Sabitlenmiş Tweet
I've written about race, genetic ancestry, analyses of large biobanks, and human history
#h.v8wagygagcry" target="_blank" rel="nofollow noopener">gusevlab.org/projects/hsq/#…
I'll summarize the key points here 🧵:
English
Sasha Gusev
21.2K posts
Sasha Gusev
@SashaGusevPosts
Statistical geneticist | Associate Prof at @DanaFarber / @harvardmed / @DFCIPopSci | Blogging at https://t.co/4D7UObBNdd
Katılım Ocak 2023
2.9K Takip Edilen21.2K Takipçiler
@krishnanrohit IMO one way this would work is to have a "Continue in ... Codex" or "Continue in Chat" that let you take your context and move into the different UIs, while still retaining their individual constraints.
English
@krishnanrohit Claude desktop does this and it is quite a mess. It can make files but it's not clear where the files actually exist and they cannot be run. You also immediately lose track of where it has access and permissions.
English
As much as I want Codex and ChatGPT to work together I'm reasonably sure this will fail, for the main reason that this is a pressing problem primarily for OpenAI and not for its users. What its users want is an actual Desktop app that can do work on their own device, not a frankenstein
English
@bigmikelevine @RuxandraTeslo I find the whole discussion pretty strange. EA is all about quantifying efficiency instead of going with gut feelings about what works but has apparently never applied the same scrutiny to itself ??
English
@SashaGusevPosts @RuxandraTeslo Knowing EA there might be surveys asking what donors would’ve done if EA didn’t exist. But presumably no hard data on the (imo largely imaginary!) people who’ve given less because EA exists. So just a “hunch” (informed by me knowing the largest EA donors personally!)
English
I can confirm this and think it's a unique thing abt them. I am not an EA, but I think they will have forever changed philantropy for the better.
Samuel Hammond 🦉@hamandcheese
The unique thing about EA / rationalist philanthropy is that, while it has its traditional "cause areas," it is broadly steerable by better arguments. That is, if someone marshalled dispositive evidence that we're headed for an AI winter or that the technical alignment problem wasn't hard or that xyz funding strategy created more costs than benefits or that shrimp are p-zombies, EA and Rat funders would turn on a dime and fund something else. You can't say that about any other big philanthropic source.
English
@bigmikelevine @RuxandraTeslo is there any analysis of this or just a hunch?
English
@SashaGusevPosts @RuxandraTeslo A lot of EA money would otherwise not have moved; some donors aren’t compelled by emotional connection. Additional money moved is almost certainly larger than any donations EA *prevented* — donors who don’t like EA can just ignore it, right? — even before considering impact
English
@bigmikelevine @RuxandraTeslo As someone who has interacted with donors I can definitely tell you that it is often "either/or"
English
@SashaGusevPosts @RuxandraTeslo “instead” doesn’t track the actual approach, at least of core EA orgs. donors can have a portfolio, some giving motivated by emotional connection and some centering beneficiaries, as determined by evidence/reason. both/and, not either/or. “in addition to”
English
@BenShindel @RuxandraTeslo It's fine if the answer is "this is very hard to quantify and EAs have not done it". I can tell you that as someone who works in a hospital and interacts with the philanthropy team a lot and asks a lot of questions about donors, Scenario #2 is not that contrived.
English
@SashaGusevPosts @RuxandraTeslo Like… obviously if you compare some contrived example to minimize one of the two with some maximalist scenario for the other you’ll find that…
This is not how we do social science.
English
@BenShindel @RuxandraTeslo They are mutually exclusive to the extent that GiveWell is pulling in donors who would otherwise have gone to their local hospital. That's why we need to know the ATE.
English
@SashaGusevPosts @RuxandraTeslo These things aren’t mutually exclusive though? Also, for example, there’s zero counterfactual world where I, a GiveWell donor, instead volunteer at a Catholic church.
English
@BenShindel @RuxandraTeslo Do we have an estimate of the average treatment effect for the existence of GiveWell that takes into account the possibility of Scenario #2?
English
@BenShindel @RuxandraTeslo their local church/shelter/hospital instead and ask if they can volunteer, this use of their time is 20% as efficient as GiveWell per $1, but they make a meaningful personal connection, they make volunteering a big part of their life, the bequeath $10 million to the hospital. ...
English
@BenShindel @RuxandraTeslo It doesn't answer my question. EA infrastructure (GiveWell, Coefficient Giving, etc) cost money and time, is there any evidence it has had a net benefit versus alternative philanthropic uses of that money?
English
@BenShindel @RuxandraTeslo I'm not criticizing anything. I'm just asking for a study that spending resources implementing EA infrastructure for giving is more effective than spending the same resources implementing e.g. Catholic Church infrastructure for giving.
English
@SashaGusevPosts @RuxandraTeslo You should probably try to understand what you’re criticizing before criticizing it!
80000hours.org/2020/08/miscon…
English
@samth @allTheYud @PDoomOrder1 A metric I would like to see is how model improvements are scaling with the amount of RL that is happening and whether there are diminishing returns.
English
@SashaGusevPosts @allTheYud @PDoomOrder1 I think the most interesting question right now is what happens if the performance specifically of autoregressive language models trained on ~all digitized English text keeps improving, since that has some distinctive properties.
English
Stumbled upon an interesting debate on AI super-intelligence from 2011. Yudkowsky makes three core claims/predictions, all of which are (to date) wrong:
1) That human intelligence is relatively simple and ASI can be achieved with a few small innovations; ...
English
@BenShindel @RuxandraTeslo So is there a study that quantifies how persuasive the EA-style approach is versus other approaches based on personal connections over data?
English
@SashaGusevPosts @RuxandraTeslo And kind of the *whole point* of EA is “trying to quantify whether their approach to philanthropy is beneficial” lol
English
Sasha Gusev retweetledi
There’s so much valuable intellectual history to be done on the voices that have (discursively) created the moment we’re in right now.
If anyone is interested, I’m always looking for coauthors.
Sasha Gusev@SashaGusevPosts
Stumbled upon an interesting debate on AI super-intelligence from 2011. Yudkowsky makes three core claims/predictions, all of which are (to date) wrong: 1) That human intelligence is relatively simple and ASI can be achieved with a few small innovations; ...
English
Since the comments are starting to tilt into a pile on direction, I'll just add that Yudkowsky has also made surprising and accurate predictions (x.com/SashaGusevPost…) and ranks about average in terms of verifiable predictions among prolific LessWrong posters.
Sasha Gusev@SashaGusevPosts
@allTheYud @PDoomOrder1 And to de-stress the discussion a bit, I'll throw in your correct 2021 prediction that AI would achieve IMO gold by 2025: x.com/GarrisonLovely…
English
@KelseyTuoc @ryancbriggs @RuxandraTeslo whereas they had previously been burning money on vanity projects and luxuries. These are probably not 100x differences in persuasiveness but they're pretty high, so EA losing out on persuasiveness in a meaningful way is not implausible to me.
English
@KelseyTuoc @ryancbriggs @RuxandraTeslo I know many parents who take their kids to volunteer at food banks (etc) and the personal experience instilled a lifelong commitment to giving. I know of a number of billionaires who donated to causes because they had a very specific personal connection with the cause ...
English
@KelseyTuoc @ryancbriggs @RuxandraTeslo And just to be clear, I'm not doing this as some kind of rhetorical game. I personally think EA makes a lot of sense, but it's also so different from the way philanthropy works in medicine that I've always had the concern that I'm missing something about persuasiveness.
English
@KelseyTuoc @ryancbriggs @RuxandraTeslo Thanks, yes I'm totally onboard with respect to effectiveness. I think it's convincing. But what about persuasiveness?
English