Will Mannon

In the evening I have a headache and I'm lying on the couch with Johanna, talking about autonomy. One way to think about human beings, at a very abstract level, is that we're information processing systems born with certain needs into a certain world. To meet our needs in this world, we have to build various mental models—both of how we work, what our needs are, and how the world works. We build mental model whose goal is to predict the world and ourselves well enough that we can make decisions that meet our needs. We try, in various ways, to seek pleasure and minimize pain, and so on. A lot of this happens through reinforcement learning. We cry, and then we notice that the people around us get stressed and can't meet our needs, so after a while we iterate our way to a different strategy, which is to suppress our tears. We crack a joke and nobody laughs and it's uncomfortable, so next time we do it a little differently. The goal of this process isn't to find "the truth" or the "optimal strategy," but simply something that's good enough for us to keep our system in some sort of reasonably satisfactory equilibrium. And the result of this kind of reinforcement learning is often something resembling those videos from 2015, when DeepMind first used reinforcement learning to teach 3D figures to walk. By giving them reward when they managed to walk further, the system could learn to walk, but there were often very strange artifacts—hands spinning around in the air, bizarre flailing of the upper body. I think many of us look like that on the inside too: we can make our way through life, but the way we walk is a bit jerky and peculiar, and it's far from ideal. Maybe we've iterated our way to alcohol as a way to regulate our negative emotions, or we do it by scrolling social media. We have various strange thought loops circling around in our heads, and so on. Different coping mechanisms that are good enough for us to avoid pain and live in relative equilibrium. They're not the strategies that would, objectively speaking, best help us meet our needs. But as I said, the algorithm we're running isn't searching for the objectively best—it's searching for what minimizes various signals (pain, prediction error, and so on), so it's easy to get stuck in these kinds of local maxima. Let me say the same thing again in slightly different words. What we call coping mechanisms are local maxima; they're solutions we've iterated our way to by trying different things and moving toward something that works better. Everything is a coping mechanism, even the most successful life strategies—the only difference is that those are coping mechanisms that are better at meeting our needs in both the short and long term, closer to the global maximum, the theoretical point where our mental models and behaviors are as good as they can be and everything we do is aligned with our needs and an accurate understanding of how the world works. To move toward that the ideal is "freely flowing information." That is: a state where we take in as much information as possible about the world—unfiltered, without biases, without avoiding things that are uncomfortable but true—so that we can efficiently update our world model and make more informed decisions; and also a free flow of information internally, where we don't suppress any feelings or needs within ourselves, but see everything as potential sources of useful information that can let us keep moving toward a deeper understanding, better decisions, and a richer life. To grow, in other words, we need to make sure this reinforcement learning process doesn't converge too early, leaving us stuck in the psychological equivalent of arms flailing in the air, but instead keeps moving forward toward a deeper understanding of ourselves and the world. And there are certain coping mechanisms, certain ways of relating to the pain that arises as you move through life, that are better at letting information flow, and those are the coping strategies worth cultivating. Instead of numbing yourself by scrolling social media, you want the coping mechanism where, when you feel stressed and bad, you get curious about the feeling. Instead of getting defensive when it turns out you're wrong, you want the coping mechanism where you get interested in your mistake and try to figure out why you were wrong. And so on. I'm not trying to say these are the optimal coping strategies—you can refine and deepen your mental models of how to handle negative emotions and feedback endlessly. What I'm describing here is just a kind of first step: coping mechanisms that lead to continued learning so you can gradually move toward even better ways of handling the situations.