View Single Post
Old 01-26-2018, 11:21 PM
SamuelA is offline
Join Date: Feb 2017
Posts: 3,903
Originally Posted by Tripler View Post
Alright, SamuelA, this is a hipshot; You've described 'Point "B"' knowing where we're at now. You're talking about the when we get there.

I'm point blank asking you that if we get there, how is it going to happen. We're at Point "A". Your 'Point "B"' is too esoteric and nebulous to argue without the 'how' to get there.

Bridge that gap, brother.
Which how do you care about? You realize that I don't realistically know. There are multiple converging paths. They all lead there. Once we get there the paths we didn't take will probably become feasible.

You know, if during the Manhattan project we had decided to go all in on just one of the 3 main methods (calutrons, centrifuge enrichment, plutonium breeding), we'd still have gotten nukes. Slightly sooner, even. And once we had nukes, going back and exploring the other methods would have been a lot easier to justify. In fact, more recently, we found a fourth method.

Right now the method that to me feels the most valid is we work on lower level systems than machine sentience. We use the shit we've already demoed and adapt it to run robots that do just limited scope tasks. Pick this weed, pick up that can, restock those shelves, pickup that rock, drill that ore vein, install that gear, drive that car.

Each task is something in the physical world that humans are currently doing. It's something where there is a correct answer, every time. It's a task you can break into smaller substeps. Where you can clearly define rules for doing the task "better". (finishing the task without dropping something and faster and without hitting the robot arm against something all make your solution better)

And it's a significant fraction of all jobs on Earth.

Once we get all that working real smooth, we get robots that blow past human ability at doing these defined tasks (they aren't just more physically capable and tireless, I expect them to be smarter. They'll find ways to do these tasks that use less motions and take less time and make less errors than a human would, even without their actuators being better) we can push it further.

Make intelligence systems that use predictive models of physical reality generated from the collective experiences of millions of robots. What I mean is that if you stick any collection of random physical objects that any of the robots in the pool have experience with in front of this new system, it'll be able to predict what will happen if you manipulate them.

It'll know from experience that the red rubber ball will bounce and by how much. That the chip bag will crumple and how. That the gear edges are sharp and can do damage to the robot's own wiring and hydraulic lines.

And then if you ask it to accomplish a task that requires building a rube goldberg machine, and write some additional task solver modules, it'll be able to do it. Not all on it's own, humans wrote the extra software to do it, but humans taking advantage of the existing knowledge and ability the machine pool has.

I think you could iterate that way until you crack things like full machine self replication and you could probably crack nanotechnology the same way.

Even non-sentient agents could predict how some carbon atoms are likely to move along a surface in a vacuum chamber when dragged around by atomic force microscope probes. Advanced agents could plan a sequence of steps to move the atoms to form some assembly. Really advanced agents could design an assembly that accomplishes a goal.

You could eventually bootstrap your way up to agents that design for you whole nanoscale assembly lines and armies of nanoscale robotic waldos, and eventually achieve self replication. (note that this is NOT what we think of as sci fi nanobots. It's these big flat plates that are very fragile and covered with I/O ports. The machinery lives in a vacuum chamber and can never see pressure or even visible light without being destroyed. There's a maze of plumbing supplying various gases to the ports. It sucks a lot of power and there's a huge flow of coolant going in and our. The products are either a fine powder or more flat plates.)

I don't know how to go from this to what we think of as full sentience. I'm not really worried about it, I think what I have described is already way beyond human ability in many areas, and I think you would be able to build various "meta" modules that self-optimize other AIs, analyze human speech, and one day you'd reach a critical mass of complexity and self-improvement loops that gives you the AI we've wanted this entire time.

Last edited by SamuelA; 01-26-2018 at 11:25 PM.