Wiki Contributions


Also see this recent podcast interview with habryka (incl. my transcript of it), which echoes some of what's written here. Unsurprisingly so, when the slack messages were from Jan 26th and the podcast was from <= Feb 5th.

See e.g. this section about the Rationality/AI Alignment/EA ecosystem.

And I suppose a counter-example to the outlier essay would be a fast food chain hiring tons of people for some seasonal increase in demand: here the idea is that workers are presumed to be replacable.

Now selection is much less useful due to the light-tailed skill distribution and high rates of churn. Instead, the organisation is designed such that work is split into simple and clearly defined roles (structural method) which anyone can be easily trained in (corrective method).

Yes, that seems right. In the Explore / Exploit tradeoff, boredom ensures we don't neglect exploration. Whereas engaging media like feeds or video games, and other superstimuli, satisfy boredom in such a way that we stay in exploitation mode.

A thing that illustrates this well, I think, is watching small kids interact with smartphones. The haptic feedback, swipe interfaces, videos etc. are incredibly engaging. So left to their own devices, kids can spend a long time on them without getting bored.

the capacity to be bored has always felt a little NPC to me.

Counterpoint: Boredom is good, and our modern amenities optimize it away at our peril.

Let's apply this framework to the Searching for Outliers essay. I suppose the lesson there is that, for sufficiently heavy-tailed outcomes (where, say, some people are a 1000x better fit than the average), selective methods (i.e. searching for outliers) dominate over corrective methods (e.g. better training). And then it talks about how to do selection well.

Omega's decision at T2 (I don't understand why you try to distinguish between T1 and T2; T1 seems irrelevant) is based on its prediction of your decision algorithm in Newcomb problems (including on what it predicts you'll do at T3). It presents you with two boxes. And if it expects you to two-box at T3, then its box B is empty. What is timing supposed to change about this?

Omega is a nigh-perfect predictor: "Omega has put a million dollars in box B iff Omega has predicted that you will take only box B."

So if you follow the kind of decision algorithm that would make you two-box, box B will be empty.

How do concepts like backwards causality make any difference here?

Weak down-vote: I feel like if one takes this position to its logical extreme, they could claim that any arbitrary AI misbehavior is not misaligned, almost by definition: you just don't know the true held values of its creators, according to which this behavior is perfectly aligned.

Laws that make customers more informed about the deals to which they agree on help with encouraging the innovation we want and improve competition.

I like the spirit of this, but want to mention GDPR as a counterpoint. One of the purposes of that law was to ensure that customers are better informed, and yet the mandatory cookie popups just made the experience of browsing websites much worse.

Load More