New Comment
Rendering 0/60 comments, sorted by
top scoring
(show all) Click to highlight new comments since: Moderation Log
Curated and popular this week
I appreciate this comment a lot. Thank you. I appreciate that it’s sharing an inside view, and your actual best guess, despite these things being the sort of thing that might get social push-back!
My own take is that people depleting their long-term resources and capacities is rarely optimal in the present context around AI safety.
My attempt to share my reasoning is pretty long, sorry; I tried to use bolding to make it skimmable.
0. A world where I'd agree with you. If all that would/could impact AI safety was a particular engineering project (e.g., Redwood’s ML experiments, for concreteness), and if the time-frame of a person’s relevance to that effort was relatively short (e.g., a year or two, either because AI was in two years, or because there would be an army of new people in two years), I agree that people focusing obsessively for 60 hours/week would probably produce more than the same people capping their work at 35 hrs/week.
But (0) is not the world we’re in, at least right now. Specific differences between a world where I'd agree with you, and the world we seem to me to be in:
1. Having a steep discount rate on labor seems like a poor predictive bet to me. I don’t think we’re within two years of the singularity; I do think labor is increasing but not at a crazy rate; and a person who keeps their wits and wisdom about them, who pays attention and cares and thinks and learns, and especially someone who is relatively new to the field and/or relatively young (which is the case for most such engineers I think), can reasonably hope to be more productive in 2 years than they are now, which can roughly counterbalance the increase (or more than counterbalance the increase) on my best guess.
E.g., if they get hired and Redwood and then stay there, you’ll want veterans a couple years later who already know your processes and skills.
(In 2009, I told myself I needed only to work hard for ~5 years, maybe 10, because after that I’d be a negligible portion of the AI safety effort, so it was okay to cut corners. I still think I’m a non-negligible portion of the effort.)
1.1. Trying a thing to see if it works (e.g. 60 hrs/week of obsession, to see how that is) might still be sensible, but more like “try it and see if it works, especially if that risk and difficulty is appealing, since “appealingness” is often an indicator that a thing will turn out to make sense / to yield useful info / to be the kind of thing one can deeply/sincerely try rather than forcing oneself to mimic, etc.” not like “you are nothing and don’t matter much after two years, run yourself into the ground while trying to make a project go.” I suppose your question is about accepting a known probability of running yourself into the ground, but I’m having trouble booting that sim; to me the two mindsets are pretty different. I do think many people are too averse to risk and discomfort; but also that valuing oneself in the long-term is correct and important. Sorry if I’m dodging the question here.
2. There is no single project that is most of what matters in AI safety today, AFAICT. Also, such projects as exist are partly managerially bottlenecked. And so it isn’t “have zero impact” vs “be above Redwood’s/project such-and-such’s hiring line,” it is “be slightly above a given hiring line” (and contribute the difference between that spot and the person who would fill it next, or between that project having one just-above-margin person and having one fewer but more managerial slack) vs “be alive and alert and curious as you take an interest in the world from some other location”, which is more continuous-ish.
3. We are confused still, and the work is often subtle, such that we need people to notice subtle mismatches between what they’re doing and what makes sense to do, and subtle adjustments to specific projects, to which projects make sense at all, and subtle updates from how the work is going that can be propagated to some larger set of things, etc. We need people who care and don’t just want to signal that they kinda look like they care. We need people who become smarter and wiser and more oriented over time and who have deep scientific aesthetics, and other aesthetics. We need people who can go for what matters even when it means backtracking or losing face. We don’t mainly need people as something like fully needing-to-be-directed subjugated labor, who try for the appearances while lacking an internal compass. I expect more of this from folks who average 35 hrs/week than 60 hrs/week in most cases (not counting brief sprints, trying things for awhile to test and stretch one’s capacities, etc. — all of which seems healthy and part of fully inhabiting this world to me). Basically because of the things pointed out in Raemon’s post about slack, or Ben’s post about the Sabbath. Also because often 60 hrs/week for long periods of time means unconsciously writing off important personal goals (cf Critch’s post about addiction to work), and IMO writing off deep goals for the long-term makes it hard to sincerely care about things.
(4. I do agree there’s something useful about being able to work on other peoples’ projects, or on mundane non-glamorous projects, that many don’t have, and that naive readings of my #3 might tend to pull away from. I think the deeper readings of #3 don’t, but it could be discussed.)
A. I still agree that in a world where all that would/could impact AI safety was a particular engineering project (e.g., Redwood’s ML experiments, for concreteness), and if the time-frame of a person’s relevance to that effort was relatively short (e.g., a year or two, or even probably even five years), people focusing obsessively for 60 hours/week would be in many ways saner-feeling, more grounding, and more likely to produce the right kind of work in the right timeframe than the same people capping their work at 35 hrs/week. (Although even here, vacations, sabbaths, or otherwise carefully maintaining enough of the right kinds of slack and leisure that deep new things can bubble up seems really valuable to me; otherwise I expect a lot of people working hard at dumb subtasks).
A2. I’m talking about “saner-feeling” and “more grounding” here, because I’m imagining that if people are somehow capping their work at 35 hrs/week, this might be via dissociating from how things matter, and dissociation sucks and has bad side-effects on the quality of work and of team conversation and such IMO. This is really the main thing I’m optimizing for ~in general; I think sane grounded contexts where people can see what causes will have what effects and can acknowledge what matters will mostly cause a lot of the right actions, and that the main question is how to cause such contexts, whether that means 60 hrs/week or 35 hrs/week or what.
A3. In this alternate world, I expect people will kinda naturally reason about themselves and one another as means to an end (to the end of us all surviving), in a way that won’t be disoriented and won’t be made out of fear and belief-in-belief and weird dissociation.
B. In the world we seem to actually be in, I think all of this is pretty different:
B1. It’s hard to know what safety strategies will or won’t help how much.
B2. Lots of people have “belief in belief” about safety strategies working. Often this is partly politically motivated/manipulated, e.g. people wanting to work at an organization and to rise there via buying into that organization’s narrative; an organization wanting its staff and potential hires to buy its narrative so they’ll work hard and organize their work in particular ways and be loyal.
B3. There are large “unknown unknowns,” large gaps in the total set of strategies being done, maybe none of this makes sense, etc.
B4. AI timelines are probably more than two years, probably also more than five years, although it’s hard to know.
C. In a context like the hypothetical one in A, people talking about how some people are worth much more than another, about what tradeoffs will have what effects, etc. will for many cash out in mechanistic reasoning and so be basically sane-making and grounding. (Likewise, I suspect battlefield triage or mechanistic reasoning from a group of firefighters considering rescuing people from a burning building is pretty sane-making.)
In a context like the one in B (which is the one I think we’re in), people talking about themselves and other people as mere means to an end, and about how much more some people are worth than another such that those other people are a waste for the first people to talk to, and so on, will tend to increase social fear, decrease sharing of actual views, and increase weird status stuff and the feeling that one ought not question current social narratives, I think. It will tend to erode trust, erode freedom to be oneself or to share data about how one is actually thinking and feeling, and increase the extent to which people cut off their own and others’ perceptual faculties. The opposite of sane-ifying/grounding.
To gesture a bit at what I mean: a friend of mine, after attending a gathering of EA elites for the first time, complained that it was like: “So, which of the 30 organizations that we all agree has no more than a 0.1% chance of saving the world do you work for?”, followed by talking shop about the specifics within that plan, with almost no attention to the rest of the probability mass.
So I think we ought mostly not to reason about ourselves and other “labor” as though we’re in simple microecon world, given the world we’re in, and given that it encourages writing off a bunch of peoples’ perceptual abilities etc. Though I also think that you, Buck (or others) speaking your mind, including when you’re reasoning this way, is extremely helpful! We of course can’t stop wrong views by taking my best guess at which views are right and doing belief-in-belief about it; we have to converse freely and see what comes out.
(Thanks to Justis for saying some of this to me in the comments prior to me posting.)
This post has been recorded as part of the LessWrong Curated Podcast, and can be listened to on Spotify, Apple Podcasts, and Libsyn.
Related to: Slack gives you the ability to notice/reflect on subtle things
Epistemic status: A possibly annoying mixture of straightforward reasoning and hard-to-justify personal opinions.
It is often stated (with some justification, IMO) that AI risk is an “emergency.” Various people have explained to me that they put various parts of their normal life’s functioning on hold on account of AI being an “emergency.” In the interest of people doing this sanely and not confusedly, I’d like to take a step back and seek principles around what kinds of changes a person might want to make in an “emergency” of different sorts.
Principle 1: It matters what time-scale the emergency is on
There are plenty of ways we can temporarily increase productivity on some narrow task or other, at the cost of our longer-term resources. For example:
If I would strongly prefer to address some situation x before time t, I may sometimes want to "borrow from the future" like this. But the time-scales matter. If I’m trying to address x as much as possible in the next five hours, skipping sleep may make sense. If I’m trying to address x as much as possible over the next year, I’ll probably do better to get my usual amount of sleep tonight. Something similar (with different, resource-specific timescales) will hold for other resources.
So, in short time-scale emergencies, it’ll often make sense to suspend a great deal of normal functioning for a short period of time. In longer time-scale emergencies, your life should mostly look closer to normal.
Principle 2: It matters how much we know how to address the emergency
Much of what we do in daily life – especially when we’re feeling “free” and “unstuck,” and as though there is nothing in particular that we “have to” do – has the effect of making us well-resourced and capable in general. For example, by default, a lot of us would spend a lot of time reading interesting books of varied sorts, nerding out about interesting topics, trying our hands at new skills and crafts, etc. Also, we often like making our living spaces nicer (and more functional), forming new friendships, and so on.
If there is a particular thingy that matters hugely, and if you have an accurate model of how exactly to change that thingy, it may make sense to sacrifice some of your general-purpose capacities in trade for increased ability to address that thingy. (E.g., if you know for a fact that you’ll lose your home unless you pay the mortgage, and if keeping your home is important, it may make sense to trade general-purpose capacities for the ability to make mortgage payments by e.g. working over-long hours at a mind-numbing job that leaves you stupid.)
However, if you don’t have an accurate map of how to address a given thingy, then, even if the thingy is very important, and even if its time-scale is short, you’ll probably mostly want to avoid sacrificing general-purpose capacities. (In a low-information context, your general-purpose capacities are perhaps more likely to turn out helpful for your very important thingy than whatever you’d be trading them off to buy.) Thus, in “emergencies” where you do not have an accurate map of how to solve the emergency, your behavior should probably be more like normal than in better-mapped emergencies.
Side-note: “Emergencies” as wake-up calls
A different way an “emergency” can sometimes rearrange priorities is by serving as a “wake-up call” that helps people peel away accidental accumulation of habits, “obligations,” complacency, etc. For example, I’m told that near encounters with death sometimes leave people suddenly in touch with what matters in life. (I’ve heard this from one friend who had cancer, and seen accounts from strangers in writing; I’m not sure how common this is or isn’t really.)
I non-confidently suspect some gravitate toward AI risk or other emergencies in the hopes that it’ll help them peel away the cruft, notice a core of caring within themselves, and choose activities that actually make sense. (See also: Something to Protect.) If “AI risk as wake-up call” works out, I could imagine AI risk helping a person rearrange their actions in a way that boosts long-term capacities. (E.g., finding courage; trying new things and paying attention to what happens; cultivating friendships right now rather than in the possibly-non-existent future; facing up to minor social conflicts; attempting research that might yield fresh insights instead of research that feels predictable.)
This sort of “post wake-up call” change is almost the opposite of the kinds of “borrowing from the future” changes that typify short-term emergency responses. You should be able to tell the difference by seeing whether a person’s actions are unusually good for their long-term broad-based capacities (e.g., do they keep their house in a pattern that is good for them, get exercise, engage in the kinds of leisure and study that boost their ability to understand the world, appear unusually willing and able to cut through comfortable rationalizations, etc.?), or unusually bad for their long-term broad-based capacities (e.g., do they live on stimulants, in a house no one would want to live in, while saying they ‘don’t have time’ for exercise or for reading textbooks and seeming kind of burn-out-y and as though they don’t have enough free energy to fully take an interest in something new, etc.?).
What sort of “emergency” is AI risk?
It seems to me that AI risk is a serious problem in the sense that it may well kill us (and on my personal models, may well not, too). In terms of time-scales, I am pretty ignorant, but I personally will not be too surprised if the highest risk period is in only a couple years, nor if it is in more than thirty years. In terms of how accurate our maps of what to do are, it seems to me that our maps are not accurate; most people who are currently burning themselves out to try to help with AI risk on some particular path might, for all I know, contribute at least as well (even on e.g. a 5-year timescale) if they built general capacities instead.
I therefore mostly suspect that we’ll have our best shot at AI risk if we, as a community, cultivate long-term, robust capacities. (For me, this is hinging more on believing we have poor maps of how to create safety, and less on beliefs about time-scale.) Our best shot probably does mean paying attention to AI and ML advances, and directing some attention that way compared to what we’d do in a world where AI did not matter. It probably does mean doing the obvious work and the obvious alignment experiments where we know what those are, and where we can do this without burning out our long-term capacities. But it mostly doesn’t mean people burning themselves out, or depleting long-term resources in order to do this.
Some guesses at long-term, robust-in-lots-of-scenarios resources that may help with AI risk
Briefly, some resources I’d like us to have, as AI approaches:
Given the above, I am excited about people in our larger orbit following their interests, trying to become scientists or writers or engineers or other cool things, exploring and cultivating. I am excited about people attempting work on AI alignment, or on other angles on AI safety, while also having hobbies and interests and friendships. For the most part I am not excited about people burning themselves out working super-long hours on alignment tasks in ways that damp their ability to notice new things or respond to later challenges, although I am excited about people pushing their limits doing work they’re excited by, and these things can blur together.
Also, I am excited about people trying to follow paths to all of their long-term goals/flourishing, including their romantic and reproductive goals, and I am actively not excited about people deciding to shelve that because they think AI risk demands it. This is probably the hardest to justify of my opinions, but, like Divia in this tweet, I am concerned (based partly on personal experience, partly on observations of others, and partly on priors/models) that when people try to table their deepest personal goals, this messes up their access to caring and consciousness in general.
Why do we see burnout in AI safety efforts and in effective altruism?
IMO, I see burnout (people working hard at the expense of their long-term resources and capacities) more often than I expect is optimal for helping with AI risk. I’m not sure why. Some of it is in people who (IMO) have a better shot and a better plan than most for reducing AI risk, which makes it more plausibly actually-helpful for those folks to be doing work at the expense of long-term capacity, though I have my doubts even there. I suspect much of it is for confused/mistaken reasons, though; I sometimes see EAs burning themselves out doing e.g. random ML internships that they think they should work hard on because of AI risk, and AFAICT this trade does not make sense. Also, when I look at the wider world (e.g. less-filtered chunks of Twitter; conversations with my Lyft drivers) I see lots of people acting as though lots of things are emergencies worth burning long-term resources for, in ways I suspect do not make sense, and I suspect that whatever dynamics lead to that may also be involved here. I’ve also heard a number of people tell me that EA or AI safety efforts caused them to lose the ability to have serious hobbies, or serious intellectual interests, and I would guess this was harmful to long-term AI safety potential in most cases. (This "lots of people lose the ability to have serious hobbies when they find EA" is something I care very much about. Both as damage to our movements' potential, and as a symptom of a larger pattern of damage. I've heard this, again, from many, though not everyone.)
I’d be curious for y’all’s takes on any of the above!