. And the subject of today’s talk: The Irrational Decision, How We Gave Computers the Power to Choose for Us.
Note that today’s event is co-sponsored by the D-Lab, the Center for Science, Technology, Medicine, and Society, the Algorithmic Fairness and Opacity Group, and the UC Berkeley Departments of Sociology and Electrical Engineering and Computer Sciences. After Ben’s talk, I will join him in a short conversation, and then we will open it up to the audience for questions. So without further ado, let me turn it over to Ben.
Thank you.
[BENJAMIN RECHT]
Thanks. Thank you so much for the wonderful introduction. Yeah, I’m excited to be here talking about this book.
A little bit intimidated to be over in social sciences talking about a book that is ostensibly a social science book. It’s terrifying sometimes to try to approach other disciplines and be humble in the face of that. I’m obviously talking as a participant observer in this case about a history of computer science.
And it’s a very particular history of computer science that was motivated by some very human questions. Notably, it’s a simple question. What is rationality?
This is a very Berkeley question. There are a lot of people who tell me that they’re very rational here. There’s a church of rationality about two blocks from my house.
But what is rationality? Rationality has come to mean something very specific in our modern context, especially in the American context, especially in the San Francisco context. And usually comes down to cost-benefit analyses.
If you’re going to have some experimental surgery, you would like to know what are the odds this is going to work, and what are the benefits that’s going to happen, and what are the potential downsides? And I will weigh those carefully in my Excel spreadsheets and come up with whether or not I should go ahead and actually have this experimental procedure, based on whether or not the cost-benefit analysis works out. And you might imagine we could apply that to other things.
You can apply that to what you’re going to eat and decide whether or not this muffin that looks very delicious in front of me is if the benefits of me eating it outweigh the costs on my calorie count for the day. And then maybe, you know, we could extend this to our extended family, and we start to think about whether or not it’s worth dealing with the protestations of our children to go to violin lessons because we know it will be good for them to get into college. And you know, even these same kind of analyses come into sports, where they, like, put up all sorts of propaganda on the screen about the probability that doing something right now is going to win the game or the Super Bowl or what have you.
Everything. In these cost-benefit analyses, somehow comes down to, like, calculating the price of insurance. Somehow, like every decision in life, the rational actor is the insurance agent.
Or even worse, the rational actor maybe is supposed to be portrayed as a gambler, right? And this kind of mindset didn’t really jive with me. It’s certainly the case that, like, if you just go and read a little bit, computer scientists don’t read, but if you were to go to read a little bit, you might find that actually rationality has lots of different definitions, right?
It traces its way all the way back to kind of the religious beliefs of the followers of Pythagoras, that you could somehow define the entire universe in terms of numbers. But I think in its modern conceptions, if you go ask a philosopher, they’ll have… You go to the Stanford Encyclopedia of Philosophy perhaps, and you’ll go find 20 different definitions of rationality, all with fancy names.
You could have the behavior within reason might be considered rational, or just using a consistent belief system could be rational, or behaving inside of the reasonable social norms could be rational. Using facts appropriately to attain goals and make the correct decisions could also be rational. And the only thing that I’ve been able to say that’s consistent is that we all believe that rationality is good and irrationality is bad.
Or at least some of us do. And so it’s this last one that was the one I was talking about at the beginning, which I think the philosophers might call instrumental rationality. But it’s more than instrumental rationality, actually, because there’s the belief that you can assign numbers to it.
It’s really a mathematical rationality that everything can be precisely given the appropriate probabilities and dollar amounts. And from those precise probabilities and dollar amounts, we can come up with a correct answer. And that was what I was hoping to kind of get at in writing this book.
It’s like, where did this idea of this very mathematical instrumental rationality come from? And all of the paths lead back to the computer. Everything comes back to when people were trying to design computers in the end of, between the rough end of World War II and the first computer, which arguably is 1950.
It all happened very quickly that we decided that we could just write everything down as maximizing expected utility and give it to machines to actually compute these numbers. It wasn’t a matter of just saying that we were going to maximize utility, it was that we could compute what the policy to achieve that utility was. And this was a rationality of computation.
And so the book kind of starts at this beginning and looks at how all of the things. Now, they’re computer scientists in here, so they might recognize some of these things. That how many of the ideas that are still prevalent today, that still kind of underlie all of our computing technology algorithmically, started in between 1945 and 1950.
For example, the neural network was invented in 1943 by McCulloch and Pitts. Information theory was invented in 1945 and published in 1948. The whole idea of Estimation and prediction of time series was kind of formalized in the late 1940s.
The first use of statistics to understand medical treatments, especially in the context of clinical trials, happened in 1946. Game theory was invented in 1944. Optimization as we now know it, even though it does seem like something that we associate maybe back with Cauchy or Newton, optimization really wasn’t formalized as a modern discipline until 1947.
And then for the people who know, the stochastic approximation method, which we now call stochastic gradient descent, was invented in 1951 and is kind of the backbone of everything that we do in AI. So all of these things happened so quickly, all between in this period of maybe eight years. And so the book just kind of goes and tries to figure out how you get these very small cast of characters, all kind of somewhere in the orbits of Bell Labs and the Institute for Advanced Studies, how they’re thinking about what makes for personal, perfect rational decisions and envisioning what such a perfect rational agent might do.
Led to developing a computer itself. The computer was kind of designed in the image of the perfect decider. And then we fast-forward 75 years, those computers get so fast, much faster than we got smarter, and we’re left with a bunch of people telling us that we should be more like computers.
And if you know who’s on the slide, you spend too much time on the internet. So the way I structure the book is I kind of split it into four different pillars, which, you know, again, is a personal story. So these are four things that kind of connect very closely to my own research and are just kind of pervasive, the kind of unavoidable, whether you be in AI or optimization research or even in statistics.
So this rationality computers, like, I would say one of the main chapters is about optimization. One is about game theory. One is about the idea of the randomized experiment, how you use that for policy, and one is about statistical prediction, which we also call machine learning, which maybe now we call AI, whatever.
It’s that last weird pillar. And it was kind of impossible for me to leave that chapter out. Each of these is unfathomably powerful when applied in the place, in particular sweet spots.
And I, in the book, talk about the specific places where it kind of unlocks all sorts of amazing capabilities. But then if you try to, like, use your hammer everywhere, seeing nails every which way you go, it doesn’t always lead to the ends that you would like. And so, like, understanding that this kind of particular rationality has a place.
Just not everywhere. So to give you an example, we’ll just go through a little tour of the optimization chapter, mostly because this one has the most fun story, but we’ll get to that in a second. So mathematical optimization, again, like I said, lots of the idea of building, writing down cost functions and maximizing them is not new to the post-war period, but what was new to the post-war period was being able to compute them, because you’ll see very quickly that if, as soon as you get more than two variables, these problems are really hard.
I think everybody has had to deal with some high school, at least high school matrices where you had to, like, solve systems of equations. It’s… Yeah, we’re all tortured.
I know we’re bringing back horrible memories. Doing that by hand, I have tortured my own students with doing that. No one ever wants to solve a matrix with four equations.
It’s just not for anybody. But the computers are really good at them, and computers, it turns out, are amazing at optimizing, and that was kind of the main idea was that we could build machines that would do these calculations for us and be able to tell us what would be optimal in a variety of different settings. And there was an interesting turn there, is that you could actually turn the optimization back on itself and use the optimizers to build computers.
And that feedback loop created the situation we’re in today, where we have computers everywhere faster than we could have possibly imagined. So that story starts with this fellow George Dantzig, who’s fascinating. And reading his own writing, both at the time in the forties and his later recollections, is quite interesting.
He was a Berkeley PhD student. He went to… In World War II, he went to the Air Force and kind of ran what we would call today their data science division, where they calculate—he made all these big statistical tables and tried to come up with plans with generals, and he found himself constantly frustrated with once they would tabulate all these statistics, what the commanding officers would do with them.
As he would reflect later, you don’t have to read all of this this quote that I just put up here. It was just hard to figure out what to put in context.
But that he was constantly frustrated with—you would go to somebody, and you would say, “What’s— what are we trying to do?” And the answer was, “Win the war.” And then depending on who you asked, they would say, “How do you do that?”
Well, the, you know, the guys in the Navy would say, “We build more boats.” The people at the Air Force would say, “We’d build more airplanes.” And so it was kind of obvious that, but Dantzig didn’t like that you would constantly have to defer to expertise or actually even defer to authority to make decisions.
What he would prefer was that you would be able to just write down some general objectives and find optimal policy solutions. And this turned out to be a quite radical development that you could come up with what was actually optimal if you could write down the rules. He presented this work for the first time in 1948 at a conference in the University of Wisconsin-Madison.
It was a workshop on econometrics. All of these big shots were there, names you would recognize, notably John von Neumann. And this was the first time where he also made the claim that not only can we do this, can we build these things and make these decisions, but we could do it if we could have machines.
We need machines that will compute these programming problems. It’s also really important to– It’s very funny that programming, this is the probably the first use of the word programming, did not mean what we mean today.
Although I don’t even know what we mean today in the era of coding agents. But neither here nor there, when we say computer programming and computer science now, we don’t mean what Dantzig meant. Dantzig meant programming in the context of military planning, coming up with a program for building stuff and shipping it places and maybe blowing things up.
That was what he was into. And also I should have noted, he was a– He remained in the Air Force for a decade before moving to the RAND Corporation and then finally ending up at Stanford.
But at the Air Force, he was what we’d consider today a program officer, gave a lot of money to a lot of different kinds of research, notably for early investment in computers. And so at this meeting, this is where we kind of get into the standard way that these conferences tend to go. Harold Hotelling, was another famous economist and statistician, told him like, “This will never work.
The world is nonlinear.” And John von Neumann said to him, “Look, the speaker said that his talk is called linear programming, and if you could write down something with linear constraints and linear costs, then you should use it. And if you don’t, don’t use it.”
Fine. I mean, Hotelling was, I think, “Why should we spend all this money on it?” It’s a good question.
But this also was kind of the start, and one example of where we start to find that you have to kind of engineer systems to make sure the axioms are true. You have to put yourself into this weird box where the axioms are true. And honestly, to Hotelling’s credit, Dantzig didn’t have a ton of examples.
He had one example of like assigning people to jobs, but he didn’t have a ton of examples at the beginning of what to do with linear programming. But there was a fun example that would arise very quickly after this meeting. Uh, he went back to Washington, DC, and people started to Ask for his help to fight with an economist named George Stigler.
Let’s get to why. So here’s an example where Dantzig found a really neat application of linear programming. I have to go back 50 years before we get to it.
So in the 1800s, all sorts of different parts of the world were afflicted with what we now know as deficiency diseases. These are kind of horrible diseases. Nobody knew how to cure them.
And the major revolution was realizing that the diseases were not coming because of a germ, but they were coming because of a lack of certain nutrients. And you could just cure them by making sure people didn’t eat monotonous diets. So by having the accessibility of just somewhat simple foods, but a diverse set of foods, you could completely eliminate all of these conditions.
And it’s kind of amazing that you have this kind of list of these different horrible afflictions, and the cure is just eat something. Just a little bit. It’s weird that we’ve kind of lost track of that, because now I think we associate vitamins with really annoying influencers on Instagram.
But they’re actually quite important, but it wasn’t at the time obvious that you should not only eat white rice all the time or not only eat hot dogs all the time. You kind of have to have a little bit of diversification in what you eat. So these were also surprisingly only discovered in the 1910s, discoveries continuing into the 1930s.
So with this knowledge in mind, you have to educate people about what to do, right? You have this new– the scientists figured this out, but now you have to tell everybody else. And so Hazel Stiebeling was at the USDA, tried to come up with plans, like communications from the USDA to educate people on how to have a balanced diet.
Stiebeling got her PhD in chemistry because, you know, vitamin theory is chemistry. And so that wasn’t that unusual at the time. She ended up at the USDA, was in charge of food economics, and she published some of their first dietary plans that included references to eating vitamins.
And she actually notably invented the term dietary allowances, which we use all the time today. So her \”Diets to Fit the Family Income\” had this nice little introduction. Again, you don’t have to read the whole thing, but had this– the important thing here is that she said that cost alone is not a measure of the desirability of a diet.
You have to hit your nutritive value, and you have to have cost, and Yeah, families with a lot of money to spend will give you a liberal diet. And then people who are cost-conscious will give you a minimum cost diet that you can go after. Now, again, this was published in 1936, in the middle of the Great Depression, so obviously costs were a high concern.
So here’s the two diets: the liberal diet and the minimum cost adequate diet. She uses this term that it’s minimum cost. And it was about $100 a year in 1939 dollars for a family of four.
Oh, sorry, that $100 a year is for one person, not the whole family of four. That’s important. It will come up in a second.
But you can see it has a bunch of stuff, milk, potatoes, whatever. George Stigler, who was an economist who had been working at Columbia at the time, he— He saw this. Now, Stigler wasn’t a— I’ll tell you a little bit more about him in a bit, but certainly not of— certainly more on the right wing of the spectrum, let’s put it this way.
He was definitely a very conservative libertarian economist, and he was upset by the paternalism of this diet. He said, “This is not minimum cost.” You shouldn’t go around telling people it’s minimum cost, and you shouldn’t go around telling people that this actually has anything to do with the appropriateness of diet.
And the way he proved that was he wrote this paper, which is from— See, Jeff would enjoy this paper because from sentence one to the very final sentence, everything is snarky and mean. Every single line of the paper, which kind of makes it clear that we’re not allowed to do that anymore. But yes, every single line of this paper is clearly sarcastic and mean.
But what he did was he tabulated what he got from the USDA, what the recommended dietary allowances were. I don’t know where they got that 3,000 calorie number from. That seems high to me for someone who weighs 154 pounds, but you know.
Obesity wouldn’t come on the radar of the USDA for a few years. But you have all the other things that you’re supposed to hit, and now you have this table. He collected this table from the Department of Agriculture, which has the dollar amounts of all of the foods and all the vitamins involved.
So you can zoom in here, and you can see, like, all the flours and the different kinds of grains, their associated price, the amount of calories per serving, the protein, calcium, and so on. So I have this giant table of costs and nutrients, and I can kind of put that together and get myself a linear program. Amazingly, solving this problem is a linear program because the cost changes linearly with the amount of food.
If I increase food by a factor of two, the cost I pay is twice. And similarly, the amount of nutrients I get also is increased by a factor of two when I double the amount of food. So it was a perfect thing for this.
This is exactly what Dantzig wanted for his model. This is the first linear program that was recorded. Stigler wrote this paper in 1945, and he solved it by hand.
You didn’t need those fancy… I know exactly what Jeff’s thinking about that. I guess he had some spare time.
And he came up with a minimum-cost diet, or at least one diet that he thought was cheap. He wasn’t sure if this was minimum cost. He went through lots of different trials and errors and added some columns together and put things in and took things out.
But this seemed reasonable to him, and it was 2.5 times cheaper than the diet proposed by Carpenter and Stiebeling. So, not bad. He was much less…
Obviously, whatever they were putting forward as minimum cost was not minimum cost at all. Now, what does this amount to? It’s actually…
this is worth knowing. So every day, you have to eat a pound of flour, two ounces of evaporated milk, five ounces of cabbage, one ounce of spinach, and five cans of navy beans. Which, yeah…
you laugh. You could feed a person for an entire year for only $90. No, this is less.
This is for $40. $40. You could feed a person for a year for $40.
How much would that cost today? If we went to Costco, I think it would be about the same. So I feel like for similar, similar prices.
Now, you might ask yourself, what would I eat? That seems really bad. Now, this is where it’s useful to have a baker in the house.
My wife is not only a world-famous professor of art history, but also an amateur baker, a really excellent baker. And she has found a recipe that actually would hit all of the requirements. So here is the bean pie.
That’s in this nice book. And I think we basically get everything on there, although I think you have to replace the bacon drippings with milk. That would be one thing you’d have to do because you’re allowed some milk.
Oh, sorry, I had to take off the butter. And maybe I think we could have kept in the baking powder, but technically speaking, that’s not allowed. And get rid of the rest of the stuff.
I think we’re still pretty close. I think that’s close enough that that modification would work, and you could eat a Georgian bean pie, many Georgian bean pies every day, one for every meal. What’s fascinating about this is that even before we started optimizing, it was pretty clear that an optimal diet was grotesque.
Dantzig would write about this again in 1990. He apparently tried to put himself on an optimal diet, for weeks, and he kept changing the model because he kept finding that whatever would spit out was too weird. And this is exactly what Stigler found too, and he would change the model and add constraints, and eventually his wife said to him, “Why don’t I just put you on my diet?”
And we stopped worrying about this, after which he lost 20 pounds. So, you know, I feel like we can each go for what we’re going for. But this is a little bit of a microcosm of optimization.
It’s now kind of a fun story. The diet problem is the first linear program we teach to our undergraduates when we teach linear programming. It’s easy to understand, but you can just start to realize just how weird it is.
Like, first of all, that. Dietary allowances are not exact sciences. The amount of nutrients in food is also not exact.
A Granny Smith apple and a Honeycrisp apple are gonna have different amounts of these vitamins. And there’s always… And also, conceptualizing everything as just minimum cost is very weird.
It’s not—you can’t really put together. All of everything into one cost function. And do you know who agreed with that?
It was George Stigler. George Stigler, of all people, was saying you cannot put all of these things into one cost function. There are fundamental objections to merging the physiological and cultural components of diet.
The first is that the particular judgments of the dietician as to minimum palatability, variety, and prestige are at present highly personal and non-scientific and should not be presented in the guise of being parts of a scientifically determined budget. That’s fair. The second reason is that these cultural judgments, while they appear modest enough to government employees and even to college professors, can never be valid in such general form.
Now, no one can say with any certainty what the cultural requirements of a particular person may be. And on its face, it will always be impossible to determine a unique cultural minimum diet for 140 million Americans of transcendental variety of background, social position, and cultural values. You know, it’s…
He has a point. I think that’s fair. What’s fascinating about this story, too, is this whole view of optimization.
Like, there, you take on two different kinds of things. Like, you are trying to please two different constituencies, and I think this would be– it’s both the right and the left in this country use this language of optimization regardless. So, it’s fascinating.
Stiebeling would go on to become the director of the Institute of Home Economics. She was responsible for some of the first school lunch programs. She was a public servant for her entire career, clearly dedicated to making people’s lives better, but also with a very technocratic view of how these things should work, of how government should be involved.
Stigler, on the other hand, is- Best known for the concept of regulatory capture. He won a Nobel Prize in economics.
He unsurprisingly moved to the University of Chicago, where he fit right in. And I think was really kind of a key thinker on the libertarian right. So you really have this tension already present and playing itself out in this really kind of amazing way with the simplest little optimization problem.
Fun fact, though. So… Dantzig was intrigued and he had a bunch of friends in the government, being a government man himself.
And he tried to actually compute this optimal problem by hand. So what he did was he went to the Mathematical Tables Project, which was still running. This was a WPA project where people were computers and would use hand calculators to compute all sorts of weird tables of numbers.
And together, in 120 person-hours, they were able to find the optimal solution, and the actual minimum cost was $39.69. Which beat out Stigler by 24 cents. So now there’s two things to take away from that.
First of all, experts can optimize, and I actually do have good intuition for how to do these optimizations, but they can be automated away. So we don’t need the experts necessarily here. We can automate them away.
But also we have the problem from the very beginning that the computer that we have in our hand is too slow and we need a better one, and I need to write a grant for it. Some people know that feeling. Today, it was true from the get-go.
And so Dantzig, with this kind of inspired result, was very inspired, made the case at the Air Force to start funding these computing projects. And he ended up being one of the main funders of a project at the National Institute of Standards and Technology called the SEAC, which is probably arguably the first von Neumann architecture machine that was up and running and actually functional. This is in 1950.
And on that one, they could compute much bigger linear programs, and this kind of kicked everything off into our modern age, and kind of we know what happened next. Computers got faster, faster than we could get smarter. And we kind of ended up…
Let me just skip over this. And we ended up in this weird situation where we went from a computer in 1954 that would do 12,000 operations per second to a computer. I lost mine.
That we lose, you know, in the podium that has two teraflops on the GPU alone. It’s quite astounding. So I think a lot of what’s happened was, and where the book takes you, is you start with this one simple idea and everybody knowing things are kind of weird and wrong, and then you just run with it, and you let the computers optimize themselves up to the front.
And at that point, we kind of forget why they’re weird and wrong, and we’re left here. And part of that, and kind of where the book closes… I’m just going to skip ahead a bit.
We have too many slides. Where the book closes, and I’ll tell you about that later, is asking a little bit about people. So, okay.
The other chapters in the book talk about all these different aspects, kind of put everything together into this idea of constantly maximizing utility up to policy. And it kind of led us to ask, “Well, if computers are good at this thing, what are people good at?” What are people optimizing?
And subject to what? And really, this was subject to a lot of heated debate, lots of heated fights, lots of back and forth. And I really think the only thing that people can, I think the only sensible takeaway from 75 years of arguing about it is that people don’t make mathematically rational decisions.
That’s the only thing. We don’t make decisions the way we design computers to. We had this idea of what good decision-making was, and then you actually ask people to do it, and we don’t do it.
We don’t do it. Is that good or bad? That’s the question.
It really just depends on this sense of depending on the prior biases of the experimenter asking the question. The question is: Is it bad for people to be deviating from rationality? Or is it good?
And in fact, that makes us different and better. I think the book tries to kind of engage with both. I think the answer is not that simple.
I think there are certainly cases, and I kind of go through them in the book, where you have very repeatable, measurable, not objective, but you have very repeatable, measurable statistical things That you would like to actually optimize. And in that case, the computers are good. And sometimes we could come to some kind of agreement together about those.
But then there are these other things which are singular and unverifiable and humanistic. And I think that it’s. A bit ridiculous to try to shoehorn our, even despite the one, the machines that give us language back now.
I think it’s a bit ridiculous to kind of shoehorn any kind of computational language on top of those. I think that we, the way that the kind of book closes is kind of engaging with the ways that people talk about what is it actually that distinguishes us from the machine and kind of what distinguishes institutions from the people that are part of those institutions. So I have a bit of a discussion on kind of the difference between rules and playing games, rules for games and playing games.
And I also touched a little bit at the end about a distinction made by Joseph Weizenbaum about what’s the difference between a decision and a choice. In Weizenbaum’s eyes, the choice is something that’s value-laden. The decision is something that can be optimized.
And I think maybe the kind of like the close, and maybe the thing that we look forward to in the future is that, you know, I think the key here is that we can choose to change the rules of all sorts of systems that kind of govern us. And that flexibility and that ambiguity is the human part that I don’t think will ever go away. All right.
With that, I’m going to turn the floor back to Marion, and we’ll have a chat.
[MARION FOURCADE]
Thank you so much. Ben, that was terrific.
[BENJAMIN RECHT]
Thank you.
[MARION FOURCADE]
As expected.
[BENJAMIN RECHT]
Oh.
[MARION FOURCADE]
So, um, reading the book, I, you know, I was struck- There’s a lot of economics in it, obviously.
[BENJAMIN RECHT]
There is.
[MARION FOURCADE]
Because computer science is largely based on cost-benefit analysis and game theory. And you have… So essentially in the book, you have what I see as sort of two main critiques of mathematical rationality.
One is simply, as you said just now, humans don’t behave rationally, right? That’s the Kahneman and Tversky, and there’s a whole literature about that. Another one, another critique which you also mentioned is that there are things that are just not measurable, verifiable, and so on.
You know, you mentioned, you said the pleasure of grandma’s chicken soup. You know, like there’s no way you can actually optimize that. But I think there’s a third critique.
That one perhaps that a sociologist would make to an economist and by extension to a computer scientist, which is that mathematical rationality tends to reduce the world to individual preferences. And so it is very hard to make sense of genuinely collective phenomena. Like solidarity, legitimacy, public goods, shared, although there’s a whole literature on public goods, but shared meanings.
And so I was wondering, you know, how your account actually speaks to sort of that critique of, you know, how do we think about sort of the social, you’re at Social Science Matrix, so I thought it was a fair question there.
[BENJAMIN RECHT]
Absolutely. Well, it’s interesting that, like, they codify so much, like, so much of what they do is codifying rules of social engagement, right? I mean, so I mean, do we treat…
I did not. This is a recent fascination, but like, if you look at org charts and you look at computer architecture charts, they don’t look that dissimilar from each other about how we even think about them. So I think that a lot of the computing mindset is trying to just encode a lot of that.
It can encode solidarity necessarily, but is encoding a certain social structure and hierarchy. And I think that that, like… One thing I didn’t talk about today, but we are encoding a lot of times in these optimization problems a benefit of an average, which is somehow a commitment to the idea that benefiting a population is benefiting the individuals.
Whether or not people say that explicitly or not, I think that’s definitely happening, at least implicitly.
[MARION FOURCADE]
So another question I thought that, you know, in the book, you also say that. You know, there’s decades of evidence that lots of decision made under mathematical rationalization don’t work, right? There’s a lot of evidence that some of them work extremely well.
But in plenty of domains, it doesn’t. And so, in those domains where it doesn’t, what do you think is it that keeps that alive? And again, here I’m trying to broaden the conversation a little bit to think perhaps about sort of the-
The structure of, uh, you know, economic structure, political expediency, I don’t know. You know, what is it? Uh, is it the prestige of math?
Is it, you know?
[BENJAMIN RECHT]
Sometimes. No. No, I think the prestige of math is part of it, but that’s not how I would answer it. Hmm. I think, again, I think it does come down to this, like, there, there’s this, like, uh, there’s this quantification trap which we always fall into.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
And, like, the computers are built on them. The, just the idea that, that in, in order… The easiest way to make something legible to a larger community is you, you try to, you try to make it mo-
You make it numerical. Mm-hmm. This just seems like a thing that we lean into.
If we want to make something that’s easy to communicate to people who aren’t experiencing our immediate experience, and we want to maybe translate that up the chain, or express what’s happening on the ground to the experts who are sitting up in the high parts of the office, or the less experts, the managers up in the offices. From the experts, we quantify. Our measurements, we write down as numbers.
And then once you… then you do have this trap, right? So once it’s quantified, then maybe it’s objective. Now, that obviously I’m stealing from social science.
So I think that’s a trope that runs through a lot of science studies, is that there is this- Natural, you just, you know, we made it quantitative because that was going to be easy, and then all of a sudden we forgot— Oh wait, no, that number is now objective because we took out a lot of discretion in the process of kind of writing down that legible thing. And now once something’s objective, that means it has authority. And then once something has authority, then it has this power behind it.
And I think even worse is that because we are always looking for those simple metrics and measurements and in these very, like– And now we have these devices that make it really trivial to do it, we just forget that we’re even doing it anymore.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
And I think that’s why it’s so everywhere. It’s like every single thing, every piece of information that our phones know about us is quantified through that weird lens. We have no idea, like, when we engage with whatever.
We engage with anything, what do you get to do? You hit like. Maybe you get more emojis these days, but mostly it’s just like or retweet or whatever.
These are the few little interactions, and that kind of goes. And then… We sit on the other side of clicks and that kind of scattered information and try to make sense of it.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
But we’re looking at it through the wrong lens. And this is not just a phenomenon of the smartphone. I mean, it happens through the electronic health record.
It happens through the kind of narrow view of testing with our kids in schools. And everything kind of gets funneled through that quantified and easily tabulated, put into a spreadsheet, computed, and then we can see if we’re hitting the cost-benefit analysis we promised.
[MARION FOURCADE]
Mm-hmm. So is there, are there any sort of, I don’t know, political interest, economic interest behind this process?
[BENJAMIN RECHT]
So that’s right. This is a great question. This is a great question.
Like, I don’t know, how much of it is deliberate? I could, maybe I could ask you, how much do you feel is deliberate? Because I, you know, if people don’t know.
Marion may have also written a book about that, about the political interest behind it. Because it’s sometimes hard to tell. I feel like sometimes it’s these cultural things, these cultural forces.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
Like people, the people who are doing them have agency.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
It’s just that collectively we kind of put ourselves into this thing where we become cogs in a machine, even though I think this is the same kind of thing. The manifestation at the whole looks a lot different than what’s happening on the ground. So it could be very well that I actually do, I am gaining a lot of pleasure from clicking on these things on the phone, and this is good, and it’s not addicting.
And then at the grand scale, we have all these other problems.
[MARION FOURCADE]
Mm-hmm. Mm-hmm.
[BENJAMIN RECHT]
What do you think?
[MARION FOURCADE]
Well, I mean, there’s a whole history of quantification that sort of shows that, you know, in part, quantification, Ted Porter and so on, that sort of shows that quantification in part arose because it was politically extremely useful to get through all kinds of political conflicts, right? You have numbers, and so, you know, and they become the trusted source that sort of settles all kinds of political disputes. So that’s one way to think about this.
But I agree that there’s also, you know, there are also ways in which some of these technologies that we see today, particularly around the tech industry, develop for reasons of economic profit.
[BENJAMIN RECHT]
Yeah.
[MARION FOURCADE]
So we optimize on essentially getting people to click more and so on and so forth. So we didn’t discuss the question of what is the goal of optimization? So you talked a lot about the process, right? But we haven’t… you haven’t really questioned what are we optimizing on?
[BENJAMIN RECHT]
So, yeah, minimizing cost. Okay. Maximizing profit.
[MARION FOURCADE]
Yeah.
[BENJAMIN RECHT]
That’s right. You could do other things.
[MARION FOURCADE]
Okay.
[BENJAMIN RECHT]
You could do other things. I think really one of the interesting chapters is what I felt was… I’ve received a few people saying, “This book is mostly behavioral economics.
There’s no computers here.” I feel like I missed those people. I’m sorry.
But there is one chapter where the computer doesn’t seem to play a huge role, although it was really responsible for so much of the data collection, which is the chapter on clinical trials.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
And clinical trials is this very interesting subcase where the argument was that you were maximizing public good.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
You were maximizing a benefit to a population directly. So should we give every child in this country a polio vaccine? We have to demonstrate that these are safe to give to children and that it actually reduces polio significantly.
And so they did this massive trial with hundreds of thousands of children to really show those benefits. Or should people, you know, there are simple questions. Should we go get screenings for every cancer?
Does that benefit everybody? And so the idea from the National Cancer Institute and others associated with the NIH was that doing these kinds of trials would take out some of the arbitrariness. And in fact, a lot of the randomized trials started explicitly because people were putting out drugs that were killing people, and the FDA was given the mandate, “You have to figure out whether or not these things are poison.”
Um, and so there was this explicit, maybe not, maybe not optimizing, but at least just guaranteeing that this is not. Yeah, you know, or hopefully not going to actually cause harm.
[MARION FOURCADE]
Mm-hmm. So that’s actually the answer to my first question about sort of collective goods. Yeah. Yeah. So I have a question, and maybe it’s a bit of a curveball, um, here, but I… I have a question about LLMs.
[BENJAMIN RECHT]
Okay, let’s do it.
[MARION FOURCADE]
Which is sort of— you know, an extension. Well, first of all, do you consider them an extension of that logic?
[BENJAMIN RECHT]
Yeah. I mean, to some extent. And then what comes out is weird.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
I’ll definitely say that because I talk about them. One of the first modern machine learning projects was the LLM. Claude Shannon did it in 1945. It’s like that’s there.
[MARION FOURCADE]
Mm-hmm.
[BENJAMIN RECHT]
And it’s like he’s doing… It’s not that different from what we’re doing today. It’s just today we do it on much bigger computers with much bigger corpora.
He did it by taking a book and going through and, actually counting how frequently this letter followed that letter in a biography of Thomas Jefferson. Now we do much bigger corpora, much bigger computers. So in some sense, the logic is the same.
Yeah. And yet what comes out is very odd, yeah?
[MARION FOURCADE]
Yeah, and so I want to see. So what is the role of LLM in decision? Yeah, because we don’t use LLMs to make decisions, but we use them increasingly as aids to decision, to sort of help us reason, help us make judgments, to write that literature review in a paper, to write a legal opinion, to summarize a case or whatever, right?
And so increasingly, they frame the terms through which we sort of- Think and actually argue with it, which sort of you say is the uniquely human aspect that we should actually keep and hold on to. So how does the framework of the book change if moral thought itself becomes shaped by the very processes that you-
[BENJAMIN RECHT]
Emphasize? Yeah, I mean, well, you could say similar things about information retrieval, right? Mm-hmm.
Which is another thing. To the points that you were making. I think there is something a little bit odd that they will also talk to you.
Yeah. Exactly. I get it.
That’s been something that they like. That’s something I’m still struggling with, trying to think, well, what’s the right way to think about those technologies? With regards to decision-making, it is a little bit complicated, because I think a lot of times, because the LLMs are the There’s so much hype behind them, and everyone’s so excited about them.
Sometimes we lose sight of the fact that a lot of the automated decision-making isn’t by the LLMs. My friend Kevin Baker just wrote a very nice piece kind of highlighting that tension in the Iran war where people were trying to attribute a lot of blame to these LLMs for making decisions about where to bomb and mistakes in bombing. But really it’s this Project Maven run by Palantir that’s doing all this kind of summarization and rapid target assessment.
So it almost less, it almost takes our eye off the target-
[MARION FOURCADE]
Yeah,
[BENJAMIN RECHT]
where the problem is Palantir. But I think we all know that. But the…
But then, just back to your other question, though, is like of how these language machines don’t seem to actually obey any logic or rationality, which I think is kind of fun, right? They’re just giving poetry back that we somehow associate in our head with being whatever we needed it to be.
[MARION FOURCADE]
Yeah, we think.
[BENJAMIN RECHT]
I know.
[MARION FOURCADE]
We see this logic, right?
[BENJAMIN RECHT]
We associate it with whatever we want it to be.
[MARION FOURCADE]
Yeah, exactly. That’s right.
[BENJAMIN RECHT]
Exactly. That’s right. That’s right.
[MARION FOURCADE]
Exactly.
[BENJAMIN RECHT]
That’s right. But it’s similar. It’s just funny how you get ungrounded language, right?
If I read your book, I treat it differently than the text that comes out of these machines. And the question is: which are we going to give authority to? That’s- That’s tricky.
I think a lot of times we are kind of just giving authority to their output and just assuming that it’s going to give us correct answers. So, yeah, maybe I’ll have to write a volume two in a couple years. We’ll see.
We’ll see.
[MARION FOURCADE]
That would be great for all of us. So, and I’ll finish with that question, perhaps. You’re an EECS professor.
[BENJAMIN RECHT]
Yeah,
[MARION FOURCADE]
writing a book that sort of tells the field that, you know, the… Well, that looks back at the history of the field and sort of thinks about the flaws in that, you know, in the way in which that history unfolded. And so I was wondering how this is landing, What’s the response in computer science?
[BENJAMIN RECHT]
There are a lot of computer science grad students who helped me write this book. So it was definitely like my group was super involved, and they were super engaged in it, but also we had a seminar a couple springs ago that had a great group of students engaging with these ideas. I think we’re all kind of open to it.
I think a lot of people in CS are kind of\u2014 You know, we all end up in CS for weird reasons— Yeah, for different reasons.
A lot of them, it’s a lot of people, really, me and Jeff, where we just played with computers a lot as kids. And so that was kind of like a natural, natural way to get towards more computing. But I think sometimes you don’t…
It’s easy because the field is moving so fast, and there’s always— a new computer, and there’s always a new thing to think about, that you never bother to be like, “Where does any of this come from?” And it is funny just to go back and I can just… it all stays.
There is obviously important things that happened before 1945 that influenced the computer, but so many of the ideas just happen all at once. Mm-hmm. And then we just fix that as the design, and we run with it afterwards, and we get to the present.
So I think, I don’t know, I think a lot of computer scientists Are interested in tracing that lineage because we’re all trying to figure out what we’re going to do next. And sometimes waiting for the next more expensive GPU is less satisfying.
[MARION FOURCADE]
Great. Thank you. So we’re going to start the Q&A part of this.
[AUDIENCE MEMBER]
So, yeah. Oh, wow. Thanks.
Loved it. Not a surprise, maybe you already know I like the book. But I was wondering, building on Mary Ellen’s question with the LLMs, if they’re not— An extreme case of, you know, here’s an optimization system optimized for one thing, right?
Next token, or next token that makes me happy, or next token that whatever, right? And then all the applied uses are not that, right? Like, when I ask it for some poetry, right?
Like, it’s not optimized to like that, right? It’s optimized for something else.
[BENJAMIN RECHT]
Sure.
[AUDIENCE MEMBER]
And so I wonder if this is just an extreme case of system optimized for one thing, then used or applied for some other thing, and maybe that’s an invitation to think of, like, well, all of these optimization systems get reappropriated and used for purposes that they weren’t originally optimized for So in that sense, maybe they’re not strange, but maybe it’s an invitation to think more about other cases that are kind of the same in that way.
[BENJAMIN RECHT]
That’s right. Yeah, yeah,
yeah. That’s interesting. I think a lot of optimization systems are just used for their optimization goal.
Like routing and supply chains. It’s just… Like, there are some classic core operations research problems where that’s just what you do.
That’s just what you do. I think the thing that’s so interesting to me about where machine learning has gone is that for a while it also was mostly about- Optimizing for one thing.
And now they still use the language of optimization. People love to talk about optimization, but they’re very complex software artifacts that have been optimized for all sorts of different things at different times. For people– I mean, we could spend a long time talking about how people build the product that you see in front of you when you see a ChatGPT window, because it’s not just the same simple
We predict one token at a time. You start there, but then you add all these things to do instruction tuning. You do all these things where you get human feedback to shape things.
They add a bunch of, like, synthetic data from software. They do all sorts of other things. They wrap it around an API.
And so now we just have this complex software artifact that sometimes works and sometimes doesn’t. That is supposed to be general purpose for whatever that means. And it’s a very odd place to be.
It’s an odd endpoint. I still don’t. I mean, they’re amazing.
Although I would say that 99% of my computer… …use is not that, right? 99% of my computer use is hoping that… … the screen comes on when I plug in the cable.
[MARION FOURCADE]
Yeah. Yeah, very true.
[AUDIENCE MEMBER]
So it’s a fascinating presentation, and one of the things that is really striking was in the beginning of the talk when you talked about how many of these intellectual developments all occur in the 1940s, and asking what’s special about that moment. You were connecting it to World War II, and I was thinking, it occurs to me that maybe there’s another important reason for what’s going on at that point in time, which was the Great Depression. And in particular, there had just been a massive crisis, particularly of the economic system, of the idea that just letting people start companies, run companies, sell whatever they can sell and buy whatever they feel like buying will lead to
[BENJAMIN RECHT]
Any kind of order that gets us to what we want seemed to have been discredited. There was, right, you know, there was a lot of receptivity to the idea that, say, having the government take more of a role in planning things.
[AUDIENCE MEMBER]
Might in principle be able to lead the world to be more efficiently organized. And I wonder how much the intellectual moment that’s leading to these developments might have been connected to that.
[BENJAMIN RECHT]
Yeah, no, that’s a great point. I mean, there’s so many things that are happening at once. I think there’s the…
There are the small group of people who are working on computers who feel like they’re isolated, and they talk like they’re isolated if you read, their own recollections. But obviously, there are these other forces around them. The fact that the administrative state grows almost as fast as Moore’s Law is interesting, right?
I mean, there was a tremendous explosion. It did seem like people probably got distracted from these fixes during the war, or maybe that, like, so much of the structure gets set up during the war. But I think 100%, going back again to the example of the Food and Drug Administration, like the first law That was passed to say that they had to make drugs safe and effective was 1938, was also in response to this free-for-all in the drug market.
But it really wasn’t until 19– Ooh, I’m forgetting the year, ’62, that it was codified that they had to start doing these kind of like very restrictive RCTs. So it took a little while to set those things up. I agree with you though, that those seeds had been building in the pre-war period.
[MARION FOURCADE]
Yeah, that’s one here, and then in the back.
[AUDIENCE MEMBER]
Huh? Thank you. I don’t know if this is on, but I was taking off from the fact that the minimum price for the best diet was only a few cents away. I wanted you to comment on… I mean, it seems recently we’re fascinated with doing the best.
[MARION FOURCADE]
Yeah.
[AUDIENCE MEMBER]
But for most of what we engage in, most of the advice we’re asked to give or the advice we take, there’s a pretty flat objective. Could you comment on the kind of I don’t know, fetishization of the optimum when it’s it’s actually usually a pretty flat. We want to stay away from the shoulders, maybe.
[BENJAMIN RECHT]
Absolutely yep. Absolutely. I think that was, especially with regards to thinking about our own lives, And people realized pretty quickly that for the most part, you just want to get something okay, good enough.
Herbert Simon had this idea of the satisfying rather than optimizing. We satisfy or satisfice is what he called it. Satisficing.
That’s all we have to do. We just have to get to good enough. But I think that the-
It’s funny how, since I could compute the optimal thing, why wouldn’t I do it? Also, it becomes this prevalent mindset that a lot of times, like, right? It’s there for me to do it, why would I do less?
You need some other kind of contingency or some other kind of something, some other kind of constraint to rear its head before people would kind of be able to. Say no to just doing what looks like the best? Well, the easy answer is that we’ve kind of forgotten about the work it takes to get to the best because now we just go for the best and we don’t even think about how long it takes, how much energy it takes.
Yeah. It seems to me there is a connection that you have people who devote themselves to these.
[AUDIENCE MEMBER]
Costless systems for optimization, the machine’s doing it. Yeah. But somehow the mindset carries over and we don’t think about the kind of development costs in figuring out what I’m gonna have for lunch. We just… Just a bad example.
[BENJAMIN RECHT]
Well, I… No, no, no, it’s a great example because I think that that’s always the perplexing thing. Especially, I mean, I think this optimization thing is on everyone’s mind right now.
And people love to claim that they can optimize everything. But like you said, just the sheer complexity that goes into actually writing down a model that reflects the reality that you live in, including these computational Limits and just the frustration of trying to go and solve an optimization problem every microsecond is just too much. So even though we have this idea, I think I misunderstood your question.
We have this idea of aspiring towards this optimality. There’s never anything as clean as that diet problem. Even the diet problem isn’t clean.
[AUDIENCE MEMBER]
Thank you for your remarks and Q&A. I was wondering about kind of going into the prehistory of the nineteenth, before World War II, because it feels to me that, yes, that World War II definitely catalyzed a certain thing, but to my way of thinking, it’s just kind of a logical development of science and remembering that science was actually called natural philosophy. It’s a philosophy of natural phenomena and great debates and theological influences, perhaps projected onto natural phenomena. So I guess I feel this actually was a civilizational advance to have a type of methodology for talking about a subset of questions that people are interested in.
And measurement of things that multiple people can look at is central to the way Western science developed. And…
[BENJAMIN RECHT]
To me, that’s actually very deep. That’s a type of intersubjectivity. It’s not just one person proclaiming God has said gravity works such and such.
So, yes, there are. There’s many social problems that can come out of this, but I think the ability to have something that people, multiple people can look at and examine in a much more forthright way rather than arguing about some abstract value such as freedom or God, that this in the domain Where it’s applicable is a huge advance, and not to minimize that. And my humble initial thought on a corrective, perhaps, to some of these things is always to present when people state, “This is the optimal solution,” you say, “Based on what?”
[AUDIENCE MEMBER]
Evaluation. Because it’s true, cost function extremely widespread, but people are optimizing in many different ways. Some people seem to be optimizing entertainment value for themselves, for example, or some people might even be optimizing avoiding looking at certain things.
So, I don’t know if you have comments on that.
[BENJAMIN RECHT]
I don’t… uh, well, there was a lot in that statement because I think there’s different things. I mean, the important thing here is that you couldn’t compute it until you had the computer.
So the objectivity, like, it’s— very small— Yeah, except for very small problems, and I think that’s the big change, yeah.
Yeah. Yeah, Eli, go ahead.
[ELI]
Yeah, so I have a kind of historical question. Going back again to World War II. D-Day was the biggest, most important optimization problem of that time.
And Eisenhower was appointed to be the general in charge of D-Day because he had a reputation of being an expert on logistics. Now, so I’m very fascinated to hear that, of course, the math for logistics was not figured out until a year after D-Day. Right?
That’s what you told us.
[BENJAMIN RECHT]
A couple years, yeah.
[ELI]
A couple years after D-Day. And the… The computers to do a real calculation were not available until 1950.
[BENJAMIN RECHT]
That’s right.
[ELI]
So the birth date for this sort of thing is 1950. But I’m wondering from you, do I have this sequence right? Because this is an incredible story: is that they managed D-Day without any of this math.
[BENJAMIN RECHT]
Yeah, yeah. I think the other kind of, like in parallel, right, the incredible story, like part of the reason why these guys were so interested in these things was that there had been so much involvement of academics in World War II in a way that just hadn’t happened before that. Vannevar Bush, who we kind of know for, for a variety of reasons, made one of the first early computers, and beyond that became a government…
He was a big advisor to a lot of the war effort. He had been in MIT, and he convinced the government to fund tons of research, throw money, like a lot at MIT, a lot at Princeton, to help on that war effort in a way that just had never been done before. And so there was all this different kinds of talk.
I don’t know how much, it’s not really. I don’t know if I should be able to figure this out for later. I will get a note for later.
I don’t know how much that actually made its way down to Eisenhower himself. But they were definitely involved in a way that was unprecedented. We know about the Manhattan Project, obviously, but also they were really deeply involved in a lot of logistics and planning that had big influences on people that, you know, we know.
Like Wiener was working on targeting systems, and that would become detection and estimation theory after the war. So it’s very interesting how there was that kind of, for the first time, like a belief in the American Academy as being crucial and critical for that effort. Yeah,
[MARION FOURCADE]
one question in the back. Yeah,
[BENJAMIN RECHT]
we’ll take a few more.
[MARION FOURCADE]
Sure, sure, sure. We’re-
[AUDIENCE MEMBER]
Hey, man. I’m curious about how you think about the way that optimization and quantification can serve, I guess, as was gotten at earlier, as a substrate to have discussions where you could actually say, “Okay, this is the part that’s wrong.” As opposed to a problem that comes up a lot in, for example, business decision-making and operations research, all these sorts of things, is that when you are in a room with four other people and you’re trying to say what’s right, who’s right, it very quickly becomes personal.
And I think… Yeah, one of the utilities, and I’m not sure where this… Comes from her if, I guess this was a sociological intention behind quantifying and datamatizing everything, is that now you have something to point at and say, “This is what should be.” I have a slider I can actually move, as opposed to someone I have to tell is actually wrong.
[BENJAMIN RECHT]
I would argue that, well, I think it could be very destructive for you to think about rational choice- Fitting your life into a rational choice framework personally and then trying to have everything be optimal. Sometimes when you have to make a decision with a bunch of people, rational choice is a nice way to kind of defer the emotional content. Like, you can all write something down about what your commitments are, and sometimes the optimal, like you can agree on this particular system of rationality as a way to disengage from the other things that are kind of keeping you from having a decision.
And a classic example that’s maybe not, I don’t know, I think it’s helpful, is the resident matching program, which is a very complicated. System for having young graduates of medical school find their residency programs. And it turned out to be something that allowed to take a lot of these kind of like heated decisions and arguments and biases out and have something that maybe isn’t perfect, but we have something that now is like we have rules that everyone agrees with and systems where everything adds a little bit extra transparency.
So, yeah.
[AUDIENCE MEMBER]
Hi, Professor Recht. Thank you so much for the talk. I was wondering whether you can spell out more of our current obsession over optimization.
Because I feel like when you mention about satisficing, to my mind, satisficing seems to be quite of a negative connotation, where it seems like we’re not doing our best. And the idea of optimization… For me, it’s kind of like the golden star that everyone wants to achieve.
So, like, is it just a matter of us? As there was, I think, previously a question on whether it’s just optimizing different values, like values that each of us have and like, or like If we give up on optimization, don’t we just give up also this satisfaction that comes from optimization? And in your talk, you mentioned that the agreed goal is the good, optimization is the good, and then something is the bad.
Isn’t that what is lost there if we give up optimization?
[BENJAMIN RECHT]
The, um, I— It is very much in the local culture. Maybe it’s worked its way over to Berkeley. But I feel like in the San Francisco culture that everybody has to be optimizing all the time, working all the time, maximizing their, I don’t know, everything.
Maximizing so many different axes, and nobody seems happy. I don’t know anybody who’s committed to these optimization programs who really feels happy. I think so much of it is about–
And so, you know, how do you optimize, how do you maximize happiness? And nobody, there’s no way. Because, you know, it’s not something that’s nice and quantifiable and neat and clean, and have all these beautiful properties that like lends itself to this optimization framework.
So I think that when it comes to our own lives, first of all, nobody– Also, if you talk to people who are optimizing, nobody even has a Number they’re going for. Even people on, like, what does it mean to be doing, making the most out of every moment? I don’t know.
Quantitatively, the most out of every moment. It’s very hard to pin down. So I think that the… but again, I don’t want to, Just to loop back though, like, I’m not against optimization in the right context.
I think there are lots of contexts where it’s fine. For yourself, it’s a little tricky for me to come up with a good one. But I think for lots of things in terms of, like, the way that we do logistics and supply chains and the way that we manage.
Flying ourselves around the world fast, like in, you know, you can get to any point on Earth in less than a day. These are amazing. So I think it’s just kind of understanding how to navigate that complex world.
And I think part of the issue that we have with a lot of the people who I see kind of throwing themselves at the whim of the personal optimization is they feel like by doing that, an optimal— it’s almost, in the light of the society that’s optimizing, it’s almost like a Calvinist ethic. That’s like the more you optimize, the better that makes you as a human being, and I just don’t think that’s true. I think sometimes you have to let go and you just have to embrace, like, what is it that I care about?
And there’s not gonna be numbers on that. And that can go up, what you care about. You can go up in some kind of way, but it’s not gonna be in a vector space.
[MARION FOURCADE]
Okay, well, so let’s take… We’re a little bit out of time. Okay, but let’s take one…
[BENJAMIN RECHT]
Wait, we have two more questions.
[MARION FOURCADE]
Three? Oh, God.
[BENJAMIN RECHT]
They can also come ask me if you want. You can ask. No, I don’t mind doing that.
[MARION FOURCADE]
So let’s take one last question.
[BENJAMIN RECHT]
Okay, I can stay. I can stay.
[MARION FOURCADE]
So we may have time for one last one.
[BENJAMIN RECHT]
I guess, like, on the heels of that subject question, it’s like I can think of kind of… an optimization or a problem that’s on the boundary of optimization in terms of it having quantifiable and qualifiable aspects. Yeah. Um, so for example, like event planning, whether it’s like a birthday party or, like a wedding.
That it’s, you know, similar to the, like, the job scheduling or logistics. There’s, like, a… You can sort of parameterize it, and then you can, you know, say, like-
[MARION FOURCADE]
Yeah.
[BENJAMIN RECHT]
It’s like, you know, similar to with the- dietary, uh, the calorie thing, right?
[MARION FOURCADE]
Yeah.
[BENJAMIN RECHT]
But it’s like, well, if you were planning, like, you know, uh, like, your colleague’s birthday, and it’s like, oh, like, I’ve worked with them for 10 years, And so I know, like, this and that. And then so, you know, there would sort of be qualitative aspects to that Absolutely.
Yeah, yeah, yeah.
I think that’s the other thing. I think that even that there are, I was just thinking now, what other things that you should plan for? Like, you know, some of the in, you know, your own personal investment portfolio is, like, good to manage from an op-
You don’t have to be optimizing, but you should be thinking about that quantitatively, and be frugal, and don’t spend all your money, especially on parties. You know, budget your parties appropriately. Have good parties, but, you know, understand that you have a lot of parties you’re gonna have to throw in your life.
I don’t know. I think it’s this complicated thing to throw around. I think that it’s, uh, there are- That quantitative doesn’t necessarily mean bad or inhuman.
I think that lots of the things that we do in our lives are best served by those numbers. It’s just the understanding that there are sweet spots for everything, right? And I think it’s just don’t get too wrapped up in one framework that’s gonna run everything.
That’s the hard part, is we just have to have that flexibility to kind of live these fulfilling lives that we get to live.
[MARION FOURCADE]
Well, that is such a perfect sentence to end this presentation. Thank you, Ben. That was great.
(applause)
(silent)

Maximilian Kasy received his PhD at UC Berkeley and joined Oxford after appointments at UCLA and Harvard University. His current research interests focus on social foundations for statistics and machine learning, going beyond traditional single-agent decision theory. He also works on economic inequality, job guarantee programs, and basic income. Kasy teaches a course on foundations of machine learning at the economics department at Oxford. Learn more at his