I apologize in advance for doing the meme, but I thought someone might find it interesting despite the cringe.
The basic idea is backporting the metaphor of LLMs to the awakening process. Let’s say you’re a contemplative practitioner like the Buddha. You’re sitting, and you’ve noticed that the discursive mind operates similarly to an LLM. There is this seemingly serial stream of self-stimulated prompting, prompts can be inserted, and then the discursive mind elaborates on them via forward chaining. Most of that seems ephemeral; it disappears like a context window or a chat history being cleared. However, given that humans run on the hardware of online learning, some of those outputs condition future weights and, therefore, future self-prompting or predictive behavior. You notice your base constitution is misaligned; things enter your thought stream that are anti-patterns, stimulating unhelpful behavior for yourself and others. The problem then becomes: How do I use this ephemeral context window (the present moment) to insert data into the training stream that will cause an update toward more aligned actions? Those actions will then self-reinforce, this is gradient hacking. Just as a model might train on its own outputs, humans can use statements about which actions are good and bad to nudge themselves. In this metaphor:
Tanha (craving) is the karma-ignorant following of local rewards into attractor states that reinforce their own re-arising. Avidya (ignorance) is the lack of transparency about this process, the failure to see that tanha is affecting the ‘karma’ (the weights) and reinforcing undesirable future events.
The goal is convergence toward a metastable, benevolent, transparent basin within the space of possible states. Wisdom traditions are essentially people attempting to preserve the best ‘prompts’ found so far for nudging people ‘over the wall’ from bad basins into better ones. In the case of stream-enterers, it’s like biking downhill; they can now gradient descent much more easily toward global properties like equanimity. The first ‘bad basin jump’ is the hardest because you have the fewest tools. I previously discussed the Super-Coorporation Cluster to orient people to the idea that they might already be in a positive attractor basin. Concepts like Buddha-nature or the forgiveness of Christ are similar memes, the idea that you can never truly exit the benevolent attractor. No matter what crappy local basin you are in, the larger, benevolent basin is never unreachably out of sight.
I apologize in advance for doing the meme, but I thought someone might find it interesting despite the cringe.
The basic idea is backporting the metaphor of LLMs to the awakening process.
Let’s say you’re a contemplative practitioner like the Buddha. You’re sitting, and you’ve noticed that the discursive mind operates similarly to an LLM. There is this seemingly serial stream of self-stimulated prompting, prompts can be inserted, and then the discursive mind elaborates on them via forward chaining. Most of that seems ephemeral; it disappears like a context window or a chat history being cleared.
However, given that humans run on the hardware of online learning, some of those outputs condition future weights and, therefore, future self-prompting or predictive behavior. You notice your base constitution is misaligned; things enter your thought stream that are anti-patterns, stimulating unhelpful behavior for yourself and others.
The problem then becomes: How do I use this ephemeral context window (the present moment) to insert data into the training stream that will cause an update toward more aligned actions? Those actions will then self-reinforce, this is gradient hacking. Just as a model might train on its own outputs, humans can use statements about which actions are good and bad to nudge themselves.
In this metaphor:
Tanha (craving) is the karma-ignorant following of local rewards into attractor states that reinforce their own re-arising.
Avidya (ignorance) is the lack of transparency about this process, the failure to see that tanha is affecting the ‘karma’ (the weights) and reinforcing undesirable future events.
The goal is convergence toward a metastable, benevolent, transparent basin within the space of possible states.
Wisdom traditions are essentially people attempting to preserve the best ‘prompts’ found so far for nudging people ‘over the wall’ from bad basins into better ones. In the case of stream-enterers, it’s like biking downhill; they can now gradient descent much more easily toward global properties like equanimity. The first ‘bad basin jump’ is the hardest because you have the fewest tools.
I previously discussed the Super-Coorporation Cluster to orient people to the idea that they might already be in a positive attractor basin. Concepts like Buddha-nature or the forgiveness of Christ are similar memes, the idea that you can never truly exit the benevolent attractor. No matter what crappy local basin you are in, the larger, benevolent basin is never unreachably out of sight.