A batch of the best highlights from what Quinn's read, .
Everyone has a firmly held belief that an equally smart and informed person disagrees with.
Rare Skills
collabfund.com
Explore v.s. Exploit: Finding Solutions Quickly Can Get You Stuck in a Local Optimum
Transcript:
Speaker 1
So when I started doing the work in AI, one of the really, very, very general ideas that comes across again and again in computer science is this idea of the explore, exploit trade on. And the idea is that you can't get a system that is simultaneously going to optimize for actually being able to do things effectively. That's the exploit part. And being able to figure out, search through all the possibilities. So let me try to describe it this way. I guess we're a podcast. So you're going to have to imagine this usually I wave my arms around a lot here. So imagine that you have some problem you want to solve or some hypothesis that you want to discover. And you can think about it as if there's a big box full of all the possible hypotheses and all the possible solutions to your problem or possible policies that you could have, for instance, Your reinforcement learning context. And now you're in a particular space in that box. That's what you know now. That's the hypotheses you have now. That's the policies you have now. Now what you want to do is get somewhere else. You want to be able to find a new idea, a new solution. And the question is how do you do that? And the idea is that there are actually two different kinds of strategies you could use. One of them is you could just search for solutions that are very similar to the ones you already have. And you could just make small changes in what you already think to accommodate new evidence or a new problem. And that has the advantage that you're going to be able to find a pretty good solution pretty quickly. But it has a disadvantage. And the disadvantage is that there might be a much better solution that's much further away in that high dimensional space. And any interesting space is going to be too large to just search completely systematically. You're always going to have to choose which kinds of possibilities you want to consider. So it could be that there's a really good solution, but it's much more different from where you currently are. And the trouble is that if you just do something like what's called hill climbing, you just look locally, you're likely to get stuck in what's called a local optimum.
Alison Gopnik on Child Development, Elderhood, Caregiving, and A.I.
COMPLEXITY: Physics of Life
The people with the most accurate models of others tend to have diverse social networks
Summary:
To correct for this handicap, we need to listen to the oppressed in the population.
This includes laborers, students, and others who are usually not given a political voice. By expanding our social networks to include more diverse perspectives, policymakers can make better decisions based on a deeper understanding of societal trends and people's desires.
Transcript:
Speaker 1
But it sounds like this gives us a really clear pointer on how to correct for this handicap. And that we really ought to be like, perhaps when it comes time to make decisions on behalf of everyone, we should really be listening to whomever the oppressed are in that population. We should be really paying attention, for example, to laborers and students and people that are ordinarily not historically, not given a lot of political voice. And what you're saying, yeah, it's in other words, what we need to do is broader our social networks include in our social networks, those people who are typically not there. So if the policymakers who are making these important decisions should know as many different people as possible. And we show in related studies that people who have most diverse social circles are also best able to predict societal trends and to understand how the overall population lives and What people want.
Mirta Galesic on Social Learning & Decision-Making