Humans can decompose complex real-world cognitive tasks into context-limited subtasks whose solutions reliably combine into a good overall solution.
Assessed
First pass with a tight budget. Assessed the claim as doubtful with credence 0.3 and added no decomposition. A later pass should record instances from the Ought and Wu et al. sources and could split the claim into weak and strong forms.
Assessed Doubtful
verdict confidence 0.60 · credence 0.30
This is the strong form of the factored cognition hypothesis, which iterated amplification and related scalable-oversight proposals rely on. It holds that a complex real-world cognitive task can be broken into subtasks, each solvable by someone who sees only a small amount of context, and that the subtask answers reliably combine into a good overall answer. A weak form is plainly true. Organizations divide large projects among people, and language-model agents that hand work to sub-agents complete some tasks well. In both cases, though, the parts usually keep rich shared context and plenty of working time. The strong form requires small context per subtask and reliable recombination. It has had few direct tests, and those were discouraging. Ought's experiments with human participants (2018–2019) found that tasks of interesting complexity needed very large trees of subtasks, and that factored evaluation of even short arguments was brittle. OpenAI's 2021 recursive book-summarization work named the assumption that parts can be handled independently as a possible fundamental limit. The tests were small and used tight time limits per subtask, so they do not refute the hypothesis. On current evidence the claim is doubtful rather than contradicted. Larger human-participant studies, or studies of agents with strictly limited context, on varied real-world tasks would settle it.
Claim entered the graph