[zozbot234]

> - cannot reliably be restricted in functionality to leave the part the students are supposed to do themselves to the student

Small local LLMs are essentially that. If an LLM can tell you to eat rocks as a tasty snack or use glue to make the cheese stick to your pizza, imagine what it says when you ask it to analyze/explain complex academic subjects, or solve fiddly problems. But it will still reliably help you polish your language, like a subject-specific dictionary/thesaurus.

[yorwba]

Small LLMs are certainly restricted, but they're not reliably restricted in a way that aligns with educational goals. Even Qwen3.5-0.8B, the smallest, dumbest LLM I have lying around, can solve quadratic equations. In math class, learning about quadratic equations for the first time, such helpfulness defeats the purpose. In physics, learning about ballistic trajectories, it might be acceptable, as long as the students themselves come up with the correct equation to model the problem, and do the reasoning to identify which of the two theoretical solutions corresponds to the physical setup.

A single model that can do many things, some less reliably than others, independently of what the requirements for the lesson are, is not a good tool to give students. You would need something that can do ancillary tasks perfectly, but won't do the part student are supposed to practice doing themselves. And what exactly that is changes with every exercise.

If you want to give students access to thesaurus functionality and nothing else, you're better off with a thesaurus.