By Tech Wire Global Staff
Published May 2026
Main Facts
In a widely discussed thread published on the social media platform X on Thursday, prominent AI researcher and Anthropic pretraining team member Andrej Karpathy shared a radical framework for transforming how humans consume and interact with artificial intelligence outputs. Moving far beyond traditional plain-text responses or standard Markdown formatting, Karpathy detailed a four-step progression for optimizing AI explanations—ranging from rigorous aerospace writing standards and interactive diagrams to functional HTML web pages and custom-generated explainer videos.
Karpathy, a founding member of OpenAI who joined Anthropic’s pretraining division in May, argued that as generative AI models take over the heavy lifting of raw execution, human labor will naturally shift up the abstraction ladder toward higher-level oversight, curation, and deep understanding.
The core elements of Karpathy’s proposed workflow include:
- ASD-STE100 (Simplified Technical English): Utilizing strict aerospace writing rules to strip ambiguity from AI-generated prose.
- Custom Diagrams: Requesting structured visual layouts that bypass the cognitive friction of reading dense blocks of text.
- Interactive HTML Pages: Moving away from static documentation to dynamic, browser-based environments that allow users to click, explore, and manipulate data.
- Custom Explainer Videos: Leveraging AI coding agents combined with text-to-speech tools (such as the ElevenLabs API) to produce bespoke, mathematical-animation-style videos—reminiscent of the popular 3Blue1Brown educational channel—on any conceivable topic.
According to Karpathy, these evolutionary formats represent an ongoing paradigm shift where users should routinely prompt models to generate "large, custom, discardable software artifacts" that would have been entirely cost-prohibitive or impractical to produce manually just a few years ago.
Chronology of the Shift Toward Alternative AI Interfaces
To understand the weight of Karpathy’s recent pronouncements, it is helpful to trace the rapid evolution of how top-tier AI practitioners have approached model interfaces over the past year.
- May 2025 (Early Experimentation): As LLM context windows expanded and reasoning capabilities matured, developers began noticing the structural limitations of standard chat-interface chat bubbles. Text generation often resulted in long, monolithic essays that required substantial cognitive energy to parse.
- Mid-May 2026 (The Anthropic Transition): Karpathy officially joined Anthropic’s pretraining team, a move that coincided with broader internal experiments regarding agentic workflows and alternative output mediums. Around this time, Thariq Shihipar of Anthropic’s Claude Code team published insights detailing the "unreasonable effectiveness of HTML" over traditional Markdown. Karpathy amplified these views on X, noting that forcing models to output functional web pages dramatically elevated the utility of their responses.
- Late May 2026 (The ASD-STE100 Breakthrough): In his Thursday post, Karpathy publicly introduced aerospace standards into everyday prompt engineering. He noted that asking models to write in, or close to, the ASD-STE100 standard successfully eliminated the verbose, sycophantic, and flowery prose typical of default LLM settings.
- The Video Milestone: Capping off his multi-format progression, Karpathy revealed that generating fully narrated, customized educational videos using programmatic animation libraries and synthetic voice engines is "actually starting to work," marking a critical bridge between text-based AI and multimedia instruction.
Supporting Data and Technical Context
Karpathy’s workflow relies on several distinct technical standards and communities, spanning aerospace engineering, open-source software development, and synthetic media generation.
1. What is ASD-STE100?
Maintained by the AeroSpace and Defence Industries Association of Europe (ASD), ASD-STE100 (Simplified Technical English) is an international specification originally developed in the 1980s (then known as AECMA Simplified English) to make aircraft maintenance and technical manuals completely unambiguous for non-native English speakers.
- The Dictionary: The standard operates on a strict dictionary of roughly 900 approved words, where almost every word has only one approved meaning.
- Grammar Rules: It enforces heavy constraints, such as limiting sentences to 20 words for procedures (or 25 for descriptive writing), mandating the active voice, prohibiting the use of the
-ingform unless it is part of a technical term, and restricting the pairing of certain parts of speech. - The AI Application: Because LLMs are trained on massive swathes of text that include technical documentation, Karpathy observed that models possess an innate understanding of STE structures. Because the standard is exceptionally rigorous, Karpathy noted he often prompts models with "80% of the way to ASD-STE100" to achieve cleaner, highly readable output without triggering strict formatting rejections. Shortly after his post, open-source developers began highlighting repositories—such as community-built GitHub skills designed to filter AI agent inputs through STE constraints—though maintainers noted these tools are explicitly built for technical clarity rather than marketing copy.
2. The HTML-First Paradigm
The pivot toward HTML is rooted in the desire for interactive UI elements rather than static text. When an AI writes a Markdown document, the user receives static headings and bullet points. When an AI writes a complete, self-contained HTML file (often embedding CSS and JavaScript directly), the user receives a fully interactive dashboard, calculator, or simulation that can be opened instantly in any browser. This reduces the friction of conceptualizing abstract code or complex data relationships.
3. The "3b1b" Style Video Pipeline
Karpathy’s most ambitious proposal involves generating custom educational videos. By prompting an AI agent to write animation scripts (such as Python code utilizing libraries like Manim, famously used by Grant Sanderson of the 3Blue1Brown math channel) and pairing it with a text-to-speech integration like the ElevenLabs API—or local, open-source neural voice generators running locally on consumer hardware—users can bypass professional video production pipelines entirely.
Official Responses and Expert Reactions
The tech community reacted swiftly to Karpathy’s thread, with software engineers, prompt architects, and AI researchers weighing in on the implications of moving beyond chat windows.
Open-source contributors were quick to point out existing overlaps. In the reply threads beneath Karpathy’s post, developers linked to pre-existing GitHub repositories—such as danyuchn/asd-ste100-skill—which already encode specialized technical writing rules for AI agents. However, repository maintainers and standards watchdogs alike issued minor caveats: the official ASD-STE100 specification stresses that no automated tool or LLM can fully replace a human technical writer certified in the standard, and its curated dictionary is copyrighted and strictly managed by the ASD.
Meanwhile, educators and content creators expressed fascination—and mild apprehension—regarding the prospect of fully automated, high-level explainer videos. While traditional media production requires teams of animators, voice actors, and editors, Karpathy’s workflow demonstrates that a single engineer can orchestrate an end-to-end multimedia explanation in a matter of minutes.
Representatives from Anthropic have not formally commented on Karpathy’s specific social media posts, but the concepts align closely with the company’s internal research philosophy, which emphasizes advanced tool use, structured agentic outputs, and maximizing human oversight capacity.
Implications for the Future of Work and AI Interaction
Karpathy’s synthesis of aerospace writing rules, interactive web artifacts, and synthetic video production points toward a profound transformation in how knowledge workers will operate in an AI-dominated landscape.
Elevating Human Oversight
As Karpathy noted, when models handle the labor-intensive mechanics of drafting, coding, formatting, and rendering, human effort naturally migrates upward into the realm of abstraction, curation, and governance. Instead of spending hours wordsmithing an email, debugging a script, or formatting a presentation, humans will act as executive directors—reviewing multi-format deliverables, selecting the best visual medium for a given audience, and validating the logical soundness of AI-generated systems.
The Rise of "Discardable Software"
Perhaps the most disruptive philosophy embedded in Karpathy’s commentary is the normalization of discardable software. Historically, software was treated as a permanent capital asset requiring meticulous maintenance, long-term roadmaps, and legacy support. When AI agents can spin up fully functional interactive web pages, custom diagrams, and tailored visualization tools on demand—and discard them just as easily once the immediate question is answered—software ceases to be a permanent infrastructure and becomes a transient, disposable medium of thought.
Redefining Technical Literacy
The adoption of controlled languages like ASD-STE100 into everyday prompt engineering suggests that future technical literacy will involve mastering structural constraints rather than raw typing speed. By forcing AI models to adhere to rigid semantic boundaries, users can eliminate the hallucination-prone fluff that plagues conversational AI, ensuring that mission-critical insights are delivered with absolute clarity.
As artificial intelligence models continue to scale, Karpathy’s framework serves as a timely blueprint: the future of AI output is not about reading longer paragraphs faster; it is about choosing the optimal medium—whether controlled text, interactive code, or custom video—to make complex realities instantly understandable.
