Andrej Karpathy ASD-STE100

Author: Data Science in your pocket · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Andrej Karpathy ASD-STE100 · Channel: Data Science in your pocket · 05:34 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI content generation, AI safety

Summary

The video presents Andrej Karpathy’s suggestion to make AI outputs easier to review by using language inspired by ASD-STE100, a controlled form of technical English. Consistent terms, clear actions, and explicit conditions help people inspect claims and assumptions, but clear writing does not guarantee that an answer is true. The video recommends preserving details and uncertainty, separating technical problem-solving from presentation, and verifying important claims and artifacts. The goal is output that supports human oversight, not output that removes the need for it.

Key points

Ask for an interface that makes an answer easier to understand and review, not only for a better answer.
ASD-STE100 uses writing rules and a controlled dictionary to promote consistent, clear technical communication.
Separate actions, conditions, and warnings so reviewers can inspect and challenge specific claims.
Use one stable term for each component to reduce ambiguity.
A clear, polished answer can still be wrong; STE improves the review surface, not the underlying truth.
Preserve numbers, units, versions, identifiers, exceptions, and uncertainty, and separate facts from assumptions and unknowns.
For high-value work, first solve and source the answer, then render it for review without changing its meaning.
Choose a representation—such as steps, diagrams, interactive pages, or video—based on what the reviewer needs to understand or decide.
Verify both important claims and the output artifact itself; richer outputs need richer verification.

Chapters

Start	Chapter	Summary
00:00	Designing AI output for review	The video introduces the problem of reviewing lengthy AI answers and the idea of designing a better interface for them.
01:08	How controlled technical English helps	ASD-STE100’s rules and consistent terminology are illustrated through a clearer, more inspectable service-improvement plan.
02:48	Preserve meaning and uncertainty	The video cautions that clear writing does not establish truth and recommends preserving details, uncertainty, and evidence.
03:53	Choose a representation for the task	A two-stage workflow and several output formats are presented, with the reviewer’s decision guiding the choice.
04:59	Verify outputs and maintain oversight	The video outlines checks for claims and artifacts and argues that AI can simplify explanations but not human responsibility.

Quotes

When generation becomes cheap, comprehension becomes expensive. — 00:00

Do not ask only for a better answer. Ask for a better interface to that answer. — 00:34

A clear wrong answer is still wrong. — 02:48

The goal is not merely shorter AI output. It is output you can inspect. — 05:34

Concepts and entities

ASD-STE100, Andrej Karpathy, controlled language, human oversight

Full transcript (original language)

00:00 The model finishes in seconds. A migration plan, 20 paper summaries, a flawless sounding explanation, then the real work lands on your desk. Thousands of words now stand between you and one simple question. Can I trust this enough to act? When generation becomes cheap, comprehension becomes expensive. Andre Karpathy’s suggestion starts with a change of perspective. Do not ask only for a better answer. Ask for a better interface to that answer. His practical

00:34 prompt is to go about 80% of the way to ASD STE100, a controlled form of technical English. But the deeper idea reaches beyond writing. Design the output around the mental work the human must do. ASD STE100 means ASD Simplified Technical English. It was created for technical documentation, especially aircraft maintenance, where readers with different levels of English still needed the same safe understanding. The project

01:08 began in the late ’70s. The first guide arrived in 1986, and issue nine became an international standard in January 2025. The standard has two parts. Writing rules control grammar, sentence structure, procedures, descriptions, and safety instructions. A controlled dictionary gives approved words specific meanings and parts of speech. This is stricter than plain English. It is not just use short sentences. Consistency is

01:40 the point. One component, one term, one action, one clear step. Imagine an AI answer about improving a slow service. One polished paragraph mixes caching, scaling, conditions, and a warning about stale data. Grammatically, nothing is wrong. Operationally, the sentence is a knot. An STE-inspired version cuts the knot. Measure latency. Add a cache when requests repeat. Set an expiration time.

02:13 Run a load test. Add instances only if the target is still missed. Then place the warning exactly where it matters. A cache can return old data. Now the reviewer has specific points to challenge. Which latency metric? Is the workload read heavy? What expiration is safe? Does the load test resemble production? Control language does more than reduce tokens. It exposes actors, actions, conditions, and assumptions as inspectable claims. Consistent

02:48 terminology removes another burden. If an answer calls the same component a worker, then a processor, then a consumer, you must decide whether it changed the subject. One stable term removes that detective work. But remember the limit. A clear wrong answer is still wrong. Polished structure can make weak evidence feel credible. STE improves the review surface, not the underlying truth. So do not prompt only for simplicity. Prompt for preservation.

03:22 Keep every number, unit, version, identifier, error message, exception, and uncertainty. Words like only, unless, before, and not verified can carry the whole decision. Ask the model to separate facts, assumptions, and unknowns. State what is missing and finish with checks you can use to verify the answer. For high-value work, split the workflow into two stages. First,

03:53 solve and source. Produce the technical answer, gather evidence, and mark uncertainty. Second, render for review. Convert that answer into STE-inspired text or another suitable view without changing its meaning. Keep the original available. For strict compliance, use a dedicated checker and expert review. AI should assist human authors, not replace them. Now the idea opens into four formats. Use controlled steps to review a deployment. Use a diagram to expose

04:29 relationships, such as every environment reading the same stale retrieval index. Use an interactive page to change assumptions about volume, ordering, replay, or cost. Use a narrated video when motion builds the mental model. The question is not STE or HTML. It is, “What must the reviewer understand or decide?” A practical workflow has six moves. Define the decision. Request

04:59 evidence. Choose the representation. Preserve the original. Check the few claims that could change your decision. Then test the artifact itself. Run the commands. Inspect every diagram edge. Validate the calculation. Richer outputs need richer verification. A beautiful diagram can invent a connection just as easily as a paragraph can invent a fact. As agents take on longer tasks, their reports become control surfaces for human oversight. So do not ask only for

05:34 a detailed summary. Ask for the system diagram, the controlled actions, every assumption, a verification checklist, and the original logs. The model can simplify the explanation. It cannot simplify away your responsibility to check it. The goal is not merely shorter AI output. It is output you can inspect.