Andrej Karpathy recently posted a tweet advising people to "spend more time designing how language models output their content." He shared several concrete approaches: ask LLMs to write following the ASD-STE100 standard, create diagrams, generate HTML pages, and even produce explanatory videos in the style of 3Blue1Brown. The tweet has since gained 50,000 likes.

But not everyone thinks these ideas are new. Elvis (@omarsar0) replied that he has been using most of what Karpathy shared for over a year. The real question worth discussing is: Is there a more general-purpose interface for human-agent collaboration?


Karpathy’s first tip sounds pretty niche: ASD-STE100, a form of controlled English originally designed for aircraft maintenance manuals. Its rules are so strict that "every word can only have one meaning"—for example, you can only use "start" and never "begin". Someone has already put this to the test: YouTuber Ege Vusal Chelebi turned the entire standard into an agent skill and a linter, then ran it across 6 categories of real-world writing tasks. The results: default LLM outputs had 4.36 violations per 100 words; the popular "ban words like delve, ban em dashes" checklist only brought it down to 4.21; Orwell’s six rules of writing brought it to 2.48; ASD-STE100 dropped it all the way to 1.12. Banning specific words doesn’t work—giving the model a complete set of grammatical constraints does. [Video](https://youtu.be/uJblcC4lKYw)

Elvis’ experiments are even more ambitious. He has been building a "general-purpose interface" that turns agent outputs into Notion-like pages, where you can embed text, videos, interactive charts, and visual interpreters—pretty much anything an LLM can generate. If the page doesn’t support a certain format, he just asks the agent to build it on the spot. This interface works for code review, note-taking, writing, research, and prototyping. He says it feels like collaborating directly with a design-savvy colleague through work artifacts, instead of reading a wall of text in a chat box.


He even tested this with GPT-6 Astra, using it to generate an interactive breakdown of the original Transformer paper:

In the thread under his reply, many others shared similar attempts. One person had their agent turn release predictions into a chart, saying "it wrote a block of text I couldn’t get through, I asked it to turn it into a chart and it instantly made sense." Another had an LLM generate a 3Blue1Brown-style explanation video on stochastic calculus. Some people are even collaborating with agents in VR.


Karpathy himself added a line in a reply: "99% of people have only used AI for less than a year now, that confuses the AI dinosaurs." That line probably explains why the tweet got 50,000 likes—for most people these are new tricks, but for long-time practitioners, this is just daily work.

Elvis says this is just a personal experiment, and he hasn’t found a perfect solution yet. But he’s noticed that Anthropic’s Artifacts is already being widely adopted, and dot’s Spaces is slowly evolving into the setup he built months ago. AI output is shifting from "a block of text" to "a collaborative work artifact". The chat box may just be an on-ramp—the actual interface is still evolving.