This post posits that the WM will have a “similar format” throughout, but that heuristics/shards may not. For example, you point out that the WM has to be able to arbitrarily interlace and do subroutine calls (e.g. “will there be a dog near me soon” circuit presumably hooks into “object permanence: track spatial surroundings”).
(I confess that I don’t quite know what it would mean for shards to have “different encodings” from each other; would they just not have ways to do API calls on each other? Would they be written under eg different “internal programming languages”? Not reuse subroutine calls to other shards?)
Keeping in mind that maybe I am only reasoning syntactically right now, here are more considerations:
I think it’s non-trivial to know why I’m worried about something / why I think e.g. someone is mad at me (I remember being quite bad at this before learning Focusing). This seems like weak evidence that “what do I predict” is not substantially more self-interpretable than “why did I act that way”, which is in turn weak evidence that WM formatting is about as consistent as shard formatting.
Shards also probably reuse machinery. In particular, “grab the cup” should not be duplicated across juice- and milk-shards. This suggests e.g. a motor coordination API will arise. Eventually the generalization of these APIs probably turns into the GPS (but that’s just my current gut guess).
The WM can literally invoke shards in rollout predictions, where it computes what (conditional on e.g. setting dog-nearby to True) the shard bids are. So I don’t see why shards should be black-boxy to the WM/GPS?
SSL updates may push shards to be more predictable to WM, so as to reduce predictive error, which seems to me like it pushes towards uniform shard encoding.
This post posits that the WM will have a “similar format” throughout, but that heuristics/shards may not. For example, you point out that the WM has to be able to arbitrarily interlace and do subroutine calls (e.g. “will there be a dog near me soon” circuit presumably hooks into “object permanence: track spatial surroundings”).
(I confess that I don’t quite know what it would mean for shards to have “different encodings” from each other; would they just not have ways to do API calls on each other? Would they be written under eg different “internal programming languages”? Not reuse subroutine calls to other shards?)
Keeping in mind that maybe I am only reasoning syntactically right now, here are more considerations:
I think it’s non-trivial to know why I’m worried about something / why I think e.g. someone is mad at me (I remember being quite bad at this before learning Focusing). This seems like weak evidence that “what do I predict” is not substantially more self-interpretable than “why did I act that way”, which is in turn weak evidence that WM formatting is about as consistent as shard formatting.
Shards also probably reuse machinery. In particular, “grab the cup” should not be duplicated across juice- and milk-shards. This suggests e.g. a motor coordination API will arise. Eventually the generalization of these APIs probably turns into the GPS (but that’s just my current gut guess).
The WM can literally invoke shards in rollout predictions, where it computes what (conditional on e.g. setting
dog-nearbytoTrue) the shard bids are. So I don’t see why shards should be black-boxy to the WM/GPS?SSL updates may push shards to be more predictable to WM, so as to reduce predictive error, which seems to me like it pushes towards uniform shard encoding.