cpollett
New member
- Joined
- Jan 13, 2005
- Messages
- 21
- Location
- San Jose
- Basic Beliefs
- Somewhere between Finitist and Ultrafinitist
I have seen a number of places where people have been polled on when true AI will arrive. For example,
there is some discussion of this in Bostrom's book SuperIntelligence. In one poll cited
there it is estimated that 90% of researcher thought there would be human-level AI (HLAI) by 2075. So I was
wondering what would be a good way to actually measure how far in the future HLAI is?
One way to get to HLAI might be using neural nets -- at its crudest maybe we can get HLAI by just simulating
human minds in computers. An upper bound on this approach would also be an upper bound
on the time to HLAI. In 1998, a state of art system of Lecun, et al for optical character recognition
was 7 layers. Google's system currently is 30 layers. Let's assume that the time to go from useful, computationally
possible topologies and training algorithms for n-layer neural nets to 10n-layer neural nets obeys some kind of
Moore's Law. Roughly, let's say in year 2000, we could do 10 layers and that by year 2040 we'll be able to do 100 layers.
Let's assume we can train recurrent networks with the same effectiveness as feed forward networks. How many
layers would we need to simulate? Well, the number of layers probably grows at less than the number of neurons in any one
direction in the brain. So probably less than the cube root of the number of neurons which is roughly 10^11.
This gives an estimate of at most 10^8 layers. As we need 7 more orders of magnitude than the 2000 level,
this gives a bound of around 2280. My guess is that 10^8 is kinda high and the true value for the depth of minds
network is more like 10^6 or 10^7 but this still gives us estimates of HLAI in about 200 years?
I was wondering what people's thoughts were on my estimates above as well as their own takes on estimating the
time to HLAI?
there is some discussion of this in Bostrom's book SuperIntelligence. In one poll cited
there it is estimated that 90% of researcher thought there would be human-level AI (HLAI) by 2075. So I was
wondering what would be a good way to actually measure how far in the future HLAI is?
One way to get to HLAI might be using neural nets -- at its crudest maybe we can get HLAI by just simulating
human minds in computers. An upper bound on this approach would also be an upper bound
on the time to HLAI. In 1998, a state of art system of Lecun, et al for optical character recognition
was 7 layers. Google's system currently is 30 layers. Let's assume that the time to go from useful, computationally
possible topologies and training algorithms for n-layer neural nets to 10n-layer neural nets obeys some kind of
Moore's Law. Roughly, let's say in year 2000, we could do 10 layers and that by year 2040 we'll be able to do 100 layers.
Let's assume we can train recurrent networks with the same effectiveness as feed forward networks. How many
layers would we need to simulate? Well, the number of layers probably grows at less than the number of neurons in any one
direction in the brain. So probably less than the cube root of the number of neurons which is roughly 10^11.
This gives an estimate of at most 10^8 layers. As we need 7 more orders of magnitude than the 2000 level,
this gives a bound of around 2280. My guess is that 10^8 is kinda high and the true value for the depth of minds
network is more like 10^6 or 10^7 but this still gives us estimates of HLAI in about 200 years?
I was wondering what people's thoughts were on my estimates above as well as their own takes on estimating the
time to HLAI?