3 ms·

I find a lot of the comments on this topic (and anytime AI comes up) very frustrating. Many of them strike me as little more than skeptical posturing: “it’s jus

by emtel 6y ago

I find a lot of the comments on this topic (and anytime AI comes up) very frustrating. Many of them strike me as little more than skeptical posturing: “it’s just statistics!” or “Current AI systems have no understanding and so can’t solve problem X” I don’t think this sort of thinking has any predictive power. If you are skeptical of deep learning because it is “just statistics”, did that skepticism allow you to predict that self driving cars would not be solved in 2020, but protein folding would? I think the best test for skeptics is the one proposed by Eliezer Yudkowsky: what is this least impressive AI achievement you are sure won’t happen in the next two years. If you really want to impress me, give me a list of N achievements that you think have a 50% chance of happening in 2 years. Or 5 years. Just make some falsifiable prediction. If you aren’t willing to do that, or something like it, why not?

WorldMaker 6y ago
That "skeptic test" reads to me like "Goal Post Moving: The Game". Your list of achievements variable N is unbound, your time window is floating (2 years, no wait 5 years now), you leave it to the interlocutor to define "some falsifiable prediction of 50% chance of happening". This test is just statistics and bad statistics at that without well defined judgment criteria. Is this an intentionally sarcastic joke of a meta-commentary on the state of ML today? Because it is almost funny.
- emtel 6y ago
  You miss my point. I'm proposing we fix goalposts, am offering flexibility as to their exact location, and then I'd like to see someone kick a field goal.
godelzilla 6y ago
If you really want to impress me, clearly define "artificial intelligence" (as opposed to statistics, data analysis, etc) and tell me the most impressive example.
hooande 6y ago
The problem with this test is that it's clear that people can commit enough resources to solve almost any one particular problem with "AI". If people are willing to pay tens or hundreds of millions of dollars they can create an AI system that can play any game (chess,jeopardy,go), fold proteins, drive a car, etc. The obvious problem is that no one AI system can do all of those tasks. Every human mind is able to do all of those things and literally infinitely many others. All for a cost of waaaaay below what DeepMind paid to learn to play Atari. If a problem involves the organization of information, "AI" can solve it. I think we've established that. I'm waiting to see something that humans can't already do
- witherk 6y ago
  If you really think that AI can solve almost any one problem, I don't think that you're the skeptic this comment is talking to.

by emtel 6y ago

WorldMaker 6y ago

That "skeptic test" reads to me like "Goal Post Moving: The Game". Your list of achievements variable N is unbound, your time window is floating (2 years, no wait 5 years now), you leave it to the interlocutor to define "some falsifiable prediction of 50% chance of happening". This test is just statistics and bad statistics at that without well defined judgment criteria. Is this an intentionally sarcastic joke of a meta-commentary on the state of ML today? Because it is almost funny.

emtel 6y ago
You miss my point. I'm proposing we fix goalposts, am offering flexibility as to their exact location, and then I'd like to see someone kick a field goal.

godelzilla 6y ago

If you really want to impress me, clearly define "artificial intelligence" (as opposed to statistics, data analysis, etc) and tell me the most impressive example.

hooande 6y ago

The problem with this test is that it's clear that people can commit enough resources to solve almost any one particular problem with "AI". If people are willing to pay tens or hundreds of millions of dollars they can create an AI system that can play any game (chess,jeopardy,go), fold proteins, drive a car, etc. The obvious problem is that no one AI system can do all of those tasks. Every human mind is able to do all of those things and literally infinitely many others. All for a cost of waaaaay below what DeepMind paid to learn to play Atari. If a problem involves the organization of information, "AI" can solve it. I think we've established that. I'm waiting to see something that humans can't already do

witherk 6y ago
If you really think that AI can solve almost any one problem, I don't think that you're the skeptic this comment is talking to.