I also wonder how it could interact with METR’s statement (upd: link) that early versions of MirrorCode produced a time horizon of 100+ hours. Back in December I made a simple model of a task being split into subtasks until each subtask became rather easy. The LLM, according to the model, would hillclimb on subtasks, then on splitting simple tasks into subtasks, then on increasingly macrostrategic ones until the coding world runs out of macrostrategy levels and each macrostrategy level would be unapproachable before the previous one is saturated. Does this imply that MirrorCode is a low-macrostrategy task?
I’m not sure if I follow this or not. But it rhymes with a general picture I have of sequential planning often being hierarchical: you get good at the small pieces, those become available abstractions to plan with, and now you’re planning in ‘option’ space (as opposed to raw ‘action’ space) which is fully recursive. You need some amount of evidence for how options work (which might be policy-/actor-dependent) and how to strategise in that ‘higher’ space. You build your library of option concepts through learning and abstraction. Starting from scratch, you need to build up from the atomic action level. If you’re seeded with lots of approximate examples across the hierarchy (as are LMs—and probably humans to some extent), the dynamics are a bit different (but perhaps not very different unless your starting base is very ‘hierarchy-level-asymmetrical’).
I also wonder how it could interact with METR’s statement (upd: link) that early versions of MirrorCode produced a time horizon of 100+ hours. Back in December I made a simple model of a task being split into subtasks until each subtask became rather easy. The LLM, according to the model, would hillclimb on subtasks, then on splitting simple tasks into subtasks, then on increasingly macrostrategic ones until the coding world runs out of macrostrategy levels and each macrostrategy level would be unapproachable before the previous one is saturated. Does this imply that MirrorCode is a low-macrostrategy task?
I’m not sure if I follow this or not. But it rhymes with a general picture I have of sequential planning often being hierarchical: you get good at the small pieces, those become available abstractions to plan with, and now you’re planning in ‘option’ space (as opposed to raw ‘action’ space) which is fully recursive. You need some amount of evidence for how options work (which might be policy-/actor-dependent) and how to strategise in that ‘higher’ space. You build your library of option concepts through learning and abstraction. Starting from scratch, you need to build up from the atomic action level. If you’re seeded with lots of approximate examples across the hierarchy (as are LMs—and probably humans to some extent), the dynamics are a bit different (but perhaps not very different unless your starting base is very ‘hierarchy-level-asymmetrical’).