Andrej Karpathy stated in a article on X Thursday that he’s had achievement asking tongue models to explain things in ASD-STE100, a controlled form of English built to create airplane care manuals easier to read.
Karpathy, a founding associate of OpenAI, joined Anthropic’s pretraining team in May.
The article moves through four formats for study example output, from controlled penning to diagrams, HTML pages, and explainer videos, calling all stage “even better.” He wrote that as models do additional of the legwork, “a lot additional of our activity volition increase up the abstractions into oversight and understanding.”
Karpathy’s Case For Controlled English
Karpathy wrote that models cognize ASD-STE100 fine and that its “heavy constraints on spotless penning style” create content he frequently finds “a lot additional readable.” Because the norm is “quite stringent,” he has sometimes softened the petition to “80% of the way to ASD-STE100.”
ASD, which represents European aerospace, security, and defence companies, maintains the standard. Its FAQ calls Simplified Technical English (STE) an global norm for specialized writing, alongside a set of penning rules and a dictionary of concerning 900 approved words, most alongside one meaning.
A answer under the article links to an open-source skill that was already on GitHub and applies STE rules to content that AI agents read. Its README says the accomplishment isn’t built for promotion copy.
Diagrams, HTML Pages, And Explainer Videos
Instead of writing, Karpathy suggests asking for a diagram, which he says can be “a lot easier to process, parse, and understand.”
He additionally recommends asking for output “in HTML” to get an interactive web page. He gave akin direction in May, concerning a week before he joined Anthropic, whenever he shared an X article by Anthropic’s Thariq Shihipar on why members of the Claude Code squad use HTML alternatively of Markdown. A type of the part ran on the Claude Blog afterward that month.
The format he is “most bullish on” is tradition explainer videos on any topic. His example immediate asks for a “3b1b style” explainer, a citation to the 3Blue1Brown math videos, and tells the example to use an ElevenLabs API key for the narration. He added that group without a key can ask the example to discover liberated alternatives that run on their own computer. “This is really starting to work!” he wrote.
Why This Matters
Karpathy’s four suggestions are distinct ways to alter how example output is presented to the individual reviewing it. He frames the writing, diagram, and HTML formats as requests to a model, during his video example additionally needs an ElevenLabs API key or a liberated local substitute for narration. For the controlled-English tip, the standard’s own FAQ says STE isn’t meant for broad writing, although principles akin abbreviated sentences, one topic per sentence, and energetic sound can transport over.
Looking Ahead
Karpathy closed the article by urging group to ask models for “large, custom, discardable application artifacts” that “would have never made awareness to create before.”
The collection rearward ASD-STE100 gives out the complete norm for free, and its FAQ says no tool can substitute it. The open-source accomplishment connected in the post’s replies says it doesn’t contain the standard’s dictionary.
Featured Image: Teodora ART/Shutterstock
.png)
