LessWrong 2.0 Reader

You can go ahead and post.

I did a check and am now more confident in my answer, and I'm not going to try to come up with an entry that uses fewer soldiers.

If you can get to Seattle for your partner's career, you can likely get a job nannying during the day, which will pay $25 to $30 an hour and doesn't require a car. 

This time last summer I was an incoming intern in Seattle and I was unable to pay less than $30 an hour for childcare during working hours, hiring by combing through Facebook groups for nannys and sending many messages. At this price, one of the nannys we worked with had a car and the other did not. I do not know what the childcare market is like near your current location.

Although GDP per capita is distinct from this expanded welfare metric, the correlation between GDP per capita and this expanded welfare metric is very strong at 0.96, though there is substantial variation across countries, and welfare is more dispersed (standard deviation of 1.51 in logs) than is income (standard deviation of 1.27 in logs).^[9]

I checked the paper and it looks like they're comparing welfare by "how much more would person X from the US have to consume to move to another country i?" Which results in equations like this:

which says what the factor ${\lambda }_i^{simple}$ ,  should be in terms of differences in life expectancy, consumption, lessure and inequality. So I suppose it isn't suprising that it's quite correlated with GDP, given the individual correlations at play here, but I am suprised that it is so strongly correlated since I'd expect e.g. life expectancy vs gdp to correlate at maybe 0.8^[1]. Which is a fair bit weaker than a 0.96 correlation!

1. ^{^}

   I checked. It's 0.67. 

I've heard people suggest that they have arguments related to RL being particularly dangerous, although I have to admit that I'm struggling to find these arguments at the moment. I don't know, perhaps that helps clarify why I've framed the question the way that I've framed it?

I think accumulate power and resources via mechanisms such as (but not limited to) hacking seems pretty central to me.

All the examples of "RL" doing interesting things that look like they involve sparse/distant reward involve enormous amounts of implicit structure of various kinds, like powerful world models.

I guess when you say “powerful world models”, you’re suggesting that model-based RL (e.g. MuZero) is not RL but rather “RL”-in-scare-quotes. Was that your intention?

I’ve always thought of model-based RL is a central subcategory within RL, as opposed to an edge-case.

For those who don't get the joke: benzos are depressants, and will (temporarily) significantly reduce your cognitive function if you take enough to have amnesia.

But Eric Neyman's post suggests that benzos don't significantly reduce performance on some cognitive tasks (e.g. Spelling Bee)

There was an era in a scientific community where they were interested in the "kinds of learning and memory that could happen in de-corticated animals" and they sort of homed in on the basal ganglia (which, to a first approximation "implements habits" (including bad ones like tooth grinding)) as the locus of this "ability to learn despite the absence of stuff you'd think was necessary for your naive theory of first-order subjectively-vivid learning".

(The cerebellum also probably has some "learning contribution" specifically for fine motor skills, but it is somewhat selectively disrupted just by alcohol: hence the stumbling and slurring. I don't know if anyone yet has a clean theory for how the cerebellum's full update loop works. I learned about alcohol/cerebellum interactions because I once taught a friend to juggle at a party, and she learned it, but apparently only because she was drunk. She lost the skill when sober.)

Yeah I actually do cite that piece in the appendix 'GDP as a proxy for welfare [LW · GW]' where I list more literature like this. So yeah, it's not a perfect measure but it's the one we have and 'all models are wrong but some are useful' and GDP is quite a powerful predictor of all kinds of outcomes: 

In a 2016 paper, Jones and Klenow used measures of consumption, leisure, inequality, and mortality, to create a consumption-equivalent welfare measure that allows comparisons across time for a given country, as well as across countries.^[6] 

This measure of human welfare suggests that the true level of welfare of some countries differs markedly from the level that might be suggested by their GDP per capita. For example, France’s GDP per capita is around 60% of US GDP per capita.^[7] However, France has lower inequality, lower mortality, and more leisure time than the US. Thus, on the Jones and Klenow measure of welfare, France’s welfare per person is 92% of US welfare per person.^[8]

Although GDP per capita is distinct from this expanded welfare metric, the correlation between GDP per capita and this expanded welfare metric is very strong at 0.96, though there is substantial variation across countries, and welfare is more dispersed (standard deviation of 1.51 in logs) than is income (standard deviation of 1.27 in logs).^[9]

GDP per capita is also very strongly correlated with the Human Development Index, another expanded welfare metric.^[10] If measures such as these are accurate, this shows that income per head explains most of the observed cross-national variation in welfare. It is a distinct question whether economic growth explains most of the observed variation across individuals in welfare. It is, however, clear that it explains a substantial fraction of the variation across individuals.

Surely asking if anything is safer is only sensible when comparing it to something. Are you comparing it to some implicit expected-if-not RL method of alignment? I don't think we have a commonly shared concept of what that would be. That's why I'm pointing to some explicit alternatives in that post.