Andrej Karpathy Explores the Future of AI-Generated Output: From Aircraft Maintenance Standards to Custom Explainer Videos

SAN FRANCISCO — In the rapidly evolving landscape of artificial intelligence, the bottleneck is no longer how much information a language model can generate, but how effectively humans can consume, digest, and oversee that vast ocean of data. Andrej Karpathy, a prominent AI researcher, founding member of OpenAI, and current member of Anthropic’s pretraining team, has proposed a provocative blueprint for solving this cognitive overload.

In a widely discussed post on X (formerly Twitter) published on Thursday, Karpathy shared a four-tiered progression for consuming model outputs—ranging from ultra-controlled technical English to fully rendered, custom-made explainer videos. According to Karpathy, as artificial intelligence systems shoulder increasingly heavy workloads, human labor will naturally transition away from raw production and “rise up the abstractions into oversight and understanding.”

This development arrives at a fascinating crossroads in AI development, highlighting how niche human communication standards—originally engineered for high-stakes aerospace environments—are unexpectedly finding new life as optimization tools for LLMs.

Main Facts

The core of Karpathy’s proposal centers on fundamentally altering how we request and interact with information generated by large language models (LLMs). Rather than accepting standard, flowing prose, Karpathy advocates for targeted, highly structured, and multi-modal presentation layers.

His four-step progression includes:

1. ASD-STE100 (Simplified Technical English): Utilizing strict aerospace maintenance writing standards to force models into delivering clean, unambiguous, and highly readable text.
2. Diagrams: Bypassing text entirely in favor of structured visual representations that are easier for the human brain to parse.
3. Interactive HTML Pages: Leveraging raw web code rather than traditional Markdown to create dynamic, interactive documents—a methodology heavily favored within Anthropic’s internal engineering ranks.
4. Custom Explainer Videos: The format Karpathy is "most bullish on," which involves prompting models to generate mathematical or conceptual explainers styled after popular educational channels like 3Blue1Brown, complete with synthesized narration via the ElevenLabs API or open-source local alternatives.

Karpathy’s overarching thesis is that generative AI enables the creation of "large, custom, discardable software artifacts" and media formats that were economically or logistically impossible to produce just a few years ago. By shifting our relationship with AI from passive readers to active directors, we can drastically elevate the quality of human oversight.

Chronology

To understand how Karpathy arrived at his current framework, it is helpful to trace the timeline of his recent public insights and industry movements:

• May 2024 (and prior): Karpathy maintains a continuous public dialogue regarding the pedagogical value of AI, frequently emphasizing the need for better interfaces between human thought and machine intelligence.
• Early May 2026: Karpathy officially joins Anthropic’s pretraining team, bringing his wealth of foundational model experience to the company behind the Claude family of assistants.
• Mid-May 2026: Karpathy shares an X post highlighting an insight from Anthropic’s Thariq Shihipar regarding the "unreasonable effectiveness of HTML" over Markdown for technical workflows. This practice is subsequently detailed on the official Claude Blog.
• Late May 2026 (Thursday): Karpathy publishes his viral four-part thread on X, detailing his success with ASD-STE100, dynamic HTML pages, and AI-generated 3Blue1Brown-style video explainers. The post instantly sparks widespread community engagement, code repository sharing, and debate across technical forums.

Supporting Data: The Mechanics of ASD-STE100 and Multi-Modal Output

To fully grasp the weight of Karpathy’s suggestions, one must examine the tools and standards he references.

The Aerospace Standard: ASD-STE100

ASD-STE100 stands for Airbus, Boeing and AECMA (Association Européenne des Constructeurs de Matériel Aérospatial) Simplified Technical English. Maintained by European aerospace, security, and defense companies, it was originally developed in the 1980s as a controlled language to make aircraft maintenance manuals easier to read for international mechanics whose native language might not be English.

Key features of the standard include:

• A strict dictionary of roughly 900 approved words, with most words restricted to a single, unambiguous meaning.
• Heavy constraints on sentence length (recommendations of maximum 20 words for procedures, 25 for descriptive writing).
• Mandatory use of the active voice.
• The elimination of gerunds and present participles where they might cause confusion.

Because the standard is exceptionally stringent, Karpathy noted that he often softens his prompts to "80% of the way to ASD-STE100" to give the model slight stylistic breathing room while maintaining clean, stripped-down output. Interestingly, the standard’s own documentation notes that STE is not designed for general creative writing or marketing copy. However, its foundational philosophy—clarity, brevity, and active phrasing—translates exceptionally well to LLM prompt engineering. Following Karpathy’s post, developers quickly surfaced open-source GitHub utilities—such as the asd-ste100-skill repository—designed to automatically apply these filtering rules to AI reading agents.

HTML and Markdown

Karpathy’s endorsement of HTML builds upon an internal workflow trend at Anthropic. Rather than settling for static Markdown files, prompting an LLM to generate a self-contained HTML page yields a rich, interactive canvas. Buttons, toggles, custom CSS styling, and local JavaScript can transform a flat list of bullet points into a dynamic dashboard tailored precisely to the user’s immediate cognitive needs.

Automated Explainer Videos

Perhaps the most forward-looking component of Karpathy’s post is his call for custom video generation. By invoking a "3b1b style" (a nod to Grant Sanderson’s legendary mathematics education channel, 3Blue1Brown), users can direct models to outline conceptual animations. Coupled with text-to-speech tools—such as an ElevenLabs API key or free, locally hosted open-source voice generators—users can spin up bespoke multimedia tutorials on demand. Karpathy noted with excitement, "This is actually starting to work!"

Official Responses and Industry Reactions

The AI research and development community responded to Karpathy’s thread with immediate enthusiasm and practical experimentation.

Open-source developers wasted no time integrating his ideas into existing ecosystems. Within hours of the post, replies on X were flooded with links to GitHub repositories, custom prompt templates, and specialized browser extensions designed to parse model outputs through simplified English filters.

Representatives from technical writing organizations have noted a bittersweet irony in the trend: a manual-writing standard designed decades ago to protect human engineers from mechanical errors is now being resurrected to help humans navigate the cognitive output of synthetic neural networks.

While the creators of ASD-STE100 maintain that no automated tool can fully replace the human editorial oversight required to certify an official aerospace document, the general consensus among AI practitioners is that Karpathy has successfully identified a profound shift in how developers will interact with synthetic text in the coming years.

Implications for the Future of Work and AI Interaction

Karpathy’s vision carries broad implications for the future of knowledge work, software engineering, and education.

1. The Shift to "Management" over "Creation"

As AI models become increasingly proficient at churning out code, essays, and data summaries, the human role naturally shifts upward. When generation is cheap and instantaneous, the primary human skill becomes curation, architecture, and critical evaluation. By demanding formats that are easier to parse—such as STE text, interactive HTML, or animated videos—humans are effectively building custom user interfaces for their own cognitive processes on the fly.

2. Disposable Software and Media Artifacts

Karpathy’s closing remark—urging people to ask models for "large, custom, discardable software artifacts that would have never made sense to create before"—points toward a paradigm where software ceases to be a permanent, heavily maintained asset. Instead, code, web pages, and explanatory videos can become single-use, highly contextual tools. If you need to understand a complex thermodynamic principle for twenty minutes, you can instruct an AI to spin up a fully interactive, narrated web app just for you, and discard it the moment you understand the concept.

3. Redefining Digital Literacy

As these multi-modal, highly tailored prompting strategies become mainstream, digital literacy will evolve past knowing how to write a basic prompt. Future knowledge workers will need to master the art of constraints—understanding how to impose strict stylistic frameworks (like ASD-STE100), leverage interactive web technologies, and orchestrate complex pipelines involving text, graphics, and voice synthesis.

Ultimately, Andrej Karpathy’s latest exploration reminds us that the true power of artificial intelligence lies not just in its raw computational output, but in its ability to adapt to the unique cognitive architecture of the human mind.