How to Talk to Grandma About AI (jstrieb.github.io)

In a recent essay, Jacob Strieb offers a practical analogy for explaining large language models (LLMs) to nontechnical audiences: treat LLMs like actors. They’re optimized to perform believable language — to “sound” right — not to establish truth. That explains why LLMs can produce convincing but false statements: they sample probable token sequences learned from vast, stylistically diverse training data rather than encoding facts. Strieb shows this has concrete, actionable implications for users: prompting models to adopt a persona (e.g., “You are a senior software engineer…” or “You are an experienced web designer…”) narrows the model’s output distribution toward the relevant slice of its training data, improving code, prose, and UI generation compared with plain or negative prompts. Beyond a helpful metaphor, the piece raises important community-level concerns. If users are “directors” who steer performances via prompts, the true “audience” shaping model behavior are the researchers and companies who design, select, and deploy models — meaning creators’ priorities and biases persist in released systems. That selection pressure, combined with marketing and hype, risks misapplication of LLMs by uninformed users. Strieb’s actor analogy thus both demystifies hallucination and gives a simple, effective prompting technique, while underscoring the need for clearer communication about models’ limits and for scrutiny of the incentives guiding model development.