A recent conversation about longtermism raised the question of an “epistemic discount rate.” Yes, the future definitely matters as much as the present. But we can’t reason with as much certainty about the future. To what extent should we discount our conclusions about future outcomes?
The first response raised a great retort that I think needs to be unpacked:
…presumably everyone agrees that generally uncertainty about outcomes is going to increase over time. But if that just means your error bars widen over time, it’s not clear that’s a reason for discounting — saving exactly 100 lives isn’t obviously far better than an even-odds lottery between saving 50 and 150 lives, etc. So I guess you might need to say why we should expect there to be an upside/downside asymmetry?
I think the simple answer to “is there upside/downside asymmetry when promoting distant future good?” is that, in the absence of a totally accurate map, most actions you can take intentionally are either useless or harmful.
The curse of dimensionality rears its head here: if you select physical outcomes at random, most of them are horrific. Try randomly arranging the atoms of your body in the space you currently occupy – the overwhelming odds are that you’ll be dead. Good outcomes, I think by definition, are almost vanishingly rare. The entire enterprise of life consists of machines fighting against entropy locally, selecting for the few good possibilities from the many fatal ones. So a “attempts to promote long term future outcomes are more likely to be asymmetrically bad when compared with improving the present” is a natural result of every other process involving life: most outcomes, selected randomly, are fatal.
When you’re using an inaccurate map to try and select distant future outcomes, it’s likely that you’re doing something more like throwing a dart into a crowd than removing a tumor with a scalpel. A tiny amount of energy, delivered with excellent precision, will almost always outperform ten times the energy delivered with 90% of the precision. A rocket where 90% of the pieces perform to spec is not going to do much besides explode on the launch pad. If you have a limited energy budget, there’s a far higher chance you are actually promoting the outcomes you want if you spend it locally. The further away from the immediate present you are planning, the less likely you’ll hit the target you want, and the greater the odds of unintended negative consequences.
For example, there was a lot of worry about nuclear war in 60’s-90’s. This worry was obviously justified! Yet one unexpected side effect of that worry was that it created intense negative social stigma around nuclear power. The end result of that stigma is more reliance on fossil fuels and thus global warming.
Can you imagine trying to make this argument to Carl Sagan? “Look, obviously nuclear war would be devastating and we don’t want that, but you might consider toning down the alarmism a bit, because you are making people hesitant about nuclear power, which is necessary to convince people to move away from more polluting forms of power such as goal and gas, which if we don’t do, will almost certainly lead to higher expected losses because they outcomes are basically guaranteed as opposed to tail risk scenarios.”
There are even second order effects here! I grew up hearing that environmentalists were scientifically illiterate hippies who stopped us from using nuclear power, damming us to use worse, more expensive power sources, and thus we could safely ignore whatever they had to say. It took me a while as a young adult to realize this was only partially true.
Now apply this logic to AI today: If we weren’t worried about the alignment thesis at all, we might rush forward to build increasingly advanced machines that save lives and lower the cost of living. It’s not too hard to imagine a future where many people discount ALL risk of alignment problems, after it turns out that some AI that was held back due to worries about alignment has been saving lives without any obvious drawbacks. Is it possible that future problems which we can’t yet imagine could be mitigated by AI, but there will be strong enough social taboos in place that only a few maniacs are willing to try?
I suspect the best thing most of us can do for the distant future is likely to be a good person, live a good life, find a partner who shares your values and complements your strengths, and then physically instantiate your love and respect for this person in the form of healthy, functional children. That algorithm has a history of success, compared to the questionable track record of different attempts to shape the distant future.
Ok, you might ask – what do I mean by good person? Isn’t defining that the entire point of the exercise? And aren’t there plenty of instances when people really have succeeded in promoting long term good outcomes?
I think these two questions have basically the same answer: your ability to promote long term good is constrained most by your ability to take excellent care of yourself, and your environment, in the present.
And to this, I would respond that you probably already have some concept of good that you are falling short of hitting 100% of the time. If you don’t, well, that’s another discussion. But my bet is that if you are honest with yourself, you have some standard that you’d like to hit, but you often fail to hit it because you are tired or stressed or worried, which tends to make you a bit of a belligerent asshole. I know that’s how I am. I don’t think this is unique to me.
Personally, I find that the more I try and just hit a simpler, more-obviously-correct standard of “be a good husband, a good father, a good son, a good brother, a good friend”, the easier and quicker it is to get feedback at how I’m doing. The more feedback I get into how I’m doing, and the more closely I can monitor my progress, the harder it is to pretend that I don’t need serious work myself.
Conversely, to the extent that I have grown and developed, the easier it is to see how that work affects others in a natural, organic way. The more I’ve worked on being mindful, patient, and present, the less my writing resembles a bizarre maniac on the internet. The more habitually I maintain calm, the more productive I can be at work, by carefully selecting priorities instead of floundering in the chaos. My body and identity seem to be the single highest points of leverage I have, and the time I spend becoming more attuned to myself seems to be an investment in advancing my agency and credibility, which play off each other heavily. The more aware I am of how I feel in each moment, the more I can stop my feelings from nosediving, which tends to turn me into an asshole.
Speaking of which, how do you feel right now? Is there tension around the muscles of your eyes? Are you slouching? Were you frowning without realizing it? Hunched forward, looking into a screen while sitting? Standing, with your chin down, scope of attention narrowed to a blinky distraction in front of your face?
You can always be appreciating the beauty around you, generating a stream of serotonin which buoys your posture and enables you to accomplish long term goals more reliability by providing buoyancy to your dopamine network. How are you going to stop a rogue AI from dissolving the world in grey goo if you can’t reliably keep your head balanced atop your spine? I think that is a noble goal and we should be working towards it, and our ability to do so is constrained mainly by our own personal limitations.
Is it possible that consistently practicing excellent self love and discipline really are the most impactful things we can do for the distant future, by strengthening our ability to reliably promote long term outcomes when doing so really is possible and desirable?