Step 10: tool results

{
  "status": "inconclusive",
  "claim_id": "a4bfe1a1-0d33-4876-b6ab-40989cb38bac",
  "assessment": "The claim is a forecast from the 2018 Iterated Amplification paper (Christiano, Shlegeris and Amodei): that the large supervision savings observed when a model learns one-step task decompositions instead of whole solutions will carry over to realistic tasks. The direct evidence for it comes from five algorithmic toy tasks, where amplification needed tens of thousands of decomposition examples against tens of millions of full-solution examples for supervised learning. Those tasks were picked because they decompose cleanly, so they say little about realistic work.\n\nThe general argument offered, that decomposing with help from subtask solvers is strictly easier than solving unaided, shows decomposition is easier to perform but not that it needs less supervision to learn, nor that good decompositions are cheap for humans to demonstrate. Evidence from realistic settings is mixed. Ought's human experiments on factored cognition (2019 to 2020) found realistic reading and argument-evaluation tasks hard to break into locally checkable steps, with frequent disagreement about step validity. OpenAI's recursive book summarization (2021) showed decomposition can make human feedback feasible on a long realistic task, but did not measure supervision against a direct baseline. Since then, most capability in practice comes from pretraining, which changes what \"learning directly\" costs and leaves the original comparison largely untested. The claim remains plausible but unestablished; a controlled comparison of label counts on a realistic task would settle it.",
  "confidence": 0.65,
  "claim_credence": 0.45,
  "marginal_yield": 0.3,
  "reasoning_trace": "Own model first: the claim is true if, for realistic tasks, the number or cost of human-provided training signals needed to learn good decompositions is much lower than for learning the end-to-end behavior. That requires (a) realistic tasks to admit decompositions whose steps are much simpler than the whole, and (b) humans to be able to demonstrate or judge those steps cheaply.\n\nFor: (1) Algorithmic experiments in arXiv 1810.08575: ~10^4 decomposition examples vs ~10^7 supervised examples. Rated likely only conditional on realistic tasks resembling the toy tasks, a premise left unstated in the source. (2) \"Decomposition is strictly easier than solving without the aid of X\": true in a performance sense, but the step from easier-to-perform to needing much less supervision is not guaranteed, and the source gives no reason for the \"much\". Rated slight.\n\nAgainst: Ought's factored-evaluation experiments (https://www.lesswrong.com/posts/pH3eKEAEupx8c2ep9/update-on-ought-s-experiments-on-factored-evaluation-of) found participants disagreed often about step validity and tree accuracy was modest (around 65 to 69% for depth 1 to 2 trees, per Alignment Newsletter 86), suggesting realistic decompositions are costly to produce and check. These were human-only experiments, not ML supervision counts, so the bearing is indirect.\n\nMixed: Wu et al. 2021, Recursively Summarizing Books with Human Feedback (arXiv 2109.10862), used decomposition to make feedback on book-length summarization tractable; it supports feasibility but not a quantified supervision saving versus direct learning.\n\nNo study found compares supervision counts for learned decomposition versus direct learning on a realistic task. Verdict: inconclusive, credence slightly below even, given the toy-task evidence is selected for clean decomposition and realistic-setting evidence is mixed. A search was brief; a wider search of process-supervision and task-decomposition literature could shift this modestly."
}