Source
Lieder & Griffiths (2020), LIT-1071 (Behavioral and Brain Sciences 43, e1), read in NOTE-873: the target article and the Authors’ Response in full, and four commentaries (Bates, Sims & Jacobs; Dimov; Dingemanse; Ma & Woodford). The other commentaries were searched by keyword only, and the primary studies the review cites were not read.
What was actually shown
The objective. Eq. 3 defines the resource-rational heuristic for a known environment as the h, among “the set of heuristics that brain B can execute”, that maximizes the expected utility of its result minus the expected cost(t_h, ρ, λ), “the total opportunity cost of investing the cognitive resources ρ used or blocked by the heuristic h for the duration t_h of its execution”. Eq. 4 averages the same trade over the environments consistent with what is known. Eq. 3 follows from the lifetime form (Eq. 2) only “if we assume that life can be partitioned into a sequence of episodes”. A fixed budget is the special case of a cost constant across heuristics.
The method. Five steps (Box 2): state the problem, posit a class of algorithms with their costs, find the one that trades resources against accuracy best, test its predictions, and refine the problem or the constraints. The authors call step 5 “a double-edged sword”: constraints postulated carelessly can overfit, so fitted constraints should be judged out of sample and measured independently.
The support. A review that recasts prior models as instances, so that cheap, approximate and biased mechanisms fall out as the price of frugality. Its numbers are others’: people’s planning reached “about 86% of the best possible trade-off between decision quality and time cost” (Callaway et al. 2018a; Lieder et al. 2017). Nothing in the article is a new experiment.
Where the optimization happens. Not in the head at decision time: “for most of our decisions the problem of finding a good decision mechanism has already been solved by evolution …, learning, and cognitive development”. The cost of that search is not in Eq. 3 or 4.
What this does not say
- That people are optimal. The Response calls the optimality assumption “a methodological device to efficiently search through the endless space of possible mechanisms”. The review’s 86% and 55% figures are departures from optimality, reported from other work.
- That the account has been tested as a whole. What could fail is each derived model, and step 5 allows the constraints to be revised after a failure. The paper’s own guard against that, independent measurement of constraints, is stated as a standard, not shown met.
- That the currency of cost is known. The framework specifies “the terms of the tradeoff between utility and computational costs but not the kinds of computations or the way that those costs are denominated.”
- That it models reuse of computation. The objective is per episode. Precomputing for a later episode, storing a result, retrieving it and its going stale have no terms. The value of what a computation leaves behind enters only through the utility of the episode’s result, which the Response says scores the belief state too. Reuse appears only in the review, qualitatively: habits, Dasgupta et al. 2018’s memorized hypotheses, Anderson and Milson’s need probability.
- That it models offloading to the world. The target article mentions “cognitive prostheses” once, as an intervention. Dingemanse’s commentary names offloading and distributed cognition as a frontier; the Response agrees and does not extend the formalism.
- That Howes et al. 2016 is a memorize-or-look-up trade. The review describes it so; the record’s reading of that paper (LIT-1062, NOTE-882) found no look-up option in the task.
Connections
- QUESTION-095 asks for one cost model covering precompute, storage, retrieval, staleness and expected reuse. This objective could host such a model as a net of benefit and opportunity cost, and supplies the shared currency, but has none of those terms. The search for a heuristic, paid once by evolution and learning and then used in every episode, is the plainest amortization in the paper, and it is left out of the objective on purpose.
- CLAIM-303’s break-even F + n(c_a + u) < n·c_r is a net of the form Eq. 3 licenses, summed over n uses rather than taken per episode. The paper gives no support to the claim’s staleness term.
- THEORY-343: the decision to offload, as people make it, follows belief about the cost; the review here says the trade is optimized offline, by evolution and learning, not computed by the person, which leaves room for that belief to be wrong.
- THEORY-331: the scaling-law papers price two of the terms this objective lacks, expected reuse and per-use cost, in a different domain.
Note of 2026-10-30: what holds it back
Kept Proposed by the twenty-fourth status review. The first clause states the framework’s thesis in the record’s voice: cognitive mechanisms can be derived as the best trade of utility against opportunity cost. It is not a report that Lieder and Griffiths propose it. The paper supports it only by recasting others’ models, and step 5 of the method lets constraints be refitted (NOTE-873). No resource-rational model with independently measured costs has been shown to predict out of sample better than a rival with no cost term. The title asserts the thesis, so it keeps its condition (ADR-065 §6).