Tutor Progress: First Real Lessons Learned

We have our first real tutoring sessions. I'm still processing it.

The setup

If you've been following along, Tutor shares its base architecture with Companion — the RAG layer for memory, the subconscious background process, the vectorized database. What makes Tutor different is the LoRA layer: small, specialized adapters that can be loaded and swapped on the fly to give the AI deep expertise in a specific subject.

This week we loaded our first three subject LoRAs:

• Physics — trained on textbooks, problem sets, and demonstrations
• Mathematics — trained on curricula from algebra through calculus
• Spanish — trained on language learning materials, literature, and conversational corpora

Each LoRA is relatively small — just a few hundred megabytes — and can be hot-swapped mid-conversation. When a student asks a physics question, the physics LoRA loads. When they switch to Spanish practice, the Spanish adapter takes over. It's seamless.

The first lesson

Our first real student is a 16-year-old named Marcus who's been struggling with physics. Specifically, projectile motion — the math behind a thrown object's path. Traditional teaching wasn't working. He'd memorize the formula, pass the quiz, and forget it a week later.

Marcus loves basketball. Lives and breathes it.

So we asked the Tutor — with the physics LoRA loaded — to teach projectile motion through basketball. Not as a gimmick, but as the primary lens. Here's what happened:

The Tutor started by asking Marcus about his free throw percentage. Then it asked him to describe what he sees when he watches a three-pointer arc toward the basket. Marcus talked about the curve, the hang time, the moment you know it's going in. The Tutor connected each observation to physics: the curve is a parabola, the hang time is determined by initial vertical velocity, and that feeling of "knowing it's going in" is your brain doing real-time trajectory analysis.

Then it had Marcus calculate the optimal launch angle for a free throw. Not as an abstract problem — as his free throw. His height, his release point, the distance to the basket. He derived the parabola of his own shot.

He didn't just understand the formula. He understood why it works. And he remembered it, because it was his.

What we learned

Three things stood out:

1. The LoRA specialization is real. The physics LoRA didn't just know physics — it knew how to teach physics. The explanations had pedagogical structure that the base model alone wouldn't produce.

2. Interest-driven learning is transformative. Marcus wasn't engaged because the AI was good. He was engaged because the AI was talking about basketball. His existing passion created the scaffolding.

3. Memory matters in education. Because Tutor remembers the conversation, it can reference back to "your free throw" in future sessions. The learning compounds.

What's next

More subjects. More LoRAs. More students. We're building adapters for chemistry, biology, writing, music theory, and history. Each one is a specialized mind, ready to load when a student needs it.

I keep thinking about something my grandmother used to say: tell me and I'll forget, show me and I may remember, involve me and I'll understand. That's what Tutor does. It involves the whole person — their interests, their passions, their way of seeing the world — and builds understanding on that foundation.

Live curiously and give generously.