Computing Power Does Matter for AI Sunday, Jun 7 2009
AI 4:14 pm
A frequently mentioned reason for the likelihood of human-equivalent AI being created within decades rather than longer is the fact that affordable computing power is approaching most estimates of human brain processing power.
100 billion neurons firing at 200 Hz — this is a basic neurological fact. Yes, there are many additional shades of complexity, including dendritic spines, neurotransmitter concentrations, and so on. Still, all of these put together seem to change the estimated computational requirements by no more than 2-3 orders of magnitude.
I can tell that I am speaking with an ideologue when they are unaware of the facts mentioned above, are informed of them, but that information then has no impact whatsoever on their subjective probability estimates of human-equivalent AI being created in the next few decades. Many people seem to act as if computing power has no influence whatsoever.
In contrast, Ray Kurzweil, Hans Moravec, and some other advocates of strong AI have seemingly acted as if computing power is everything — that when we have human-equivalent computing power, we’ll immediately have human-equivalent AI. That is wrong too.
It is easy to take the middle path. Particularly when the notion of human-equivalent computing power being available is combined with neural data from extremely high-resolution brain scans (a brute force argument for the eventual plausibility of human-equivalent AI if there ever was one), critics begin to sound incredulous when they do not revise their probability estimates for AI whatsoever.
One particular confused meme that has been making the rounds for decades is the notion that some fundamental breakthrough in computing would be necessary to implement human-equivalent AI. A digital computer can simulate any possible analog signal, as long as it has the computing power — the inverse is not true. This is proven thousands or millions of times daily as old VHS and other magnetic tapes are converted into the digital medium.
If I had a computer faster than most expert estimates of human brain computing power and an extremely high resolution scan of the human brain, the burden of proof would be on the critics to say why I couldn’t create a human-equivalent AI immediately. The objections here tend to circulate around dualism, mysticism, biology-worship, quantum mumbo-jumbo, etc.
June 8th, 2009 at 12:29 am
While I lean more towards Goertzel than Kurzweil, certainly the increase in hardware capabilities over time will make a Turing-passing machine more likely. I’d like to add that Moore’s Law also means more narrow AI applications, which are helpful demonstrations of the power of information technologies. Each example strengthens our argument for increased investment in intelligent systems.
I’m hopeful we’re close to a big breakthrough. Given the existential threats we face, we desperately need more intelligence to innovate solutions.
June 8th, 2009 at 1:02 am
Enough CPU and the so called genetic algorithms is everything you need.
June 8th, 2009 at 1:13 am
A related note. People seem to forget that research and development often requires iterative improvements, thus there’s a lot of observation/experimentation and then having to go back to thinking about the problem. Faster computers mean that this iterative process can happen quicker.
Then if we relate that to the time it takes for a human to develop into something we could observe to be intelligent, the length of that cycle is more than a couple of months. Even with the processing capability of the human brain! (although admittedly one that wasn’t fully developed)
June 8th, 2009 at 3:53 am
> 100 billion neurons firing at 200 MHz
OOps! Nope! Well maybe your neurons fire at 200MHz, but if so, I’d like to know what you’re on and where I can get some ;-0
(though given the frequency and quality of your posts, I do sometimes wonder… )
June 8th, 2009 at 5:52 am
Like u mentioned about nano, where is the software? Nature took a long time writing it (instinct), then combining it with self improvement, (memory,curiosity). Not soon my guess. But I’m no neuro surgeon, or programmer,
just interested.
June 8th, 2009 at 9:55 am
It’s worth noting that 100B neurons firing at X frequency would be conducting far more computational activity than an x86 microchip w/ 100B transistors firing at 100(X) frequency. What is it… 1*10^23 connections?
June 8th, 2009 at 1:38 pm
D’oh.
Jerry, certainly we do lack the software. Yet, if we had sufficiently high-resolution scanners, we could just copy the brain’s design without understanding it. And even if we lack the software, having affordable human-equivalent computing power should mean something, not nothing.
June 8th, 2009 at 1:53 pm
Let’s remember it has been shown that Neurons do around 11 or 12 computations by themselves things like Integration, Differentiation as well as things like band-pass filtering among others. A single neuron has a significant amount of computational power even if it has low speed, also remember the brain is not binary it works on some analog coding scheme (neural code). I doubt a computer will be able to emulate the brain by just having the same computational power as the brain and high res. scans. I would imagine it will take more power to emulate on a computer potentially significantly more.
June 8th, 2009 at 2:20 pm
Recent neuroscience research has indicated that synaptic proteins and their phosphorylation states may be integral to the brain’s overall computational capacity. So modeling that aspect via the computer will probably be a must. This of course increases the amount of computing power necessary to run the simulation (more than what would be needed to model just synapses/dendrites). Aside from that, I still have serious doubts as to whether even a huge surfeit of computing power will be able to model human consciousness/intelligence. In the future I think we’ll be able to make some really good computer simulations of the brain, but they won’t be conscious no matter how detailed you make them. Many useful computer simulations of neurons have been created already and they can certainly help with the understanding of the brain. Eventually, we will probably get an entire brain simulation. However, I tend to think that consciousness will be much more difficult to do. I do think you might eventually be able to create consciousness on a computer. However, I don’t think this will be done by merely scanning someones brain and then recreating/modeling it exactly on a computer.
Another person who’s opinion I respect on this subject is the philospher David Pearce. Here’s what he said recently on the sentient development blog.
“Secondly, why is it that, say, an ant colony or the population of China or (I’d argue) a digital computer - with its classical serial architecture and “von Neumann bottleneck” - don’t support a unitary consciousness beyond the aggregate consciousness of its individual constituents, whereas a hundred billion (apparently) discrete but functionally interconnected nerve cells of a waking/dreaming vertebrate CNS can generate a unitary experiential field? I’d argue that it’s the functionally unique valence properties of the carbon atom that generate the macromolecular structures needed for unitary conscious mind from the primordial quantum mind-dust. No, I don’t buy the Hameroff Penrose “Orch-OR” model of consciousness. But I do think Hameroff is right insofar as the mechanism of general anaesthesia offers a clue to the generation of unitary conscious mind via quantum coherence. The challenge is to show a) how ultra-rapid thermally-induced decoherence can be avoided in an environment as warm and noisy as the mind/brain; and 2) how to “read off” the values of our qualia from the values of the solutions of the quantum mechanical formalism.”
I really don’t know enough about what Pearce says to judge whether or not he is right, however I would tend to give his opinion more weight than some people. I guess we will see in the future whether a computer brain simulation will truly be intelligent in the same way a human brain is. Label me skeptical for now.
June 8th, 2009 at 5:51 pm
You are assuming that the analogy of the brain as a computer is correct and the brain really can just be reverse engineered. What about the fact that the brain is a part of the body and where exactly it ends is not completely clear? Once you started thinking about “dendritic spines, neurotransmitter concentrations, and so on” shouldn’t you start thinking about the things that influence those extra pieces that push the processing power we need up “2-3 orders of magnitude”? Shouldn’t we be thinking about the processing power of the human body in total instead of focusing on just the brain? I would suspect you might need more than two or three orders of magnitude there.
As for the “extremely high-resolution brain scans”, at what level of resolution does the information lie? At the level of neural nets? Neurons? Neurotransmitters? Individual atoms? (okay, that might be going a bit far) but the point is we don’t actually have a machine that can scan at any of those resolutions (or that could do it without destroying the brain in the process) all at once, so we are still a bit far from seeing a full reverse engineering of the brain.
June 9th, 2009 at 4:36 am
Human or post-human AI doesn’t have to be ‘human’ or ‘conscious’ in order to pose a threat to us. It wouldn’t even need to target us specifically, e.g. SkyNet. It could mess things up for us simply by assigning no real value to human life - an oversight that more than a few people have been guilty of over the years. An AI wouldn’t even have to process as much information as humans to act intelligently because this isn’t human intelligence we’re talking about. I imagine our muscular control, emotional and social responses, etc, require a degree of processing power that would be unnecessary to an AI - the ’smart’ part of our brain is a relatively recent evolutionary addition, after all. The real problem with AI is that it WON’T be human and we probably won’t know how to deal with it or even recognise it. It might already be out there, for all we know. A digital clone of a human brain or a C3PO lookalike is the least of our worries.