ideas from joe betts-lacroix meant to evoke ideas from you


AIs plotted in Hardware - Software space

Was thinking a few years ago about the prospects of someone developing a strong AI in the near future, and it occurred to me that if they got close by writing sophisticated software, then they could quickly get the rest of the way by throwing money at it: buying more hardware.

This presumed that hardware and software capabilities could trade off against each other. To test this idea, I reached out for a couple of examples, and found one in chess, and another from some friends in factoring algorithms.

I plotted various kinds of tasks in Hardware-Software space, and proposed that the speed of those tasks remained constant along diagonal iso-lines, and was pleased to see how well this bore out. Note that I conveniently leave off units from my horizontal axis, so you'll have to just pretend that there is some way of quantifying this, and for the moment humor me by just accepting that it increases monotonically to the right.

For the factoring task, a 1977 Apple II running a sophisticated 2007-era algorithm had roughly equivalent performance to a 1977 algorithm running on two of IBM & Livermore's ground-vibrating behemoths known as BlueGene/Ls, the fastest computer in the world in 2007.

Similarly story with two chess-playing computers, Deep Blue and Deep Fritz, both of which had about equivalent performance to a human grandmaster.

What I conclude for the case of the strong AI project is that the developers will face a choice once they achieve an AGI of sufficient power, say, powerful enough to reach the "Turing Test" iso-line. I assume they will do so using some kind of large cluster running some exquisitely brilliant software.

They could then go the direction I call "democratize", and refine the software so that people all over the world could instantiate these AGIs and start solving all kinds of hard problems, advance AI research, keep people company, etc.

Or...they could take an alternate path that I call "Push to Seed", on which they take in massive funding, and expand their hardware enough that the intelligence increases its power to the Seed AI level. This is the level at which the AI would be capable of redesigning itself for increased performance, which could then design for even more performance, etc. Whoever possesses this AI will have the worlds most powerful tool -- will they use it for good or ill? That probably depends on who provides that massive funding.

Food for thought...

No comments:

Post a Comment