Doing calls, lead gen, application reviewing, and public speaking for the 80k advising team
Mjreard
I think you’ve failed to think on the margin here. I agree that the broad classes of regulation you point to here have *netted out* badly, but this says little about what the most thoughtful and determined actors in these spaces have achieved.
Classically, Germany’s early 2000s investments in solar R&D had enormous positive externalities on climate and the people who pushed for those didn’t have to support restricting nuclear power also. The option space for them was not “the net-bad energy policy that emerged” vs “libertarian paradise;” it was: “the existing/inevitable bad policies with a bet on solar R&D” vs “the existing/inevitable bad policies with no bet on solar R&D.”
I believe most EAs treat their engagement with AI policy as researching and advocating for narrow policies tailored to mitigate catastrophic risk. In this sense, they’re acting as an organized/expert interest group motivated by a good, even popular per some polls, view of the public interest. They are competing with rather than complementing the more selfishly motivated interest groups seeking the kind of influence the oil & gas industry did in the climate context. On your model of regulation, this seems like a wise strategy, perhaps the only viable one. Again the alternative is not no regulation, but regulation that leaves out the best, most prosocial ideas.
To the extent you’re trying to warn EAs not to indiscriminately cheer any AI policy proposal assuming it will help with x-risk, I agree with you. I don’t however agree that’s reflective of how they’re treating the issue.
Tiny nit: I didn’t and don’t read much into the 80k comment on liking nice apartments. It struck me as the easiest way to disclose (imply?) that he lived in a nice place without dwelling on it too much.
Yes, in general it’s good to remember that people are far from 1:1 substitutes for each other for a given job title. I think the “1 into 2” reasoning is a decent intuition pump for how wide the option space becomes when you think laterally though and that lateral thinking of course shouldn’t stop at earning to give.
A minor, not fully-endorsed object level point: I think people who do ~one-on-one service work like (most) doctors and lawyers are much less likely to 10x the median than e.g. software engineers. With rare exceptions, their work just isn’t that scalable and in many cases output is a linear return to effort. I think this might be especially true in public defense where you sort of wear prosecutors down over a volume of cases.
Looks like the UK hardcover release isn’t until 21 May, but it’s available on Kindle? Is that right?
If the lives of pests are net negative,* I think a healthy attitude is to frame your natural threat/disgust reaction to them as useful. The pests you see now are a threat to all the future pests they will create. It’s imperative to the suffering of those future creatures that the first ones don’t live to create them. Our homes are fertile breeding grounds for enormous suffering. I think creating these potential breeding grounds gives us a responsibility to prevent them from realizing that potential.
I take the central (practical) lesson of this post to be that that responsibility should spark some urgency to act and overcome guilt when we notice the first moth or mouse. We’ve already done the guilty thing of creating this space and not isolating it. The only options left are between more suffering and less.
Thank you for the post!
*I mean this broadly to include both cases where their lives are net negative in the intervention-never scenario and in (more likely) scenarios like these where the ~inevitable human intervention might make them that way.
Nice punchy writing! I hope this sparks some interesting, good faith discussions with classmates.
I think a powerful thing to bring up re earning to give is how it can strictly dominate some other options. e.g. a 4th or 5th year biglaw associate could very reasonably fund two fully paid public defender positions with something like 25-30% of their salary. A well-paid plastic surgeon could fund lots of critical medical workers in the developing world with less.
One important thing to keep in mind when you have these chats is that there are better options; they’re just harder to carve out and evaluate. One toy model I play with is entrepreneurship. Most people inclined towards working for social good have a modesty/meekness about them where they just want to be a line-worker standing shoulder-to-shoulder with people doing the hard work of solving problems. This suggests there might be a dearth of people with this outlook looking to build, scale, and importantly sell novel solutions.
As you point out, there are a lot of rich people out there. Many/most of them just want to get richer, sure, but lots of them have foundations or would fund exciting/clever projects with exciting leaders, even if there wasn’t enormous (or any) profitability in it. The problem is a dearth of good prosocial ideas – which Harvard students seem well positioned to spin up: you have four years to just think and learn about the world, right? What projects could exist that need to? Figure it out instead of soldiering away for existing things.
Curious if you’ve seen or could share botecs on the all-in cost per retreat?
Naïvely, people like to benchmark 5% of property value per year as the all-in-cost of ownership alone (so ~$750k/yr here? Really not sure how this scales to properties like Wytham).
I wonder how that compares to the savings in variable retreat costs? Like if you had 20 retreats/yr are you saving (close to?) $32,500 per retreat (assuming 750k is the right number). Accommodation for 25 people for 4 nights in Oxford could plausibly be ~$20k itself, so it seems like with a given number of retreats or attendees, you could get quite close, but the numbers matter here.
For what it’s worth, I think you shouldn’t worry about the first two bullets. The way you as an individual or EA as a community will have big impact is through specialization. Being an excellent communicator of EA ideas is going to have way bigger and potentially compounding returns than your personal dietary or donation choices (assuming you aren’t very wealthy). If stressing about the latter takes away from the former, that seems like a mistake to be worried about.
I also shouldn’t comment without answering the question:
I balk at thorny or under-scoped research/problems that could be very valuable
It feels aversive to dig into something without a sense of where I’ll end up or whether I’ll even get anywhere
If there’s a way I can bend what I already know/am good at into the shape of the problem, I’ll do that instead
One way this happens is that I only seek out information/arguments/context that are legible to me, specifically more big picture/social science-oriented things like Holden, Joe Carlsmith or Carl Shulman, even though understanding whether technical aspects of AI alignment/evals make sense is a bigger and more unduly under-explored crux for understanding what matters
I fail to be a team player in a lot of ways.
I have my own sense of what my team/org’s priorities should be
I expect others around me to intuit, adopt these priorities with no or minimal communication
When we don’t agree or reach consensus and there’s a route for me to avoid resolving the tension, I take the avoidant route. Things that I don’t think are important, but others do, don’t happen
I think this is a version of a more general form of motivated reasoning where one seeks out a variable in an argument which is:
imprecise,
ambiguous,
dependent on multiple other hard-to-track variables, or
a variable over which they can claim unique knowledge (here, ‘what I am good at personally and how good at it I am’)
which they can then ratchet up to the maximum value for things they want to believe and the minimum value for things they don’t want to believe.
I noticed this acutely in the comments on the 80k/Rational Animations crossover video, namely things like “If you become a doctor, you don’t know how many life-saving situations you run into” (imprecision about likelihoods) or “Dr. Nalin couldn’t have achieved what he did without the help of many others, down to the bricklayers and garbagemen who provided the essentials he needed to focus” (ambiguity/dependencies about credit).
Finding low-confrontation ways to point such things out seems valuable. Maybe the Scout Mindset remains the best work here.
It is scary and painful for people to admit they were mistaken, especially about their basic narratives concerning what’s valuable or what they intended to do with their lives. I’d guess highlighting that truth-seeking is a broader, more-endorsed narrative – that also implies lots of changing your mind – is one way to shake people out of these more contingent narratives.
I think this characterizes the disagreement between pause advocates and Anthropic as it stood before the Claude 3 release with some pause-advocacy-favorable assumptions about the politics of maintaining one’s position in the industry. Full-throated, public pause advocacy doesn’t seem like a good way to induce investment in your company, for example.
More broadly I think Anthropic, like many, hasn’t come to final views on these topics and is working on developing views, probably with more information and talent than most alternatives by virtue of being a well-funded company.
As I understand it, [part of] Anthropic’s theory of change is to be a meaningful industry player so its safety agenda can become a potential standard to voluntarily emulate or adopt in policy. Being a meaningful industry player in 2024 means having desirable consumer products and advertising them as such.
It’s also worth remembering that this is advertising. Claiming to be a little bit better on some cherry picked metrics a year after GPT-4 was released is hardly a major accelerant in the overall AI race.
Too high. I thought there were huge scaling barriers based on something Linch wrote ~2 years ago. Maybe that’s wrong or been retracted.
I found the cultivated meat one a little surprising so made a market:
I was confused by this:
There were lots of explicit statements that no, of course we did not mean that and of course we do not endorse any of that, no one should be doing any of that. And yes, I think everyone means it. But it’s based on, essentially, unprincipled hacks on top of the system, rather than fixing the root problem, and the smartest kids in the world are going to keep noticing this.
You seem to be saying that Sam was among the smartest and so saw that these were unprincipled hacks and ignored them, yet the rest of the post goes into great detail on how profoundly stupid Sam was for buying into such a naive, simplistic world model that very predictably failed on its own terms. I would expect the smartest people in the world to make better predictions and use better models.
I interpret EA as common sense morality+. Abiding by common sense morality will often leave you with a lot of time and energy left over. If you want to do more good, you should use those to increase welfare, and do so in a scientific and scope-sensitive way. Is that clearly not EA or an unprincipled hack?
I appreciate the care and detail here, but would guess that wild animals dwarf everything considered here and present a much more difficult + important question.
How bad are forests per unit of land vs corn/soy/wheat fields or cattle ranches that have been replacing them seems like a key question.
I think a chatbot fails the cost-benefit analysis pretty badly at this point. There are big reputational hits organizations can take for giving bad advice and potential hallucinations just create a lot of surface area there. Importantly, the upside is quite minimal too. If a user wants to, they can pull up ChatGPT and ask it to act as an 80k advisor. It might do okay (or similarly to how okay it would do if we tried to develop one), only it’d be much clearer that we didn’t sanction its output.
People are often surprised that full time advisors only do ~400 calls/year as opposed to something like 5 calls/day (i.e.1,300/yr). For one thing, my BOTEC on the average focus time for an individual advisee is 2.25 hours (between, call prep, the call itself, post-call notes/research on new questions, introduction admin, and answering follow up emails). Beyond that, we have to keep up with what’s going on in the world and the job markets we track, as well as skilling up as generalist advisors. There’s also more formal systems we need to contribute to like marketing, impact assessment, and maintaining the systems that get us all the information we use to help advisees and keep that 2.25 hours at 2.25 hours.
Perhaps surprisingly (and perhaps not as relevant to this audience): take cause prioritization seriously, or more generally, have clarity about your ultimate goals/what you’ll look to to know whether you’ve made good decisions after the fact.
It’s very common that someone wants to do X, I ask them why, they give an answer that doesn’t point to their ultimate priorities in life, I ask them “why [thing you pointed to]?” and they more or less draw a blank/fumble around uncertainly. Granted it’s a big question, but it’s your life, have a sense of what you’re trying to do at a fundamental level.
Don’t be too fixated on instant impact. Take good opportunities as they come of course, but people are often drawn towards things that sound good/ambitious for the problems of the moment even though they might not be best positioned to tackle those things and might burn a lot of future opportunities by doing so. Details will vary by situation of course.
If you haven’t, you should talk to the 80k advising team with regard to feedback. We obviously aren’t the hiring orgs ourselves, but I think we have reasonable priors on how they read certain experiences and proposals. We’ve also been through a bunch of EA hiring rounds ourselves and spoken to many, many people on both sides of them.