Recently Martin Steffen passed around a link to "a Symposium on the Promises and Challenges of the Revolutions in Genomics and Computer Science". See LJ:REF_TALKS:3239. Our supervisor, George Church, spoke at the symposium. His PPT slides are here. I didn't have a chance to attend the symposium but George mentioned that another speaker, Marvin Minsky, came and left very abruptly. I've attended a few classes of 6.868 led by Marvin; in the last class of the semester he mentioned that his comments at the "Future of Human Nature" symposium sparked some controversy: Slashdot, Wired etc.

Marvin's actual thoughts from the symposium are in a Usenet thread here.

Excerpt: Most early researchers in artificial intelligence had the goal to build machines that would be very intelligent.

However, it soon turned out that solving hard problems usually needs a lot of knowledge. This was recognized in the early years— by researchers in the 60s and 70s who built knowledge-based systems. Indeed, in the 1980's, the so-called expert systems became widely productive and popular. However there was a problem with them: For each different kind of problem the construction of such systems had to start all over again, because they didn't have, or accumulate what we called commonsense knowledge. To be sure, each new system could use the same ‘shell' — but I think it turned out, at least in my view, that this was more of a fault than a virtue.

Only one researcher recognized that this was so serious as to commit himself entirely to it. That was Douglas Lenat, who has directed the multi year project called CYC—which has resulted in solving some problems in this area.

Unfortunately, in my view, the rest of the artificial intelligence community tried, instead, to find alternative medicines to deal with this problem. For example, many projects were aimed at what I call building baby machines, which were supposed to learn from experience, eventually to become as smart as people. These all failed to make much progress because (in my view) they lacked architectural features to equip them to think about the causes of their successes and failures— and then to make appropriate changes.

Instead, most researchers went in other direction— of trying to build an evolutionary system, that would start with very simple machines and then, by one or another mutation scheme, evolve more architecture. None of those projects ever have gotten very far.

The story is very much the same in the field of building large neural networks. These can frequently solved interesting problems However those networks don't have the capability, of reflecting on what they have learned, and then making appropriate changes.

Other researchers have aimed their work toward making make a ‘unified theory' of thinking. Each of those projects has proposed some good architectural ideas, but not enough to these to support good self-reflective processes, with which they could improve their own operations.

Evolving the simple into the more complex is receiving attention in high-prestige scientific journals and not just the computer science community. See LJ:REF_NATURE:1245. What happens if that simple starting point is quantum mechanical? I'm working on that now...