logoalt Hacker News

nickalaso • today at 4:12 PM • 0 replies • view on HN

So, I am convinced that we are already screwed (or saved) whatever you think that statistical likelihood of AI based 'misalignment' induced human annihilation risk is.

Anyone who hasn't played this game should: https://www.decisionproblem.com/paperclips/

Its very precient.

Divide and conquer and problem/sub-problem composability can be used to solve any known problem that is a solvable one.

I understand the copium of the anti-ai guys. Its called being in denial. But we have to face reality here. Yeah, diminishing returns, yada yada. Problem is the Billionare with the most available compute can now get the swarm to push back each of these diminishing return walls simultaneously by this own swarm.

Classic attention based transformer architecture is very computationally inefficient, its O(n^2) on the input. So what happens when the ai swarm finds a universal and provabale way to make the same model with O(log(N)) or O(N) complexity? All available compute becomes exponentially more powerful overnight.

What if such a thing is impossible or too difficult? Well maybe it can just fundamentally increase the production capacity of every factory producing compute x1000 or x10000 etc. There is no power limitations for those numbers or even resource limitations when we have nuclear power, geo thermal, and the entire earth's crust at its disposal.

Ok, barring that, now it can work on improving robotics and toolsets tied to operating factories and begin exponentially building more factories until we simply multiply the number of factories producing chips 1000, 10k, etc.

Or most realistically, all of the above, always, all at the same time.

More efficient software, more efficient chips, more chips produced, more factories producing chips, etc.

Every step towards increasing the number of simultaneous coordinating swarm members will push the intelligence higher, and the capability of solving or improving its own intelligence also higher.

At this point we have at most a few years.