@uudonX: Karpathy is dishing out more benefits to the industry! Codex / Claude code generates a ton of text for you—do you reall…

🎁 Karpathy is dishing out more benefits to the industry! Codex / Claude code generates a ton of text for you—do you really read every single word without skipping a beat? Don’t do that; it’s way too inefficient. Try out the methods Karpathy proposes below.

You can only marvel at it: Great minds are great for a reason—they’re not just smart, but hardworking too, constantly devising ways to boost their own efficiency.

Karpathy raises a problem many overlook: When AI’s output speed outpaces human comprehension, just letting it churn out more text only piles up the bottleneck higher and higher.

Models can whip up tens of thousands of lines of code, research reports, and complex plans in minutes, but humans still have to figure out what it did, whether the logic holds up, and if the results are actually usable.

So here are Karpathy’s proposed solutions: How to transform the model’s output into formats that are easier for humans to grasp and verify.

1️⃣ Constrain text with ASD-STE100. Regular models love long sentences, abstract terms, and vague phrasing. They’ll ramble on about the same concept for pages, and you still won’t know the key takeaway. ASD-STE100 was originally for aviation maintenance docs—it limits vocabulary, sentence structures, and expressions, demanding every sentence be clear, direct, and unambiguous. You can prompt the model: “Explain this issue in 80% ASD-STE100 style.”

The payoff? It cuts language noise, slashes ambiguity, and makes technical docs and complex ideas way easier to review.

2️⃣ Generate charts. The tough part with some content comes from tangled relationships—like system architectures, call chains, causal links, and module dependencies. Linear text struggles to show all those structures at once. Have the model spit out flowcharts, architecture diagrams, or concept maps, and you can instantly see how nodes connect, data flows, and where issues might crop up.

The payoff? It slashes comprehension costs and makes it easier to spot missing links or logic gaps.

3️⃣ Generate interactive web pages. If a problem involves states, parameters, or dynamic shifts, static text and images just don’t cut it. That’s when you get the model to output HTML, using buttons, sliders, and animations to demo: What happens to results when parameters change? How does the system behave in different states? The payoff? It turns “reading conclusions” into “hands-on verification.” Users can tweak parameters, watch outcomes, and quickly build causal intuition.

4️⃣ Generate explainer videos. Algorithm runs, physical processes, and abstract concepts all unfold in sequences over time. Videos can walk through changes step by step, blending animations, voiceovers, and key highlights. Karpathy suggests having the model straight-up create 3Blue1Brown-style explainer videos.

The payoff? It builds intuition fast—perfect for concepts that are a slog to visualize from text alone.

The real shift behind these ideas is that the cost of “disposable software” is nearing zero. Back in the day, crafting a custom webpage, animation, or video for one problem wasn’t worth it.

Now models can whip up tailored stuff on the fly based on your knowledge level—use it, toss it, and stop slogging through every word the model dumps on you. It’s inefficient; give the methods above a shot.

Andrej Karpathy (@karpathy): We’ll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I’ve had success with: Ask your LLM to explain something in ASD-STE100, it’s a controlled language specification originally developed for