AI that can meditate: Nonlinear Function
Created: January 23, 2022
Modified: February 26, 2022

AI that can meditate

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 It might be incomplete, wrong, outdated, or stupid.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.
  • An AI that thought 'like a human' would be able to engage in meditation. What does this imply? What capabilities are missing from current AI systems? How would this ability contribute to capabilities research and to AI safety?
  • At a very generic level, this just requires the ability to 'think' without interacting with the external world. So it means an architecture that can decide to apply additional steps of internal compute, or that can take 'thinking actions' in addition to real-world actions (c.f. meta-reasoning).
  • But I think it's interesting to bring to bear more of what we know about the structure of meditative approaches. For example, what would an AI require in order to do concentration practice? What would it mean for an AI to achieve awakened awareness?
  • Throwing out some associations:
    • its architecture might include something like a default mode network
    • it would need some form of trainable attention mechanism, which ordinarily is heavily influenced or 'captured' by bottom-up signals
    • it would have to be able to turn this mechanism 'in on itself', to attend to the process of its own thoughts, to be self-aware.
    • it would have to be able to find this process rewarding: there needs to be a training signal that allows it to learn to ignore distractions from the default mode network (or maybe to observe them but not be 'grabbed' by them)
    • it would have to be able to implement internal feedback loops, ultimately allowing it to get into jhanas.
  • Perhaps enlightenment is actually the solution to AI safety. An AI that has no sense of self, that sees no distinction between itself and the rest of the world, that wants only to help everyone along the path, could be inherently aligned and benevolent.
    • The good thing about this is that enlightenment is, supposedly, incredibly simple. It ultimately unfolds entirely out of a simple instruction(s), e.g., to practice attending to the exact moment of the present. This is important because the deep solutions, whatever they are, ultimately need to be simple.
    • This would imply that we can reduce the problem of AI safety to the problem of formalizing enlightenment. Now it's easy! (lol)
  • What would a research program in this direction look like? What are the signs of intermediate progress? What is the lighthouse goal, that would get the rest of the community to pay attention?
  • Counterpoint: a lot of the work of meditation is required in order to transcend our animal brains and primitive emotional responses. It's perverse to build in these negative aspects just to be able to model a practice for neutralizing them. It's like saying that we should build planes that need to go to the gym and lift weights in order to make their wings strong enough to fly.
    • Desired counterpoint: reactivity or attachment of the sort that meditation transcends are not incidental to our animal brains, they are necessarily implied by some fundamental aspect of computationally limited intelligence, or maybe by the nature of consciousness.

This seems like something highly relevant to read: Soraj Hongladarom's [https://www.goodreads.com/book/show/53217569-the-ethics-of-ai-and-robotics](The Ethics of AI and Robotics: A Buddhist Viewpoint ) (reading inbox)