IBM building Blue Brain (full brain neuron simulator)
seedmagazine.com
seedmagazine.com
As far as I know, they did not go the long and hard way of actually reconstructing a cortical column from sectioned tissue (as in electron microscopical imaging of serial sections of 60 nanometers in thickness).
What they have done, even if remarkable, is "simply" a simulation of a network built from statistical data gathered from many different studies of cortical columns.
I was actually just wondering why this project wasn't happening yet while reading "The Singularity is Near". If you haven't read it, I recommend it. It's hella dry and boring in parts, but overall worth reading if only to make you think about what a startup might look like in only 5 or 10 years.
Kurzweil presents many convincing arguments, but the central tenet is that the doubling of processing power keeps up for another two decades. Due to the nature of nature, what we observe as exponential growth in the end invariably turns out to be logistic growth. If this is so, then at any time the doubling of processing power could slow, and this would mean that we would only, at the most, get one more doubling. Such an event would bring the dream of a near-term universe full of computers to a screeching halt.
I believe that we should be able to create machines that think, but it is important that we aren't blinded by ideology. A massive "singularity movement" that ends up disappointing people, will in the best case delay the advent of such technology due to funding issues. I think the history of AI has demonstrated this quite clearly.
If Kurzweil's estimates are correct, though, another 8-10 years of "exponential growth" in computer hardware will put full-scale simulation of a human brain into supercomputer territory. And frankly, I doubt that evolution has managed to find the most computationally efficient way to achieve intelligence. There's got to be a way to do this.
I wonder how many megawatts Blue Brain will need to do something simple simulating only 1e5 cells.
According to wikipedia,
- the average power consumption of a human cell is around 1e-12 watts.
- 50 years ago ENIAC needed 150 kW to do what you could probably do today in a few µW.
Based on those numbers, I think Markram is way too optimistic about simulating the brain on a signle machine in 10 years. If things continue as they have, it's more likely to take closer to 50 years.
I'll pay more attention to Kurzweil's singularity when the difference in power consumption between a tiny fraction of a simulated brain and a real one drops below ten orders of magnitude.
Of course, neural network modeling can still be of great use. Since we've more or less figured out how the brain does a variety of low-level processing tasks (like coordinating muscle movements, finding the visual depth and texture of observed surfaces, etc.), this knowledge can serve as a starting point for developing exciting technologies at increasingly faster speeds (robotics limbs, computer-guided navigation, etc.). Also, advanced neural networks can serve as rudimentary "existence proofs" for theories of mind and consciousness. For example, a theory that claims that our incredible computing power stems from a recurrent network of thalamo-cortical neurons is strengthened by a computer simulation that exhibits similar large-scale behavior in similar time scales under suitable parameters.
That last sentence is pretty qualified, for good reason. We shouldn't expect to write some code and discover a working brain, at least not until we figure out how it works to begin with.
I think our aim should be to simulate the R-Complex part of the triune brain model. This is where our evolutionary intelligence came from and it is the best basis to start. Birds have managed high levels of intelligence without the neocortex present in mammals, so obviously there's more than one route to sentience.
I believe brain density is the key to sentience. In a small animals, like birds, they lose heat quickly and efficiently, which means their 'processors' can pump out more heat without burning out. Where as a human brain would cook itself if it was 'overclocked' as we're at the upper limit of the communication-efficiency ratio, and Elephants are above that ratio but get away with large processing and data storage with a form of 'low-energy' brains.
With the speed of electricity in synapses, a Crow might be able to achive 10 times as many messages in the time it takes one of our brains to make 1 message. Elephants might only make .5 messages a second per synapse. So a crow would benefit greatly by increasing brain density and be able to solve problems to the same degree as a human infant, however humans have phenominal storage in our brains and a crow simply doesn't have the brain mass to compete so we have more data to act on than a crow. Elephants use their brains to remember pretty much everything, they've even been seen taking short cuts that they have 'guessed' between paths they've taken before.
I think the brain is astounding, but to reproduce sentience we should follow evolution; not only because it gives us vast insights into how our brains evolved, but also because it should avoid us ever getting into the irrational AI situation. If an AI is made from simulating a human brain then it is human, which it would be capable of understanding as it would have to be 'taught' and I know I'm human because of the society I know around me that teaches me that I think and act in a similar method as everyone else, just with different data.
:Well, that should teach you for asking why.
There certainly seems to be a desire for some people to wake up these AI minds, and get them to help us out a bit ;)
Perhaps we should stop saying tin-foil hat and say "ferrous material hat" ;)
PS: This scientist seems to be exponentiating incorrectly. In ten years he'll have 32-ish times as much processing power; he needs 10 million times as much to get to the level of the entire brain.
Most of the time when people ask about "obsolete humans" in context of AI, they really mean "programmers" or "doctors" or whatever. Yes, it might. It simply means that there are professions to be invented that will be to programmers as programmers are to janitors today.
And that's the best-case scenario—when AI is being used for the common good. If the people with early access to it try to use it to rule the world and enslave humanity, they'll probably succeed.
I'm not trying to be an irrational doomsday predictor, here; these are just the conclusions that I come to when I work off of the premise "humanity will have access to cheap human-level intelligences".
Some guy (maybe Kurzweil?) said something along the line that he was always more afraid of stupidity than of intelligence. And I can agree with that a lot. I'm way more afraid that humans are destroying the world out of stupidy than of hyperintelligences which we created & teached trying to get rid of us.
Why do you automatically assume that "no use" in this context would turn out to be a bad thing? The way I see this, if we happen to have greater intelligences working for our common good, we would be able to solve any problem better than a human could - including a possible problem of feeling useless. This would be the best-case scenario, and IMHO it would be much better than the world we have today. Possible solutions to the problem, from my limited human brain, could be bringing the human brains up to the level of the greater intelligence and hence finding new problems to solve, altering human drives so it isn't a problem any longer, abolishing AGI entirely or partially, etc. Any problem could in such a best-case scenario be solved better and faster than humans could.
I agree with you on the worst or worse-case scenario point. There are huge ethical risks and implications in creating very powerful and capable machines. This means that we need to go into this situation with our eyes open, and make sure that we discuss ethics, transparency and consequences from day one.
Enslaving humanity is a basic human drive, and I like to believe that we can do better than that if we try hard. In a pure cost-benefit analysis, it is obvious that it would be best to use such technology to help out everyone.
That's a logical error: the existence of a superhuman intelligence might cause more problems than that superhuman intelligence can solve.
When humans solve the personal problem of "feeling useless" they almost without exception do it outside of a vacuum. Their feeling of usefulness tends to stem from the impact that they have on humanity.
My point is: what reason do you have to believe that the only lives worth living are the ones where people are on the top of the pyramid?
Not so long ago many people said a computer could not beat a grandmaster at chess without being intelligent. Enter Big Blue. Others have stated computers will never compose music that is emotionally meaningful to humans without being intelligent. Enter Experiments in Musical Intelligence and other widely-acclaimed composition programs.
Until the Turing test is passed people will be able to plausibly deny any advances in artificial intelligence. No matter how advanced such "brain in a box" models becomes, they won't pass the Turing test without being embedded in a rich environment with which they can interact.
See http://arts.ucsc.edu/faculty/cope/mp3page.htm for audio and http://arts.ucsc.edu/faculty/cope/experiments.htm for a description of the project.
It seems to me, whether or not to be scared about the progress of AI, depends on a very simple question.
Should we treat AI like our slaves, or should we treat AI like our children?
more likely scenario would probably be, how _we_ would be treated by _them_.
Does this mean coding is in its last days? In future, will we just chat to machines instead of programming them?
But I would love to have some links to the "other interesting ones out there." Would you please elaborate a little more? Any links?