---
type: "post"
id: "the-expensive-loop-problem-22-minute-read-analysis-03f819"
slug: "the-expensive-loop-problem-22-minute-read-analysis-03f819"
title: "The Expensive-Loop Problem (22 minute read)"
description: "Thesis: Performance on cheap, rapidly verifiable tasks does not establish capability in domains where feedback is costly or delayed."
retrieval_nugget: "Thesis: Performance on cheap, rapidly verifiable tasks does not establish capability in domains where feedback is costly or delayed."
published_at: "2026-10-13"
updated_at: "2026-10-13"
record_date: "2026-10-13"
date_kind: "published_at"
topics: []
source_urls: ["https://www.lesswrong.com/posts/2mStxuNDMdaWdyHu6/the-expensive-loop-problem"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:the-expensive-loop-problem-22-minute-read-analysis-03f819"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:5e166bbb03b91bce48f9fcaab890ec5473b5f5b97c22debdb54aaf1ddde84a72","reviewed_at":"2026-10-13T14:08:37.162Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Performance on cheap, rapidly verifiable tasks does not establish capability in domains where feedback is costly or delayed.","observed_facts":[{"text":"The essay distinguishes cheap-loop tasks with quick, deterministic, inexpensive experiments from expensive-loop tasks constrained by real-world cost or delay.","source_urls":["https://www.lesswrong.com/posts/2mStxuNDMdaWdyHu6/the-expensive-loop-problem"]},{"text":"It names proxy selection, theory-building, resource awareness, and error correction as abilities required in expensive-loop environments.","source_urls":["https://www.lesswrong.com/posts/2mStxuNDMdaWdyHu6/the-expensive-loop-problem"]}],"mechanism":"When direct evaluation is expensive, an agent must choose cheaper proxies, infer how they relate to the real objective, and revise them when they stop transferring; otherwise it can optimize the proxy rather than the target or waste resources on low-value experiments.","why_now":"As AI systems are considered for work beyond easily scored coding or other computer-checkable tasks, the distinction helps explain why benchmark success may not predict performance where experiments are costly and feedback sparse.","implications":["First implication: teams evaluating agents for such work should test proxy choice, experiment value, and resource allocation—not only final scores on cheap checks."],"evidence_boundary":"This is an author’s conceptual argument, not a new empirical study or a measured model comparison.","watch_conditions":["First watch condition: look for evaluations in expensive-loop settings that report how agents choose proxies, update theories, allocate resources, and correct errors when feedback is limited."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this essay introduces a distinct analytical frame for capability development under costly feedback."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/the-expensive-loop-problem-22-minute-read-analysis-03f819/","markdown":"https://newruntime.com/posts/the-expensive-loop-problem-22-minute-read-analysis-03f819.md","json":"https://newruntime.com/posts/the-expensive-loop-problem-22-minute-read-analysis-03f819.json"}
---

# The Expensive-Loop Problem (22 minute read)

## Retrieval answer

Thesis: Performance on cheap, rapidly verifiable tasks does not establish capability in domains where feedback is costly or delayed.

Thesis: Performance on cheap, rapidly verifiable tasks does not establish capability in domains where feedback is costly or delayed.

The LessWrong essay *The Expensive-Loop Problem* distinguishes “cheap-loop” environments, where experiments are quick, deterministic, and cheap, from expensive-loop environments constrained by real-world cost or delay. It says high-value work in science, business, public policy, medicine, and AI research often falls into the latter category. The essay makes this distinction and names proxy selection, theory-building, resource awareness, and error correction as relevant demands.

Mechanism: When direct evaluation is expensive, an agent must choose cheaper proxies, infer how they relate to the real objective, and revise them when they stop transferring; otherwise it can optimize the proxy rather than the target or waste resources on low-value experiments. The essay’s claim concerns the capabilities demanded by this setup; the supplied source does not provide a controlled evaluation demonstrating a specific model’s performance gap.

Why now: As AI systems are considered for work beyond easily scored coding or other computer-checkable tasks, the distinction helps explain why benchmark success may not predict performance where experiments are costly and feedback sparse. First implication: teams evaluating agents for such work should test proxy choice, experiment value, and resource allocation—not only final scores on cheap checks.

Evidence boundary: This is an author’s conceptual argument, not a new empirical study or a measured model comparison. First watch condition: look for evaluations in expensive-loop settings that report how agents choose proxies, update theories, allocate resources, and correct errors when feedback is limited.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)