Developing my worldview. Interested in meta-ethics, epistemics, psychology, AI safety, and AI strategy.
Jack R
Could you try to expand upon what you mean by “equal intensities”?
What is the SFE response to the following point, which is mostly made by Carl Shulman here? Pain/pleasure asymmetry would be really weird in the technological limit (Occam’s razor), and that it makes sense that evolution would develop downside-skewed nervous systems when you think about the kinds of events that can occur in the evolutionary environment (e.g. death, sex) and the delta “reproductive fitness points” they incur (i.e. the worst single things that can happen to you, as a coincidental fact about evolution, the evolutionary environment, and what kinds of “algorithms” are simple for your nervous system to develop, are way worse from evolution’s perspective than the best single things that can happen to you), but that our nervous systems aren’t much evidence of the technological possibilities of the far-future?
Thanks! This is the exact kind of thing I was interested in hearing about. If you don’t mind sharing, is there any significant way in which the 25 people were selected for? E.g. “people who expressed interest in a program about doing good” vs “people who had engaged with EA for at least N hours and were the top 25 most promising from our perspective out of 100 who applied.” I’m hoping for the sake of meta-EA tractability that it was closer to the former :)
[Edited] How important do you think it is to have ML research projects be lead by researchers who have had a lot of previous success in ML? Maybe it’s the case that the most useful ML research is done by the top ML researchers, or that the ML community won’t take Redwood very seriously (e.g. won’t consider using your algorithms) if the research projects aren’t lead by people with strong track records in ML.
Additionally, what are/how strong are the track records of Redwood’s researchers/advisors?
Around May 2022, TFG will host a week-long retreat, training corporate executives with 10+ years experience
Interesting. Do you have any data/anecdata about the tractability of getting 30+ year-olds to switch into EA careers? My current guess is that on the margin, although this seems valuable, week-long retreats teaching people about EA should be done for high-achieving high-schoolers (mostly because they would be more willing to change their career paths). Targeting high-schoolers makes less sense if you want to solve the management gap, though you could, for instance, target high-achieving entrepreneurial high-schoolers to help solve the entrepreneur gap (that I perceive there to be).
Thanks for the response! I found the second set of bullet points especially interesting/novel.
Also, how important does it seem like governance is here versus other kinds of coordination? Any historical examples that inform your beliefs?
It’s 2035, Redwood has built an array of alignment tools that make SOTA models far less existentially risky without sacrificing hardly any performance. But these tools don’t end up being used by enough of the richest labs such that we still face doom. What happened?
What meaning of the phrase “hardcore EA” are you meaning to use here?
This seems like really good advice, thanks for writing this!
Also, I’m compiling a list of CS/ML bootcamps here (anyone should feel free to add items).
I don’t think your point about Toby’s GDP recommendation is inconsistent with David’s claim that Toby/Will seem to imply “Effective Altruism should focus entirely on longtermism” since EA is not in control of all of the world’s GDP. It’s consistent to recommend EA focus entirely on longtermism and that the world spend .1% of GDP on x-risk (or longtermism).
Off-topic, but can you give an example of irreducible uncertainty? I’ve been thinking that, technically, all uncertainty is epistemic uncertainty and that what people call aleatoric uncertainty is really just epistemic uncertainty that is quite expensive to reduce.
Thanks for this Rob—I was going to post this myself but you beat me to it :)
Also, wow—I was systematically wrong. I think my (relative) x-risk optimism affected my predictions majorly.
SPOILER: My predictions for the mean answers from each org. The first number is for Q2, the second is for Q1 (EDIT: originally had the order of the questions wrong):
OpenAI: 15%, 11%
FHI: 11%, 7%
DeepMind: 8%, 6%
CHAI/Berkeley: 18%, 15%
MIRI: 60%, 50%
Open Philanthropy: 8%, 6%
I think it’s hard to automate things
Can you elaborate on why you think this?
Instead of “counterfactually” should we say “Shapily” now?
I feel like my question wasn’t answered. For instance, Carl suggests using units such that when a painful experience and pleasurable experience are said to be of “equal intensity” then you are morally indifferent between the two experiences. This seems like a super useful way to define the units (the units can then be directly used in decision calculus). Using this kind of definition, you can then try to answer for yourself things like “do I think a day-long headache is more units of pain than a wedding day is units of pleasure?” or “do I think in the technological limit, creating 1 unit of pain will be easier than creating 1 unit of pleasure?”
What I meant by my original question was: do you have an alternative definition of what it means for pain/pleasure experiences to be of “equal intensity” that is analogous to this one?