(crossposted from my substack)

Being a freshman at university, I seem to have been bestowed the great privilege of infinite possibilities. There is this strange feeling of trying to plan a career in a world that might not exist in five years. Not in a doomer sense but that the world of 2031 might be so radically different from today that most attempts of planning are incoherent. I contemplate machine god super intelligence arriving before I graduate, intelligence explosions compressing 10,000 years of progress into one, the possibility of being among the last humans to die before death itself is solved; then I do my laundry and pick electives. I am surprised I am not entirely losing my mind.

Robin Hanson writes, in 2009, This is The Dream Time:

Perhaps most important, our descendants may remember how history hung by a precarious thread on a few crucial coordination choices that our highly integrated rapidly changing world did or might have allowed us to achieve, and the strange delusions that influenced such choices. These choices might have been about global warming, rampaging robots, nuclear weapons, bioterror, etc. Our delusions may have led us to do something quite wonderful, or quite horrible, that permanently changed the options available to our descendants.

…

Our dreamtime will be a time of legend, a favorite setting for grand fiction, when low-delusion heroes and the strange rich clowns around them could most plausibly have changed the course of history. Perhaps most dramatic will be tragedies about dreamtime advocates who could foresee and were horrified by the coming slow stable adaptive eons, and tried passionately, but unsuccessfully, to prevent them.

This post includes some of my thoughts for other people in a similar position, who take these possibilities seriously enough to let them reshape their decisions, but aren’t sure how. Specifically: how to translate a probability distribution over timelines into an actual plan for what to do with the next few years of your life, when the wrong choice might round your impact to zero. Even the right choice might. But doing nothing guarantees it. 

I'm relatively uncertain about most of what follows. Tell me where I'm wrong.

Expected value calculation under timeline uncertainty

A lot of career decisions under AI timeline uncertainty seem to reduce to weird, complicated calculations that are running in the background. Here I try to write one down. Real career choices are certainly not binary, but let’s take a simplified version that illustrates the framework of aptitude-adjusted expected value. 

Say you’re choosing between two paths. Path A: get into technical research as fast as possible right now, upskill full-time in ML, try to contribute to alignment as fast as possible. Path B: stay in school, build foundations, maybe do field-building.

You think short and medium timelines are about equally as likely. Under short timelines the sprint maybe gets you 5 units of impact. You are starting behind, competing with people with varying skills and dispositions, but you are contributing something. In this short timeline world the slow investment gets you roughly 0, everything changed before anything paid off. Under medium timelines, the sprint gets you maybe 2, you contribute but at the cost of burning through your runway without deep foundations. The slower investment gets you 50,

EV(sprint-heavy) = 0.5 × 5 + 0.5 × 2 = 3.5

EV(invest-heavy) = 0.5 × 0 + 0.5 × 50 = 25

The invest-heavy orientation wins, not because medium timelines are more likely, but because the magnitude of your impact there is so much larger. The ratio matters more than the probability. Even if you’re 80% confident in short timelines, the invest-heavy approach still wins in this example. You’d need to be around 91% confident before the sprint orientation becomes the better bet. And 91% confidence in anything this uncertain should be pretty concerning.

These numbers are made up and yours should differ. Maybe you think the sprint would get you more than 5 in a short-timeline world. Maybe you think field-building is worth less than 50 in a medium-timeline world. You should run it yourself. The strategy of betting on the world where you have the most leverage, rather than the one you deem most likely, focuses on maximizing asymmetric upside—situations where your potential gains significantly outweigh your potential losses. This approach prioritizes expected value over mere probability, focusing on high-impact scenarios where you possess a distinct advantage. 

What should I do if everything is learnable:

EA career advice has historically emphasized “personal fit”, find what you’re naturally good at and apply it to important problems. I think this view is sometimes wrong and slightly in contradiction with the idea that most things are learnable.

Robert Musil writes in The Man Without Qualities:

But if there is a sense of reality, and no one will doubt that it has its justification for existing, then there must also be something we can call a sense of possibility. Whoever has it does not say, for instance: Here this or that has happened, will happen, must happen; but he invents: Here this or that might, could, or ought to happen. If he is told that something is the way it is, he will think: Well, it could probably just as well be otherwise. So the sense of possibility could be defined outright as the ability to conceive of everything there might be just as well, and to attach no more importance to what is than to what is not. The consequences of so creative a disposition can be remarkable, and may, regrettably, often make what people admire seem wrong, and what is taboo permissible, or, also, make both a matter of indifference. Such possibilists are said to inhabit a more delicate medium, a hazy medium of mist, fantasy, daydreams, and the subjunctive mood. Children who show this tendency are dealt with firmly and warned that such persons are cranks, dreamers, weaklings, know-it-alls, or troublemakers.

Some evidence for malleability:

• Deliberate practice research shows people’s ceilings are far higher than their intuitions suggest
• You can actually just learn things
• The set of things you could plausibly become good at is enormous and largely unconstrained by your skill profile now
• There are lots of people who successfully pivoted into AI safety from fields with no obvious direct pipeline
• Whatever skills you have, there is probably a way to apply it to AI safety.
• …

However, if you genuinely believe you can become and learn almost anything, the search space turns into infinity. This seems not conducive to decision making. Full malleability and infinitely possibilities can be pretty paralyzing. 

Personally, I seem to have a comparative advantage in being able to speak Chinese, which is plausibly useful for AI governance and international coordination. But I’ve also barely done anything and have little evidence for what I might be good at. Even if I’m not maximally gifted at math, I can probably learn a good amount if I really tried. I could probably also learn a good amount of policy analysis, or ML engineering, or institutional design, or whatever. Or can I just do all of these? 

Here, timeline views provide some constraints to the search space that disposition alone cannot. Short timelines suggest one to build the most immediately deployable skill and medium timelines favor investing in foundations that compound. For the same reason that the world in whole should have people betting on different timelines, you can diversify your skill profile. You don’t have to pick one timeline and go all in.

Should I have my own thoughts about timelines?

My thoughts about this are here. 

Summary: The question of whether to defer or think for yourself relies on a false binary that ignores better options. You probably cannot out-predict Daniel Kokotajlo. But you should still actually understand why experts believe what they believe, and what the object-level grounds that inform timeline forecasts are. This is instrumentally important for being able to produce impactful work, maybe developing research taste, and being able to make informed decisions.

Questions and answers synchronize like meshing gears; everyone has only certain fixed tasks to do; professions are located in special areas and organized by group; meals are taken on the run. Tension and relaxation; activity and love, are precisely timed and weighed on the basis of exhaustive laboratory studies. If anything goes wrong in any of these activities the whole thing is simply dropped; something else or sometimes a better way will be found or someone else will find the way one has missed; there’s nothing wrong with that, while on the other hand nothing is so wasteful of the social energies as the presumption that an individual is called upon to cling for all he is worth to some specific personal goal. In a community coursed through by energies every road leads to a worthwhile goal, provided one doesn’t hesitate or reflect too long. Targets are short-term, but since life is short too, results are maximized, which is all people need to be happy, because the soul is formed by what you accomplish, whereas what you desire without achieving it merely warps the soul. Happiness depends very little on what we want, but only on achieving whatever it is. Besides, zoology teaches that a number of flawed individuals can often add up to a brilliant social unit.

— Robert Musil, The Man Without Qualities

Doing something is probably better than doing nothing. Doing the best thing requires both luck and calculation. You can increase your surface area for luck, but you cannot control it. AGI is super scary but don’t go cheap on actually doing object level world modeling! Revisit your assumptions frequently. Do some healthy deference when needed. There has never been a better time to get into AI safety. But before signing up for any of these things actually maybe sit down and run your own numbers. Make decisions that are good enough across the range of plausible futures, and then to act, because acting on an imperfect view beats perfecting a view you never act on.