HN Explorer

Ask HN: What could we do today with a 1,000,000X increase in computing power?

7 points by netcan ↗ HN
I was recently discussing Ray Kurzweil with friends. Many of his arguments extend out Moore's Law, speculating about what these future computers will power. I argued that Moore's law alone doesn't get us "intelligence" without the discovery of algorithms (or other mechanisms) that power intelligence in organisms. If we could mimic fungal intelligence with today's resources, then we would just need more resources to make a super human (or super fungal) like brain. But since we can't, Moore's Law isn't predictive of anything and all Ray's timelines are bunk.

I think Kurzweil is working backwards. AI is coming. How do I predict it using current trends? I think we should go the other way: Here are the trends. What do they predict?

So here is the question: What are the things we could do with a 1,000,000X increase in computing power that don't depend on unrelated future discoveries. What programs are just waiting for more power? Are our current evolutionary algorithms or some other AI approaches constrained in this way? Any less AI-ish programs constrained this way? What could they do if they weren't?

13 comments

[ 3.5 ms ] story [ 44.9 ms ] thread
netcan
For an example of what I mean about Kurzweil's timeline:

2018 10^13 bits (=10 TB) of computer memory—roughly the equivalent of the memory space in a single human brain—will cost $1000.

2023 10^16 calculations per second—roughly the equivalent of one human brain—will cost $1,000.

http://en.wikipedia.org/wiki/Predictions_made_by_Ray_Kurzwei...

artemio
IMHO play games (matrix style)
thret
We might be able to model certain experiments with greater accuracy than real life (because there would be no noise).
feld
Break crypto.
debacle
Our AI was designed by millions of years of trial and error. Any AI of similar capacity will be useless unless similarly trained. You can't go from zero to human straight away, you'll just wind up with nebulous goo.

But on the bright side, in a few (20-30) years we'll be able to copy a human brain dendrite for dendrite. Once a few different technologies converge, we'll be able to do much more with AI than ever before.

api
A few ideas:

Evolve nontrivial software with genetic programming.

Voxel worlds at retina resolution.

Fold proteins and simulate entire cell bio chemistries in silico.

Find bugs in complex apps by exploring large chunks of their combinatorial space. (Think fuzzing at scale.)

Lots and lots of other very scientifically interesting huge combinatorial searches: superconducting materials, novel theories of physics (mathematical combinatorial space), etc.

netcan
How resource constrained is simulated protein folding? Do you have any idea what we could do if it wasn't?
api
I know it's immensely compute-intensive and takes on the order of days to weeks to compute a single protein fold.

If we could do that in seconds, we could do things like take a genome and "compile" a cell and then simulate it to some degree of accuracy. That might finally unlock the latent potential of genetics.

netcan
Really? With enough computing power we could compile a genome? If this is for real, this is definitely in the world changing things that Moore's Law will produce category.

Is computing power really the main thing missing?

3beard
For many computationally hard problems six extra orders of magnitude isn't really going to buy you much.
morphle
What will I do with massive parallel computing? Better simulations and writing non-optimized code in 1/10,000 lines of code (see Steps project at Alan Kay's VieuwPoints Research Institute http://www.vpri.org/pdf/tr2007008_steps.pdf and the steps papers at http://www.vpri.org/html/writings.php ). See the lectures on the top of our web page http://morphle.com for some demo's by Alan Kay)

Since 2008 we at MetaMorph research institute are seriously attempting to build a single Wafer Scale Integration (WSI) with up to 800,000 cores (at 14nm) called SiliconSqueak that is a bytecode machine (like the B5500) running Squeak. In 2015 we sell 10, 60 and 300 core prototypes on an FPGA. The final WSI will cost around $6000 and a 5000 core version (at 90nm) would cost around $500. We will need around 1500 people to buy it to become profitable. You could cut it up into chips costing around $1 each.

The main effort is writing software for it. For example a Lisp bytecode machine, a port of the RoarVM and a version of OpenCroquet needs to be enhanced for it. However, running Linux on it would be pointless, it needs a new type of operating system based on its hardware based message passing.

You can also use it as an openflow switch with 100 50 Gbps optical links or as a RamCloud with half a terabyte of 3D memory. It is done in our startup that still needs two extra co-founders. a scientific paper will be released by our research group in a few months on the WSI.

MalcolmDiggs
From my understanding of Kurzweil's arguments: Raw increases in computing power when applied to FMRI (and other brain-imaging systems) will allow greater resolution of the brain and allow us to reconstruct it (once we can see exactly what every neuron is doing at once, we, in theory, should be able to reverse engineer the system). I don't know how true this is, but I believe he's arguing that that is the technological progression worth watching. Not any particular new discoveries just higher and higher resolution imaging.