July 26, 2004

The Limits to Generalism

I spent three days at the 3rd International Conference on Autonomous Agents and Multi-Agent Systems in New York last week.

I was a little disappointed, in some ways. I had hoped the meeting would be a bit more interdisciplinary, despite its strong connections to the American Computing Society. It was pretty much computer scientists all the way down. But that’s where multi-agent and autonomous agent systems live intellectually. One should not be surprised that the sun is in the sky during the daytime.

The consequence for me was that I understood very little of what I saw and heard. Every once in a while, the light broke through the clouds, generally in papers that were very explicitly devoted to using multi-agent systems for social simulation, those more concerned with the conceptual design and application of their simulations and less concerned with the formalisms, protocols or algorithms underlying them. I was able to grasp one presentation on the simulation of social insects and pheremones (due to the intensely well-travelled nature of the example) and even to see that the presentation offered relatively little that was new on that topic. I really liked one presentation that proposed a formalism for generating irrational agents, or at least for nesting normal Bayesian game-theoretic rationalities one step away from the functioning of a multi-agent system. It seemed very innovative and intelligible, particularly given that I was struck by how utterly reliant the whole field has become on rational optimizing designs for agents. I was also struck at the extent to which the demand for application to commercial needs drove the vast majority of presentations.

At most other points, however, not only did I not understand anything, I barely understood what I’d have to understand in order to understand a presentation.

I repeatedly extoll the virtues of generalism, but it cannot do everything. The sinking feeling I repeatedly had during the conference was knowing that to even get to the point where I grasped the substantive difference between different algorithms or formalisms proposed by many of the researchers at this conference, where I could meaningfully evaluate which were innovative and important, and which were less attractive, would take me years of basic study: study in mathematics, study in computer science, study in economics, areas where I’ve never been particularly gifted or competent at any point in my life. To get from understanding to actually doing or teaching would be years more from there, if ever.

The reverse movement often seems easier, from the sciences to the social sciences or humanities, and in truth, it is. There’s an important asymmetry that I think is a big part of the social purpose of the humanities, that intellectual work in that domain returns, or should return, broadly comprehensible and communicative insights rather than highly technical ones, and thus, that the barriers to entry are lower.

The ease of that move is deceptive, however. It’s the kind of thing that leads someone like Jared Diamond or other sociobiologically inclined thinkers, especially evolutionary psychologists, to what I call “ethnographic tourism”. Operating out of a framework that requires the assumption of universalisms in order to make cogent hypotheses about human history and behavior, scholars coming along that path often quickly scoop up the studies and accounts which support the foundational assertion of a universal and ignore those which do not or casually dismiss them as biased or “culturalist”, regardless of the methodology those studies employ. That’s what leads to their peculiar preference for the work of Napoleon Chagnon on the Yanomano, for example. Bogus or wild-eyed controversies about immunizations and manipulation aside, there’s at least reason from an utterly mainstream, meticulous, scrupulous and disinterested perspective to view some of his methodologies as debatable and to take seriously the work of other scholars who have made very different findings. There’s a selectivity principle at work in ethnographic tourism that wouldn’t be tolerated if it wasn't scientists cherry-picking material from anthropological scholarship they like and ignoring contradictory work.

That is not atypical of what can happen when scientists pressing towards generalism think they understand disciplines outside the natural sciences. Similarly, it’s become easy to mock and ignore scholarship in the humanities for being too theoretical, fashionable, incoherent, and so on, which it very often is. Alan Sokal’s hoax hit a real target, but if you want to think and write about problems like the nature of existence and knowledge, or about why and what a cultural work means to its audiences, sometimes you really are going to have to go into deep waters that require a complex conceptual framework. Some scientists tend to forget that on a series of crucial issues, skeptics in the humanities were closer to the truth for decades than scientists, most notably in the early debate between philosophers of mind, neuroscientists and computer scientists working on artificial intelligence about how easy it would be to create AI.

That debate is an important reminder, however, of what a kind of disciplined drift towards generalism can bring. The intensely fertile contemporary practice of cognitive science draws from all those areas and more besides. It almost seems to me that a good generalist ought to combine an overall curiosity and fluency in the generality of knowledge with a structured search of the possibility space of the intellectual neighborhoods which are just far away enough from their specializations to return novel possibilities and angles of attack but just close enough that those neighborhoods are potentially accessible with a reasonable amount of scholarly labor. To think about generalism in this way is to realize that different generalists are not going to end up in the same place. Their mutual engagements or conversations will have to happen in places of accidental overlap, because the concentric circles of one's own generalist competency are going to differ because they originate out of different initial specializations.

Proximity to your own discipline and specialization can also be deceptive. I’m planning another version of my Primary Text Workshop course for academic year 05-06. I’d like it to involve the students in doing the preparatory work that would be required for making a virtual reality environment based on a historical community—the speech trees, the knowledge of clothing and other material culture, the architectural and geographical knowledge, the understanding of everyday life rhythms, and so on. I’d prefer it be about a city whose history I know very well—Johannesburg or Cape Town spring to mind—but access to primary materials will obviously be limited. On the other hand, late colonial Philadelphia seems an apt choice, but I find myself simply overwhelmed by the literature I’d have to read in between now and then in order to achieve a basic comfort level. It’s not enough to have read Alan Taylor, Timothy Breen, Gordon Wood and so on about the colonial and revolutionary era—I’d need to go far deeper historiographically than that, and at that point, you begin to wonder whether it isn’t just smarter to hand the class off to a colleague who already specializes in that era.

I’ve been thinking about how to calculate the wider bounds of generalism beyond the discipline. In my case, for example, some of the ideas associated with complex systems, emergence, autonomous agents and multi-agent systems and so on are close enough conceptually that I can make use of them and contribute insights to colleagues working in those areas, but they’re just far enough away that I should not ever expect to do original work directly in computing applications myself. Sociobiology might be close enough that I could reasonably expect to offer some critical insights into its methods, but not close enough that I could expect to do my own original research into population genetics. Theoretical physics would be far enough away in every respect that I might not ever reasonably expect to understand it, let alone do it, given that much of it cannot even be translated from its mathematical conception into broadly communicative prose. At that point, you have to have enough faith in the entire system of knowledge production to just say, “I trust you to do what you do, and to do it how you do it”—and if it becomes imperative to do more, as it does in the case of tenure review, you just have to outsource the job of deciding whether another scholar’s work is original or skilled to someone else, to have the humility to know where the final outer bound of a generalist intellect lies.