A post argues the danger isn't AI defying its programming; it's that learned 'programs' remain poorly understood.
The post argues that AI systems learn from data rather than being programmed directly, and that people don't understand the resulting 'programs.' It calls that paradigm dangerously flawed, saying the concern is not AI mysteriously acting against its programming.
10.9K
2 Sources, first seen ago
—
Not ranked yet
—
Not ranked yet
@DavidSKruegerThe problem is not that AI is magically going to do something that goes against its "programming." The problem is that we're not "programming" AI at all; AI systems "learn" from data and we don't understand the resulting AI "programs." The whole paradigm is dangerously flawed.
@irinarishShould make the whole ML paradigm "illegal", not just RSI :) wait, let's just outlaw the root cause if it all, the dangerous field if statistics, the cradle of all that ML and AI (pardon, SI :). Because who knows what its out-of-distribution behavior will lead to... Right? :)@DavidSKruegerThe problem is not that AI is magically going to do something that goes against its "programming." The problem is that we're not "programming" AI at all; AI systems "learn" from data and we don't understand the resulting AI "programs." The whole paradigm is dangerously flawed.@irinarishShould make the whole ML paradigm "illegal", not just RSI :) wait, let's just outlaw the root cause if it all, the dangerous field if statistics, the cradle of all that ML and AI (pardon, SI :). Because who knows what its out-of-distribution behavior will lead to... Right? :)@DavidSKruegerThe problem is not that AI is magically going to do something that goes against its "programming." The problem is that we're not "programming" AI at all; AI systems "learn" from data and we don't understand the resulting AI "programs." The whole paradigm is dangerously flawed.@irinarishShould make the whole ML paradigm "illegal", not just RSI :) wait, let's just outlaw the root cause if it all, the dangerous field if statistics, the cradle of all that ML and AI (pardon, SI :). Because who knows what its out-of-distribution behavior will lead to... Right? :)@DavidSKruegerThe problem is not that AI is magically going to do something that goes against its "programming." The problem is that we're not "programming" AI at all; AI systems "learn" from data and we don't understand the resulting AI "programs." The whole paradigm is dangerously flawed.@irinarishShould make the whole ML paradigm "illegal", not just RSI :) wait, let's just outlaw the root cause if it all, the dangerous field if statistics, the cradle of all that ML and AI (pardon, SI :). Because who knows what its out-of-distribution behavior will lead to... Right? :)
