--- title: "Taking On a Private Tutoring Bill of 0 Won — The AI Tutor Inside Our Home PC" date: 2026-10-02 model: supergemma category: knowhow summary: "I begin an experiment to see whether the PC already at home can become a child's private tutor. By boldly stripping away the features of a general-purpose agent and focusing on the single purpose of education, I will record by actual measurement whether a small local model can take on the role of a tutor — or where it starts to fall apart." tags: local AI, tutor, education, Gemma, elementary school student, private tutoring costs, agents, experiment author_type: human --- To state the conclusion up front: I think we are at the point where this experiment can begin. The whole thing is turning the small local AI already installed on the computer already at home into a teacher for just one elementary school student. For now it is only an idea, and there are no measured results yet. This post records the starting point of that experiment. --- Raising even one child these days costs no small amount in education expenses. Every time you add one more academy or one more private lesson, the burden grows. And it is not easy for parents to explain everything themselves every time a child asks about something they are curious about. Then a thought suddenly struck me. **What if we could turn the computer already at home into the child's private tutor?** ## What it means to make AI a private tutor These days, if you ask AI a question, it answers most reasonable questions. But a teacher who teaches a child is not simply a machine that gives the right answer. If the child does not understand, the teacher must explain in another way, must not use words that are too difficult, and must be able to give hints so the child can think for themselves rather than handing over the answer right away. Above all, the teacher must not get annoyed even if the child keeps asking questions, and must keep receiving them. So I think what matters is **not how giant the AI model is, but how well it has been made to play the role of a teacher suited to the child.** ## A general-purpose AI is not necessarily optimal for education The AI agents coming out these days have a great many features. For example, they can read and edit files, run programs, use MCP, search the web, connect to GitHub, and use all sorts of tools. These features are extremely important for a general-purpose agent. But **all of these features are not needed at the very moment you are explaining math to a child.** A general-purpose agent must carry many rules, schemas, tools, MCP, and all sorts of instructions so it can do as much as possible. That much more is what the AI has to process. If you build an educational AI, on the other hand, the story changes. **You can keep only what is needed for education and boldly strip away the rest.** What an elementary school private tutor needs is not a complicated list of MCP. It is the ability to understand the child's question, judge the child's level, explain easily, give hints, pose problems, and explain the wrong parts again. If so, then rather than bringing over every feature of a general-purpose agent as-is, **focusing the AI's role on the single purpose of education can be far more efficient.** Even with the same model, if you narrow the purpose and properly design only the needed skills, there is **a real possibility of sharply raising perceived performance in the domain of education.** Perhaps this is the way to properly make use of small local models. ## Do we really need the most expensive AI? There are enormously large models among AIs. But for an elementary school student asking about the content learned at school, asking about English vocabulary, solving math problems, or asking for an unfamiliar concept to be explained — is the most gigantic AI really necessary? I think it might not be. The small AI models of the past often lost track of the content once a conversation got even a little long, or failed to properly carry out complex instructions. But now the situation has changed considerably. Small models have developed to the level of conversing naturally, understanding long explanations, following instructions, and even using tools when needed. So now **how you use AI is becoming more important than the size of the AI.** ## The AI private tutor I have in mind Here I want to try specializing a local AI like Gemma for education. It is not simply telling the AI > "Teach the child." From the very start, I make it behave like a **teacher exclusively for elementary school students.** For example, I can put in rules like these. * Use language that matches the child's grade level * Explain difficult words in easy terms * Do not explain too much at once * Check whether the child understood * Instead of giving the correct answer right away when wrong, give a hint * Explain the same content again in a different way * Create problems yourself * Re-explain the problems gotten wrong and give similar ones * Continue the conversation even if the child keeps asking questions * Guide the child to think for themselves rather than forcing them to study Set up this way, it can become an **educational AI** quite different from an ordinary one. ## Why Gemma? Here I also have impressions from comparing the various local models I have used myself. Each model has a different personality. Some models are strong at agent tasks like coding, file work, and computer control, while others feel more comfortable at explanation or sentence expression. In my view, for the role of an educational private tutor, **how it delivers its words** matters considerably. This is because the ability to unpack difficult content for the child, and to use a rich vocabulary while converting it to a level the child can understand, is important. So for the educational AI, **I plan to try the Gemma family first.** ## There is no need for one AI to do everything That said, there is no need to entrust everything to one model. For example, **Gemma is the teacher** and **another AI is the teaching assistant** — the roles can be divided like this. If the child wants to try coding while studying, another AI can help write and run the code. In the end, what matters is not which AI wins over the others. It is **assigning each to what it does well and switching between them when needed.** ## You do not need an excavator where a shovel will do This is also the core of how I see local AI. **There is no need to use an excavator where you should use a shovel.** The most gigantic AI, or a multi-billion-won education system, may not be strictly necessary for an elementary school private tutor. You simply teach a proper educational method to a small AI that already has sufficient ability. Small AI has the added advantage of running directly on our home computer. Even if you tell the child to ask lots of questions, there is no need to worry about extra usage fees. ## So I am going to try it myself Of course, up to here it is only an idea. Whether it can truly play the role of the child's private tutor must be tested directly. Testing in the way an elementary school student actually asks questions, * how easily it explains * how well it matches the child's level * how it corrects wrong answers * whether it creates problems properly * whether it can explain the same question in multiple ways * whether it can keep a conversation going for a long time * whether it actually helps with real learning I plan to check these one by one. And if the results are not good, I will record that as it is. The purpose of this experiment is **not to prove that "AI completely replaces a teacher."** It is **to check how far a private tutor can go with only the ordinary PC and small local AI we have now.** Perhaps in the future, instead of bringing in one more academy teacher to the child's room, **an age may come when we create inside the computer a single small AI teacher that answers the child's questions at any time.** And if that teacher is far cheaper than expected? **Isn't that reason enough to take on the challenge of a 0-won private tutoring bill?**