What Does Cilantro Taste Like?
Bunches of coriander. Photo by Kpsudeep, 2020. CC BY-SA 4.0, via Wikimedia Commons.
Bunches of coriander. Photo by Kpsudeep, 2020. CC BY-SA 4.0, via Wikimedia Commons.
What does cilantro taste like? It’s a question students can answer for themselves, and a useful one to put to several language models. In this activity, they compare answers in the Allen AI Playground, then change the settings and see what happens to the responses.
No technical background is required. Check the playground before class and prepare saved answers if the relevant controls are unavailable. Choose two named models and record their visible settings; their parameter counts need not be the only difference.
Suggested submission: a small comparison table recording model, changed setting, response, and one observation. Distinguish a truncated answer from an inaccurate one, and repeat a setting before treating one answer as typical.
Students open playground.allenai.org on their devices, or you run the comparison on a projector.
First, try a small model. Choose an older, lower-parameter model and submit the question. Look for circular descriptions, repetition, and details that say little about the taste. Save the answer for comparison.
Then try a larger model. Submit the same question. Larger models may offer a fuller description: citrus, herbs, or the soapy taste some people experience. Ask students which details are useful and which they would want to check. Mentioning the soap connection is a difference worth discussing, but it doesn’t by itself demonstrate how well a model understands the subject.
Finally, change the settings. Use the same model and prompt for this round, changing one setting at a time:
prompt
What does cilantro taste like?
Constructed example — invented outputs, not a playground result. With one model and unchanged sampling settings, a short output limit yields “Cilantro tastes fresh and”; a longer limit allows the sentence to finish. This supports an observation about truncation. It doesn’t show that the longer answer is more knowledgeable.
If a second model mentions a soapy taste, students can record that added detail. They cannot attribute it solely to size when training and other settings also differ.
Ask students to point to a specific difference between two answers and identify what changed between the runs. Model size, training, and settings are different things; comparing two models doesn’t isolate just one of them. The settings round gives the class a more controlled comparison.
It also helps to compare responses across students. Even when they use the same model and question, the answers may vary. That gives the class something more precise to discuss than whether “AI” gives a good answer.
Check which models and settings the playground currently offers, and run the comparison once yourself. The available models change, so the examples above may not be the ones your class sees. If two answers are similar, keep that result in the discussion rather than treating the contrast as guaranteed.
This sketch is licensed under Creative Commons BY-NC-SA 4.0. You're welcome to share and adapt it for noncommercial purposes, with credit, as long as you share your adaptation under the same license.
Image Bunches of coriander. Photo by Kpsudeep, 2020. CC BY-SA 4.0, via Wikimedia Commons. The image's own license applies to it, separately from the sketch.
Suggested citation
Fred Gibbs. "What Does Cilantro Taste Like?." AI Sketchbook, Amaranth, University of New Mexico, March 1, 2026. https://amaranth.unm.edu/ai-sketchbook/teaching/model-parameters.html. Licensed under CC BY-NC-SA 4.0.