on the AI tutor, I originally allow a user ask a few qeustions without requiring login, but Claude advised against it for security reasons (possible abuse of unauthenticated model relay/call)
on the AI tutor, I originally allow a user ask a few qeustions without requiring login, but Claude advised against it for security reasons (possible abuse of unauthenticated model relay/call)
Because I'm using pretty small models (biggest option I've been working with is a 14B) and primary age kids writing is really different from adults, there's a lot of very targeted prompts around mechanics, vocabulary, sentence and paragraph structure, general encouragement and so on. I've also got LanguageTool feeding spelling and grammar analysis into it, since small language models just make up so much dumb stuff when it comes to spelling (it helps that we have an Amazon Echo and they're used to dumb AI so don't expect too much of it).
The kids can "submit" their draft to get a bit of summative feedback which runs it through the whole suite and they can go through the different feedback contexts (highlights relevant passages, suggests word choices, gives general feedback related to the writing purpose and prompt). Each draft takes a snapshot so my wife and I can see what the kids change about the draft after the feedback. This part actually works pretty well.
There's also a "nudge" mode where occasionally the LLM tutor will surface a bit of feedback with a subtle notification while they are writing. This aspect sounded pretty good in my head but since it's analysing a writing fragment sometimes it's not super helpful and even though the notification isn't too disruptive, it still breaks into the flow.
The more I think about it, the more useful this sort of model sounds useful for code too, since it doesn't need to just be about "does the code run, are the outputs correct" but could reason a bit more about program structure, naming of variables and functions and whether they are appropriate to the purpose, quality of comments and whether they indicate an understanding of the code and so on. Most of that you could get with a decent AST analysis, but relevance of comments etc is something that could be quite useful.
Good luck with teaching Python to your own kiddo. I'm finding my own are more interested in programming microcontrollers like the micro:bit than graphics. Gets you really comfortable with fairly simple code and immediate feedback about whether you gets things right. See the micro:bit Python editor for a really well scaffolded IDE too (https://python.microbit.org/v/3).