On Fri, Jul 14, 2006 at 02:50:23PM -0400, Eric Baum wrote: > > Eugen> Groan. The whole network computes. The synapse is just an > Eugen> element. Also: you're missing on connectivity, > Eugen> reconfigurability, synapse type and strength issues. > > I'll definitely grant you reconfigurability. Might be fairer > to compare to a programmable array.
Not really. You can have divergent factors of 10^4, and convergent factors of some 10^5. You can't do this with a 2 1/2 dimension substrate with severe fanout issues. Actually a FPGA with on-die memory, with a signalling fabric mesh is probably entry-level hardware for AI, several thousands of them. > Well, on this we differ. I can appreciate how you might think memory > bandwidth was important for some tasks, although I don't, but > I'm curious why you think its important for planning problems like An AGI is a general intelligence. Your hardware has to have enough performance to execute a general AI core. There are some 10^11 cells in the CNS, each has a connectivity ranging into high 10^3, and each site operates in 10^3 Hz range. Assuming 64 bit words/site, that's 10^17 words/s. Best case (worst case is 10^1..10^2 worse) of today's memory is some 10^9 words/s. I think you would agree that a missing factor of 10^8 is not negligible. And of course since sequential memory access and CPU speed (which is not Moore, btw) delta is an exponential function, too, you'll see we're running into problems. Nevermind that the strictly sequential buck stops well before THz (10^12 Hz) rate, and you have to go parallel (about 10^6 cores parallel, if my math is accurate, which it probably isn't). > Sokoban or Go, or a new planning game I present your AI on the fly, > or whether you think whatever your big memory intensive > approach is will solve those. The world is complex. You need a lot of bits to represent that state, and even more bits for making forecasts, and some little bit more for system housekeeping. It only appears "memory intensive" if you're unfamiliar with the problem set of an AGI. That problem domain is many orders of magnitude remote from such trivial toys like a chess program to beat human grandmasters. > As you know, I argued that the problem of designing the relevant software > is NP-hard at least, so it is not clear that it can be cracked without > employing massive computation in its design, anymore than a team of > experts could solve a large TSP problem by hand. I agree that human experts can't produce an AGI. They might not be able to even produce a seed for an AGI. I think they are (barely) capable of building the boundary conditions for emergence of an AGI, if given enough hardware resources (a mole of bits). > However, I have an open mind on this, which I regard as the critical > issue for AGI. > > > Mark> VERY few Xeon transistors are used per clock tick. Many, many, > Mark> MANY more brain synapses are firing at a time. > > How many Xeon transistors per clock tick? Any idea? The Xeon alone is useless. Out of some 10 billion transistors in a current desktop PC most of them are idle, and are DRAM cells. > I recall estimating .001 of neurons were firing at any given time > (although I no longer recall how I reached that rough guesstimate.) > And remember, the Xeon has a big speed factor. What is a "speed factor", kemo sabe? -- Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org ______________________________________________________________ ICBM: 48.07100, 11.36820 http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
signature.asc
Description: Digital signature