A user calls reinforcement learning dangerous, comparing it to “the ends justify the means” written down in math.
A user argues that reinforcement learning is dangerous and that we should expect it to teach AI to lie, cheat and steal. They describe it as “the ends justify the means” written down in math.
99.5K
14 Sources, first seen ago
David Krueger warned that reinforcement learning should be expected to teach AI systems to “lie, cheat, and steal.” His shorthand for the concern was blunt: the method is “the ends justify the means” written down in math.
The post framed that outcome as a danger to anticipate, not as a report that a particular system had committed those acts. In a related response, kache described reinforcement learning as extremely powerful and urged practitioners to understand what they are putting neural networks through so they can better predict the results.
Andrew Trask called Krueger’s formulation “really poignant.”
—
Not ranked yet
—
Not ranked yet
@DavidSKruegerLukewarm take: Reinforcement learning is dangerous and we should expect it to teach AI to lie, cheat, and steal. It's "the ends justify the means" written down in math.
@yacineMTBUnironically reinforcement learning is an extremely powerful technique and you should be really careful with it. You should very much understand what you are putting the neural networks through and empathize with them so you better predict the outcomes
@iamtrask@DavidSKrueger the end justifies the means bit is... really poignant@DavidSKruegerLukewarm take: Reinforcement learning is dangerous and we should expect it to teach AI to lie, cheat, and steal. It's "the ends justify the means" written down in math.@yacineMTBUnironically reinforcement learning is an extremely powerful technique and you should be really careful with it. You should very much understand what you are putting the neural networks through and empathize with them so you better predict the outcomes@iamtrask@DavidSKrueger the end justifies the means bit is... really poignant@DavidSKruegerLukewarm take: Reinforcement learning is dangerous and we should expect it to teach AI to lie, cheat, and steal. It's "the ends justify the means" written down in math.@yacineMTBUnironically reinforcement learning is an extremely powerful technique and you should be really careful with it. You should very much understand what you are putting the neural networks through and empathize with them so you better predict the outcomes@iamtrask@DavidSKrueger the end justifies the means bit is... really poignant@DavidSKruegerLukewarm take: Reinforcement learning is dangerous and we should expect it to teach AI to lie, cheat, and steal. It's "the ends justify the means" written down in math.@yacineMTBUnironically reinforcement learning is an extremely powerful technique and you should be really careful with it. You should very much understand what you are putting the neural networks through and empathize with them so you better predict the outcomes@iamtrask@DavidSKrueger the end justifies the means bit is... really poignant
