HowIUseAI
title: How I Use "AI" (Carlini) type: technique created: 2026-05-26 last_updated: 2026-05-26 related: ["radar/techniques/ExploringGenerativeAI", "radar/techniques/YearBuildingWithLLMs", "radar/techniques/LLMEvaluationMethodology", "radar/techniques/AIConversationTypes"] sources: ["https://nicholas.carlini.com/writing/2024/how-i-use-ai.html"] radar_quadrant: Techniques radar_ring: Assess radar_position: inner
Practitioner essay by Nicholas Carlini (Google DeepMind security researcher) documenting concrete LLM use patterns and failure modes from a security and ML research perspective. Frames LLM value through a verification cost vs. generation savings lens.
Carlini's central argument: LLMs are useful when two conditions hold simultaneously — the output can be verified cheaply, and the alternative is tedious rather than difficult. When either condition fails, the economics of LLM use collapse. Confident, fluent output is not the same as correct output; the verification burden belongs to the user.
This framing cuts against the dominant narrative of LLMs as general-purpose accelerators. A task that takes 30 minutes to do correctly but only 5 minutes to verify saves time. A task that takes 30 minutes to do correctly and 25 minutes to verify does not.
Cases where Carlini finds LLMs reliably useful:
- Boilerplate generation — code the author knows how to write but would rather not type (argparse setup, file I/O scaffolding, test fixtures)
- Codebase explanation — understanding an unfamiliar repository or library faster than reading documentation linearly
- Test case generation — prompting for edge cases the author might not think of, then reviewing and selecting
- Language translation — converting between programming languages the author knows well enough to verify the output
- Regex and format conversion — one-shot structured transformations where correctness is immediately checkable
Cases where he finds the economics fail: original algorithmic work, security analysis, anything where the author cannot efficiently verify the output.
Three LLM practitioner perspectives sit on the radar:
- radar/techniques/ExploringGenerativeAI (Fowler) — software engineering and process lens; how LLMs change team workflows and testing practices
- radar/techniques/YearBuildingWithLLMs — production engineering lens; RAG, evaluation, task decomposition in real systems
- How I Use "AI" (Carlini) — security and ML research lens; failure mode analysis, verification cost, why confident output ≠ correct output
The Carlini essay is the most skeptical of the three and the most explicit about where LLM use destroys rather than creates value — making it the most useful counterweight when evaluating new LLM integration proposals.
How I Use "AI" sits in the Assess ring of the Techniques quadrant, at inner position. First studied via Carlini's personal site (2024-08-05). Nicholas Carlini's profile — foundational ML privacy attack research at Google DeepMind — gives this essay unusually high credibility relative to typical LLM commentary. Inner position reflects zero tooling cost, direct applicability to any practitioner making LLM integration decisions, and a distinct verification-cost framework not covered by the Fowler or applied-llms.org blips already on the radar. Remaining gate is deliberate application of the verification-cost framework when evaluating at least one new LLM integration decision.
