Read on Twitter
Can we please, please not do this?
Like, the days when I worked on AI research are long past. But this is still nauseatingly familiar to me.
Like, the days when I worked on AI research are long past. But this is still nauseatingly familiar to me.
https://twitter.com/ArmandDoma/status/1664331870564147200
It's practically a law of heuristic or evolving algorithms that the edges of the test harness are weaker than the problem you're trying to solve, so the AI seeks them out instead. And then smoothing out all the edge cases and glitches turns out to be harder than the problem!
This isn't like "weird unanticipated behavior", it is *absolutely standard*. Even incredibly early genetic algorithm style tests have it -- think of the GA that's designing creatures, looking for new forms of locomotion, but instead it finds *every bug in the simulator.*
Back in the old days, we were using a constraint-solving AI to plan an academic conference. Every panel had a list of requirements: microphones, projector, a certain number of seats, length, etc. There were too many to fully satisfy all of them, so the goal was to optimize.
The planner happily reported back that it had an excellent solution! Every panel had every requirement satisfied except one.
When we looked at the plan, all the panels were in the largest room, with all the gear. They were all one minute long, one after another.
When we looked at the plan, all the panels were in the largest room, with all the gear. They were all one minute long, one after another.
This led to the (re)discovery of a very common problem in this space, which is the *expressing the true constraints of the problem is equivalent to or harder than solving the problem.*
The constraint language we'd built to tell the AI what it was trying to do wasn't expressive enough. It couldn't say, "Okay, cutting a 60 minute panel to 40 minutes might be plausible, but cutting it to 1 minute is completely useless."
The more realistic your simulation is -- the more options the AI has -- the worse this problem becomes.
In the drone scenario, there are tons of options that would literally never occur to a human pilot, because they're carrying a TON of implicit knowledge!
In the drone scenario, there are tons of options that would literally never occur to a human pilot, because they're carrying a TON of implicit knowledge!
But the drone isn't, so you have to put it in the constraint language. And its one of those situations that I think every programmer has encountered, where what you thought was like a minor little task at the end turns out to be an endless sucking void.
You write the simulator and you think your done. It works, look at the little drone flying around. High five!
And you don't realize that about 99% of the work is still ahead. And it may turn out to be actually impossible.
And you don't realize that about 99% of the work is still ahead. And it may turn out to be actually impossible.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
