Stanley: Direct pursuit of human-level AI is a naive objective
Ken Stanley · Kenneth Stanley: Set The Right Objectives · Oct 4, 2022 · at 39:29
AI researcher Ken Stanley discusses the realistic limits of goal-directed AI research with host Shane Parrish.
0:00 / 0:17exact quote · 17.1s
720p mp4 · rendered on demand · StarZero watermark
“Like, at some level, there's, like, this really grandiose conception of, like, some human-like computer, and, like, that is, I think, a naive objective right now. Like, you just don't want to know how to do that. We don't know what the stepping stones are that lead to that, although they're getting closer, perhaps, but they're still not close, I would say.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Ken Stanley
Stanley: Dismissing experts' subjective sense of 'interesting' cripples innovation
“It's actually the subjective judgments are the interesting ones because, like, the objective judgments are easy. You don't need a degree to just measure something. You know, some kid takes some test, you get a score. You can average them across everybody at th…”
Stanley: AI research proves objective-driven achievement is deeply flawed
“And in the course of doing that research, we just started to see undeniable evidence that that approach to achievement has some serious flaws.”
Stanley: Major breakthroughs require positioning for serendipity rather than setting objectives
“And actually like a lot of what we do that facilitates making these kinds of important discoveries is to actually set us up, set ourselves up for having effective serendipity, which is not the way that we, you know, talk about things when you talk about settin…”
Stanley: Complex problems like education cannot be solved by direct metric optimization
“We cannot make progress in certain kinds of extremely complex problems, which education is one of those. Simply by just laying out some assessment system and then trying to follow it towards this global objective, which is just incredibly complex to get to.”
Stanley: Metrics are useless for hard problems because required stepping stones are counterintuitive
“If the actual stepping stones that lead to where we want to go are counterintuitive, in other words, they're not what you would expect, then the metrics are useless, right? Because they won't detect those stepping stones because they don't look like what the m…”
Stanley: Evolution solved all problems in one run, unlike typical ML
“Evolution, it's a very unique thing in the sense that it's kind of like a search or like a learning algorithm that discovered everything that was ever created in nature in a single run. This is very different from like what you see in typical machine learning,…”