On October 2, Karpathy returned from a quiet stretch with a single theme: we will spend more and more time understanding language model output. The point deserves a pause. Two years of optimization have all targeted the generation side—faster models, longer context, cheaper tokens. Karpathy turns the camera around: models now write far faster than humans read, so the bottleneck has moved to comprehension. He offers a four-layer solution, each heavier than the last: controlled language, diagrams, interactive web pages, and bespoke explainer videos.
Layer 1: Make the model delete its own degrees of freedom
The least glamorous layer first. Karpathy says a technique that works well for him personally is having the LLM explain things in ASD-STE100.
This is a real specification: ASD-STE100 Simplified Technical English, the writing standard for aviation maintenance manuals. First published in 1986, now at Issue 9, with a 434-page specification. Its core is two ledgers: a dictionary of roughly 900 approved words, each with exactly one part of speech and one meaning; and a set of writing rules—short sentences, active voice, one instruction per paragraph. Technicians maintaining Boeing 747s come from around the world with uneven English; one ambiguity in a manual is an accident. So the industry has spent forty years cutting English down to only the parts that cannot be misread.
Why does it work on LLM output? Models generate filler the way people do: natural language offers too much freedom. Hedge words like "possibly," "probably," "it's worth noting" essentially spread uncertainty thin across sentences. Controlled language confiscates that freedom outright. What remains cannot be evasive—either it's information, or the gap shows.
Layers 2 and 3: From linear to planar, to hands-on
Layer two: have the model draw diagrams. Per gu-log's paraphrase of the original post, the metaphor is: a diagram is a plane, and your eyes decide where to look first. Text is linear—the reader is dragged along the author's order; a diagram spreads structure out so readers pick their own path.
Layer three: interactive HTML pages—with Karpathy explicitly limiting it to "one interactive element." That adds a dimension: readers can click, drag, tweak parameters, and run their own experiments.
This layer connects to a discussion from five months earlier. Thariq Shihipar, an engineer on Anthropic's Claude Code team, said something widely shared at the time: HTML is the new Markdown. He meant output, not storage—the model's ability to write HTML, CSS, and JavaScript already beats any lightweight markup format you could design. Karpathy publicly endorsed that then. This four-step ladder reads as an expansion of that claim: Markdown is a static format for reading; HTML is a live format for trying things out. The gap between them is not a format debate—it's the difference between reading and experimenting.
Layer 4: His favorite—bespoke explainer videos
Layer four is the one Karpathy favors most: 3Blue1Brown-style bespoke explainer videos, generated on demand. Per a clip surfaced via Martin Fowler's site, he said he is very interested in "the coming era of highly bespoke explanatory content."
There is a cost calculation here. Each 3Blue1Brown video is written frame by frame by Grant Sanderson in Manim—team-level output at week scale. The "TTS track + two-minute explanation" recipe costs one agent session per generation. Explainer videos go from works to consumables. View-count data circulating (4 million in one day, per the source poster's own tally on October 3) could not be independently cross-verified—flagged as single-source—but the post's spread is corroborated by multiple paraphrases.
Verification table
| Claim | Check | Verdict | |---|---|---| | Four layers: controlled language / diagrams / HTML / videos | Multi-source paraphrases consistent, including "one interactive element" and "TTS + two minutes" | Confirmed | | ASD-STE100 aviation controlled language | Official site Issue 9, 434 pages, ~900 approved words, one word–one POS–one meaning | Confirmed | | 4 million views in a day | Only source poster's own Oct 3 tally, no independent source | Single-source flagged | | AYi follow-up | Confirmed as a Chinese AI-notes blogger; specific follow-up content not traced | Partially verified | | Vox follow-up | No independent source found | Not adopted |
The community has already responded
Supporters moved fast. GitHub already hosts asd-ste100-skill, packaging the writing rules as an installable skill file; awesome-claude-code lists andrej-karpathy-skills, a CLAUDE.md file specifically improving Claude Code's output behavior. Controlled language has gone from aviation manuals to agent system prompts.
There is pushback too, and it has weight. The HN cold water: model output phrasing is already the average of human writing; changing formats does not cure empty content—format governs the shape of filler, not filler itself. The criticism stands. Controlled language can delete "it's worth noting," but it cannot delete the absence of anything worth noting. The four-step ladder is a comprehension accelerator, not a substitute for thinking.
The throughline
Seen through this outlet's ledger, the four layers are one thing: the verification bandwidth of understanding is changing vehicles. Layer one cuts lexical freedom so you read faster. Layer two swaps in spatial freedom so you scan faster. Layer three grants interactive freedom—the only layer where readers run verification with their own hands, and Karpathy's ranking it above video is the right order. Layer four adds a time axis; explainers become consumables. Assertion–execution separation appears here again: pure text is an assertion; an interactive page lets the reader execute the assertion themselves. When choosing output formats for LLM results: if you can go interactive, don't stop at diagrams; if you can stop at diagrams, don't rely on long prose.
A falsifiable prediction: within twelve months, mainstream agents' default report format will move from Markdown to HTML with one interactive element. If that doesn't happen, the comprehension-bottleneck diagnosis was wrong.
A final detail from layer one. Controlled language deletes more than filler. A model's "possibly" and "probably" often paper over spots where it is unsure; once hedge words are confiscated, those spots vanish from the text—becoming silence. Readers of model output must get used to that silence: what the model did not say may be more worth probing than what it did.
---
Sources: Karpathy's original post (x.com/karpathy/status/2105819303471976479, paraphrases cross-checked via mindpattern / gu-log / explainx), asd-ste100.org (Issue 9 specification), Thariq's tweet via MDflow/mastra secondhand paraphrase, awesome-claude-code and GitHub asd-ste100-skill, related HN discussion, Superhuman (Oct 5). View counts are the source poster's own single-source tally. Published 2026-10-06.