At first glance, Hubert Dreyfus’ 1992 book What Computers Still Can’t Do (WCSCD, originally published in 1972 as What Computers Can’t Do) seems untimely in the current business climate, which favours massive and widespread investment in AI (these days, often understood as being synonymous with machine learning and neural networks). However, being untimely may in fact allows us to act “against our time and thus hopefully also on our time, for the benefit of a time to come” (Nietzsche). And the book’s argument might in fact not be outdated, but simply forgotten in the frenzy of activity that is our present AI summer.
Dreyfus outlines four assumptions that he believes were (in many cases, still are) implicitly made by AI optimists.
The biological assumption. On some level, the (human) brain functions like a digital computer, processing discrete information.
The psychological assumption. The mind, rather than the brain, functions like a digital computer, even if the brain doesn’t happen to do so.
The epistemological assumption. Even if neither minds nor brains function like digital computers, then this formalism is still sufficient to explain and generate intelligent behaviour. An analogy would be that planets moving in orbits are perhaps not solving differential equations, but differential equations are adequate tools for describing and understanding their movement.
The ontological assumption. Everything essential to intelligent behaviour — such as information about the environment — can in principle be formalised as a set of discrete facts.
These assumptions all relate to the limitations of computation (as we currently understand it) and of propositional logic.
Dreyfus is famous for interpreting thinkers such as Heidegger and Merleau-Ponty, and consistently draws upon these thinkers in his arguments. In fact, as he points out in WCSCD, the phenomenological school attacks the very long philosophical tradition that sees mind and world as strictly separate, and that assumes that the mind functions by way of a model that somehow can be reduced to logical operations (we can see why the field of AI has implicitly, and in many cases unwittingly, taken over this tradition). Historically, this tradition reached perhaps one of its purest expressions with Descartes. Indeed Being and Time, Heidegger’s major work, is very anti-Cartesian. Heidegger’s account of intelligibility demands that one (Dasein) is in a world which appears primarily as meaningful interrelated beings (and not primarily as atomic facts, or sources thereof, to be interpreted), and is historically in a situation, making projections on the basis of one’s identity. Here, calculation and correspondence-based theories of truth are derived and secondary things. There is no clear separation between world and “model” since there is no model, just the world and our ability to relate to it.
I will hazard a guess that most neuroscientists today would not take the first two assumptions seriously. In all kinds of biology and medicine, we regularly encounter new phenomena and mechanisms that could not be captured by the simple models we originally came up with, forcing us to revise our models. Making brains (bodies) and/or minds somehow isomorphic to symbolic manipulation seems wholly inadequate. More interesting, and much harder to settle unambiguously, are the epistemological and the ontological assumptions. If the epistemological assumption is false, then we will not be able to generate “intelligent behaviour” entirely in software. If the ontological assumption is false, then we will not be able to construct meaningful (discrete and isolated) models of the world.
The two latter assumptions are indeed the stronger ones out of these four. If the epistemological assumption turns out to be invalid, then the biological and psychological assumptions would necessarily also be invalid. The ontological assumption is closely related and similarly strong.
By contrast, Nick Bostrom‘s Superintelligence: Paths, Dangers, Strategies is a more recent (2014) and very different book. While they are certainly worth serious investigation, theories about a possible technological singularity can be somewhat hyperbolic in tone. But Bostrom comes across as very level-headed as he investigates how a superintelligence might be formed (as an AI, or otherwise), how it might or might not be controlled, and the political implications of such an entity coming into existence. For the most part, the book is engrossing and interesting, though clearly grounded in the “analytical” tradition of philosophy. It becomes more compelling because of the potential generality of its argument. Does a superintelligence already exist? Would we know if it did? Could it exist as a cybernetic actor, a composite of software, machines, and people? It is interesting to read the book, in parallel, as a speculation on (social, economic, geopolitical, technological, psychological or composites thereof) actors that may already exist but that are beyond our comprehension.
Bostrom’s arguments resemble how one might think about a nuclear arms race. He argues that the first superintelligence to emerge might have a decisive strategic advantage and, once in place, prevent (or be used to prevent) the emergence of competing superintelligences. At the same time it would bestow upon those who control it (if it can be controlled) a huge tactical advantage.
Even though Bostrom’s argument is mostly very general, at times it is obvious that much of the thinking is inspired by or based on the idea of AI as software running on a digital computer. To me this seemed implicit in many of the chapters. For example, Bostrom talks about being able to inspect the state of a (software agent’s) goal model, to be able to suspend, resume, and copy agents without information loss, to measure hedonic value, and so on. Bostrom in many cases implies that we would be able to read, configure and copy an agent’s state precisely, and sometimes also that we would be able to understand this state clearly and unambiguously, for example in order to evaluate whether our control mechanisms are working. Thus many of Bostrom’s arguments seem tightly coupled to the Church-Turing model of computation (or at least to a calculus/operational substrate that allows for inspection, modification and duplication of state). Some of his other arguments are, however, sufficiently general that we do not need to assume any specific substrate.
Bostrom, it seems to me, implicitly endorses at least the epistemological assumption throughout the book (and possibly also the ontological one). Even as he rightly takes pains to avoid stating specifically how technologies such as superintelligences or whole brain emulation would be implemented, it is clear that he imagines the formalism of digital computers as “sufficient to explain and generate intelligent behaviour”. In this, but perhaps not in everything he writes, he is a representative of current mainstream AI thinking. (I would like to add that even if he has wrongly taken over these assumptions, the extreme caution he advises us to proceed with regarding strong AI deserves to be taken seriously – the risks in practice are sufficiently great for us to be quite worried. I do not wish to undermine his main argument.)
It is thinkable but unlikely that in the near future, through a resounding success (which could be an academic, industrial or commercial one, for example), the epistemological assumption will be proven true. What I hold to be more likely (for reasons that have been gradually developed on this blog) is that current AI work will converge on something that may well be extremely impressive and that may affect society greatly, but that we will not consider to be human-like intelligence. The exact form that this will take remains to be discovered.
Hubert Dreyfus passed away in April 2017, while I was in the middle of writing this post. Although I never had the privilege of attending his lectures in person, his podcasted lectures and writings have been extremely inspirational and valuable to me. Thank you.