Thanks that's interesting!
Can I double check, do you think this affects the bottom lines?
The bottom line is supposed to be that FLOP/s vs. FLOP per forward pass can be used as an upper bound, and memory bandwidth vs. model size can be used as an lower bound, and real life efficiency falls somewhere in the middle depending on a many factors (inc. length of KV cache), which I don't try to get into, but is plausibly around 15% of the upper bound for GPT-4 on H100s.
Are you saying that the lower bound for output tokens should maybe be even lower, because the KV cache can be larger than the model weights?
You need to add in the endowments of the colleges as well. The richest college at Cambridge (Trinity) has an endowment of about $1.5bn; whereas the richest college at Oxford has only about $300m.
What are the chances of being a billionaire or getting $30m plus if you go to Harvard rather than an elite uni?
And then what about HBS rather than Harvard?
Agree - Glassdoor is mainly designed to appeal to job seekers. The way they get their data is by only granting access if you reveal your salary. So the salary data ends up tilted towards the people who are seeking jobs.
There's also a sampling problem. Google has ~10,000 engineers, but there's probably only ~100 who earn $1mn+. Large companies normally only have a couple of responses, so even if you sampled everyone randomly, you'd only get ~1 top earner in the sample.
Hi Jonah,
Great posts.
I agree these figures show it's plausible that the value of donations in finance are significantly larger than the direct economic contribution of many jobs, though I see it as highly uncertain. When you're working in highly socially valuable sectors like research or some entrepreneurship, it seems to me that the two are roughly comparable, with big error bars.
However, I don't think this shows it's plausible that earning to give is likely to be the path towards doing the most good. There are many careers that seem to offer influence over budgets significantly larger than what you could expect to donate. For instance, the average budget per employee at DfiD is about $6mn per year, and you get similar figures at the World Bank, and many major foundations. It seems possible to move this money into something similarly effective or better than cash transfers. We've also just done an estimate of party politics showing that the expected budget influenced towards your preferred causes is 1-80mn if you're an Oxford graduate over a career, and that takes account of chances of success.
You'd expect there to be less competition to influence the budgets of foundations for the better than to earn money, so these figures make sense.
(And then there's all the meta things, like persuading people to do earning to give :) )
One point to note with Carl's 30x figure - that's only when comparing the short-run welfare impact of a GDP boost with a transfer to GiveDirectly. If you also care about the long-run effects, then it becomes much less clear.
Glassdoor rarely properly includes the top paid employees (those people don't fill out the survey). According to Goldman's own figures, mean compensation per employee (across all employees) is ~$400k. It'll be significantly higher if you're in front office. Your expected earnings from a Goldman job are roughly the mean earnings multiplied by the expected number of years you'll stay at the firm.
I think both research and advocacy (both to governments and among individuals) are highly important, and it's very unclear which is more important at the margin.
It's too simple to say basic research is more important, because advocacy could lead to hugely increased funding for basic research.
We've collated a list of all the approaches that seem to be on the table in the effective altruism community for improving the long-run future. There's some other options, including funding GiveWell and GCRI. This doc also explains a little more of the reasoning behind the approaches. If you like more detail on how 80k might help reduce the risk of extinction, drop me an email at ben@80000hours.org.
In general, I think the question of how best to improve the long-run future is highly uncertain, but has high value of information, so the most important activities are: (i) more prioritisation research (ii) building flexible capacity which can act on whatever turns out to be best in the future.
MIRI, FHI, GW, 80k, CEA, CFAR, GCRI all aim to further these causes, and are mutually supporting, so are particularly hard to disentangle. My guess is that if you buy the basic picture, the key issues will be things like 'which organisation has the most pressing room for more funding at the moment?' rather than questions about the value of the particular strategies.
Another option would be to fund research into which org can best use donations. There's a chance this could be commissioned through CEA, though we'll need to think of some ways to reduce bias!
Disclaimer: I'm the Executive Director of 80,000 Hours, which is part of CEA.
Note that Toby is a trustee of CEA and did most of his government consulting due to GWWC, not the FHI, so it's not clear that FHI wins out in terms of influence over government.
Moreover, if your concern is influence over government, CEA could still beat FHI (even if FHI is doing very high level advocacy) by acting as a multiplier on the FHI's efforts (and similar orgs): $1 donated to CEA could lead to more than $1 of financial or human capital delivered to the FHI or similar. I'm not claiming this is happening, but just pointing out that it's too simple to say FHI wins out just because they're doing some really good advocacy.
Disclaimer: I'm the Executive Director of 80,000 Hours, which is part of CEA.
I agree. I'd be very interested in anyone's forecasts for how they might evolve.
I've been working with (very roughly) another ~10x or so improvement in "inference efficiency" by 2030 (or how to measure this and make sure it's independent from other factors).
By this I mean that if we were able to train a model with 10^26 FLOP this year, achieving a fixed level of learning efficiency, it would require 10X FLOP to generate useful output, while by 2030 it would only require X FLOP to get the same output.