Possible Insight

I attended the Singularity Summit today. Overall, it was worth the time spent.  I did not attend the workshop on Friday because it didn’t look substantive when I reviewed the program.  Today, I spoke to several people who were there and they confirmed my prediction. I took 7 pages of notes at the summit and hope to have some insightful synthesis of the material in a few days [Edit: first thought here, more here].  In the meantime, here is a short review of the talks.

To set the stage, Vernor Vinge submitted to a wide ranging interview. Unsurprisingly given that I’m a fan of his science fiction, I found the interview entertaining. He made one particularly excellent point and one particularly glaring error that are worth mentioning. The excellent point was that embedded, network processors are so useful that they will become ubiquitous (he mentioned that there will be trillions of them). Unfortunately, this popularity will also make them a critical point of failure.  The glaring error was to assert that as humans outsource their cognition to machines, the number of jobs suitable for humans will narrow. Economic history contradicts this theory, but more on this topic in a bit.

Then Nova Spivak talked about collective intelligence. While it took a while to get there, I found his ultimate point insightful: in order to truly achieve collective intelligence, we need some sort of “meta-self” that maintains models of the internal state of the collective and how the collective relates to the external world, as well as structures the goals of the collective and tracks progress towards the goals.

Esther Dyson was caught wrong-footed. Apparently, she thought someone was going to interview here and had not prepared any material. She talked a little bit about genetics, but nothing I found even remotely new.

James Miller was the hit of the morning. He spoke (somewhat tongue in cheek) about the economic implications of a significant portion of the population anticipating the singularity: more money spent on safer cars, construction workers becoming more expensive, people saving less for retirement, the market for office buildings crashing, students not wanting to study anything boring.

Justin Rattner gave a fascinating (to me) talk about the nuts and bolts of Intel’s approach to maintaining the inexorable march of computing power growth. I was blown away by the fact that Moore’s Law stopped last year and nobody noticed. The original formulation of Moore’s Law was about CMOS transistors. But CMOS reached its limits and Intel switched to HiK-MG without a blip in rest of the supply line. They have technologies mapped out to maintain exponential growth for another 8 years, which is about how long they’ve historically had visibility into future production technology.

Eric Baum had some good points about what “understanding” really means. He emphasized the ability to rapidly assemble programs to solve problems and illustrated this point with a comparison between how evolution and engineers design limbs. Evolution has a general representation for a limb.  Mutate this representation and you still have a limb. The difference in design instructions for an arm, a wing, and a flipper aren’t very much. Obviously, this isn’t currently the case for human-engineered prosthetic. I’m not sure I buy his conclusion that this implies we need some sort of hybrid programming tool that combines human-directed design with computer-generated programming. Seemed like a big inferential leap.

Like Rattner, Dharmendra Modha surprised me with some nuts and bolts. Apparently, Almaden Labs already has a simulation of a rat’s brain that runs at 1/10,000th real-time. Assuming the lower bound of the effective complexity of a neuron/synapse (there’s a lot of uncertainty about how much computation goes on here), they say they’ll have the infrastructure to simulate a human brain in real time by 2018. He noted that “software” to run on the brain is an open issue, but I’m still going to have to revise downward my estimation of the time until high quality brain uploading.

Ben Goertzel discussed OpenCog, an open platform for building AGI programs. I was impressed because he clearly understood the failures of past AGI projects and seemed like a smart guy. However, I’m not convinced this path will work, though this framework may accelerate the pace at which researchers narrow down the hard problems.

The only truly bad talk was by Marshall Brain. It’s not a good idea to discuss the economic implications of AI and robotics when you don’t understand anything about economics. He thinks the rise of interactive automatons will cause 50% unemployment. His use of economic statistics was worse than amateurish. He turned the aforementioned glaring error by Vinge into a painful 20 minute presentation.

Cynthia Breazeal cleansed the palate after the bad taste left by Marshall. She demonstrated how she’s working on imbuing computers with emotional intelligence. She showed some reasonably impressive videos of her emotive automaton. This avenue seems mostly like crank turning to me, necessary but not ground breaking because we pretty much understand the cognitive psychology already. However, I was somewhat impressed that they’ve managed to architect their software so the automaton uses the model of its own relationship to the world to model the state of someone else. As a result, the automaton can operate on false-beliefs in others just like a human child.

After lunch, the summit opened with a debate between John Horgan and Ray Kurzweil on whether the singularity would occur in the near future. John said that complexity is too high while Ray said that exponential growth would overcome the complexity.  John was badly overmatched.

Pete Estep told us that knowledge was expanding too fast for meat brains to keep pace so we needed to augment our intelligence. Yeah, yeah.  Preaching to the choir. He claimed that Innerspace was already working on a fully integrated memory prosethtic. Cool if true, but it appears to just be a prize at this point.

The most mind blowing presentation was by Neil Gershenfeld. I already thought the Fab Lab was pretty cool. But the long term stuff he’s working on is breathtaking. There’s a duality between computing and physics. For example, we use physics to build computers that we then use to model physics. The duality is much more fundamental than that (e.g., the equivalence of thermodynamic entropy and Shannon entropy). They have discovered/created a programming paradigm called asynchronous logic automata (ALA: so new there’s not a good reference on the Web; see also Conformal Computing: no good references on that either [edit 04/08/09: this term was coined by James Reynolds and Lenore Mullin in this paper]) that he says is based on fundamental physical properties. They can use ALA to PROGRAM MATTER. Such matter is made of identical cells that assemble themselves like proteins, based on the ALA instructions. He had some animations and it’s unclear from my notes whether these were merely simulations or visualizations of something they’d actually built.  My memory is that they were actual, but at a large scale.  Neil said they should be able to get exponential scaling and they don’t really rely on quantum effects. The bottom line was: 20 years to the Star Trek replicator.  This is the number one thing on my list to keep track of now.

Peter Diamandis talked about the space tourism and the X Prize.  Cool, but not that relevant.

Finally, Ray commented on all the talks. The most important comment was to dispell the notion that technology destroys jobs. He gave the example of gathering all the farmers and manufacturing workers in 1900 and telling them that farming and manufacturing jobs would only be a few percent of the total jobs in 2000. There’s just know way they could imagine all the new jobs like ASIC engineer, Web designer, and network programmer. Technology creates more opportunities than it destroys.  Hallelujah brother.