research worth doing

(see also: impact)

I've been feeling depressed partly because the actual PhD research I did was (in my view) pointless, and more broadly I'm not convinced that research in AI more generally is good for the world (we might hope that it will improve the world but most new technically is morally neutral at best). Plus of course I'm not really sure how to advance AI research, although if I were really excited about the cause I might be more enthusiastic about various routes to get there.

• A lot of work in tech really does miss what seem like the high-leverage points of utility. Personal assistants for rich people, food delivery, even better movie recommendations or ad-click predictions, even self-driving cars, even political targeting aimed at winning elections, even education, have claims for how they will make the world better along certain metrics. But once they make the world better, people will still not be happy. depression will still be a thing. People will still have lives we view as not worth living. And having been depressed, I would trade all of the physical inconvenience in the world, for a sense of meaning, purpose, community, fulfillment. (I would hesitate to literally trade a limb, but research shows that amputees end up as happy as everyone so objectively I probably should make that trade).
• So what are the high-leverage points? Probably curing sources of stupid physical pain (i.e., crippling back pain that never normalizes, or cancer that leads to slow, terrible deaths full of nausea and fear and defeat). But also figuring out how to make the perceived experience of life not suck. Since my material circumstance is as close to as good as it can be -- rich, well-credentialed, in good physical health and shape, not as young as I could be but still relatively young by adult standards -- the difference between my lonely, depressed, despondent self, and a thriving, optimistic, joyful version of myself has to be pretty much just a state of mind (though changing apartments would also help a lot). Addressing depression should involve understanding how the brain deals with perception, goals, pleasure, and rewards. I doubt there's a magic bullet, but a better understanding of these phenomena could help us make people's lives a lot better, and build more human-like AIs (that can also be trained to align with human values and make people's lives better).
• (of course the means by which an 'understanding' could make people's lives better are super vague to me. better drugs? better talk therapy? other interventions like games or movies? directly allowing patients like me to better introspect their own depression, and plan for happiness in a more correct model? or designing social institutions like academia to avoid causing depressive symptoms in the first place. ideally I would like to understand what is really going on in say buddhist gurus achieving enlightenment -- a sense of calm, acceptance, living in the present, seeing the beauty in life -- and translate that into modern scientific terms so that these mental states become more accessible to a modern audience)
• Inspired by Scott Alexander's discussion of the predictive processing model
• http://slatestarcodex.com/2017/09/05/book-review-surfing-uncertainty/
• I feel like there is something to be learned here from neuroscience. This book seems like a nice confluence of ideas: top-down and bottom-up processing, generative modeling, control as prediction, simple probabilistic explanations for complex disorders like autism and Parkinson's, detailed understanding of brain operation including neurotransmitters, etc. Some of the ideas are not at all novel in ML, some are things Geoff Hinton knows but I want to learn (eg neurotransmitters), some might be totally novel.
• A question related to these thoughts: which level of processing has moral salience? Suppose I believe the brain is a hierarchy: there are 'elemental' reward fns like physical pain and pleasure, but there are different levels of something like TD learning, where the perceived 'reward' is actually the residual between the predicted reward at the lower level, and the reward actually observed. In an MDP it's obvious mathematically that the true reward is what counts, any value function you might compute is just an instrument towards maximizing that reward. But in the brain, we might attend to the reward channel at a higher level, or at a range of levels, so that physical pleasure can still be bad if it does not meet our expectations, or that even a relatively painful life can be fulfilling if we perceive it as advancing higher-level goals. This latter would be the mechanism co-opted by religion, to give us a purpose to survive earthly discomforts. If this attention is fixed or limited, we learn something about morality (and how to create happiness, by either improving bottom-up reward perceptions or creating more pessimistic predictive models at a given level). If this attention is varying, we maybe learn something else: we can create happiness by shifting the attention itself.
• Of course this model is very hand-wavy, speculative, unproven. And even so I have no idea how to answer the question, or what answer would be satisfying (if we can show that the brain does attend specifically to certain levels of the hierarchy, does that mean those levels have the greatest moral salience? the is/ought dichotomy makes me hesitate, but it seems possible…)
• But one thing I like about the model is it answers for the fundamental badness of life. It can literally be true that most object-level rewards might be negative, people's lives are generally bad. But if we attend to the good things, then the agent inside the Cartesian theater (this isn't quite right--I mean the theater of rewards rather than visions, and the 'theater' does not reproduce the world but shows only higher-level abstractions and in this case residuals) might perceive the world as mostly positive -- and then it would actually be mostly positive assuming that agent is the moral entity that matters.
• A somewhat different point. Some work is 'not worth doing' because it's either morally neutral (AI capabiities, maybe) or at least hard to forsee the second-order effects of (self-driving cars discourage transit? running for office crowds out other voices. facebook targets community and relationships, but steals our attention and agency and we obsess over comparing our real lives to curated lives, causing massive FOMO, inferiority complexes, etc). But other work has felt 'not worth doing' because even though its effects are liekly positive they still feel trivial. For example, individual ministry or therapy, teaching math, improving speech recognition, building responsive analytics databases w/ dataflow, etc.
• That fact that I admit the second category exists means it is possible to do good in the world. Then the question is how much good. Of course I don't want to work on things that are trivial. But if I see a purpose that I think is important (like the one above - AI and neuro for understanding depression) and if I get excited about working on it, then of course I should follow that. Whether or not the thing is really important, as long as it makes me enjoy life and doesn't hurt anyone then that is a life worth living. (there is no cost to a life except for my own pain -- so as long as I'm making progress on goals, ie my TD deltas are positive at whatever the relative level of attention, then the life is worth living.)