Go to Allen AI Playground, pick the smallest available model, and ask it one question: What does cilantro taste like? Then pick the largest model and ask the same question. The difference between those two answers is a 20-minute lesson in how language models actually work.
Have students open playground.allenai.org on their own devices, or run it as a class demo on a projector.
Round 1 — small model: Select an older, lower-parameter model from the dropdown. Submit the prompt. Note what comes back: responses from small models tend to be short, repetitive, or circular — describing cilantro in terms of cilantro, producing thin or generic sentences, sometimes looping.
Round 2 — large model: Switch to the largest available model and submit the same prompt. The difference is usually striking. Larger models typically describe cilantro’s fresh, citrusy, herbal quality — and often mention the well-known genetic variation that makes it taste like soap to roughly 10% of people. That detail is a good marker of genuine knowledge depth.
Round 3 — model settings: With the large model selected, experiment with the available parameters:
Cilantro is an ideal demo question because it has a concrete, sensory answer students can verify from their own experience — and because the soap-taste detail is a meaningful test of whether a model has encountered real knowledge about the world, or just plausible-sounding filler. The genetic variation angle also opens a brief but productive conversation about what “training data” actually means: the model knows about cilantro-as-soap because enough people wrote about it and it “knew” enough to not filter it out.
The parameter demo grounds abstract concepts — temperature, tokens, sampling — in something students can observe in real time rather than read about in a textbook. Most students arrive thinking AI is a single, fixed thing. This activity gives them a concrete counter-example in under 30 minutes.
The specific models available on the Allen AI Playground change over time — the contrast between small and large models is the point, not any particular model’s result. If the playground changes its interface or model lineup, the activity still works; just find the smallest and largest options available.
The activity works best when students generate their own responses so they can see variation across the class — even the same prompt to the same model will produce slightly different output each time, which is itself worth discussing. Everyone generalizes about “AI output,” which makes sense, but students benefit from seeing how much variation lives inside that phrase.