,

3 min read

Karpathy’s AI Literacy Standard: From Prompts to Agency

An anonymous professional assembles an AI-powered media workflow while a repetitive recommendation stream fades behind them.

Your team can complete an AI course, use a copilot every day, and still be unable to improve a broken workflow without waiting for an expert. That is why adoption is a weak proxy for AI literacy. The question is not whether people can make a model produce text. It is whether they can use models to regain control over a real outcome.

Karpathy’s HackerTube experiment offers product and AI leaders a more demanding standard. Faced with a YouTube feed that did not match his idea of valuable content, he chose a different quality signal, used models to process it, and shipped a working alternative. In this view, AI literacy is applied agency: diagnose the problem, specify the desired result, delegate bounded work, verify the output, and take responsibility for what gets deployed.

AI literacy begins with questioning the default

A recommendation feed can work as designed and still work badly for you. YouTube ranks content for engagement at population scale. Your definition of a worthwhile video may be narrower, more technical, less sensational, or simply different from the behavior its system rewards.

The literate response is not merely to complain about the output or ask an AI assistant for a better list. It is to examine the objective behind the output. Before choosing a model or writing a prompt, ask:

  • What is the existing system optimizing? Name the behavior, not the feature. A feed may optimize continued viewing even when you care about learning.
  • Where does that objective diverge from the result you want? Describe the mismatch in observable terms. Avoid vague labels such as “bad recommendations.”
  • What evidence would represent value more faithfully? Look for a signal created by people whose judgment is relevant to the job.
  • Which parts require judgment, and which are processing work? Keep the value decision with a person. Give the model the repetitive reading, classification, coding, or transformation.

HackerTube followed that sequence. Instead of asking a model to decide which videos were good, it treated recommendations that programmers made to one another as the quality signal. A smaller Anthropic model sorted Hacker News comments spanning 2007 through 2026 and retained roughly 40,000 comments that recommended YouTube videos. The resulting collection contained 6,454 videos across 150 topics. Claude Opus 5.5 wrote the website, placed it on a rented server, and secured it.

The important design choice came before the code: people supplied the judgment, while AI supplied the leverage. That distinction matters for almost every AI product. If you cannot state who defines quality and what evidence represents it, faster generation will only produce uncertain results faster.

The build reveals four capabilities worth teaching

The visible part of this kind of work is often called vibe coding. The less visible part is product judgment. Prompt fluency helped, but it was only one component of a larger capability.

<!– wp:list {

Want this applied to your product org?

A free 45-minute consultation: AI product strategy, GTM, transformation and PM hiring — practical next steps, no pitch.