Sergio Tenenbaum’s Rational Powers in Action presents a new theory of instrumental rationality – the Extended Theory of Rationality (ETR) – that is tailored to accommodate a series of problem cases for standard Rational Choice Theory (RCT). Tenenbaum writes in a breezy and engaging style, and the book does a good job of putting forward his theory as an attractive candidate for a ‘theory of instrumental rationality’. In short, Tenenbaum notes that the ends rational agents pursue are often long-term and not fully determinate. Ends are ‘long-term’ in the sense that they are achieved not by performing one particular action at one particular time, but rather by performing some acceptable sequence of actions over time (where there are many acceptable such sequences). And ends are not fully determinate in the sense that some possible outcomes of my actions are borderline cases of my having achieved my ends.

If I want to make a sandwich, I have to slice the bread, get the butter, ham and mustard from the fridge, spread the butter, cut the ham, daub on the mustard and then cut the sandwich in two. If, while retrieving the butter I do a quick inventory of what food we have in store for dinner this evening, I am not thereby preventing myself from successfully making a sandwich. If I stop before I put the mustard on to check on the sleeping baby, I am likewise not preventing a successful sandwich from being made. So my end of making a sandwich is not one act, but a series of acts, possibly interspersed with other acts that do not contribute to my achieving my end. And at all scales, our ends are like this. Tenenbaum returns many times to the example of the end of writing a book: one does not have to write every day in order to achieve this end, although the more one writes, the more likely one is to achieve it.

My end of making a sandwich does not necessitate that I get out the ham or mustard: cheese and pickle would also constitute an acceptable sandwich filling. My end of making a sandwich does not fully specify what ingredients I should use. And, indeed, certain choices of ingredients provide me with more or less clear cut cases of a sandwich. Is two slices of buttered bread a sandwich? Is a hot dog a sandwich? Is a burrito a sandwich? So my end of making a sandwich is not fully determinate and admits of better or worse realisations of the end.

Tenenbaum points out that standard RCT handles these complications quite poorly and that his alternative theory does better. One could frame Tenenbaum’s project as follows. The first phase of the project is to demonstrate a build-up of anomalies that afflict the dominant paradigm in the theory of rationality (RCT). These anomalies stem from some idealizations that RCT makes: that ends are fully determinate, and that an act (a means to an end) is implemented at one time. As we saw above, neither of these things are true of real acts and real ends. This is, broadly, the topic of the first four chapters. Next Tenenbaum takes aim at some who seek to repair the mainstream theory by adding epicycles to the basic model. Tenenbaum focuses on those who take a ‘planning’ approach to sequential choice, and argues that such ‘future-directed intentions’ are not as good a solution to the problem as his preferred alternative. This occupies Chapters 5 and 6. Chapters 7 and 8 then elaborate Tenenbaum’s ETR, his preferred alternative new paradigm to replace the degenerating research programme of RCT. And finally, Chapter 9 shows how and when the results of the old paradigm can be recovered with the resources of the new. In particular, the remarkable (if circumscribed) successes of RCT – of representing agents as expected utility maximizers – can be replicated within ETR.

My choice of the language of scientific revolutions – ‘paradigm’, ‘build up of anomalies’, ‘degenerating research programme’ etc. – is not accidental. I think this is a fruitful way to think about what Tenenbaum is doing. Tenenbaum is advocating replacing the old paradigm of RCT with the new paradigm of ETR: Tenenbaum is advocating for a revolution. But who, specifically, is he trying to convince to join him in revolting against the dominant paradigm? This is, to my mind, a book written with a philosophical audience in mind. Philosophers who want a theory of instrumental rationality are the target audience for this book, and for that audience, I think Tenenbaum is successful at providing a convincing case in favour of this paradigm shift. People for whom accommodating indeterminacy of ends and the temporally extended nature of our actions is important will be swayed by this book’s arguments.

However, for other users of RCT this book perhaps won’t be as convincing. RCT is a rigorously mathematized theory integrated into the mainstream of theoretical economics, and I think many economists will be unmoved by the fact that this theory has these flaws stemming from a particular kind of idealisation. ‘That’s how you do science’: they will shrug ‘you idealise’. These users of RCT are coming from a paradigm where a false assumption is not a killer blow to your model, where consistency with nearby theories, and a certain level of mathematical sophistication are theoretical virtues to be valued.

This is where we appear to reach a kind of bedrock disagreement about what theoretical virtues one’s theory of rationality ought to satisfy: an example of the kind of incommensurability of paradigms that Thomas Kuhn is famous for.Footnote ¹ On the one hand, we have the philosophers who care about the indeterminacy of ends, and don’t care about integration with theoretical economics; and on the other we have the economists who care about integration with economics but not about the indeterminacy of ends.Footnote ² The philosophers will point at RCT and say ‘look this is a false theory of rational agency’ and the economists will respond ‘no, it’s merely a model with a false assumption, but is no less worthy of study for that’. All models are wrong, but some models are useful, as the saying goes. I am not convinced there is actually any serious disagreement to resolve here. If you set out to solve different problems and you rank theoretical virtues differently, then you’ll end up using different theories and there’s no real puzzle to resolve here.

I think it’s worth pointing out that the term ‘idealization’ is not particularly apt. There are two kinds of idealization involved in RCT. One thing theorists do is they ignore the ways in which actual agents make mistakes or otherwise fail to live up to the standards of ideal rationality. And the other thing theorists do is they abstract away distracting or complicating features of the real world in order to focus on the effects of interest. This latter kind of idealization is the one that is common currency in philosophy of science discussions of ‘idealization’: think frictionless planes, point masses, and so on. There’s no normative judgement involved in this latter kind of idealization. To some extent this distinction tracks what get described as ‘rationality axioms’ and ‘structural assumptions’ in the literature on representation theorems (see e.g. Joyce Reference Joyce1999).

So there are two ways for a defender of RCT to push back on Tenenbaum’s criticism. One would be to argue that ideal rational agents do in fact make decisions at a time, and in a way that is not subject to any indeterminacy. And the other would be to argue that determinacy of action and the lack of temporal extension are legitimate abstractions in analogy to the frictionless plane. I think the former line of argument is hopeless. We are typically not in complete control over the outcomes of our acts, and so at least some indeterminacy is inescapable. I may want to cut a slice of bread of a specific thickness and evenness, but I am incapable of doing so. I cannot bring about very specific bread-slicing outcomes, so it would not help me to have perfectly specific bread-slice ends. The latter line of argument seems significantly more plausible to me. On this latter approach, RCT is not faulty or mistaken, it is simply incomplete in the same way that Newtonian mechanics is incomplete because it fails to properly accommodate the complications of things being very small, very heavy or very fast.

The difference between the groups I earlier labelled ‘the economists’ and ‘the philosophers’ is how important they find it to accommodate those complications. Interestingly, on Tenenbaum’s view this faultline also shows up in another way. RCT has the attractive feature that rational agency consists in conforming to a set of principles (precisely which set of principles is a matter of lively debate). Satisfying the principles is sufficient for rationality. In ETR, by contrast, some aspects of rational agency cannot be accommodated as principles to be satisfied: there are some so-called ‘rational virtues’ which, while desirable are not necessarily satisfied by all rational agents. Tenenbaum suggests that rational virtue is in certain respects like being brave: it’s something that comes in degrees, with some measure of indeterminacy as to whether you have succeeded. Is it a desideratum to have one’s theory of rationality be entirely principle-based, or is the introduction of non-necessary rational virtues an advantage? If we are in the business of producing a rigorously mathematical theory, then being based entirely on principles we can formulate as axioms is an advantage. If, on the other hand, we want to accommodate the intuition that one can be more or less rational then the virtue-based theory has some attractive elasticity.Footnote ³

Rational virtue is, I think, a close cousin of two other frustratingly nebulous ideas that also deserve more attention in philosophy. The first of these is the idea of ‘common sense’, in particular, the common sense that computer programs appear to lack. If I am designing a self-driving car, the ends I want to achieve are indeterminate: ‘have the car drive smoothly but safely’ or something like that. But on some level, the software driving the car can’t help but be principle-based: all computers can really do is logic and arithmetic. Computers can’t exercise good judgement. Given that our ends – and the ends we want to use computers for – are indeterminate, there will always be cases where the computer gets it wrong, and obviously wrong from the perspective of any competent human observer. The field of evolutionary computation provides a number of excellent examples of how even apparently quite carefully specified ends can be achieved in unconventional borderline ways by computers that appear to lack common sense. For example, an algorithm ‘learned’ that it could win games of tic-tac-toe by causing its opponent to run out of memory, crash and thus forfeit the game, rather than by actually playing legal moves (Lehman et al. Reference Lehman, Clune, Misevic, Adami, Altenberg, Beaulieu and Bentley2020). The end the algorithm is aiming at is still clear, but its manner of reaching that end falls outside what we’d consider acceptable in normal circumstances.

In deep learning, things can get significantly weirder. We’ve made computers remarkably good at mimicking human performance on a number of tasks such as image recognition by effectively giving up control of which principle the computer embodies: machine learning essentially involves letting the computer search for the principle to conform to that best allows it to score well on the data it’s trained on. This can lead to even more surprising failures. Image recognition algorithms can be fooled by changes to the image that make no difference to human performance (Heaven Reference Heaven2019). And some can even be fooled by attaching a text label: a picture of an apple with a label that says ‘iPod’ on it gets classified as an iPod (Goh et al. Reference Goh, Cammarata, Voss, Carter, Petrov, Schubert, Radford and Olah2021). These examples remind us that whatever it is computers are doing, and however well it is that they do it, they are operating in a very different way to humans (this despite the fact that artificial neural networks are supposed to mimic the low level functioning of real neurons).

The second kindred idea to rational virtue is what Pierre Duhem called ‘good sense’. Returning to scientific revolutions, many people have noted that there is no principled algorithm for choosing among scientific theories (Kuhn Reference Kuhn1977). Indeed Kuhn argues that there should not be such an algorithm. It seems like there is some element of judgement or good sense that is required to make a decision about which theories to believe or pursue (Ivanova Reference Ivanova2010).

Despite the many appearances of this kind of idea – common sense, Duhem’s good sense, Tenenbaum’s rational virtues – the concept remains difficult to theorize about (although obviously important). Theorizing about determinate principles rather than squishy virtues remains easier. What is the upshot of recognizing this family resemblance? In a sense, useful artificial intelligence is going to involve training computers to exhibit (or at least mimic) rational virtues rather than merely conforming to rational principles: this is so because we want AI to be means to our indeterminate ends. And the only methods we seem to have available to do this involve approximating virtuous good sense through getting the computer to learn which function best approximates the ‘good sense’ correct answers in the training data. Hardcore materialists might want to go a step further and say that actual human agents are just complicated meat machines, and we are actually not doing anything different from the AI. The inescapable conclusion of this line of thought seems to be that Tenenbaum’s rational virtues are different only in degree from rational principles, not different in kind.Footnote ⁴ Even those without strong materialist commitments might want to take a moment to consider whether the things they consider inviolable principles are not better thought of as rational virtues with some high degree of determinacy or strength.

So Tenenbaum’s book presents a theory of instrumental rationality. It is geared towards convincing an audience that takes capturing indeterminate and extended ends as important desiderata for such a project. The cost of Tenenbaum’s ETR is that we lose some of RCT’s advantages: its integration with economic practice and its rigorously mathematized foundations. Tenenbaum’s theory, unlike RCT, is not wholly principle-based: there is scope for rational virtues. Whether this is a cost or a benefit will depend on one’s perspective. I think it is obvious that the more philosophically minded readers of this journal should read Rational Powers in Action; I also think that even the more economically minded readers should too. Even if one doesn’t expect to be convinced by the argument, even if one expects one’s endorsement of the theoretical virtues exhibited by RCT to outweigh its failure to accommodate extended and indeterminate ends; this book will help to clarify the costs of sticking with the old paradigm rather than joining in with the revolution.