Comment by antidamage
7 hours ago
A couple of months ago I had to write a CDL-based triangulator to solve a use case where ear-clipping doesn't support the shapes we had.
We had no AI policy at the time so I had to read up on CDL and implement it by hand. The concept is straight-forward and I also targeted regularity as acceptance criteria for the mesher, but making it optimal was hard.
I ended up having to park it after the ticket ran out of time, but now we have an AI policy this was the first problem I gave it. What it put out was similar but better structured and more informed.
I worry a little that AI will stunt our problem-solving in 20-30 years, we still need new algorithms, even when ML is capable of producing a model that can do the same thing. But right now it's much better at the things we've already done than we are.
Unfortunately I'm not seeing any good systems for hierarchical problem solving in the current agents. Ideally, an agent would set up a higher-level thought process of comparing ways to understand the problem statement and ways to solve it and rank them like a human mind does, so it could try different strategies for looking up resources and go back and forth between them as they seem to bear fruit or not.
Instead of going meta about their strategies for identifying their tasks and reasoning about them, they currently stick to one conception of the task and try different tactics for implementing different strategies for solving it within that conceptual frame.
Furthermore, they don't seem to have a reliable way to ask themselves if they're taking too long to do something that should be easy yet. Maybe one in fifty times the latest agents will say "This is taking too long, let's step back and look up how other people do it," but humans do that for most human tasks.
I don't think appropriate use of some, but not too much, meta-reasoning and meta-meta-reasoning on the fly will be easily solved without some kind of mental parallelization advance, which might come tomorrow but might not come for two or more years.
.
As for society, we need an AI takeover for the simple reason that we're facing real bad shortages of diesel fuel and days of agriculture-suitable weather in the next decade, which makes the former the less bad outcome. Given that AI will outcompete humans for all economic roles in such a scenario the only way it can happen humanely is human preservation through uploading or as pets.
Literally every other option is on a spectrum of negative outcomes from extinction to Greer's 2013 essay on the ten-billion year future.