Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!wuarchive!rex!ukma!seismo!dimacs.rutgers.edu!aramis.rutgers.edu!athos.rutgers.edu!nanotech From: Hanson@charon.arc.nasa.gov (Robin Hanson) Newsgroups: sci.nanotech Subject: Re: Down and Out in Nanoland Message-ID: Date: 12 Jan 91 05:09:52 GMT Sender: nanotech@athos.rutgers.edu Lines: 82 Approved: nanotech@aramis.rutgers.edu In JoSH writes: > It'll take at least a MIPS to simulate a synapse with any claim > to fidelity, so we need 10 trillion MIPS (that's 10 million tera-ops > or 10^19 instructions per second) to run the simulation. This can > be compared with Moravec's estimate of 10^13 IPS (10 million MIPS) > for human equivalence AI style (i.e. not simulating neuron by neuron). > Now AI seems to be moving slowly if you're sitting behind it in > traffic, but my own 15 year's association with the field convinces > me that it's moving fast enough to keep up with the machines it > has to run on. With the rules of thumb 1990=10 MIPS and a decade > gives 1000x computing power, we get full AI in 2010 but don't have > a machine you can upload into until 2030. (Nanotech doesn't change > the rules of thumb, it simply helps them stay on the curve after > electronics give out.) --JoSH] Thanks for the overview. In "Mind Children" Moravec uses the estimate of *two* decades to give 1000x computing power, and hence estimates the hardware to support human-level AI will be available in 2030, not 2010. And I presume everyone understands that all these MIPS estimates are *very* crude! Of course having sufficient hardware for a human level AI does *not* mean we will know how to write programs to take full advantage of that hardware. The software is the hard part. The idea that we will know how to write human-level AI software as soon as the hardware is available seems quite suspicious to me. For example, I think it is highly unlikely that we now know how to write the most intelligent agent consistent with today's hardware. It may very well be that present hardware is sufficient to support a human-level "teletype" AI (i.e. without full vision, sight, and tactile abilities). Regarding our relative progress in hardware and software, the Nanotech transition *does* make a difference. Hardware capabilities may go through a tremendous increase in a relatively short time, while software would not. Thus I estimate that while, as you say, AI systems of comparable ability would take much less computing power, we will not know how to program them when the hardware sufficient for uploading arrives. Robin P.S. This is a fun discussion! [It's also one of the higher quality discussions we've had here, thank you for starting it! I think that AI is a lot simpler than most people believe, in one sense, and harder in another. To invoke HPM again, most of the machinery of the brain is involved in recognizing objects, not bumping into trees, etc. My guess is that when we have good algorithms for those things, the re-application of the algorithms to the higher-level thought processes that give us so much trouble now will be fairly simple. That's because (I claim) that's how evolution did it in the first place, i.e., copy and modify existing functionality. That's the "easy" part. The hard part is getting to the two-year-old chimp stage. Now what Moravec et al. discovered, in obstacle avoidance anyway, is that there is a level of brute force at which, still using plenty of ingenuity to be sure, things suddenly seem to work a lot better even though they're running simpler algorithms than before (in some conceptual sense). A big 2D array is an easier-to-manage map than a dynamically balanced quadtree. Indeed, you can do some mathematically more sophisticated things with your array and still have a simpler program overall. Well, lo and behold, it turns out the big array implementation is a lot less "brittle" than the old sophisticated ones; but it takes more horsepower to run. So more MIPS not only lets you run your AI program faster; it makes it easier to write. More than half the hair of AI programs is in the pursuit of efficiency, I opine: not only does it soak up effort but it makes the other half harder to write. Try to explain the Rete algorithm in 25 words or less. In an appropriate associative processor the equivalent is "do a pattern-match between expression A and every expression in set B." Highly inefficient, says the algorithmicist. But the heavy hardware version allows for dynamic rulebases, not just adding and removing rules but on-the-fly subsetting and rulebase merging. Again brute force has made the algorithm simpler but also produced an unexpected bonus in functionality and/or robustness. I will give you 5-to-1 odds there is a tera-ops machine by 2000; 100-to-1 odds there's one by 2010--exact wording to be worked out if you're interested. --JoSH]