I have a few questions to the subset of readers who:
Believe technical AI alignment research is both important and hard to make significant progress in
Have a personal connection with a person who doesn’t know much about AI alignment, but who you think would have a real chance to make valuable contributions to the field if they entered it (or perhaps you know someone who cares about AI risk and have such a personal connection, and you have enough knowledge to talk on their behalf). It may be your friend, colleague, supervisor, etc.
I would love to hear your thoughts on some of the following questions:
What reasons prevent you from introducing them to AI alignment by e.g. scheduling time with them and talking about some of the motivations and open problems in the field?
If you’ve tried something like this, how did it go?
What factors do you think would increase your willingness to bring AI alignment to their attention and/or the potential value resulting from it? Bonus points for reasonably low-hanging fruit here.
When I worked a FAANG research job, my experience was that it was socially punishable to bring up AI alignment research in just about any context, with exceptions as it was relevant to the team’s immediate mission, for example robustness on the scale required for medical decisions (a much smaller scale than AGI ruin, but a notably larger scale, in the sense of errors being costly, than most deep learning systems in production use at the time).
I find that in some social spaces, Rationality/EA-adjacent ones in particular, it’s seen as distracting, rude, and low status to emphasize a hobby horse social justice issue at the expense of whatever else is being discussed. This is straightforward when “whatever else is being discussed” is AI alignment, which the inside view privileges roughly as “more important than everything else, with vague exceptions when the mental health of high-value people who might otherwise do productive work on the topic is at stake.”
On a medical research team, I took a little too long to realize that I’d implicitly bought into a shared vision of what’s important. We were going to save lives! We weren’t going to cure cancer–everyone falls for that trap, aiming too high. We’re working on the ground, saving real people, on real timescales. Computer vision can solve the disagreement-among-experts problem in all sorts of medical classification problems, and we’re here to fight that fight and win.
So you’ve gathered a team of AI researchers, some expert, some early-career, to finally take a powerful stab at the alignment problem. A new angle, or more funding, or the right people in the room, whatever belief of comparative advantage you have that inspires hope beyond death with dignity. And you have someone on your team who deeply cares about a complicated social issue you don’t understand. Maybe this is their deepest mission, and they see this early-engineer position at your new research org as a stepping stone toward the fairness and accessibility team at Brain that’s doing the real work. They do their best to contribute in the team’s terms of what’s valuable, and they censor themselves constantly, waiting for the right moment to make the pivotal observation that there’s not a single cis woman in the room, or that the work we’re doing here may be building a future that’s even more hostile toward people with developmental disabilities, or this adversarial training scheme has some alarming implications when you consider that the system could learn race as a feature even if we exclude it from the dataset, or something.
I think this is a fair analogue to my situation, and I expect more broadly among people already doing AI research toward a goal other than alignment. It’s
Distracting: We have something else we’re working on, and that is a deep question, and you probably could push hard enough on me to nerd snipe me with it if I don’t put up barriers.
Rude: It implies that the work we’re doing here, which we all care deeply about (right?) is problematic for reasons well outside our models of who we are and what we’re responsible for, and challenging that necessitates a bunch of complicated shadow work.
Low status: Wait, are you one of those LessWrong people? I bet you’re anti-woke and think James Damore shouldn’t have been fired, huh? And you’re so wound up in your privilege bubble that you think this AGI alarmism is more important than the struggles of real underprivileged people who we know actually exist, here, now? Got it.
I’m being slightly unfair in implying that these are literally interactions I had with real people in the industry. This is more representative of my experiences online and in other spaces with less of a backdrop of professional courtesy. At [FAANG company] these interactions were subtler.
This story is meant to provide answers to your questions 1 and 2. As far as question 3 and making a change, I’m bullish on narratives, aesthetics, anthropology and the like as genuine interventions upstream of AI safety. We’re in a social equilibrium where only certain sorts of people can move into AI safety without seriously disrupting the means by which their social needs are met. There are many wonderful people in that set, but it is relatively quite small compared to the set of people who, if they were convinced to genuinely try, could contribute meaningfully.
I would guess this doesn’t appear to qualify for bonus points for being reasonably low-hanging. I come from an odd place though: personally sufficiently traumatized by my experiences in AI research that in practical terms contributing there is more or less off limits for me for the time being, yet compelled by AGI ruin narratives and experienced with substantial relevant technical background. So at least for me, this is the way forward.
Scott Aaronson recently wrote something relevant to these issues:
There is also a question on EA Forum about the same issue: What are the coolest topics in AI safety, to a hopelessly pure mathematician?
I wonder how valuable it would be to have a high quality post or sequence on open problems in AI alignment that is substantially optimized for nerd sniping. Is it even possible to make something like this?
Extremely valuable I’d guess, but the whole problem is that alignment is still preparadigmatic. We don’t actually know yet what the well-defined nerd snipe questions we should be asking are.
I think that preparadigmatic research and paradigmatic research are two different skill sets, and most Highly Impressive People in mainstream STEM are masters at the later, not the former.
I do think we’re more paradigmatic than we were a year ago, and that we might transition fully some time soon. I’ve got a list of concrete experiments on modularity in ML systems I’d like run for example, and I think any ML savvy person could probably do those, no skill at thinking about fuzzy far mode things required.
So I’m not sure a sequence like this could be written today, but maybe in six months?
I stream-of-consciousness’d this out and I’m not happy with how it turned out, but it’s probably better I post this than delete it for not being polished and eloquent. Can clarify with responses in comments.
Glad you posted this and I’m also interested in hearing what others say. I’ve had these questions for myself in tiny bursts throughout the last few months.
When I get the chance to speak to people earlier in their career stage than myself (starting undergrad, or is a high schooler attending a mathcamp I went to) who are undecided about their careers, I bring up my interest in AI Alignment and why I think it’s important, and share resources for them after the call in case they’re interested in learning more about it. I don’t have very many opportunities like this because I don’t actively seek to identify and “recruit” them. I only bring it up by happenstance (e.g. joining a random discord server for homotopy type theory, seeing an intro by someone who went to the same mathcamp as me and is interested in cogsci, and scheduling a call to talk about my research background in cogsci and how my interests have evolved/led me to alignment over time).
I know very talented people who are around my age at MIT and from a math program I attended; students who are breezing by technical double majors with perfect GPAs, IMO participants, good competitive programmers, etc. Some things that make it hard for me:
If I know them well, I can talk about my research interests and try to get them to see my motivation, but if I’m only catching up with them 1-2x a year, it feels very unnatural and synthetic for me to be spending that time trying to convert them into doing alignment work. If I am still very close to them / talk to them frequently, there’s still an issue of bringing it up naturally and having a chance to convince them. Most of these people are doing Math PhDs, or trading in finance, or working on a startup, or… The point is that they are fresh on their sprint down the path that they have chosen. They are all the type who are very focused and determined to succeed on the goals they have settled on. It is not “easy” to get them (or for this matter, almost any college student) to halt their “exploit” mode, take 10 steps back and lots of time from their busy lives, and then “explore” another option that I’m seemingly imposing onto them. FWIW, the people I know who are in trading seem to be the most likely to switch out (explicitly have told me in conversations that they just enjoy the challenge of the work, but want to find more fulfilling things down the road. And to these people I share ideas and resources about AI Safety.)
I shared resources after the call, talked about why I’m interested in alignment, and that’s the furthest I’ve gone wrt potentially converting someone who is already in a separate career track, to consider alignment.
If it was MUCH easier to convince people that ai alignment is worth thinking about in under an hour, and I could reach out to people to talk to me about this for a hour without looking like a nutjob and potentially damaging our relationship because it seems like I’m just trying to convert them into something else, AND the field of AI Alignment was more naturally compelling for them to join, I’d do much more of this outreach. On that last point, what I mean is: for one moment, let’s suspend the object level importance of solving AI Alignment. In reality, there are things that are incredibly important/attractive for people when pursuing a career. Status, monetary compensation, and recognition (and not being labeled a nutjob) are some big ones. If these things were better (and I think they are getting much better recently), it would be easier to get people to spend more time at least thinking about the possibility of working on AI Alignment, and eventually some would work on it because I don’t think the arguments for x-risk from AI are hard to understand. If I personally didn’t have so much support by way of programs the community had started (SERI, AISC, EA 1-1s, EAG AI Safety researchers making time to talk to me), or it felt like the EA/X-risk group was not at all “prestigious”, I don’t know how engaged I would’ve been in the beginning, when I started my own journey learning about all this. As much as I wish it weren’t true, I would not be surprised at all if the first instinctual thing that led me down this road was noticing that EAs/LW users were intelligent and had a solidly respectable community, before choosing to spend my time engaging with the content (a lot of which was about X-risks).
This is imo the biggest factor holding back (people going into) AI safety research by a wide margin. I personally know at least one very talented engineer who would currently be working on AI safety if the pay was anywhere near what they could make working for big tech companies.
I heard recently that at least some of the AI safety groups are making offers competitive with major tech companies. (Is this not the case?)
Could be! The highest I recall is ~$220k/year at Anthropic, with the stipulation that you must live in the Bay (they offer partial remote). Compared to big tech, there is less career capital/informal status due to reduced name recognition. Additionally, at big tech, there exists the possibility of very lucrative internal promotion.
See: Have You Tried Hiring People?
I sympathise with this view, but I think there’s a huge issue with treating “AI safety research” as a kind of fixed value here. We need more of the right kinds of AI safety research. Not everything labelled “AIS research” qualifies.
Caveats—the below applies:
To roles to the extent that they’re influencing research directions. (so less for engineers than scientists—but still somewhat)
To people who broadly understand the AIS problem.
“AIS research” will look very different depending on whether motivations are primarily intrinsic or extrinsic. The danger with extrinsic motivation is that you increase the odds that the people so motivated will pursue [look like someone doing important work] rather than [do important work]. This is a much bigger problem in areas where we don’t have ground truth feedback, and our estimates are poor, e.g. AI safety.
The caricature is that by incentivizing monetarily we get a load of people who have “AI Safety Research” written on their doors, can write convincing grant proposals and publish papers, but are doing approximately nothing to solve important problems.
It’s not clear to me how far the reality departs from such a caricature.
It’s also unclear to me whether this kind of argument fails for engineers: heading twice as fast in even slightly the wrong direction may be a poor trade.
I think it’s very important to distinguish financial support from financial incentive here.
By all means pay people in the bay area $100k or $125k
If someone needs >$250k before they’ll turn up (without a tremendously good reason), then I don’t want them in a position to significantly influence research directions.
This is less clear when people filter out non-competitive salary areas before learning about them. However, I’d want to solve this problem in a different way than [pay high salaries so that people learn about the area].
[Edit: I initially thought of this purely tongue in cheek, but maybe there is something here that is worth examining further?]
You have cognitively powerful agents (highly competent researchers) who have incentives (250k+ salaries) to do things that you don’t want them to do (create AGIs that are likely unaligned), and you want them to instead do things that benefit humanity (work on alignment) instead.
It seems to me that offering $100k salaries to work for you instead is not an effective solution to this alignment problem. It relies on the agents being already aligned to the extent that a $150k/yr loss is outweighed by other incentives.
If money were not a tight constraint, it seems to me that offering $250k/yr would be worthwhile even if for no other reason than having them not work on racing to AGI.
The “pay to not race to AGI” would only make sense if there were a smallish pool of replacements ready to step in and work on racing to AGI. This doesn’t seem to be the case. The difference it’d make might not be zero, but close enough to be obviously inefficient.
In particular, there are presumably effective ways to use money to create a greater number of aligned AIS researchers—it’s just that [give people a lot of money to work on it] probably isn’t one of them.
In those terms the point is that paying $250k+ does not align them—it simply hides the problem of their misalignment.
[work on alignment] does not necessarily benefit humanity, even in expectation. That requires [the right kind of work on alignment] - and we don’t have good tests for that. Aiming for the right kind of work is no guarantee you’re doing it—but it beats not aiming for it. (again, argument is admittedly less clear for engineers)
Paying $100k doesn’t solve this alignment problem—it just allows us to see it. We want defection to be obvious. [here I emphasize that I don’t have any problem with people who would work for $100k happening to get $250k, if that seems efficient]
Worth noting that we don’t require a “benefit humanity” motivation—intellectual curiosity will do fine (and I imagine this is already a major motivation for most researchers: how many would be working on the problem if it were painfully dull?).
We only require that they’re actually solving the problem. If we knew how to get them to do that for other reasons that’d be fine—but I don’t think money or status are good levers here. (or at the very least, they’re levers with large downsides)
I’ve tried to raise the topic with smart physics people I know or encounter whenever the opportunity presents itself. So far, the only ones who actually went on to take steps to try and enter alignment already had prior involvement with EA or LW.
For the others, the main reactions I got seemed to be:
Sounds interesting, but this is all too hypothetical for me to really take seriously. It hinges on all these concepts and ideas you propose about how AGI is going to work, and I don’t buy yet that all of them are correct
Sounds concerning, but I’d rather work on physics
Sounds depressing. I already thought climate change will kill us all, now there’s also this? Let me just work on physics and not think about this any more.
I’m not a mind reader of course, so maybe their real reaction was “Quick, say something conciliatory to make this person shut up about the pet topic they are insane about.”
I think there’s a bit of a social barrier to asking people with established directions to change careers (outside e.g. EA). People get invested in their current directions and may not perceive “please change careers” well even if tactfully put.
On the flip side, people who are considering changing careers are often pretty open to being told about new opportunities, and I have introduced people to AI safety who were already thinking about a change. I’m not sure that they were sold though...
A conventional approach might lead one to consider that inside the LW / AI safety bubble it borders on taboo to discount the existential threat posed by unaligned AI, but this is almost an inversion of the outside world, even if limited to to 25⁄75 of what LW users might consider “really impressive people.”
This is one gateway to one collection of problems associated with spreading awareness of AI alignment, but let’s go in a different direction: somewhere more personal.
Fundamentally, it seems a mistake to frame alignment as an AI issue. While unaligned AGI appears to be rapidly approaching and we have good reasons to believe this will probably result in the extinction of our species, there is another, more important alignment problem that underlies, and somewhat parallels the AI alignment problem. Of course, this larger issue is the alignment problem as faced by humanity at large.
Humans are famously unaligned on many levels: with respect to the self, interpersonally, and micro / macro-socially. No good solution to any tier of this problem has been discovered over thousands of years of inquiry. In the 20th century, humans developed technology useful for acquiring a great deal of information about the universe beyond our world, and “coincidentally” our capability of concentrated destruction increased in effectiveness by orders of magnitude, to the scale where killing at least large portions of the species in a short time is plausible. Thus, the question of why we don’t see others like us even though there appears to be ample space tended to find answers along the lines of intelligent life destroying itself. Of course, this is the result of an alignment “problem.”
Dull humans forecasted that nuclear arms would end the world and slightly smarter humans suggested that we might wait for antimatter, nanotech, genetically engineered pathogens or some other high-impact dangerous technology. As we’re seeing now, these problems are difficult. What appears to be less difficult is AGI.
So, even though it’s not in the interest of the continuity of the species, humanity can’t help but to race redundantly at breakneck pace toward this new technological capability, embodying a slightly disguised, concentrated and lethal version of one of the oldest and most fundamental problems our species has ever faced. That AI alignment is not taken more seriously could be seen as a reflection of “really impressive people” actually not having paid much mind to the alignment problems embedded in and endemic to who we are.
Should one introduce really impressive people to AI alignment? Maybe, but one must remember that magic appears unavailable and that for various reasons, it is predictably the case that most people, even “really impressive” people, will not consider the problem to be more than an abstract curiosity with even the best presentation. So to evangelize about AI alignment seems most useful as a fulfillment of one’s personal / social interests rather than much of a useful tool to increase work to save the species.
Full disclosure: it’s not clear that alignment is a meaningful concept, it’s not clear that humans have meaningful or consistent values, it’s very much not clear that continuing the human species is a good thing (at any point in our history, past, present or future) from an S-risk perspective, and it’s not clear that humans have any business rationally evaluating the utility in survival and reproduction as these are goals we’re apparently optimized for. So it should be the case that this post is written with less motivation to evangelize.