I’m currently a Visiting Scholar at Forethought. Previously, I worked at the AI Safety unit of the European AI Office and as a capacity building grant-maker at Open Phil (now Coefficient Giving). Prior to that I co-founded ARENA with Callum McDougall. I studied Engineering Science with a specialization in ML.
(optionally anonymous) feedback form
last updated in June 2026
Interesting! Continuing this chain of thought…
But then, entertainment sims might oversample entertaining outcomes (e.g. rerun an event 10 times and use the branch with the most entertaining result). And then, as a result, acausal trade folks’ marginal willingness to pay for info about how those entertaining worlds turn out would go down. And for similar reasons, the stakes of our actions would then generally be lower in any given entertainment sim.
I don’t know why but I usually think in terms of cost per unit good, not good per unit cost. I said “1% future-improvement per $5B” but I really think like “$5B per 1% future-improvement.”
These two old GWWC posts (1, 2) argue for using good per unit cost since that’s arithmetically more convenient/intuitive in relevant scenarios. Most relevant excerpts below:
[Unlike consumers making day-to-day purchases (e.g. when you want to buy exactly one chocolate bar, and you are interested in comparing how much money you can save by choosing between various shops),] donors are usually interested in the differences in how much good they can do with a given amount of money, rather than the differences in how much money they can save for a given amount of good they want to do.
And:
What the considerations in post I and this post suggest is that it is usually wise to report cost-effectiveness in terms of DALY/$ rather than $/DALY, because then
the comparison of differences in cost-effectiveness that we care about are more conspicuous, as the differences could be obtained by simple subtraction and;
the most often relevant averages of cost-effectiveness could be reached simply by taking the familiar arithmetic mean, rather than the more cumbersome harmonic mean.
Relevant classic paper from Steven Levitt. Abstract [emphasis mine]:
Little is known about whether people make good choices when facing important decisions. This paper reports on a large-scale randomized field experiment in which research subjects having difficulty making a decision flipped a coin to help determine their choice. For important decisions (e.g. quitting a job or ending a relationship), those who make a change (regardless of the outcome of the coin toss) report being substantially happier two months and six months later. This correlation, however, need not reflect a causal impact. To assess causality, I use the outcome of a coin toss. Individuals who are told by the coin toss to make a change are much more likely to make a change and are happier six months later than those who were told by the coin to maintain the status quo. The results of this paper suggest that people may be excessively cautious when facing life-changing choices.
I’ve updated toward the views Daniel expresses here and I’m now about half way between Ajeya’s views in this post and Daniel’s (in geometric mean).
I’m curious what the biggest factors were that made you update?
Regarding our career development and transition funding (CDTF) program:
The default expectation for CDTF grants is that they’re one-off grants. My impression is that this is currently clear to most CDTF grantees (e.g., I think most of them don’t reapply after the end of their grant period, and the program title explicitly says that it’s “transition funding”).
(When funding independent research through this program, we sometimes explicitly clarify that we’re unlikely to renew by default).
Most of the CDTF grants we make have grant periods that are shorter than a year (with the main exception that comes to mind being PhD programs). I think that’s reasonable (esp. given that the grantees know this when they accept the funding). I’d guess most of the people we fund through this program are able to find paid positions after <1 year.
(I probably won’t have time to engage further.)
Just wanted to flag quickly that Open Philanthropy’s GCR Capacity Building team (where I work) has a career development and transition funding program.
The program aims to provide support—in the form of funding for graduate study, unpaid internships, self-study, career transition and exploration periods, and other activities relevant to building career capital—for individuals at any career stage who want to pursue careers that could help reduce global catastrophic risks (esp. AI risks). It’s open globally and operates on a rolling basis.
I realize that this is quite different from what lemonhope is advocating for here, but nevertheless thought it would be useful context for this discussion (and potential applicants).
Thanks for the feedback! I’ll forward it to our team.
I think I basically agree with you that from reading the RFP page, this project doesn’t seem like a central example of the projects we’re describing (and indeed, many of the projects we do fund through this RFP are more like the examples given on the RFP page).
Some quick reactions:
FWIW, our team generally makes a lot of grants that are <$100k (much more so than other Open Phil teams).
I agree the application would probably take most people longer than the description that Gavin gave on Manifund. That said, I think it’s still relatively lean considering the distribution of projects we fund, though I agree it’s slightly long for projects as small as this one (but I think Gavin could have filled it out in <<2 days). For reference, this is our form.
Regarding turnaround time, my guess is for this project, we would have taken significantly less than 3 months, especially if they had indicated that receiving a decision was time-sensitive. For reference, the form currently says:
We expect to make most funding decisions in 3 months or less (assuming prompt responses to any follow-up questions we may have), and we may or may not be able to accommodate requests for greater time-sensitivity. Applicants asking for over $500K should expect a decision to take the full 3 months (or more, in particularly complex cases), and apply in advance accordingly. We’ll let you know as soon as we can if we anticipate a longer than 3-month decision timeline. [emphasis in original]
For $500k+ projects, I think a 3-month turnaround time is more defensible, though I do personally wish we generally had faster response times.
I work at Open Philanthropy, and I recently let Gavin know that Open Phil is planning to recommend a grant of $5k to Arb for the second project on your list: Overview of AI Safety in 2024 (they had already raised ~$10k by the time we came across it). Thanks for writing this post Austin — it brought the funding opportunity to our attention.
Like other commenters on Manifund, I believe this kind of overview is a valuable reference for the field, especially for newcomers.
I wanted to flag that this project would have been eligible for our RFP for work that builds capacity to address risks from transformative AI. I worry that not all potential applicants are aware of the RFP or its scope, so I’ll take this opportunity to mention that this RFP’s scope is quite broad, including funding for:
Training and mentorship programs
Events
Groups
Resources, media, and communications
Almost any other type of project that builds capacity for advanced AI risks (in the sense of increasing the number of careers devoted to these problems, supporting people doing this work, and sharing knowledge related to this work).
More details at the link above. People might also find this page helpful, which lists all currently open application programs at Open Phil.
There are two very similar pages. This one and https://www.lesswrong.com/tag/scoring-rules/
By “refining pure human feedback”, do you mean refining RLHF ML techniques?
I assume you still view enhancing human feedback as valuable? And also more straightforwardly just increasing the quality of the best human feedback?
Amazing! Thanks so much for making this happen so quickly.
To anyone who’s trying to figure out how to get it to work on Google Podcasts, here’s what worked for me (searching the name didn’t, maybe this will change?):
Go to the Libsyn link. Click the RSS symbol. Copy the link. Go to Google Podcasts. Click the Library tab (bottom right). Go to Subscriptions. Click symbol that looks like adding a link in the upper right. Paste link, confirm.
Hey Paul, thanks for taking the time to write that up, that’s very helpful!
Hey Rohin, thanks a lot, that’s genuinely super helpful. Drawing analogies to “normal science” seems both reasonable and like it clears the picture up a lot.
I would be interested to hear opinions about what fraction of people could possibly produce useful alignment work?
Ignoring the hurdle of “knowing about AI safety at all”, i.e. assuming they took some time to engage with it (e.g. they took the AGI Safety Fundamentals course). Also assume they got some good mentorship (e.g. from one of you) and then decided to commit full-time (and got funding for that). The thing I’m trying to get at is more about having the mental horsepower + epistemics + creativity + whatever other qualities are useful, or likely being able to get there after some years of training.
Also note that I mean direct useful work, not indirect meta things like outreach or being a PA to a good alignment researcher etc. (these can be super important, but I think it’s productive to think of them as a distinct class). E.g. I would include being a software engineer at Anthropic, but exclude doing grocery-shopping for your favorite alignment researcher.
An answer could look like “X% of the general population” or “half the people who could get a STEM degree at Ivy League schools if they tried” or “a tenth of the people who win the Fields medal”.
I think it’s useful to have a sense of this for many purposes, incl. questions about community growth and the value of outreach in different contexts, as well as priors about one’s own ability to contribute. Hence, I think it’s worth discussing honestly, even though it can obviously be controversial (with some possible answers implying that most current AI safety people are not being useful).
That’s a very detailed answer, thanks! I’ll have a look at some of those tools. Currently I’m limiting my use to a particular 10-minute window per day with freedom.to + the app BlockSite. It often costs me way more than 10 minutes (checking links after, procrastinating before...) of focus though, so I might try to find an alternative.
Sorry for the tangent, but how do you recommend engaging with Twitter, without it being net bad?
Thanks! Great to hear that it’s going well!
Maybe I’m being stupid here. On page 42 of the write-up, it says:
In order to ensure we learned the human simulator, we would need to change the training strategy to ensure that it contains sufficiently challenging inference problems, and that doing direct translation was a cost-effective way to improve speed (i.e. that there aren’t other changes to the human simulator that would save even more time). [emphasis mine]
Shouldn’t that be?
In order to ensure we learned the direct translator, …
Most importantly for this context, market cap is about expectations for *future* earnings. So, it’s not really reflecting the economy today.
And tech stocks tend to have more growth baked into their valuations.