Re: Strong AI Takeoff Scenarios

From: Matt Mahoney (
Date: Fri Sep 21 2007 - 19:01:34 MDT

--- CyTG <> wrote:

> Off topic from a ghost :) .. but you guys are frequently talking about
> beeing somewhat far away from the takeoff scenario both in terms of software
> and hardware.. are there any lists/sites that deals with a more precise
> micro detailed evaluation of these specifics? Just how much hardware do we
> think we need in regards to wich framework and what does curret frameworks
> lack? What bottlenecks, hurdles, are we looking at and what sort of R&D are
> going into solving them ?
> If anyone knows!
> Thanks.

We have historically underestimated the difficulty of AI. Take your pick:

- Turing predicted in 1950 that in 50 years a machine with 10^9 bits of
memory, but no faster than current hardware at the time could pass the Turing
test [1].

- Landauer estimated that human long term memory capacity is 10^9 bits [2].

- The Blue Brain project simulated a mouse cortex (1/1000 the size of a human
cortex) at 1 ms resolution in 1/10 real time on the 4096 processor Blue Gene/L
with 1 TB memory [3].

- Google is about as close as we have to AI right now, and they use a few
hundred thousand processors.

- The human brain has about 10^15 synapses. Neurons have an information rate
of about 10 bits per second. Therefore the equivalent computation is 10^16
OPS and 10^15 bits of memory.

Of course simulating one brain is not enough. For computers to surpass the
human race as a whole, you need to simulate about 10^10 brains.


1. Turing, A. M., (1950) Computing Machinery and Intelligence, Mind,

2. Landauer, Tom (1986), How much do people remember? Some estimates of the
quantity of learned information in long term memory, Cognitive Science (10)
pp. 477-493.

3. Frye, James, R. Ananthanarayanan, D. S. Modha (2007), Toward Real-Time,
Mouse-Scale Cortical Simulations, IBM Research Report RJ10404 (A0702-001) Feb.
5, 2007.

-- Matt Mahoney,

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:58 MDT