I really don’t see how stuff like this isn’t an op

[legend]

23 minutes ago, Signifyin(g)Monkey said:

Question for AI gurus: If we view society as an old, outdated neural net of massive size, Isn’t the climate activist claim that ‘people are taking too long to recognize the threat of climate change?’ essentially a sort of recapitulation of the vanishing gradient problem?

 

That's an interesting analogy! Like any analogy, I don't think it holds up the more you scrutinize it, but I can see why you might say that. I think the parts where it doesn't quite work is it's unclear what the topology of society w.r.t. that loss signal and whether it is sequential (where it's the sequential  squashing operations that can induce vanishing gradients). There is some degree of sequential topology, but society is a bit more wide than deep, and the lack of action happens even on the highest level, not just the lowest.

 

If I were to make an AI related analogy, I might go for saying it's a value function approximation error. In AI (RL), the value function estimates the cumulative future expected reward. In practice, this function has to be approximated (often with a neural net), but for various reasons it can be quite hard to approximate this stably and accurately, and small errors in the value function estimate can result in catastrophic failure of policy improvement. So in the case of climate change, people are too stupid to sufficiently estimate the value function