mirror neurons evolved in animals with diverse action spaces as a computationally efficient mechanism which allows the brain to reuse egocentric motor circuitry for modeling any observed action in general. these latent representations generalize across viewpoints and embodiments, allowing for imitation-based supervision to emerge at scale even outside the confines of ones on body or species. the cheapest way to model a sufficiently similar agent is to reuse existing abstractions linked to behavior. despite not observing any human data throughout pretraining, the sheer diversity of the robot-centric data observed by this policy model has pushed the representations away from those that are niche, appearance-specific, and superficial, to ones that are more abstracted, embodiment-agnostic, and goal-based (rather than method-based) in a way that mirrors the "grokking" phenomenon. such representations are exactly the kind of emergent abstraction one can expect from a well-fit foundation model, and resemble the emergence of perspective-agnostic mirror neurons for generalized action modeling in some advanced primates. by the time the policy model has reached the finetuning stage, it's cross-embodiment action representations from pretraining are sufficiently abstracted that adapting them to the humanoid embodiment becomes arguably trivial for the model compared to alternatives. true "imitation learning", in the original biomimetic sense, involves learning all sorts of behavior not from cloned egocentric recordings of experts at scale but simply from observing them behave in the wild and indirectly extracting supervision from that.
@WystanTBS Thanks for sharing!
I'm currently in the camp of "unshakable recognition of your mind's nature is enough".
Can you say more about why "constancy of continued development of samadhi applied to cutting through what remains" is necessary?
“Communities should have one thing that is common and the same for all members, whether they share in it equally or unequally,” - Aristotle.
Nice definition!
An honorable agent is one that sticks to the rules despite having the option to win by cheating.
Having an honorable society of agents allows for good games to be played
Recently I've been interested in the idea of "games" as a concept of deep importance.
I just read @FakeNousBlog 's article about honor.
Could it be that honor is sticking to the rules of the game when one could win by cheating?
Feels like a lot of his examples fit this
We are agents playing our own games (pursuing goals in an environment).
When agents play shared games, these games often have rules/restrictions for coordination purposes (to make things fun, to prevent tragedy of the commons, etc).
A society building agi is like a human undergoing physical augmentation. Would a human refuse augmenting their brain and body out of compassion for their biological cells? What about when competing against other humans in the workforce? Would you do it?
(https://t.co/4wz9zangUi)
I want to reflect on whether this base-game issue can be overcome via technology. Maybe if all relevant agents have to exist on some kind of substrate that bakes in "fixed rules" into it. No agent/mind on that substrate can override the rules by design.
I'm reading Nick Land's essay: Rules. Very interesting.
First thought is that the base game of being a physical body living in a physical universe means that rules that have been set up can always be challenged by beings that have access to kinetic weapons/sufficient force
I wonder: can we map out the space of games? Probably! I guess we'd find that many games are isomorphic to each other ... they just differ by decoration (Different shooting games have the same fundamental dynamic even though the settings might be different)
Video games provide a perfect opportunity to step into another game. I feel like there is a deep sense in which they give us access to a new library of ways of being agents that we previously didn't have access to, given that we have been restricted to playing human games