being able to access your own prompts & history as a symbolic object is half of what makes an RLM such a powerful design
llms should really learn to copy code with tool instead of using tokens it's insane
Users praise symbolic access to prompts and history in RLMs because it turns thought traces into manipulable objects that aid debugging and reduce black-box problems.
the second half is being able to write little programs that invoke RLMs, aka symbolic recursion
being able to access your own prompts & history as a symbolic object is half of what makes an RLM such a powerful design
@lateinteraction this was one of my favorite parts about RLMs - i had thought about using special copy/paste/edit tokens but the bitter lesson strikes again
@lateinteraction being able to yank ur own prompts back like a rubber band is what makes it stick
without that, its just a black box with a pretty coat
@lateinteraction the history-as-object design is the missing piece nobody talks about
something about having the metadata physically accessible changes the entire mental model
@lateinteraction the idea of your own thought traces being a manipulable object is huge for debugging where the black box is the problem
@lateinteraction saying it out loud makes it obvious why raw token prediction alone feels half baked
whats the deeper take here agency over memory?
@lateinteraction turning ur own thought log into a tangible thing changes how u interact with the whole system
kinda like a mirror u can actually grab