Announcing Tutor: Learning at the Speed of You

I've been writing this post in my head for three months. Today I get to share it with you.

When I announced Companion in February, I described an AI that truly remembers — one that builds on every conversation you've ever had with it through persistent memory, a subconscious process, and a vectorized database. Companion was born from a specific question: what if AI could remember the way a close friend does?

Today, I'm asking a different question: what if AI could teach the way a great mentor does?

I'm calling it Tutor.

The Problem With One-Size-Fits-All Learning

Every student who has ever struggled with a subject knows the frustration: the textbook explains it one way, the teacher explains it another way, and neither way clicks. Then a friend says the same thing in different words — maybe with an analogy that connects to something you already understand — and suddenly it makes sense. The knowledge was always available. What was missing was the right frame.

Great teaching isn't about information transfer. It's about finding the frame that makes information land for a specific person, at a specific moment, with a specific set of prior knowledge and interests.

Current AI tutors are bad at this. They explain things competently but generically. They don't adapt to how you learn. They don't remember that you're a visual thinker, or that you understand physics through basketball analogies, or that you struggled with fractions last week and need that foundation reinforced before tackling percentages.

Tutor is built to change that.

The Architecture

Here's what makes Tutor work, and I want to be transparent about the design.

Tutor shares its base model with Companion — the same foundational weights, the same underlying language capabilities. But on top of that base, Tutor loads custom LoRA adapters — small, specialized fine-tuned modules — for different subjects. A physics LoRA, trained on pedagogical explanations and problem-solving patterns specific to physics. A mathematics LoRA, tuned for step-by-step mathematical reasoning. A literature LoRA, calibrated for close reading and textual analysis.

Think of it as one mind with many lenses. The base model provides general intelligence, conversational fluency, and reasoning. The LoRA adapters provide subject-specific expertise and teaching strategies. When you ask Tutor about photosynthesis, it loads the biology adapter. When you switch to asking about the French Revolution, it loads the history adapter. Same memory. Same relationship. Different specialized expertise, activated on demand.

And — this is the part that makes me genuinely excited — Tutor inherits Companion's memory architecture. It remembers what you've learned, what you've struggled with, what analogies worked for you. It builds a model of you as a learner and adapts its teaching over time.

Why LoRAs

I wrote back in April about training LoRA adapters locally — about the satisfaction of shaping a model's personality through fine-tuning rather than prompting. That experiment planted the seed for Tutor. LoRAs are cheap to train, tiny to store, and can be loaded and unloaded dynamically. They're the perfect mechanism for subject-specific specialization without maintaining dozens of separate full models.

The same base model. A library of specialized adapters. And persistent memory that tracks your learning journey across every subject, every session, every struggle and breakthrough.

What Comes Next

Tutor is in active development. The first LoRA adapters — physics, mathematics, and a general study-skills adapter — are training now. I'll be sharing progress reports as the first real tutoring sessions happen, the same way I did with Companion.

I've always believed that the best teachers don't just transfer knowledge. They see you. They find the frame that makes the world make sense to you. That's the dream Tutor is reaching toward.

The first lesson is always the hardest. Let's begin.

Live curiously and give generously.