Created: January 27, 2022
Modified: February 10, 2022
Modified: February 10, 2022
computational life coach
This page is from my personal notes, and has not been specifically reviewed for public consumption. It might be incomplete, wrong, outdated, or stupid. Caveat lector.- How do you start building and selling computational therapy? It can't just be a medical product, because that's a hugely regulated and yet niche market. We want something that would get a core group of dedicated users early on. We would want people to genuinely love it, not just use it because they're depressed and the doctor prescribed it.
- Instead of therapy, the first product would be life coaching.
- We could do much better than traditional coaching by offering to record and analyze all of your day-to-day interactions. The system knows who all of your friends and acquaintances are, can recognize their voices, and can keep a record of your conversations with them, what was said, how it was said, and it can correlate these records with your own later recollections. (of course it won't be perfect and will miss some interactions, but it will have at least as much context in your life as a significant other or close friend who spends a lot of time with you)
- The system knows how you actually spend your time. It sees your sleeping patterns, eating habits, social habits, time spent productively, time wasted. It can give real feedback. It can see through your bullshit.
- There will be federated learning: your assistant will learn not just from all of literature, TV, the internet, etc, but also from interactions with all other users of the system. There will have to be privacy safeguards. Ideally we could 'anonymize' and then ship the full richness of your raw data, in case the system notices things that we don't. Maybe differential privacy guarantees would help here. But as a fallback we could analyze the raw data locally and only ship off gradient updates for some high-level abstract representation. We could also do this trade only if you want your experiences to help other people.
- The system has to really want to help you. It needs to be value aligned with you.
- This AI would, in a sense, be constantly running experiments. If it recommends to someone that they focus more on their relationship, and then that person turns out to be happier later in life, that provides feedback that will help the AI learn that this was a good recommendation in those circumstances.
- Of course recommendations themselves can be hard to follow. If the AI says, "you should fast every third day", I'm probably still not going to do it unless I can gather the critical mass of willpower. That requires the right circumstances to come along, and it also helps quite a lot if the choice to do the fasting comes from me. I can't be doing it because the AI wanted me to or bullied me into it. It has to be because I want to, because I see the purpose in it.
- If this works well, it would be incredibly high value for a lot of rich (and not rich) people. It will help them be more fulfilled. It could work to prevent depression. It would stimulate the economy by helping people make better life choices. It would grow through word of mouth from people who love it and benefit from it, just like Roam Research has. In fact, a service like this could read your Roam notebook, or other diary. It could even be an interactive part of the diarizing process, like a Github Copilot for diarizing. As you write, it pops up questions, asks you to clarify, comments on surprising points, and gives suggestions for how to deal with the situations you're writing about.
- I could prototype this today with access to GPT-3 or similar and my own Roam notebook. when I quit things to build