What if AI supercomputer already has been created? But they cannot continue working on it, until we build cities on mars ? so that incase if it goes full rogue against humanity (skynet style) we can nuke earth as our last chance and live on mars? Because i bet underground cities wont be good enough to run away from an AI that has robot and machines that can even dig underground. Which is why we need Elon to work faster
You're viewing a single comment thread. View all comments, or full comment thread.
Comments (8)
sorted by:
An xbox series x runs at a massive 12 terraflops. It is a supercomputer .the US airforce had 200 ps3 systems in an array to make a supercomputer that was only 2.3 terraflops! I know its nothing its to do with AI ,but just look at the power for under 500 dollars
I remember when a imac computer came out 20 years ago and they made a big deal about how bad it was a consumer could get thier hands on this computer that was powerful enough to control a guided missile.
I think in terms of hardware we are already there. Eggheads have been saying that since the early days of computing [1], but this time I am saying it. The brain has 100 billion neurons, a number that has kept increasing for decades, but now we've hit a brickwall, because more neurons don't fit into the cranium. (23 micrometer length/width/depth for a single neuron if I did not fuck that up).
If we assume that a neuron is a small microprocessor with perhaps 10000 transistors, we arrive at 1e4 * 1e11 transistors for a human brain, i.e. 1e15 (1000 trillions). This is a farily generous budget. Eggheads of yore assumed that a neuron was something like a single transistor. The 6502, seminal processor in the Nintendo NES, had ca. 6000 transistors. The Nvidia Ampere contains 59 billion transistors, approximately 100 billion, ca as many as the brain has neurons. This means I could build something like a brain from 10000 Nvidia Amperes, which would probably be a good engineering decision. 10000 graphics cards are expensive, of course, but the Oakridge Summit contains 27,648 Nvidia Tesla V100 GPUs, so that's NOT science fiction.
These are WORST case assumptions, because neurons are fairly slow, the NES 6502 ran a thousand times slower than a modern graphics card and harddisks are huge and can swap neurons in and out quickly, so you don't need 10000 of those but a fraction, perhaps a 1000 or even less would be enough. Another point is that most of the neurons in the brain are for backup. Demented people or people with serious brain-trauma can function with as little as 20%-30% of neurons. Children who got seriously brain-damaged early in their lives can become A-students and lead productive lives with a large part of their brain removed.
TL;DR: electronics for complete real-time (or faster) brain simulation is absolutely feasible for a mid-sized company or university research lab. (Paying for all the required wattage might be a different matter.)
The remaining challenge is connecting the neurons, but that might not be as difficult as one might think, because the process is largely self-guided and essentially what you do with with your files and folders. Stuff you need more often goes into folders that are more easily accessible and closer to home. Your browser gives you the URLs you need more often with fewer keystrokes by counting usage. Another aspect is datamining, i.e. detecting correlations by applying probability theory. You don't need to understand the correlations datamining gives you for exploiting them, and you don't need to understand why neurons are wired the way they are to arrive at a fully functioning mind.
TL;DR: Nobody can tell what a brain has learned by looking at how the neurons are wired, but google probably can't tell why you are served a particular cat-video next, either. You can leave learning to the simulated brain and go with the result without understanding the details, just as you do with your own.
Pure speculation at this point are consciousness, volition and motives in such an artificial intelligence. Consciousness might be an emergent phenomenon, but volition and motives are not. Those are instincts deliberately evolved for survival. You have to put those into a machine deliberately if you want them. On the other hand: If you build a machine to put profit above everything else and neglect adult supervision, you might end up with a catastrophe, but we have arrived at THAT point long ago with datamining in management and advertising and fairly dumb computers. Even more disturbing are the implications for superior AIs in war. Do you want to win or do you want to remain in full control of the AI and its decisions? Note that your enemy also has superior AIs...
[1] We had a psychology book in school stating that the capacity of the human brain was 2 Megabit [~1980]. I found that lmao because at the time we used the book, a toy computer had 0.5 Megabit RAM and more than 2 Megabit on floppy disks.