I struggle a lot trying to figure out what I want to say, or if it’s even appropriate for me to say. But I guess I’ll say it, since you are saying it’s good to do.
I think the bar is too high for what I want to do. I want to contribute to AI x-risk reduction somehow. Ideally, I’d do this in some kind of a role. I have pretty awful productivity unemployed, and without some kind of meaningful/costly signal from others that my output actually matters, I default to thinking anything I do doesn’t matter.
But when I go look at the people that are actually out there doing things that matter, they are all way better than me. I’ve looked at the backgrounds of dozens of them by this point. It’s generally at least one of:
-Top tier university graduate (Harvard, Oxford, Stanford, etc.)
-Started a PhD
-Self-taught six figure programmer
-Started a business
-Some kind of amazing autodidact in other ways
And all this by the time they are 25, 30 at the latest. Even the people getting entry-level ops roles at Constellation are fresh private liberal arts college grads.
I don’t have any of that. I’m half-assing my third try to get a degree at a third-tier online school at 30, and the degree (CS) will probably be obsolete soon anyways. I don’t have any applicable work experience. I don’t have any notable talents or knowledge.
So to get a role, or to get costly signals that what I do matters, and that I’m worth investing in (even if it’s just me investing in myself), and hence be able to try to help stop the end of the world, it seems like I’d have to fulfill one of those criteria—You Must Be This Tall To Ride. And I am not tall enough.
I plan to move to the Bay Area in a few months. Live in a closet, or something, to make it work on my savings, or be homeless if I have to. Try to find anything at all I can do to help.
But even being there, it’s hard for me to imagine what I could do. It would take several years for me to get good enough at technical alignment to hope to do anything meaningful, and that’s assuming it wouldn’t just serve as safetywashing for the AI race. Governance requires a degree, a lot of background knowledge and skill, and also tends to require being an elite. I could… post on LW? Talk to people? Organize? Protest? My generally low capabilities would seem to indicate I wouldn’t be impactful at those on the margin, and it wouldn’t lead to a role. I don’t know what else I can do that doesn’t seem to amount to “bootstrap yourself into an agentic, autodidactic elite.”
Hopefully I am wrong about all this, and all I am missing is a supportive environment of in-person like-minded peers and networking opportunities in order to do something. Or maybe my model of the world is deeply flawed in some way. I am depressed, so maybe my mind will change after I get electroconvulsive therapy. Maybe I’ve been confirmation-biasing myself into believing that it’s all just elites? Or there’s been some other selection effect going on in what I see vs. what is real? I hope so.
Thank you for writing the post, it must have been pretty tough to write that all down. First and foremost, I want to encourage you to think about financial sustainability when considering moving to the bay. Personal growth and learning complex technical topics are energy-intensive and expensive activities (in terms of time and attention), and being constantly short on money and resources will not make either easier. My own experiences with mental health in the past showed me that doing good research gets exponentially harder if I can’t manage my own situation first.
With regards to your concerns about not learning fast enough/having enough skills to contribute to AI safety or related problems, I have several thoughts. First of all, while it may be true that you have limited skills in “traditional” technical research, that doesn’t mean that you therefore have to work to fit yourself into that mould. Right now I see a lot of people who are interested in AI safety taking the same correlated bets (moving to bay area, specialising into ML/technical AI research, joining a lab or an AI safety startup, relying heavily on Claude to accelerate progress), and overall I think this is quite a narrow minded angle with which to address a literally global problem. For example, even if MIRI secures its landmark policy goal (an international treaty to pause AI development), that still means that the US government is one of the two or three linchpins in ensuring that the treaty is carried out safely and competently. Thus local governance and campaigning for the ’26 and ’28 election seasons would be very important to shape who is in power at critical turning points. Being a non-US citizen, I have limited context here, but if you are American then this is something you have a unique power to shape. There are other similar opportunities here but I will not go on endlessly.
Finally, I want to push back a little against the idea of the “agentic, autodidactic elite” and the implied “plebs/NPCs/normies”. IMO most of what is called “agentic” is risk taking combined with leverage (e.g. pre-existing connections, resources etc.). As you said yourself many of the hotshots in this community already had some degree of backing, attended elite institutions etc., so its not really a case of “those with the will succeed and those without do not”. Insofar as people can grow and attain better skills, it is hard to do so when we constantly see ourselves as inferior to the titans and giants (although tbh I feel that way most of the time as well, I have to work quite hard to suppress that circuit in my brain). I hope that this helps somewhat and I hope that you can find firmer footing in your near future!
Well, you make good points about the Bay… it would be kinda hard to make the money work (though I’ve dealt with privation before), and it would be a correlated bet… I think part of why I want to be there, is to have any sort of like-minded community at all, instead of mouldering in my house all day, every day. I am highly motivated by other people.
I do think you’re right that technical research is not really the right path. There’s too much investment in that, compared to advocacy.
I am an American, but it’s hard to see a theory of change that actually routes through me? I can… vote, call my Congressperson, show up to protests… I’m going to PauseCon in DC, but I don’t see how that leads to a sustainable trajectory of change, including one that hopefully could get me paid. I am not skilled or resourced enough, I think, to go organizing all on my own, and I don’t have anywhere near the credentials to be an AI safety lobbyist or an assistant to one, not that there’s much funding for that in the US anyways. Unless you propose I move to a swing state and district specifically to influence the representative there...?
I am curious about the other opportunities you have in mind.
A lot of elites have impostor syndrome. You’re an elite, you went to Cambridge. I saw an article on EA Forum where someone said they weren’t good enough because they “weren’t in the top half of Oxford.” As someone who is, I’m pretty sure, an actual impostor (outside of being in intellectual sync with LW and the like), this is rather frustrating to listen to. You have the credentials, accomplishments, and connections to back up your ambitions, and the talent and family background to have gotten to where you are in the first place, and I do not. I don’t see how I can grow when I don’t have that kind of wealth.
“bootstrap yourself into an agentic, autodidactic elite.”
I find it likely that you are overestimating the difficulty of doing this. However, given that you said you’re depressed, you probably do have a harder time that most people would. So, maybe first focus you energies on getting creative about not being depressed. You mentioned electroconvulsive therapy, which makes me think you’ve already tried a bunch of easier stuff, but just in case something here is new: standard drugs (get advice from doctors, from reddit, from forums like this), unstandard drugs (no comment about breaking laws), treatments (light therapy? electroconvulsive as mentioned. transcranial magnetic stimulation?), therapies (I hear CBT is unusually good, and also that you should shop around for a good therapist and follow your gut if it feels like it one isn’t working), and scene changes (you mentioned moving cities—presumably you’ve done encouraging-friend filtering where you already are).
I think that reading someone else’s research program and then trying to tackle a relatively small unit of their open problems is the way to go, but some probably disagree, and considering that I haven’t succeeded at doing this yet I am not confident. I think the people who disagree would recommend that you think about the problem by yourself for a while, and then follow your queries and ideas where they lead you.
Personally, looking at agent foundations work made me go “Oh Cthulhu, look at this low hanging fruit, even I might me able to do this, why aren’t all the people-much-better-than-me already doing this?”
I think it’s possible that getting yourself to just… really feel a drive to do something, enough to make you work on it regularly, would be a sufficient superpower. If you are depressed, I would guess you have problems doing that. I also have some severe problems in that department, working on it.
I do not know how fast it would take for someone that isn’t me to learn the prerequisite math (up to the level that just trying the problem is a better use of time then tackling prerequisites). Personally I feel that if you spent a couple hours every day, then you could probably solve some minor problem or win a bounty after a couple years. I’m not sure whether I’m failing to account properly for you learning being slower than me.
As an example: how much background does someone need to have a shot at solving this bounty problem (cherry picked for background and concreteness)? I feel like I originally learnt most of the prerequisites from the information theory in the sequences and, like, seeing the definition of the Kullback-Leibler divergence and the intuition that it’s like extra bits after compression compared to perfect compression. There’s the less legible background of “mathematical maturity”, which I don’t actually know how to train besides saying “try learning some rigorous math for a while, doesn’t really matter which, and then hope you get the skill of how to learn and do math”. I’m thinking the sort of thing that lets you have something to do when you see a math problem you don’t know how to solve in front of you. I think math competition problems (which shoot for low prerequisite) for gradeschoolers train parts of this well, but perhaps you’ll just be better served picking it up while trying to learn a useful prerequisite?
Okay, I’m reading your post on light glasses (I like it!) and I think you should think that if you can gather the knowledge needed to write that post, and then go and do the things you wrote up, and then write it up, then you have demonstrated ‘smart and agentic’ capability, and I suggest that the various AI x-risk contribution paths ‘only’ require you to do something like that, but more often.
Thank you for your reply! I’ll respond to both comments here.
>I find it likely that you are overestimating the difficulty of doing this.
Could you make an argument for this? I have a hard time understanding this perspective. I don’t think most people can just become agentic autodidactic elites if they want to?
>getting creative about depression
I have heard of all of these, yes. I’ve tried a few legal drugs, I don’t plan on taking illegal drugs, multiple therapists haven’t worked, I already do light therapy for other reasons as you noticed, if I’m going to do TMS I may as well do ECT, and part of my motivation for moving to the Bay is to have an in-person community I feel like I could belong in.
>research problems
I don’t think technical safety research is a particularly impactful thing to do. That would be a bet that solving every aspect of x-risk relevant AI safety can happen before the frontier labs get to AGI with mere product safety, and that doesn’t seem like a very good bet. It would, as you note, take at least a couple years to make even the smallest contribution in the first place.
I don’t believe in myself enough (I don’t have academic connections, or a large body of internalized knowledge, etc.) to think that going off in my own direction would work very well either.
>math
I do not understand any of that bounty. I did not read much of the Sequences, and I don’t have the math chops to understand the more mathier ones. Perhaps if I went and studied math for a while I would, but, is that the best use of my time, compared to e.g. trying to figure out a way to do AI safety advocacy?
>post on light glasses
Thank you : ) Most of that knowledge was not my own, it was from the VLiDACMel document, whose author actually did the research. I just implemented it in my particular case—being highly motivated to do so to fix my sleep—and then wrote up the resulting protocol, and gave some easily googleable context about the nature of sleep and related products?
The motivation to write it came from a bunch of people telling me in person they’d read an article about the light glasses if I wrote it. That is, I got a clear external signal that my output mattered. But to be in the kind of situation where I could get signals like that for AI safety, I have to first be accomplished enough for people to notice and push on me. And I didn’t get accomplishments from my background, and to make new accomplishments, I have to get an external signal… it’s a catch-22. Hence why I am skeptical of my ability to bootstrap.
I don’t think technical safety research is a particularly impactful thing to do
I spoke of that because it’s the only thing I’ve looked into at all. I suspect the lesson is generalizable to, say, technical governance or politicking (which it sounded like you think are more useful). As an example: if you live near a college, I would consider showing up to the relevant shelling point location with a plastic chair, plastic table, and whiteboard that says something like “I think superintelligent AI will kill us all, and we need to stop development immediately—Change My Mind”. e.g. my campus has a place where people tend to show up to argue about abortion or religion or whatever (though usually one of those two...)
Most of that knowledge was not my own, it was from the VLiDACMel document, whose author actually did the research. I just implemented it in my particular case—being highly motivated to do so to fix my sleep—and then wrote up the resulting protocol, and gave some easily googleable context about the nature of sleep and related products?
This sure sounds like what someone with some special capabilities would say? Plenty of (imo unusually smart and hard working!) people I know would think “impressive, I couldn’t do that—where would I even start?”. I kinda feel that way too, I suspect that if I was in your shoes I wouldn’t have ended up thinking to do that, or doing it as well.
>depression
Yeah, it looked like you were being pretty reasonable in trying to fix that. I’d still say to keep pushing even when it feels like you’ve exhausted everything, but you probably have already heard this or already do that. I still think the approach is a good one.
I struggle a lot trying to figure out what I want to say, or if it’s even appropriate for me to say. But I guess I’ll say it, since you are saying it’s good to do.
I think the bar is too high for what I want to do. I want to contribute to AI x-risk reduction somehow. Ideally, I’d do this in some kind of a role. I have pretty awful productivity unemployed, and without some kind of meaningful/costly signal from others that my output actually matters, I default to thinking anything I do doesn’t matter.
But when I go look at the people that are actually out there doing things that matter, they are all way better than me. I’ve looked at the backgrounds of dozens of them by this point. It’s generally at least one of:
-Top tier university graduate (Harvard, Oxford, Stanford, etc.)
-Started a PhD
-Self-taught six figure programmer
-Started a business
-Some kind of amazing autodidact in other ways
And all this by the time they are 25, 30 at the latest. Even the people getting entry-level ops roles at Constellation are fresh private liberal arts college grads.
I don’t have any of that. I’m half-assing my third try to get a degree at a third-tier online school at 30, and the degree (CS) will probably be obsolete soon anyways. I don’t have any applicable work experience. I don’t have any notable talents or knowledge.
So to get a role, or to get costly signals that what I do matters, and that I’m worth investing in (even if it’s just me investing in myself), and hence be able to try to help stop the end of the world, it seems like I’d have to fulfill one of those criteria—You Must Be This Tall To Ride. And I am not tall enough.
I plan to move to the Bay Area in a few months. Live in a closet, or something, to make it work on my savings, or be homeless if I have to. Try to find anything at all I can do to help.
But even being there, it’s hard for me to imagine what I could do. It would take several years for me to get good enough at technical alignment to hope to do anything meaningful, and that’s assuming it wouldn’t just serve as safetywashing for the AI race. Governance requires a degree, a lot of background knowledge and skill, and also tends to require being an elite. I could… post on LW? Talk to people? Organize? Protest? My generally low capabilities would seem to indicate I wouldn’t be impactful at those on the margin, and it wouldn’t lead to a role. I don’t know what else I can do that doesn’t seem to amount to “bootstrap yourself into an agentic, autodidactic elite.”
Hopefully I am wrong about all this, and all I am missing is a supportive environment of in-person like-minded peers and networking opportunities in order to do something. Or maybe my model of the world is deeply flawed in some way. I am depressed, so maybe my mind will change after I get electroconvulsive therapy. Maybe I’ve been confirmation-biasing myself into believing that it’s all just elites? Or there’s been some other selection effect going on in what I see vs. what is real? I hope so.
Thank you for writing the post, it must have been pretty tough to write that all down. First and foremost, I want to encourage you to think about financial sustainability when considering moving to the bay. Personal growth and learning complex technical topics are energy-intensive and expensive activities (in terms of time and attention), and being constantly short on money and resources will not make either easier. My own experiences with mental health in the past showed me that doing good research gets exponentially harder if I can’t manage my own situation first.
With regards to your concerns about not learning fast enough/having enough skills to contribute to AI safety or related problems, I have several thoughts. First of all, while it may be true that you have limited skills in “traditional” technical research, that doesn’t mean that you therefore have to work to fit yourself into that mould. Right now I see a lot of people who are interested in AI safety taking the same correlated bets (moving to bay area, specialising into ML/technical AI research, joining a lab or an AI safety startup, relying heavily on Claude to accelerate progress), and overall I think this is quite a narrow minded angle with which to address a literally global problem. For example, even if MIRI secures its landmark policy goal (an international treaty to pause AI development), that still means that the US government is one of the two or three linchpins in ensuring that the treaty is carried out safely and competently. Thus local governance and campaigning for the ’26 and ’28 election seasons would be very important to shape who is in power at critical turning points. Being a non-US citizen, I have limited context here, but if you are American then this is something you have a unique power to shape. There are other similar opportunities here but I will not go on endlessly.
Finally, I want to push back a little against the idea of the “agentic, autodidactic elite” and the implied “plebs/NPCs/normies”. IMO most of what is called “agentic” is risk taking combined with leverage (e.g. pre-existing connections, resources etc.). As you said yourself many of the hotshots in this community already had some degree of backing, attended elite institutions etc., so its not really a case of “those with the will succeed and those without do not”. Insofar as people can grow and attain better skills, it is hard to do so when we constantly see ourselves as inferior to the titans and giants (although tbh I feel that way most of the time as well, I have to work quite hard to suppress that circuit in my brain). I hope that this helps somewhat and I hope that you can find firmer footing in your near future!
Well, you make good points about the Bay… it would be kinda hard to make the money work (though I’ve dealt with privation before), and it would be a correlated bet… I think part of why I want to be there, is to have any sort of like-minded community at all, instead of mouldering in my house all day, every day. I am highly motivated by other people.
I do think you’re right that technical research is not really the right path. There’s too much investment in that, compared to advocacy.
I am an American, but it’s hard to see a theory of change that actually routes through me? I can… vote, call my Congressperson, show up to protests… I’m going to PauseCon in DC, but I don’t see how that leads to a sustainable trajectory of change, including one that hopefully could get me paid. I am not skilled or resourced enough, I think, to go organizing all on my own, and I don’t have anywhere near the credentials to be an AI safety lobbyist or an assistant to one, not that there’s much funding for that in the US anyways. Unless you propose I move to a swing state and district specifically to influence the representative there...?
I am curious about the other opportunities you have in mind.
A lot of elites have impostor syndrome. You’re an elite, you went to Cambridge. I saw an article on EA Forum where someone said they weren’t good enough because they “weren’t in the top half of Oxford.” As someone who is, I’m pretty sure, an actual impostor (outside of being in intellectual sync with LW and the like), this is rather frustrating to listen to. You have the credentials, accomplishments, and connections to back up your ambitions, and the talent and family background to have gotten to where you are in the first place, and I do not. I don’t see how I can grow when I don’t have that kind of wealth.
I’ll send you a DM!
I find it likely that you are overestimating the difficulty of doing this. However, given that you said you’re depressed, you probably do have a harder time that most people would. So, maybe first focus you energies on getting creative about not being depressed. You mentioned electroconvulsive therapy, which makes me think you’ve already tried a bunch of easier stuff, but just in case something here is new: standard drugs (get advice from doctors, from reddit, from forums like this), unstandard drugs (no comment about breaking laws), treatments (light therapy? electroconvulsive as mentioned. transcranial magnetic stimulation?), therapies (I hear CBT is unusually good, and also that you should shop around for a good therapist and follow your gut if it feels like it one isn’t working), and scene changes (you mentioned moving cities—presumably you’ve done encouraging-friend filtering where you already are).
I think that reading someone else’s research program and then trying to tackle a relatively small unit of their open problems is the way to go, but some probably disagree, and considering that I haven’t succeeded at doing this yet I am not confident. I think the people who disagree would recommend that you think about the problem by yourself for a while, and then follow your queries and ideas where they lead you.
Personally, looking at agent foundations work made me go “Oh Cthulhu, look at this low hanging fruit, even I might me able to do this, why aren’t all the people-much-better-than-me already doing this?”
I think it’s possible that getting yourself to just… really feel a drive to do something, enough to make you work on it regularly, would be a sufficient superpower. If you are depressed, I would guess you have problems doing that. I also have some severe problems in that department, working on it.
I do not know how fast it would take for someone that isn’t me to learn the prerequisite math (up to the level that just trying the problem is a better use of time then tackling prerequisites). Personally I feel that if you spent a couple hours every day, then you could probably solve some minor problem or win a bounty after a couple years. I’m not sure whether I’m failing to account properly for you learning being slower than me.
As an example: how much background does someone need to have a shot at solving this bounty problem (cherry picked for background and concreteness)? I feel like I originally learnt most of the prerequisites from the information theory in the sequences and, like, seeing the definition of the Kullback-Leibler divergence and the intuition that it’s like extra bits after compression compared to perfect compression. There’s the less legible background of “mathematical maturity”, which I don’t actually know how to train besides saying “try learning some rigorous math for a while, doesn’t really matter which, and then hope you get the skill of how to learn and do math”. I’m thinking the sort of thing that lets you have something to do when you see a math problem you don’t know how to solve in front of you. I think math competition problems (which shoot for low prerequisite) for gradeschoolers train parts of this well, but perhaps you’ll just be better served picking it up while trying to learn a useful prerequisite?
Okay, I’m reading your post on light glasses (I like it!) and I think you should think that if you can gather the knowledge needed to write that post, and then go and do the things you wrote up, and then write it up, then you have demonstrated ‘smart and agentic’ capability, and I suggest that the various AI x-risk contribution paths ‘only’ require you to do something like that, but more often.
Thank you for your reply! I’ll respond to both comments here.
>I find it likely that you are overestimating the difficulty of doing this.
Could you make an argument for this? I have a hard time understanding this perspective. I don’t think most people can just become agentic autodidactic elites if they want to?
>getting creative about depression
I have heard of all of these, yes. I’ve tried a few legal drugs, I don’t plan on taking illegal drugs, multiple therapists haven’t worked, I already do light therapy for other reasons as you noticed, if I’m going to do TMS I may as well do ECT, and part of my motivation for moving to the Bay is to have an in-person community I feel like I could belong in.
>research problems
I don’t think technical safety research is a particularly impactful thing to do. That would be a bet that solving every aspect of x-risk relevant AI safety can happen before the frontier labs get to AGI with mere product safety, and that doesn’t seem like a very good bet. It would, as you note, take at least a couple years to make even the smallest contribution in the first place.
I don’t believe in myself enough (I don’t have academic connections, or a large body of internalized knowledge, etc.) to think that going off in my own direction would work very well either.
>math
I do not understand any of that bounty. I did not read much of the Sequences, and I don’t have the math chops to understand the more mathier ones. Perhaps if I went and studied math for a while I would, but, is that the best use of my time, compared to e.g. trying to figure out a way to do AI safety advocacy?
>post on light glasses
Thank you : ) Most of that knowledge was not my own, it was from the VLiDACMel document, whose author actually did the research. I just implemented it in my particular case—being highly motivated to do so to fix my sleep—and then wrote up the resulting protocol, and gave some easily googleable context about the nature of sleep and related products?
The motivation to write it came from a bunch of people telling me in person they’d read an article about the light glasses if I wrote it. That is, I got a clear external signal that my output mattered. But to be in the kind of situation where I could get signals like that for AI safety, I have to first be accomplished enough for people to notice and push on me. And I didn’t get accomplishments from my background, and to make new accomplishments, I have to get an external signal… it’s a catch-22. Hence why I am skeptical of my ability to bootstrap.
(only currently replying to a couple subpoints)
I spoke of that because it’s the only thing I’ve looked into at all. I suspect the lesson is generalizable to, say, technical governance or politicking (which it sounded like you think are more useful). As an example: if you live near a college, I would consider showing up to the relevant shelling point location with a plastic chair, plastic table, and whiteboard that says something like “I think superintelligent AI will kill us all, and we need to stop development immediately—Change My Mind”. e.g. my campus has a place where people tend to show up to argue about abortion or religion or whatever (though usually one of those two...)
This sure sounds like what someone with some special capabilities would say? Plenty of (imo unusually smart and hard working!) people I know would think “impressive, I couldn’t do that—where would I even start?”. I kinda feel that way too, I suspect that if I was in your shoes I wouldn’t have ended up thinking to do that, or doing it as well.
>depression
Yeah, it looked like you were being pretty reasonable in trying to fix that. I’d still say to keep pushing even when it feels like you’ve exhausted everything, but you probably have already heard this or already do that. I still think the approach is a good one.