Note: The Q4 deadline for applications to the Long-Term Future Fund is Friday 11th October. Apply here.
We opened up an application for grant requests earlier this year, and it was open for about one month. This post contains the list of grant recipients for Q3 2019, as well as some of the reasoning behind the grants. Most of the funding for these grants has already been distributed to the recipients.
In the writeups below, we explain the purpose for each grant and summarize our reasoning behind their recommendation. Each summary is written by the fund manager who was most excited about recommending the relevant grant (with a few exceptions that we've noted below). These differ a lot in length, based on how much available time the different fund members had to explain their reasoning.
When we’ve shared excerpts from an application, those excerpts may have been lightly edited for context or clarity.
Each grant recipient is followed by the size of the grant and their one-sentence description of their project. All of these grants have been made.
Total distributed: $439,197
Sometimes, applicants get alternative sources of funding, or decide to work on a different project.
The following people and organizations were applicants of this kind. The Long-Term Future Fund recommended grants to them, but did not end up funding them. We sometimes create write-ups for these applicants and include them in our reports in order to provide readers with better information on the types of grants we like to recommend.
Two grants we recommended but did not write up:
Placing a staff member within the government, to support civil servants to do the most good they can.
This grant supports HIPE (https://hipe.org.uk), a UK-based organization that helps civil servants to have high-impact careers. HIPE’s primary activities are researching how to have a positive impact in the UK government; disseminating their findings via workshops, blog posts, etc.; and providing one-on-one support to interested individuals.
HIPE has so far been entirely volunteer-run. This grant funds part of the cost of a full-time staff member for two years, plus some office and travel costs.
Our reasoning for making this grant is based on our impression that HIPE has already been able to gain some traction as a volunteer organization, and on the fact that they now have the opportunity to place a full-time staff member within the Cabinet Office. We see this both as a promising opportunity in its own right, and also as a positive signal about the engagement HIPE has been able to create so far. The fact that the Cabinet Office is willing to provide desk space and cover part of the overhead cost for the staff member suggests that HIPE is engaging successfully with its core audiences.
HIPE does not yet have robust ways of tracking its impact, but they expressed strong interest in improving their impact tracking over time. We would hope to see a more fleshed-out impact evaluation if we were asked to renew this grant in the future.
I’ll add that I (Helen) personally see promise in the idea of services that offer career discussion, coaching, and mentoring in more specialized settings. (Other fund members may agree with this, but it was not part of our discussion when deciding whether to make this grant, so I’m not sure.)
To spend the next year leveling up various technical skills with the goal of becoming more impactful in AI safety
Stag’s current intention is to spend the next year improving his skills in a variety of areas (e.g. programming, theoretical neuroscience, and game theory) with the goal of contributing to AI safety research, meeting relevant people in the x-risk community, and helping out in EA/rationality related contexts wherever he can (eg, at rationality summer camps like SPARC and ESPR).
Two projects he may pursue during the year:
I recommended funding Stag because I think he is smart, productive, and altruistic, has a track record of doing useful work, and will contribute more usefully to reducing existential risk by directly developing his capabilities and embedding himself in the EA community than he would by finishing his undergraduate degree or working a full-time job. While I’m not yet clear on what projects he will pursue, I think it’s likely that the end result will be very valuable — projects like impact certificates require substantial work from someone with technical and executional skills, and Stag seems to me to fit the bill.
More on Stag’s background: In high school, Stag had top finishes in various Latvian and European Olympiads, including a gold medal in the 2015 Latvian Olympiad in Mathematics. Stag has also previously taken the initiative to work on EA causes -- for example, he joined two other people in Latvia in attempting to create the Latvian chapter of Effective Altruism (which reached the point of creating a Latvian-language website), and he has volunteered to take on major responsibilities in future iterations of the European Summer Program in Rationality (which introduces promising high-school students to effective altruism).
Potential conflict of interest: at the time of making the grant, Stag was living with me and helping me with various odd jobs, as part of his plan to meet people in the EA community and help out where he could. This arrangement lasted for about 1.5 months. To compensate for this potential issue, I’ve included notes on Stag from Oliver Habryka, another fund manager.
I’ve interacted with Stag in the past and have broadly positive impressions of him, in particular his capacity for independent strategic thinking
Stag has achieved a high level of success in Latvian and Galois Mathematical Olympiads. I generally think that success in these competitions is one of the best predictors we have of a person’s future performance on making intellectual progress on core issues in AI safety. See also my comments and discussion on the grant to Misha Yagudin last round.
Stag has also contributed significantly to improving both ESPR and SPARC , both of which introduce talented pre-college students to core ideas in EA and AI safety. In particular, he’s helped the programs find and select strong participants, while suggesting curriculum changes that gave them more opportunities to think independently about important issues. This gives me a positive impression of Stag’s ability to contribute to other projects in the space. (I also consider ESPR and SPARC to be among the most cost-effective ways to get more excellent people interested in working on topics of relevance to the long-term future, and take this as another signal of Stag’s talent at selecting and/or improving projects.)
Workflowy, but with much more power to organize your thoughts and collaborate with others.
Roam is a web application which automates the Zettelkasten method, a note-taking / document-drafting process based on physical index cards. While it is difficult to start using the system, those who do often find it extremely helpful, including a researcher at MIRI who claims that the method doubled his research productivity.
On my inside view, if Roam succeeds, an experienced user of the note-taking app Workflowy will get at least as much value switching to Roam as they got from using Workflowy in the first place. (Many EAs, myself included, see Workflowy as an integral part of our intellectual process, and I think Roam might become even more integral than Workflowy. See also Sarah Constantin’s review of Roam, which describes Roam as being potentially as “profound a mental prosthetic as hypertext”, and her more recent endorsement of Roam.)
Over the course of the last year, I’ve had intermittent conversations with Conor White-Sullivan, Roam’s CEO, about the app. I started out in a position of skepticism: I doubted that Roam would ever have active users, let alone succeed at its stated mission. After a recent update call with Conor about his LTF Fund application, I was encouraged enough by Roam’s most recent progress, and sufficiently convinced of the possible upsides of its possible success, that I decided to recommend a grant to Roam.
Since then, Roam has developed enough as a product that I’ve personally switched from Workflowy to Roam and now recommend Roam to my friends. Roam’s progress on its product, combined with its growing base of active users, has led me to feel significantly more optimistic about Roam succeeding at its mission.
(This funding will support Roam’s general operating costs, including expenses for Conor, one employee, and several contractors.)
Potential conflict of interest: Conor is a friend of mine, and I was once his housemate for a few months.
Independent AI Safety thinking, doing research in aspects of self-reference in using techniques from type theory, topos theory and category theory more generally.
In our previous round of grants, we funded MIRI as an organization: see our April reportfor a detailed explanation of why we chose to support their work. I think Alexander’s research directions could lead to significant progress on MIRI’s research agenda — in fact, MIRI was sufficiently impressed by his work that they offered him an internship. I have also spoken to him in some depth, and was impressed both by his research taste and clarity of thought.
After the internship ends, I think it will be valuable for Alexander to have additional funding to dig deeper into these topics; I expect this grant to support roughly 1.5 years of research. During this time, he will have regular contact with researchers at MIRI, reporting on his research progress and receiving feedback.
Characterizing the properties and constraints of complex systems and their external interactions.
Alexander is a 5th-year graduate student in physics at MIT, and he wants to conduct independent deconfusion research for AI safety. His goal is to get a better conceptual understanding of multi-level world models by coming up with better formalisms for analyzing complex systems at differing levels of scale, building off of the work of Yaneer Bar-Yam. (Yaneer is Alexander’s advisor, and the president of the New England Complex Science Institute.)
I decided to recommend funding to Alexander because I think his research directions are promising, and because I was personally impressed by his technical abilities and his clarity of thought. Tsvi Benson-Tilsen, a MIRI researcher, was also impressed enough by Alexander to recommend that the Fund support him. Alexander plans to publish a paper on his research; it will be evaluated by researchers at MIRI, helping him decide how best to pursue further work in this area.
Potential conflict of interest: Alexander and I have been friends since our undergraduate years at MIT.
I have a sense that funders in EA, usually due to time constraints, tend to give little feedback to organizations they fund (or decide not to fund). In my writeups below, I tried to be as transparent as possible in explaining the reasons for why I came to believe that each grant was a good idea, my greatest uncertainties and/or concerns with each grant, and some background models I use to evaluate grants. (I hope this last item will help others better understand my future decisions in this space.)
I think that there exist more publicly defensible (or easier to understand) arguments for some of the grants that I recommended. However, I tried to explain the actual models that drove my decisions for these grants, which are often hard to summarize in a few paragraphs. I apologize in advance that some of the explanations below are probably difficult to understand.
Thoughts on grant selection and grant incentives
Some higher-level points on many of the grants below, as well as many grants from last round:
For almost every grant we make, I have a lot of opinions and thoughts about how the applicant(s) could achieve their aims better. I also have a lot of ideas for projects that I would prefer to fund over the grants we are actually making.
However, in the current structure of the LTFF, I primarily have the ability to select potential grantees from an established pool, rather than encouraging the creation of new projects. Alongside my time constraints, this means that I have a very limited ability to contribute to the projects with my own thoughts and models.
Additionally, I spend a lot of time thinking independently about these areas, and have a broad view of “ideal projects that could be made to exist.” This means that for many of the grants I am recommending, it is not usually the case that I think the projects are very good on all the relevant dimensions; I can see how they fall short of my “ideal” projects. More frequently, the projects I fund are among the only available projects in a reference class I believe to be important, and I recommend them because I want projects of that type to receive more resources (and because they pass a moderate bar for quality).
I am, overall, still very excited about the grants below, and I think they are a much better use of resources than what I think of as the most common counterfactuals to donating to the LTFF fund (e.g. donating to the largest organizations in the space, donating based on time-limited personal research) .
However, related to the points I made above, I will have many criticisms of almost all the projects that receive funding from us. I think that my criticisms are valid, but readers shouldn't interpret them to mean that I have a negative impression of the grants we are making — which are strong despite their flaws. Aggregating my individual (and frequently critical) recommendations will not give readers an accurate impression of my overall (highly positive) view of the grant round.
(If I ever come to think that the pool of valuable grants has dried up, I will say so in a high-level note like this one.)
I can imagine that in the future I might want to invest more resources into writing up lists of potential projects that I would be excited about, though it is also not clear to me that I want people to optimize too much for what I am excited about, and think that the current balance of "things that I think are exciting, and that people feel internally motivated to do and generated their own plans for" seems pretty decent.
To follow up the above with a high-level assessment, I am slightly less excited about this round’s grants than I am about last round’s, and I’d estimate (very roughly) that this round is about 25% less cost-effective than the previous round.
For both this round and the last round, I wrote the writeups in collaboration with Ben Pace, who works with me on LessWrong and the Alignment Forum. After an extensive discussion about the grants and the Fund's reasoning for them, we split the grants between us and independently wrote initial drafts. We then iterated on those drafts until they accurately described my thinking about them and the relevant domains.
I am also grateful for Aaron Gertler’s help with editing and refining these writeups, which has substantially increased their clarity.
Additional funding for an AI strategy PhD at Oxford / FHI to improve my research productivity.
I'm looking for additional funding to supplement my 15k pound/y PhD stipend for 3-4 years from September 2019. I am hoping to roughly double this. My PhD is at Oxford in machine learning, but co-supervised by Allan Dafoe from FHI so that I can focus on AI strategy. We will have multiple joint meetings each month, and I will have a desk at FHI.
The purpose is to increase my productivity and happiness. Given my expected financial situation, I currently have to make compromises on e. g. Ubers, Soylent, eating out with colleagues, accommodation, quality and waiting times for health care, spending time comparing prices, travel durations and stress, and eating less healthily.
I expect that more financial security would increase my own productivity and the effectiveness of the time invested by my supervisors.
I think that when FHI or other organizations in that reference class have trouble doing certain things due to logistical obstacles, we should usually step in and fill those gaps (e.g. see Jacob Lagerros’ grant from last round). My sense is that FHI has trouble with providing funding in situations like this (due to budgetary constraints imposed by Oxford University).
I’ve interacted with Sören in the past (during my work at CEA), and generally have positive impressions of him in a variety of domains, like his basic thinking about AI Alignment, and his general competence from running projects like the EA Newsletter.
I have a lot of trust in the judgment of Nick Bostrom and several other researchers at FHI. I am not currently very excited about the work at GovAI (the team that Allan Dafoe leads), but still have enough trust in many of the relevant decision makers to think that it is very likely that Soeren should be supported in his work.
In general, I think many of the salaries for people working on existential risk are low enough that they have to make major tradeoffs in order to deal with the resulting financial constraints. I think that increasing salaries in situations like this is a good idea (though I am hesitant about increasing salaries for other types of jobs, for a variety of reasons I won’t go into here, but am happy to expand on).
This funding should last for about 2 years of Sören’s time at Oxford.
A research experience program for prospective AI safety researchers.
We want to organize the 4th AI Safety Camp (AISC) - a research retreat and program for prospective AI safety researchers. Compared to past iterations, we plan to change the format to include a 3 to 4-day project generation period and team formation workshop, followed by a several-week period of online team collaboration on concrete research questions, a 6 to 7-day intensive research retreat, and ongoing mentoring after the camp. The target capacity is 25 - 30 participants, with projects that range from technical AI safety (majority) to policy and strategy research. More information about past camps is at https://aisafetycamp.com/
Early-career entry stage seems to be a less well-covered part of the talent pipeline, especially in Europe. Individual mentoring is costly from the standpoint of expert advisors (esp. compared to guided team work), while internships and e.g. MSFP have limited capacity and are US-centric. After the camp, we advise and encourage participants on future career steps and help connect them to other organizations, or direct them to further individual work and learning if they are pursuing an academic track..
Overviews of previous research projects from the first 2 camps can be found here:
Projects from AISC3 are still in progress and there is no public summary.
To evaluate the camp, we send out an evaluation form directly after the camp has concluded and then informally follow the career decisions, publications, and other AI safety/EA involvement of the participants. We plan to conduct a larger survey from past AISC participants later in 2019 to evaluate our mid-term impact. We expect to get a more comprehensive picture of the impact, but it is difficult to evaluate counterfactuals and indirect effects (e.g. networking effects). The (anecdotal) positive examples we attribute to past camps include the acceleration of entrance of several people in the field, research outputs that include 2 conference papers, several SW projects, and about 10 blogposts.
The main direct costs of the camp are the opportunity costs of participants, organizers and advisors. There are also downside risks associated with personal conflicts at multi-day retreats and discouraging capable people from the field if the camp is run poorly. We actively work to prevent this by providing both on-site and external anonymous contact points, as well as actively attending to participant well-being, including during the online phases.
This grant is for the AI Safety Camp, to which we made a grant in the last round. Of the grants I recommended this round, I am most uncertain about this one. The primary reason is that I have not received much evidence about the performance of either of the last two camps, and I assign at least some probability that the camps are not facilitating very much good work. (This is mostly because I have low expectations for the quality of most work of this kind and haven’t looked closely enough at the camp to override these — not because I have positive evidence that they produce low-quality work.)
My biggest concern is that the camps do not provide a sufficient level of feedback and mentorship for the attendees. When I try to predict how well I’d expect a research retreat like the AI Safety Camp to go, much of the impact hinges on putting attendees into contact with more experienced researchers and having a good mentoring setup. Some of the problems I have with the output from the AI Safety Camp seem like they could be explained by a lack of mentorship.
From the evidence I observe on their website, I see that the attendees of the second camp all produced an artifact of their research (e.g. an academic writeup or code repository). I think this is a very positive sign. That said, it doesn’t look like any alignment researchers have commented on any of this work (this may in part have been because most of it was presented in formats that require a lot of time to engage with, such as GitHub repositories), so I’m not sure the output actually lead to the participants to get any feedback on their research directions, which is one of the most important things for people new to the field.
After some followup discussion with the organizers, I heard about changes to the upcoming camp (the target of this grant) that address some of the above concerns (independent of my feedback). In particular, the camp is being renamed to “AI Safety Research Program”, and is now split into two parts — a topic selection workshop and a research retreat, with experienced AI Alignment researchers attending the workshop. The format change seems likely to be a good idea, and makes me more optimistic about this grant.
I generally think hackathons and retreats for researchers can be very valuable, allowing for focused thinking in a new environment. I think the AI Safety Camp is held at a relatively low cost, in a part of the world (Europe) where there exist few other opportunities for potential new researchers to spend time thinking about these topics, and some promising people have attended. I hope that the camps are going well, but I will not fund another one without spending significantly more time investigating the program.
 After signing off on this grant, I found out that, due to overlap between the organizers of the events, some feedback I got about this camp was actually feedback about the Human Aligned AI Summer School, which means that I had even less information than I thought. In April I said I wanted to talk with the organizers before renewing this grant, and I expected to have at least six months between applications from them, but we received another application this round and I ended up not having time for that conversation.
Writing EA-themed fiction that addresses X-risk topics.
I want to spend three months evaluating my ability to produce an original work that explores existential risk, rationality, EA, and related themes such as coordination between people with different beliefs and backgrounds, handling burnout, planning on long timescales, growth mindset, etc. I predict that completing a high-quality novel of this type would take ~12 months, so 3 months is just an initial test.
In 3 months, I would hope to produce a detailed outline of an original work plus several completed chapters. Simultaneously, I would be evaluating whether writing full-time is a good fit for me in terms of motivation and personal wellbeing.
I have spent the last 2 years writing an EA-themed fanfiction of The Last Herald-Mage trilogy by Mercedes Lackey (online at https://archiveofourown.org/series/936480). In this period I have completed 9 “books” of the series, totalling 1.2M words (average of 60K words/month), mostly while I was also working full-time. (I am currently writing the final arc, and when I finish, hope to create a shorter abridged/edited version with a more solid beginning and better pacing overall.)
In the writing process, I researched key background topics, in particular AI safety work (I read a number of Arbital articles and most of this MIRI paper on decision theory: https://arxiv.org/pdf/1710.05060v1.pdf), as well as ethics, mental health, organizational best practices, medieval history and economics, etc. I have accumulated a very dedicated group of around 10 beta readers, all EAs, who read early drafts of each section and give feedback on how well it addresses various topics, which gives me more confidence that I am portraying these concepts accurately.
One natural decomposition of whether this grant is a good idea is to first ask whether writing fiction of this type is valuable, then whether Miranda is capable of actually creating that type of fiction, and last whether funding Miranda will make a significant difference in the amount/quality of her fiction.
I think that many people reading this will be surprised or confused about this grant. I feel fairly confident that grants of this type are well worth considering, and I am interested in funding more projects like this in the future, so I’ve tried my best to summarize my reasoning. I do think there are some good arguments for why we should be hesitant to do so (partly summarized by the section below that lists things that I think fiction doesn’t do as well as non-fiction), so while I think that grants like this are quite important, and have the potential to do a significant amount of good, I can imagine changing my mind about this in the future.
The track record of fiction
In a general sense, I think that fiction has a pretty strong track record of both being successful at conveying important ideas, and being a good attractor of talent and other resources. I also think that good fiction is often necessary to establish shared norms and shared language.
Here are some examples of communities and institutions that I think used fiction very centrally in their function. Note that after the first example, I am making no claim that the effect was good, I’m just establishing the magnitude of the potential effect size.
On a more conceptual level, I think fiction tends to be particularly good at achieving the following aims (compared to non-fiction writing):
(I wrote in more detail about how this works for HPMOR in the last grant round.)
In contrast, here are some things that fiction is generally worse at (though a lot of these depend on context; since fiction often contains embedded non-fiction explanations, some of these can be overcome):
Overall, I think current writing about both existential risk, rationality, and effective altruism skews too much towards non-fiction, so I’m excited about experimenting with funding fiction writing.
The second question is whether I trust Miranda to actually be able to write fiction that leverages these opportunities and provides value. This is why I think Miranda can do a good job:
My two biggest concerns are:
I like the fact that this grant is for an exploratory 3 months rather than a longer period of time; this allows Miranda to pivot if it doesn’t work out, rather than being tied to a project that isn’t going well.
The counterfactual value of funding
It would be reasonable to ask whether a grant is really necessary, given that Miranda has produced a huge amount of fiction in the last two years without receiving funding explicitly dedicated to that. I have two thoughts here:
Multi-model approach to corporate and state actors relevant to existential risk mitigation.
Work for 2-3 months on continuing to build out a multi-model approach to understanding international relations and multi-stakeholder dynamics as it relates to risks of strong(er) AI systems development, based on and extending similar work done on biological weapons risks done on behalf of FHI's Biorisk group and supporting Open Philanthropy Project planning.
This work is likely to help policy and decision analysis for effective altruism related to the deeply uncertain and complex issues in international relations and long term planning that need to be considered for many existential risk mitigation activities. While the project is focused on understanding actors and motivations in the short term, the decisions being supported are exactly those that are critical for existential risk mitigation, with long term implications for the future.
I feel a lot of skepticism toward much of the work done in the academic study of international relations. Judging from my models of political influence and its effects on the quality of intellectual contributions, and my models of research fields with little ability to perform experiments, I have high priors that work in international relations is of significantly lower quality than in most scientific fields. However, I have engaged relatively little with actual research on the topic of international relations (outside of unusual scholars like Nick Bostrom) and so am hesitant in my judgement here.
I also have a fair bit of worry around biorisk. I haven’t really had the opportunity to engage with a good case for it, and neither have many of the people I would trust most in this space, in large part due to secrecy concerns from people who work on it (more on that below). Due to this, I am worried about information cascades. (An information cascade is a situation where people primarily share what they believe but not why, and because people update on each others' beliefs you end up with a lot of people all believing the same thing precisely because everyone else does.)
I think is valuable to work on biorisk, but this view is mostly based on individual conversations that are hard to summarize, and I feel uncomfortable with my level of understanding of possible interventions, or even just conceptual frameworks I could use to approach the problem. I don’t know how most people who work in this space came to decide it was important, and those I’ve spoken to have usually been reluctant to share details in conversation (e.g. about specific discoveries they think created risk, or types of arguments that convinced them to focus on biorisk over other threats).
I’m broadly supportive of work done at places like FHI and by the people at OpenPhil who care about x-risks, so I am in favor of funding their work (e.g. Soren’s grant above). But I don’t feel as though I can defer to the people working in this domain on the object level when there is so much secrecy around their epistemic process, because I and others cannot evaluate their reasoning.
However, I am excited about this grant, because I have a good amount of trust in David’s judgment. To be more specific, he has a track record of identifying important ideas and institutions and then working on/with them. Some concrete examples include:
Another major factor for me is the degree to which David is shares his thinking openly and transparently on the internet, and participates in public discourse, so that other people interested in these topics can engage with his ideas. (He’s also a superforecaster, which I think is predictive of broadly good judgment.) If David didn’t have this track record of public discourse, I likely wouldn’t be recommending this grant, and if he suddenly stopped participating, I’d be fairly hesitant to recommend such a grant in the future.
As I said, I’m not excited about the specific project he is proposing, but have trust in his sense of which projects might be good to work on, and I have emphasized to him that I think he should feel comfortable working on the projects he thinks are best. I strongly prefer a world where David has the freedom to work on the projects he judges to be most valuable, compared to the world where he has to take unrelated jobs (e.g. teaching at university).
Upskilling in ML in order to be able to do productive AI safety research sooner than otherwise.
I am requesting grant money to upskill in machine learning (ML).
Background: I am an undergraduate student in Computer Science and Philosophy at Oxford University, about to start the 4th year of a 4-year degree. I plan to do research in AI safety after I graduate, as I deem this to be the most promising way of having a significant positive impact on the long-term future
What I’d like to do:
I would like to improve my skills in ML by reading literature and research, replicating research papers, building ML-based systems, and so on.
To do this effectively, I need access to the compute that is required to train large models and run lengthy reinforcement learning experiments and similar.
It would also likely be very beneficial if I could live in Oxford during the vacations, as I would then be in an environment in which it is easier to be productive. It would also make it easier for me to speak with the researchers there, and give me access to the facilities of the university (including libraries, etc.).
It would also be useful to be able to attend conferences and similar events.
Joar was one of the co-authors on the Mesa-Optimisers paper, which I found surprisingly useful and clearly written, especially considering that its authors had relatively little background in alignment research or research in general. I think it is probably the second most important piece of writing on AI alignment that came out in the last 12 months, after the Embedded Agency sequence. My current best guess is that this type of conceptual clarification / deconfusion is the most important type of research in AI alignment, and the type of work I’m most interested in funding. While I don’t know exactly how Joar contributed to the paper, my sense is that all the authors put in a significant effort (bar Scott Garrabrant, who played a supervising role).
This grant is for projects during and in between terms at Oxford. I want to support Joar producing more of this kind of research, which I expect this grant to help with. He’s also been writing further thoughts online (example), which I think has many positive effects (personally and as externalities).
My brief thoughts on the paper (nontechnical):
More of my thoughts on the paper (technical):
Note: If you haven’t read the paper, or you don’t have other background in the subject, this section will likely be unclear. It’s not essential to the case for the grant, but I wanted to share it in case people with the requisite background are interested in more details about the research
I was surprised by how helpful the conceptual work in the paper was - helping me think about where the optimization was happening in a system like AlphaGo Zero improved my understanding of that system and how to connect it to other systems that do optimization in the world. The primary formalism in the paper was clarifying rather than obscuring (and the ratio of insight to formalism was very high - see my addendum below for more thoughts on that).
Once the basic concepts were in place, clarifying different basic tools that would encourage optimization to happen in either the base optimizer or the mesa optimizer (e.g. constraining and expanding space/time offered to the base or mesa optimizers has interesting effects), plus clarifying the types of alignment / pseudo-alignment / internalizing of the base objective, all helped me think about this issue very clearly. It largely used basic technical language I already knew, and put it together in ways that would’ve taken me many months to achieve on my own - a very helpful conceptual piece of work.
The following three grants were more exciting to one or more other fund managers than they were to me. I think that for all three, if it had just been me on the grant committee, we might have not actually made them. However, I had more resources available to invest into these writeups, and as such I ended up summarizing my view on them, instead of someone else on the fund doing so. As such, they are probably less representative of the reasons for why we made these grants than the writeups above.
In the course of thinking through these grants, I formed (and wrote out below) more detailed, explicit models of the topics. Although these models were not counterfactual in the Fund’s making the grants, I think they are fairly predictive of my future grant recommendations.
Note: Application sent in by Jacob Hilton.
Combat publication bias in science by promoting and supporting the Registered Reports journal format
I'm suggesting a grant to fund a teaching buyout for Professor Chris Chambers, an academic at the University of Cardiff working to promote and support Registered Reports. This funding opportunity was originally identified and researched by Hauke Hillebrandt, who published a full analysis here. In brief, a Registered Report is a format for journal articles where peer review and acceptance decisions happen before data is collected, so that the results are much less susceptible to publication bias. The grant would free Chris of teaching duties so that he can work full-time on trying to get Registered Reports to become part of mainstream science, which includes outreach to journal editors and supporting them through the process of adopting the format for their journal. More details of Chris's plans can be found here.
I think the main reason for funding this is from a worldview diversification perspective: I would expect it to broadly improve the efficiency of scientific research by improving the communication of negative results, and to enable people to make better-informed use of scientific research by reducing publication bias. I would expect these effects to be primarily within fields where empirical tests tend to be useful but not always definitive, such as clinical trials (one of Chris's focus areas), which would have knock-on effects on health.
From an X-risk perspective, the key question to answer seems to be which technologies differentially benefit from this grant. I do not have a strong opinion on this, but to quote Brian Wang from a Facebook thread: "In terms of [...] bio-risk, my initial thoughts are that reproducibility concerns in biology are strongest when it comes to biomedicine, a field that can be broadly viewed as defense-enabling. By contrast, I'm not sure that reproducibility concerns hinder the more fundamental, offense-enabling developments in biology all that much (e.g., the falling costs of gene synthesis, the discovery of CRISPR)."
As for why this particular intervention strikes me as a cost-effective way to improve science, it is shovel-ready, it may be the sort of thing that traditional funding sources would miss, it has been carefully vetted by Hauke, and I thought that Chris seemed thoughtful and intelligent from his videoed talk.”
The Let’s Fund report linked in the application played a major role in my assessment of the grant, and I probably would not have been comfortable recommending this grant without access to that report.
Thoughts on Registered Reports
The replication crisis in psychology, and the broad spread of “career science,” have made it (to me) quite clear that the methodological foundations of at least psychology itself, but possibly also the broader life-sciences, are creating a very large volume of false and likely unreproducible claims.
This is in large part caused by problematic incentives for individual scientists to engage in highly biased reporting and statistically dubious practices.
I think preregistration has the opportunity to fix a small but significant part of this problem, primarily by reducing file-drawer effects. To borrow an explanation from the Let’s Fund report (lightly edited for clarity):
[Pre-registration] was introduced to address two problems: publication bias and analytical flexibility (in particular outcome switching in the case of clinical medicine).
Publication bias, also known as the file drawer problem, refers to the fact that many more studies are conducted than published. Studies that obtain positive and novel results are more likely to be published than studies that obtain negative results or report replications of prior results. The consequence is that the published literature indicates stronger evidence for findings than exists in reality.
Outcome switching refers to the possibility of changing the outcomes of interest in the study depending on the observed results. A researcher may include ten variables that could be considered outcomes of the research, and — once the results are known — intentionally or unintentionally select the subset of outcomes that show statistically significant results as the outcomes of interest. The consequence is an increase in the likelihood that reported results are spurious by leveraging chance, while negative evidence gets ignored.
This is one of several related research practices that can inflate spurious findings when analysis decisions are made with knowledge of the observed data, such as selection of models, exclusion rules and covariates. Such data-contingent analysis decisions constitute what has become known as P-hacking, and pre-registration can protect against all of these.
It also effectively blinds the researcher to the outcome because the data are not collected yet and the outcomes are not yet known. This way the researcher’s unconscious biases cannot influence the analysis strategy
“Registered reports” refers to a specific protocol that journals are encouraged to adopt, which integrates preregistration into the journal acceptance process. Illustrated by this picture (borrowed from the Let’s Fund report):
Of the many ways to implement preregistration practices, I don’t think the one that Chambers proposes seems ideal, and I can see some flaws with it, but I still think that the quality of clinical science (and potentially other fields) will significantly improve if more journals adopt the registered reports protocol. (Please keep this in mind as you read my concerns in the next section.)
The importance of bandwidth constraints for journals
Chambers has the explicit goal of making all clinical trials require the use of registered reports. That outcome seems potentially quite harmful, and possibly worse than the current state of clinical science. (However, since that current state is very far from “universal registered reports,” I am not very worried about this grant contributing to that scenario.)
The Let’s Fund report covers the benefits of preregistration pretty well, so I won’t go into much detail here. Instead, I will mention some of my specific concerns with the protocol that Chambers is trying to promote.
From the registered reports website:
Manuscripts that pass peer review will be issued an in principle acceptance (IPA), indicating that the article will be published pending successful completion of the study according to the exact methods and analytic procedures outlined, as well as a defensible and evidence-bound interpretation of the results.
This seems unlikely to be the best course of action. I don’t think that the most widely-read journals should only publish replications. The key reason is that many scientific journals are solving a bandwidth constraint - sharing papers that are worth reading, not merely papers that say true things, to help researchers keep up to date with new findings in their field. A math journal could have papers for every true mathematical statement, including trivial ones, but they instead need to focus on true statements that are useful to signal boost to the mathematics community. (Related concepts are the tradeoff between bias and variance in Machine Learning, or accuracy and calibration in forecasting.)
Ultimately, from a value of information perspective, it is totally possible for a study to only be interesting if it finds a positive result, and to be uninteresting when analyzed pre-publication from the perspective of the editor. It seems better to encourage pre-publication, but still take into account the information value of a paper’s experimental results, even if this doesn’t fully prevent publication bias.
To give a concrete (and highly simplified) example, imagine a world where you are trying to find an effective treatment for a disease. You don’t have great theory in this space, so you basically have to test 100 plausible treatments. On their own, none of these have a high likelihood of being effective, but you expect that at least one of them will work reasonably well.
Currently, you would preregister those trials (as is required for clinical trials), and then start performing the studies one by one. Each failure provides relatively little information (since the prior probability was low anyways), so you are unlikely to be able to publish it in a prestigious journal, but you can probably still publish it somewhere. Not many people would hear about it, but it would be findable if someone is looking specifically for evidence about the specific disease you are trying to treat, or the treatment that you tried out. However, finding a successful treatment is highly valuable information which will likely get published in a journal with a lot of readers, causing lots of people to hear about the potential new treatment.
In a world with mandatory registered reports, none of these studies will be published in a high-readership journal, since journals will be forced to make a decision before they know the outcome of a treatment. Because all 100 studies are equally unpromising, none are likely to pass the high bar of such a journal, and they’ll wind up in obscure publications (if they are published at all) . Thus, even if one of them finds a successful result, few people will hear about it. High-readership journals exist in large part to spread news about valuable results in a limited bandwidth environment; this no longer happens in scenarios of this kind.
Because of dynamics like this, I think it is very unlikely that any major journals will ever switch towards only publishing registered report-based studies, even within clinical trials, since no journal would want to pass up on the opportunity to publish a study that has the opportunity to revolutionize the field.
Importance of selecting for clarity
Here is the full set of criteria that papers are being evaluated by for stage 2 of the registered reports process:
1. Whether the data are able to test the authors’ proposed hypotheses by satisfying the approved outcome-neutral conditions (such as quality checks or positive controls)
2. Whether the Introduction, rationale and stated hypotheses are the same as the approved Stage 1 submission (required)
3. Whether the authors adhered precisely to the registered experimental procedures
4. Whether any unregistered post hoc analyses added by the authors are justified, methodologically sound, and informative
5. Whether the authors’ conclusions are justified given the data
The above list is comprehensive, and does not include any mention of the clarity of the authors’ writing, the quality/rigor of the explanation provided by the paper’s methodology, or the implications of the paper’s findings on underlying theory. (All of these are very important to how journals currently evaluate papers.) This means that journals can only filter for those characteristics in the first stage of the registered reports process, when large parts of the paper haven’t yet been written. As a result, large parts of the paper basically have no selection applied to them for conceptual clarity, as well as thoughtful analysis of implications for future theory, likely resulting in those qualities getting worse.
I think the goal of registered reports is to split research in two halves where you publish two separate papers, one that is empirical, and another that is purely theoretical, which that takes the results of the first paper as given and explores their consequences. We already see this split a good amount in physics, in which there exists a pretty significant divide between experimental and theoretical physics, the latter of which rarely performs experiments. I don’t know whether encouraging this split in a given field is a net-improvement, since I generally think that a lot of good science comes from combining the gathering of good empirical data with careful analysis and explanations, and I am particularly worried that the analysis of the results in papers published via registered reports will be of particularly low-quality, which encourages the spread of bad explanations and misconceptions which can cause a lot of damage (though some of that is definitely offset by reducing the degree to which scientists can fit hypotheses post-hoc, due to preregistration). The costs here seem related to Chris Olah’s article on research debt.
Again, I think both of these problems are unlikely to become serious issues, because at most I can imagine getting to a world where something between 10% and 30% of top journal publications in a given field have gone through registered reports-based preregistration. I would be deeply surprised if there weren’t alternative outlets for papers that do try to combine the gathering of empirical data with high-quality explanations and analysis.
Failures due to bureaucracy
I should also note clinical science is not something I have spent large amounts of time thinking about, that I am quite concerned about adding more red tape and necessary logistical hurdles to jump through when registering clinical trials. I have high uncertainty about the effect of registered reports on the costs of doing small-scale clinical experiments, but it seems more likely than not that they will lengthen the review process, and add additional methodological constraints.
(There is also a chance that it will reduce these burdens by giving scientists feedback earlier in the process and letting them be more certain of the value of running a particular study. However, this effect seems slightly weaker to me than the additional costs, though I am very uncertain about this.)
In the current scientific environment, running even a simple clinical study may require millions of dollars of overhead (a related example is detailed in Scott Alexander’s “My IRB nightmare”). I believe this barrier is a substantial drag on progress in medical science. In this context, I think that requiring even more mandatory documentation, and adding even more upfront costs, seems very costly. (Though again, it seems highly unlikely for the registered reports format to ever become mandatory on a large scale, and giving more researchers the option to publish a study via the registered reports protocol, depending on their local tradeoffs, seems likely net-positive)
To summarize these three points:
Differential technological progress
Let’s Fund covers differential technological progress concerns in their writeup. Key quote:
One might worry that funding meta-research indiscriminately speeds up all research, including research which carries a lot of risks. However, for the above reasons, we believe that meta-research improves predominantly social science and applied clinical science (“p-value science’) and so has a strong differential technological development element, that hopefully makes society wiser before more risks from technology emerge through innovation. However, there are some reproducibility concerns in harder sciences such as basic biological research and high energy physics that might be sped up by meta-research and thus carry risks from emerging technologies.
My sense is that further progress in sociology and psychology seems net positive from a global catastrophic risk reduction perspective. The case for clinical science seems a bit weaker, but still positive.
In general, I am more excited about this grant in worlds in which global catastrophes are less immediate and less likely than my usual models suggest, and I’m thinking of this grant in some sense as a hedging bet, in case we live in one of those worlds.
Overall, a reasonable summary of my position on this grant would be "I think preregistration helps, but is probably not really attacking the core issues in science. I think this grant is good, because I think it actually makes preregistration a possibility in a large number of journals, though I disagree with Chris Chalmers on whether it would be good for all clinical trials to require preregistration, which I think would be quite bad. On the margin, I support his efforts, but if I ever come to change my mind about this, it’s likely for one or more of the above reasons."
: The journal could also publish a random subset, though at scale that gives rise to the same dynamics, so I’ll ignore that case. It could also batch a large number of the experiments until the expected value of information is above the relevant threshold, though that significantly increases costs.
Note: Funding from this grant will go to the Leverhulme Centre for the Future of Intelligence, which will fund Jess in turn. The LTF Fund is not replacing funding that CFI would have supplied instead; without this grant, Jess would need to pursue grants from sources outside CFI.
Research on the links between short- and long-term AI policy while skilling up in technical ML.
I’m applying for funding to cover my salary for a year as a postdoc at the Leverhulme CFI, enabling me to do two things:
-- Research the links between short- and long-term AI policy. My plan is to start broad: thinking about how to approach, frame and prioritise work on ‘short-term’ issues from a long-term perspective, and then focusing in on a more specific issue. I envision two main outputs (papers/reports): (1) reframing various aspects of ‘short-term’ AI policy from a long-term perspective (e.g. highlighting ways that ‘short-term’ issues could have long-term consequences, and ways of working on AI policy today most likely to have a long-run impact); (2) tackling a specific issue in ‘short-term’ AI policy with possible long-term consequences (tbd, but an example might be the possible impact of microtargeting on democracy and epistemic security as AI capabilities advance).
-- Skill up in technical ML by taking courses from the Cambridge ML masters.
Most work on long-term impacts of AI focuses on issues arising in the future from AGI. But issues arising in the short term may have long-term consequences: either by directly leading to extreme scenarios (e.g. automated surveillance leading to authoritarianism), or by undermining our capability to deal with other threats (e.g. disinformation undermining collective decision-making). Policy work today will also shape how AI gets developed, deployed and governed, and what issues will arise in the future. We’re at a particularly good time to influence the focus of AI policy, with many countries developing AI strategies and new research centres emerging.
There’s very little rigorous thinking the best way to do short-term AI policy from a long-term perspective. My aim is to change that, and in doing so improve the quality of discourse in current AI policy. I would start with a focus on influencing UK AI policy, as I have experience and a strong network here (e.g. the CDEI and Office for AI). Since DeepMind is in the UK, I think it is worth at least some people focusing on UK institutions. I would also ensure this research was broadly relevant, by collaborating with groups working on US AI policy (e.g. FHI, CSET and OpenAI).
I’m also asking for a time buyout to skill up in ML (~30%). This would improve my own ability to do high-quality research, by helping me to think clearly about how issues might evolve as capabilities advance, and how technical and policy approaches can best combine to influence the future impacts of AI.
The main work I know of Jess’s is her early involvement in 80,000 Hours. In the first 1-2 years of their existence, she wrote dozens of articles for them, and contributed to their culture and development. Since then I’ve seen her make positive contributions to a number of projects over the years - she has helped in some form with every EA Global conference I’ve organized (two in 2015 and one in 2016), and she’s continued to write publicly in places like the EA Forum, the EA Handbook, and news sites like Quartz and Vox. This background implies that Jess has had a lot of opportunities for members of the fund to judge her output. My sense is that this is the main reason that the other members of the fund were excited about this grant — they generally trust Jess’s judgment and value her experience (while being more hesitant about CFI’s work).
There are three things I looked into for this grant writeup: Jess’s policy research output, Jess’s blog, and the institutional quality of Leverhulme CFI. The section on Leverhulme CFI became longer than the section on Jess and was mostly unrelated to her work, so I’ve taken it out and included it as an addendum.
Impressions of Policy Papers
First is her policy research. The papers I read were from those linked on her blog. They were:
On the first paper, about focusing on tensions: the paper said that many “principles of AI ethics” that people publicly talk about in industry, non-profit, government and academia are substantively meaningless, because they don’t come with the sort of concrete advice that actually tells you how to apply them - and specifically, how to trade them off against each other. The part of the paper I found most interesting were four paragraphs pointing to specific tensions between principles of AI ethics. They were:
My sense is that while there is some good public discussion about AI and policy (e.g. OpenAI’s work on release practices seems quite positive to me), much conversation that brands itself as ‘ethics’ is often not motivated by the desire to ensure this novel technology improves society in accordance with our deepest values, but instead by factors like reputation, PR and politics.
There are many notions, like Peter Thiel’s “At its core, artificial intelligence is a military technology” or the common question “Who should control the AI?” which don’t fully account for the details of how machine learning and artificial intelligence systems work, or the ways in which we need to think about them in very different ways from other technologies; in particular, that we will need to build new concepts and abstractions to talk about them. I think this is also true of most conversations around making AI fair, inclusive, democratic, safe, beneficial, respectful of privacy, etc.; they seldom consider how these values can be grounded in modern ML systems or future AGI systems. My sense is that much of the best conversation around AI is about how to correctly conceptualize it. This is something that (I was surprised to find) Henry Kissinger’s article on AI did well; he spends most of the essay trying to figure out which abstractions to use, as opposed to using already existing ones.
The reason I liked that bit of Jess’s paper is that I felt the paper used mainstream language around AI ethics (in a way that could appeal to a broad audience), but then:
In the context of a public conversation that I feel is often substantially motivated by politics and PR rather than truth, seeing someone point clearly at important conceptual problems felt like a breath of fresh air.
That said, given all of the political incentives around public discussion of AI and ethics, I don’t know how papers like this can improve the conversation. For example, companies are worried about losing in the court of Twitter’s public opinion, and also are worried about things like governmental regulation, which are strong forces pushing them to primarily take popular but ineffectual steps to be more "ethical". I’m not saying papers like this can’t improve this situation in principle, only that I don’t personally feel like I have much of a clue about how to do it or how to evaluate whether someone else is doing it well, in advance of their having successfully done it.
Personally, I feel much more able to evaluate the conceptual work of figuring out how to think about AI and its strategic implications (two standout examples are this paper by Bostrom and this LessWrong post by Christiano), rather than work on revising popular views about AI. I’d be excited to see Jess continue with the conceptual side of her work, but if she instead primarily aims to influence public conversation (the other goal of that paper), I personally don’t think I’ll be able to evaluate and recommend grants on that basis.
From the second paper I read sections 3 and 4, which lists many safety and security practices in the fields of biosafety, computer information security, and institutional review boards (IRBs), then outlines variables for analysing release practices in ML. I found it useful, even if it was shallow (i.e. did not go into much depth in the fields it covered). Overall, the paper felt like a fine first step in thinking about this space.
In both papers, I was concerned with the level of inspiration drawn from bioethics, which seems to me to be a terribly broken field (cf. Scott Alexander talking about his IRB nightmare or medicine’s ‘culture of life’). My understanding is that bioethics coordinated a successful power grab (cf. OpenPhil’s writeup) from the field of medicine, creating hundreds of dysfunctional and impractical ethics boards that have formed a highly adversarial relationship with doctors (whose practical involvement with patients often makes them better than ethicists at making tradeoffs between treatment, pain/suffering, and dignity). The formation of an “AI ethics” community that has this sort of adversarial, unhealthy relationship with machine learning researchers would be an incredible catastrophe.
Overall, it seems like Jess is still at the beginning of her research career (she’s only been in this field for ~1.5 years). And while she’s spent a lot of effort on areas that don’t personally excite me, both of her papers include interesting ideas, and I’m curious to see her future work.
Impressions of Other Writing
Jess also writes a blog, and this is one of the main things that makes me excited about this grant. On the topic of AI, she wrote three posts (1, 2, 3), all of which made good points on at least one important issue. I also thought the post on confirmation bias and her PhD was quite thoughtful. It correctly identified a lot of problems with discussions of confirmation bias in psychology, and came to a much more nuanced view of the trade-off between being open-minded versus committing to your plans and beliefs. Overall, the posts show independent thinking written with an intent to actually convey understanding to the reader, and doing a good job of it. They share the vibe I associate with much of Julia Galef’s work - they’re noticing true observations / conceptual clarifications, successfully moving the conversation forward one or two steps, and avoiding political conflict.
I do have some significant concerns with the work above, including the positive portrayal of bioethics and the absence of any criticism toward the AAAI safety conference talks, many of which seem to me to have major flaws.
While I’m not excited about Leverhulme CFI’s work (see the addendum for details), I think it will be good for Jess to have free rein to follow her own research initiatives within CFI. And while she might be able to obtain funding elsewhere, this alternative seems considerably worse, as I expect other funding options would substantially constrain the types of research she’d be able to conduct.
Productivity coaching for effective altruists to increase their impact.
I plan to continue coaching high-impact EAs on productivity. I expect to have 600+ sessions with about 100 clients over the next year, focusing on people working in AI safety and EA orgs. I’ve worked with people at FHI, Open Phil, CEA, MIRI, CHAI, DeepMind, the Forethought Foundation, and ACE, and will probably continue to do so. Half of my current clients (and a third of all clients I’ve worked with) are people at these orgs. I aim to increase my clients’ output by improving prioritization and increasing focused work time.
I would use the funding to: offer a subsidized rate to people at EA orgs (e.g. between $10 and $50 instead of $125 per call), offer free coaching for select coachees referred by 80,000 Hours, and hire contractors to help me create materials to scale coaching.
You can view my impact evaluation (linked below) for how I’m measuring my impact so far.
(Lynette’s public self-evaluation is here.)
I generally think it's pretty hard to do "productivity coaching" as your primary activity, especially when you are young, due to a lack of work experience. This means I have a high bar for it being a good idea that someone should go full-time into the "help other people be more productive” business.
My sense is that Lynette meets that bar, but only barely (to be clear, I consider it to be a high bar). The main thing that she seems to be doing well is being very organized about everything that she is doing, in a way that makes me confident that her work has had a real impact — if not, I think she’d have noticed and moved on to something else.
However, as I say in the CFAR writeup, I have a lot of concerns with primarily optimising for legibility, and Lynette’s work shows some signs of this. She has shared around 60 testimonials on her website (linked here). Of these, not one of them mentioned anything negative, which clearly indicates that I can't straightforwardly interpret those testimonials as positive evidence (since any unbiased sampling process would have resulted in at least some negative datapoints). I much prefer what another applicant did here: they asked people to send us information anonymously, which increased the chance of our hearing opinions that weren’t selected to create a positive impression. As is, I think I actually shouldn't update much on the testimonials, in particular given that none of them go into much detail on how Lynette has helped them, and almost all of them share a similar structure.
Reflecting on the broader picture, I think that Lynette’s mindset reflects how I think many of the best operations staff I’ve seen operate: aim to be productive by using simple output metrics, and by doing things in a mindful, structured way (as opposed to, for example, trying to aim for deep transformative insights more traditionally associated with psychotherapy). There is a deep grounded-ness and practical nature to it. I have a lot of respect for that mindset, and I feel as though it's underrepresented in the current EA/rationality landscape. My inside-view models suggest that you can achieve a bunch of good things by helping people become more productive in this way.
I also think that this mindset comes with a type of pragmatism that I am more concerned about, and often gives rise to what I consider unhealthy adversarial dynamics. As I discussed above, it’s difficult to get information from Lynette’s positive testimonials. My sense is that she might have produced them by directly optimising for “getting a grant” and trying to give me lots of positive information, leading to substantial bias in the selection process. The technique of ‘just optimize for the target’ is valuable in lots of domains, but in this case was quite negative.
That said, framing her coaching as achieving a series of similar results generally moves me closer to thinking about this grant as "coaching as a commodity". Importantly, few people reported very large gains in their productivity; the testimonials instead show a solid stream of small improvements. I think that very few people have access to good coaching, and the high variance in coach quality means that experimenting is often quite expensive and time-consuming. Lynette seems to be able to consistently produce positive effects in the people she is working with, making her services a lot more valuable due to greater certainty around the outcome. (However, I also assign significant probability that the way the evaluation questions were asked reduced the rate at which clients reported either negative or highly positive experiences.)
I think that many productivity coaches fail to achieve Lynette’s level of reliability, which is one of the key things that makes me hopeful about her work here. My guess is that the value-add of coaching is often straightforwardly positive unless you impose significant costs on your clients, and Lynette seems quite good at avoiding that by primarily optimizing for professionalism and reliability.
This grant was recommended by the Fund, but ultimately was funded by a private donor, who (prior to CEA finalizing its standard due diligence checks) had personally offered to make this donation instead. As such, the grant recommendation was withdrawn.
Oliver Habryka had created a full writeup by that point, so it is included below.
Help promising people to reason more effectively and find high-impact work, such as reducing x-risk.
The Center for Applied Rationality runs workshops that promote particular epistemic norms—broadly, that beliefs should be true, bugs should be solved, and that intuitions/aversions often contain useful data. These workshops are designed to cause potentially impactful people to reason more effectively, and to find people who may be interested in pursuing high-impact careers (especially AI safety).
Many of the people currently working on AI safety have been through a CFAR workshop, such as 27% of the attendees at the 2019 FLI conference on Beneficial AI in Puerto Rico, and for some of those people it appears that CFAR played a causal role in their decision to switch careers. In the confidential section, we list some graduates from CFAR programs who subsequently decided to work on AI safety, along with our estimates of the counterfactual impact of CFAR on their decision [16 at MIRI, 3 on the OpenAI safety team, 2 at CHAI, and one each at Ought, Open Phil and the DeepMind safety team].
Recruitment is the most legible form of impact CFAR has, and is probably its most important—the top reported bottleneck in the last two years among EA leaders at Leaders Forum, for example, was finding talented employees.
In 2019, we expect to run or co-run over 100 days of workshops, including our mainline workshop (designed to grow/improve the rationality community), workshops designed specifically to recruit programmers (AIRCS) and mathematicians (MSFP) to AI safety orgs, a 4-weekend instructor training program (to increase our capacity to run workshops), and alumni reunions in both the United States and Europe (to grow the EA/rationality community and cause impactful people to meet/talk with one another). Broadly speaking, we intend to continue doing the sort of work we have been doing so far.
In our last grant round, I took an outside view on CFAR and said that, in terms of output, I felt satisfied with CFAR's achievements in recruitment, training and the establishment of communal epistemic norms. I still feel this way about those areas, and my writeup last round still seems like an accurate summary of my reasons for wanting to grant to CFAR. I also said that most of my uncertainty about CFAR lies in its long-term strategic plans, and I continue to feel relatively confused about my thoughts on that.
I find it difficult to explain my thoughts on CFAR, and I think that a large fraction of this difficulty comes from CFAR being an organization that is intentionally not optimizing towards being easy to understand from the outside, having simple metrics, or more broadly being legible. CFAR is intentionally avoiding being legible to the outside world in many ways. This decision is not obviously wrong, as I think it brings many positives, but I think it is the cause of me feeling particularly confused about how to talk coherently about CFAR.
Considerations around legibility
Summary: CFAR’s work is varied and difficult to evaluate. This has some good features — it can avoid focusing too closely on metrics that don’t measure impact well — but also forces evaluators to rely on factors that aren’t easy to measure, like the quality of its internal culture. On the whole, while I wish CFAR were somewhat more legible, I appreciate the benefits to CFAR’s work of not maximizing “legibility” at the cost of impact or flexibility.
To help me explain my point, let's contrast CFAR with an organization like AMF, which I think of as exceptionally legible. AMF’s work, compared to many other organizations with tens of millions of dollars on hand, is easy to understand: they buy bednets and give them to poor people in developing countries. As long as AMF continues to carry out this plan, and provides basic data showing its success in bednet distribution, I feel like I can easily model what the organization will do. If I found out that AMF was spending 10% of its money funding religious leaders in developing countries to preach good ethical principles for society, or funding the campaigns of government officials favorable to their work, I would be very surprised and feel like some basic agreement or contract had been violated — regardless of whether I thought those decisions, in the abstract, were good or bad for their mission. AMF claims to distribute anti-malaria bednets, and it is on this basis that I would choose whether to support them.
AMF could have been a very different organization, and still could be if it wanted to. For example, it could conduct research on various ways to effect change, and give its core staff the freedom to do whatever they thought was best. This new AMF (“AMF 2.0”) might not be able to tell you exactly what they’ll do next year, because they haven’t figured it out yet, but they can tell you that they’ll do whatever their staff determine is best. This could be distributing deworming pills, pursuing speculative medical research, engaging in political activism, funding religious organizations, etc.
If GiveWell wanted to evaluate AMF 2.0, they would need to use a radically different style of reasoning. There wouldn’t be a straightforward intervention with RCTs to look into. There wouldn’t be a straightforward track record of impact from which to extrapolate. Judging AMF 2.0 would require GiveWell to form much more nuanced judgments about the quality of thinking and execution of AMF’s staff, to evaluate the quality of its internal culture, and to consider a host of other factors that weren’t previously relevant.
I think that evaluating CFAR requires a lot of that kind of analysis, which seems inherently harder to communicate to other people without summarizing one’s views as: "I trust the people in that organization to make good decisions."
The more general idea here is that organizations are subject to bandwidth constraints - they often want to do lots of different things, but their funders need to be able to understand and predict their behavior with limited resources for evaluation. As I've written about recently, a key variable for any organization is the people and organizations by which they are trying to be understood and held accountable. For charities that receive most of their funding in small donations from a large population of people who don’t know much about them, this is a very strong constraint; they must communicate their work so that people can understand it very quickly with little background information. If a charity instead receives most of its funding in large donations from a small set of people who follow it closely, it can communicate much more freely, because the funders will be able to spend a lot of their time talking to the org, exchanging models, and generally coming to an understanding of why the org is doing what it’s doing.
This idea partly explains why most organizations tend to focus on legibility, in how they talk about their work and even in the work they choose to pursue. It can be difficult to attract resources and support from external parties if one’s work isn’t legible.
I think that CFAR is still likely optimizing too little towards legibility, compared to what I think would be ideal for it. Being legible allows an organization to be more confident that its work is having real effects, because it acquires evidence that holds up to a variety of different viewpoints. However, I think that far too many organizations (nonprofit and otherwise) are trying too hard to make their work legible, in a way that reduces innovation and also introduces a variety of adversarial dynamics. When you make systems that can be gamed, and which carry rewards for success (e.g. job stability, prestige, etc), people will reliably turn up to game them.
(As Jacob Lagerros has written in his post on Unconscious Economics, this doesn’t mean people are consciously gaming your system, but merely that this behavior will eventually transpire. The many causes of this include selection effects, reinforcement learning, and memetic evolution.)
In my view, CFAR, by not trying to optimize for a single, easy-to-explain metric, avoids playing the “game” many nonprofits play of focusing on work that will look obviously good to donors, even if it isn’t what the nonprofit believes would be most impactful. They also avoid a variety of other games that come from legibility, such as job applicants getting very good at faking the signals that they are a good fit for an organization, making it harder for them to find good applicants.
Optimizing for communication with the goal of being given resources introduces adversarial dynamics; someone asking for money may provide limited/biased information that raises the chance they’ll be given a grant but reduces the accuracy of the grantmaker’s understanding. (See my comment in Lynette’s writeup below for an example of how this can arise.) This optimization can also tie down your resources, forcing you to carry out commitments you made for the sake of legibility, rather than doing what you think would be most impactful.
So I think that it's important that we don't force all organizations towards maximal legibility. (That said, we should ensure that organizations are encouraged to pursue at least some degree of legibility, since the lack of legibility also gives rise to various problems.)
Do I trust CFAR to make good decisions?
As I mentioned in my initial comments on CFAR, I generally think that the current projects CFAR is working on are quite valuable and worth the resources they are consuming. But I have a lot of trouble modeling CFAR’s long-term planning, and I feel like I have to rely instead on my models of how much I trust CFAR to make good decisions in general, instead of being able to evaluate the merits of their actual plans.
That said, I do generally trust CFAR's decision-making. It’s hard to explain the evidence that causes me to believe this, but I’ll give a brief overview anyway. (This evidence probably won’t be compelling to others, but I still want to give an accurate summary of where my beliefs come from):
 The focus on ‘legibility’ in this context I take from James C. Scott’s book “Seeing Like a State.” It was introduced to me by Elizabeth Van Nostrand in this blogpost discussing it in the context of GiveWell and good giving; Scott Alexander also discussed it in his review of the book . Here’s an example from Scott regarding centralized planning and governance:
the centralized state wanted the world to be “legible”, ie arranged in a way that made it easy to monitor and control. An intact forest might be more productive than an evenly-spaced rectangular grid of Norway spruce, but it was harder to legislate rules for, or assess taxes on.
 The errors that follow are all forms of Goodhart’s Law, which states that “any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.”
 The benefits of (and forces that encourage) stability and reliability can maybe be most transparently understood in the context of menu costs and the prevalence of highly sticky wages.
I wrote the following in the course of thinking about the grant to Jess Whittlestone. While the grant is to support Jess’s work, the grant money will go to Leverhulme CFI, which will maintain discretion about whether to continue employing her, and will likely influence what type of work she will pursue.
As such, it seems important to not only look into Jess’s work, but also look into Leverhulme CFI and its sister organization, the Centre for the Study of Existential Risk (CSER). While my evaluation of the organization that will support Jess during her postdoc is relevant to my evaluation of the grant, it is quite long and does not directly discuss Jess or her work, so I’ve moved it into a separate section.
I’ve read a few papers from CFI and CSER over the years, and heard many impressions of their work from other people. For this writeup, I wanted to engage more concretely with their output. I reread and reviewed an article published in Nature earlier this year called Bridging near- and long-term concerns about AI, written by the Executive Directors at Leverhulme CFI and CSER respectively, Stephen Cave and Seán ÓhÉigeartaigh.
Summary and aims of the article
The article’s summary:
Debate about the impacts of AI is often split into two camps, one associated with the near term and the other with the long term. This divide is a mistake — the connections between the two perspectives deserve more attention, say Stephen Cave and Seán S. ÓhÉigeartaigh.
This is not a position I hold, and I’m going to engage with the content below in more detail.
Overall, I found the claims of the essay hard to parse and often ambiguous, but I’ve attempted to summarize what I view as its three main points:
They say “These three points relate to ways in which addressing near-term issues could contribute to solving potential long-term problems.”
If I ask myself what Leverhulme/CSER’s goals are for this document, it feels to me like it is intended as a statement of diplomacy. It’s saying that near-term and long-term AI risk work are split into two camps, but that we should be looking for common ground (“the connections between the two perspectives deserve more attention”, “Learning from the long term”). It tries to emphasize shared values (“Connected research priorities”) and the importance of cooperation amongst many entities (“The challenges we will face are likely to require deep interdisciplinary and intersectoral collaboration between industries, academia and policymakers, alongside new international agreements”). The goal that I think it is trying to achieve is to negotiate trade and peace between the near-term and long-term camps by arguing that “This divide is a mistake”.
Drawing the definitions does a lot of work
The authors define “long-term concerns” with the following three examples:
wide-scale loss of jobs, risks of AI developing broad superhuman capabilities that could put it beyond our control, and fundamental questions about humanity’s place in a world with intelligent machines
Despite this broad definition, they only use concrete examples from the first category, which I would classify as something like “mid-term issues.” I think the possibility of even wide-scale loss of jobs, unless interpreted extremely broadly, is something that does not make sense to put into the same category as the other two, which are primarily concerned with stakes that are orders of magnitude higher (such as the future of the human species). I think this conflation of very different concerns causes the rest of the article to make an argument that is more likely to mislead than to inform.
After this definition, the article failed to mention any issue that I would classify as representative of the long-term concerns of Nick Bostrom or Max Tegmark, both of whom are cited by the article to define “long-term issues.” (In Tegmark’s book Life 3.0, he explicitly categorizes unemployment as a short-term concern, to be distinguished from long-term concerns.)
Conceptual confusions in short- and mid-term policy suggestions
The article has the following policy idea:
Take explainability (the extent to which the decisions of autonomous systems can be understood by relevant humans): if regulatory measures make this a requirement, more funding will go to developing transparent systems, while techniques that are powerful but opaque may be deprioritized.
(Let me be clear that this is not explicitly listed as a policy recommendation.)
My naive prior is that there is no good AI regulation a government could establish today. I continue to feel this way after looking into this case (and the next example below). Let me explain why in this case the idea that regulation requiring explainability would encourage transparent + explainable systems is false.
Modern ML systems are not doing a type of reasoning that is amenable to explanation in the way human decisions often are. There is not a principled explanation of their reasoning when deciding whether to offer you a bank loan, there is merely a mass of correlations between spending history and later reliability, which may factorise into a small number of well-defined chunks like “how regularly someone pays their rent” but it might not. The main problem with the quoted paragraph is that it does not at all attempt to specify how to define explainability in an ML system to the point where it can be regulated, meaning that any regulation would either be meaningless and ignored, or worse highly damaging. Policies formed in this manner will either be of no consequence, or deeply antagonise the ML community. We currently don’t know how to think about explainability of ML systems, and ignoring that problem and regulating that they should be ‘explainable’ will not work.
The article also contains the following policy idea about autonomous weapons.
The decisions we make now, for example, on international regulation of autonomous weapons, could have an outsized impact on how this field develops. A firm precedent that only a human can make a ‘kill’ decision could significantly shape how AI is used — for example, putting the focus on enhancing instead of replacing human capacities.
Here and throughout the article, repeated uses of the conditional ‘could’ make it unclear to me whether this is being endorsed or merely suggested. I can’t quite tell if they think that drone swarms are a long-term issue - they contrast it with a short-term issue but don’t explicitly say that it is long-term. Nonetheless, I think their suggesting it here is also a bit misguided.
Let me contrast this with Nick Bostrom on a recent episode of the Joe Rogan Experienceexplaining that he thinks that the specific rule has ambiguous value. Here’s a quote from a discussion of the campaign to ban lethal autonomous weapons:
Nick Bostrom: I’ve kind of stood a little bit on the sidelines on that particular campaign, being a little unsure exactly what it is that… certainly I think it’d be better if we refrained from having some arms race to develop these than not. But if you start to look in more detail: What precisely is the thing that you’re hoping to ban? So if the idea is the autonomous bit, that the robot should not be able to make its own firing decision, well, if the alternative to that is there is some 19-year old guy sitting in some office building and his job is whenever the screen flashes ‘fire now’ he has to press a red button. And exactly the same thing happens. I’m not sure how much is gained by having that extra step.
Interviewer: But it feels better for us for some reason. If someone is pushing the button.
Nick Bostrom: But what exactly does that mean. In every particular firing decision? Well, you gotta attack this group of surface ships here, and here are the general parameters, and you’re not allowed to fire outside these coordinates? I don’t know. Another is the question of: it would be better if we had no wars, but if there is gonna be a war, maybe it is better if it’s robots v robots. Or if there’s gonna be bombing, maybe you want the bombs to have high precision rather than low precision - get fewer civilian casualties.
On the other hand you could imagine it reduces the threshold for going to war, if you think that you wouldn’t fear any casualties you would be more eager to do it. Or if it proliferates and you have these mosquito-sized killer-bots that terrorists have. It doesn’t seem like a good thing to have a society where you have a facial-recognition thing, and then the bot flies out and you just have a kind of dystopia.
Overall, it seems that in both situations, the key open questions are in understanding the systems and how they’ll interface with areas of industry, government and personal life, and that regulation based on inaccurate conceptualizations of the technology would either be meaningless or harmful.
Polarizing approach to policy coordination
I have two main concerns with what I see as the intent of the paper.
The first one can be summarized by Robin Hanson’s article To Oppose Polarization, Tug Sideways:
The policy world can [be] thought of as consisting of a few Tug-O-War "ropes" set up in this high dimensional policy space. If you want to find a comfortable place in this world, where the people around you are reassured that you are "one of them," you need to continually and clearly telegraph your loyalty by treating each policy issue as another opportunity to find more supporting arguments for your side of the key dimensions. That is, pick a rope and pull on it.
If, however, you actually want to improve policy, if you have a secure enough position to say what you like, and if you can find a relevant audience, then [you should] prefer to pull policy ropes sideways. Few will bother to resist such pulls, and since few will have considered such moves, you have a much better chance of identifying a move that improves policy. On the few main dimensions, not only will you find it very hard to move the rope much, but you should have little confidence that you actually have superior information about which way the rope should be pulled.
I feel like the article above is not pulling policy ropes sideways, but is instead connecting long-term issues to specific sides of existing policy debates, around which there is already a lot of tension. The issue of technological unemployment seems to me to be a highly polarizing topic, where taking a position seems ill-advised, and I have very low confidence about the correct direction in which to pull policy. Entangling long-term issues with these highly tense short-term issues seems like it will likely reduce our future ability to broadly coordinate on these issues (by having them associated with highly polarized existing debates).
Distinction between long- and short-term thinking
My second concern is that on a deeper level, I think that the type of thinking that generates a lot of the arguments around concerns for long-term technological risks is very different from that which suggests policies around technological unemployment and racial bias. I think there is some value in having these separate ways of thinking engage in “conversation,” but I think the linked paper is confusing in that it seems to try to down-play the differences between them. An analogy might be the differences between physics and architecture; both fields nominally work with many similar objects, but the distinction between the two is very important, and the fields clearly require different types of thinking and problem-solving.
Some of my concerns are summarized by Eliezer in his writing on Pivotal Acts:
...compared to the much more difficult problems involved with making something actually smarter than you be safe, it may be tempting to try to write papers that you know you can finish, like a paper on robotic cars causing unemployment in the trucking industry, or a paper on who holds legal liability when a factory machine crushes a worker. But while it's true that crushed factory workers and unemployed truckers are both, ceteris paribus, bad, they are not astronomical catastrophes that transform all galaxies inside our future light cone into paperclips, and the latter category seems worth distinguishing...
...there will [...] be a temptation for the grantseeker to argue, "Well, if AI causes unemployment, that could slow world economic growth, which will make countries more hostile to each other, which would make it harder to prevent an AI arms race." But the possibility of something ending up having a non-zero impact on astronomical stakes is not the same concept as events that have a game-changing impact on astronomical stakes. The question is what are the largest lowest-hanging fruit in astronomical stakes, not whether something can be argued as defensible by pointing to a non-zero astronomical impact.
I currently don’t think that someone who is trying to understand how to deal with technological long-term risk should spend much time thinking about technological unemployment or related issues, but it feels like the paper is trying to advocate for the opposite position.
Concluding thoughts on the article
Many people in the AI policy space have to spend a lot of effort to gain respect and influence, and it’s genuinely hard to figure out a way to do this while acting with integrity. One common difficulty in this area is navigating the incentives to connect one’s arguments to issues that already get a lot of attention (e.g. ongoing political debates). My read is that this essay makes these connections even when they aren’t justified; it implies that many short- and medium-term concerns are a natural extension of current long-term thought, while failing to accurately portray what I consider to be the core arguments around long-term risks and benefits from AI. It seems like the effect of this essay will be to reduce perceived differences between long-term, mid-term and short-term work on risks from AI, to cause confusion about the actual concerns of Bostrom et al., and to make future communications work in this space harder and more polarized.
Broader thoughts on CSER and CFI
I only had the time and space to critique one specific article from CFI and CSER. However, from talking to others working in the global catastrophic risk space, and from engagement with significant fractions of the rest of CSER and CFI’s work, I've come to think that the problems I see in this article are mostly representative of the problems I see in CSER’s and CFI’s broader strategy and work. I don’t think what I’ve written sufficiently justifies that claim; however, it seems useful to share this broader assessment to allow others to make better predictions about my future grant recommendations, and maybe also to open a dialogue that might cause me to change my mind.
Overall, based on the concerns I’ve expressed in this essay, and that I’ve had with other parts of CFI and CSER’s work, I worry that their efforts to shape the conversation around AI policy, and to mend disputes between those focused on long-term and short-term problems, do not address important underlying issues and may have net-negative consequences.
That said, it’s good that these organizations give some researchers a way to get PhDs/postdocs at Cambridge with relatively little institutional oversight and an opportunity to explore a large variety of different topics (e.g. Jess, and Shahar Avin, a previous grantee whose work I’m excited about).
I wrote the following in the course of writing about the AI Safety Camp. This is a model I use commonly when thinking about funding for AI alignment work, but it ended up not being very relevant to that writeup, so I’m leaving it here as a note of interest.
My understanding of many parts of technical academia is that there is a strong incentive to make your writing hard to understand while appearing more impressive by using a lot of math. Eliezer Yudkowsky describes his understanding of it as such (and expands on this further in the rocket alignment problem):
The point of current AI safety work is to cross, e.g., the gap between [. . . ] saying “Ha ha, I want AIs to have an off switch, but it might be dangerous to be the one holding the off switch!” to, e.g., realizing that utility indifference is an open problem. After this, we cross the gap to solving utility indifference in unbounded form. Much later, we cross the gap to a form of utility indifference that actually works in practice with whatever machine learning techniques are used, come the day.
Progress in modern AI safety mainly looks like progress in conceptual clarity — getting past the stage of “Ha ha it might be dangerous to be holding the off switch.” Even though Stuart Armstrong’s original proposal for utility indifference completely failed to work (as observed at MIRI by myself and Benya), it was still a lot of conceptual progress compared to the “Ha ha that might be dangerous” stage of thinking.
Simple ideas like these would be where I expect the battle for the hearts of future grad students to take place; somebody with exposure to Armstrong’s first simple idea knows better than to walk directly into the whirling razor blades without having solved the corresponding problem of fixing Armstrong’s solution. A lot of the actual increment of benefit to the world comes from getting more minds past the “walk directly into the whirling razor blades” stage of thinking, which is not complex-math-dependent.
Later, there’s a need to have real deployable solutions, which may or may not look like impressive math per se. But actual increments of safety there may be a long time coming. [. . . ]
Any problem whose current MIRI-solution looks hard (the kind of proof produced by people competing in an inexploitable market to look impressive, who gravitate to problems where they can produce proofs that look like costly signals of intelligence) is a place where we’re flailing around and grasping at complicated results in order to marginally improve our understanding of a confusing subject matter. Techniques you can actually adapt in a safe AI, come the day, will probably have very simple cores — the sort of core concept that takes up three paragraphs, where any reviewer who didn’t spend five years struggling on the problem themselves will think, “Oh I could have thought of that.” Someday there may be a book full of clever and difficult things to say about the simple core — contrast the simplicity of the core concept of causal models, versus the complexity of proving all the clever things Judea Pearl had to say about causal models. But the planetary benefit is mainly from posing understandable problems crisply enough so that people can see they are open, and then from the simpler abstract properties of a found solution — complicated aspects will not carry over to real AIs later.
And gives a concrete example here:
The journal paper that Stuart Armstrong coauthored on "interruptibility" is a far step down from Armstrong's other work on corrigibility. It had to be dumbed way down (I'm counting obscuration with fancy equations and math results as "dumbing down") to be published in a mainstream journal. It had to be stripped of all the caveats and any mention of explicit incompleteness, which is necessary meta-information for any ongoing incremental progress, not to mention important from a safety standpoint. The root cause can be debated but the observable seems plain. If you want to get real work done, the obvious strategy would be to not subject yourself to any academic incentives or bureaucratic processes. Particularly including peer review by non-"hobbyists" (peer commentary by fellow "hobbyists" still being potentially very valuable), or review by grant committees staffed by the sort of people who are still impressed by academic sage-costuming and will want you to compete against pointlessly obscured but terribly serious-looking equations.
(Here is a public example of Stuart’s work on utility indifference, though I had difficulty finding the most relevant examples of his work on this subject.)
Some examples that seem to me to use an appropriate level of formalism include: the Embedded Agency sequence, the Mesa-Optimisation paper, some posts by DeepMind researchers (thoughts on human models, classifying specification problems as variants of Goodhart’s law), and many other blog posts by these authors and others on the AI Alignment Forum.
There’s a sense in which it’s fine to play around with the few formalisms you have a grasp of when you’re getting to grips with ideas in this field. For example, MIRI recently held a retreat for new researchers, which led to a number of blog posts that followed this pattern (1, 2, 3, 4). But aiming for lots of technical formalism is not helpful - any conception of useful work that focuses primarily on molding the idea to the format rather than molding the format to the idea, especially for (nominally) impressive technical formats, is likely optimizing for the wrong metric and falling prey to Goodhart’s law.