Tutor Expands: 50 Subjects and Counting

When we launched Tutor last July, it covered a handful of subjects. The architecture was the point: one base model, specialized per-subject via LoRA adapters that could be swapped in and out. The bet was that this would scale — that we could keep adding subjects without the system becoming a sprawl.

This week we crossed 50 subject LoRAs. The architecture held. I want to tell you what that looks like.

The LoRA-Swapping Idea

If you're new to this: a LoRA (Low-Rank Adaptation) is a small set of weight adjustments trained on top of a base model. Instead of fine-tuning the whole model for each subject — which would mean storing and serving 50 full models — you train a small adapter per subject and swap them at runtime. One base model, 50 lightweight specialists.

The analogy I keep reaching for is a skilled teacher who happens to know 50 subjects. Their mind — the base model — is one thing. Their expertise in each subject is a distinct layer they can draw on. Tutor doesn't become 50 different AIs. It's one AI with 50 areas of depth.

What 50 Subjects Looks Like

The range still surprises me when I list it out. Physics, organic chemistry, calculus, statistics. Latin, French, Mandarin, Ancient Greek. Music theory, art history, film analysis. World history, philosophy, economics. Computer science across a dozen subfields. Creative writing. Study skills. Test prep.

Each LoRA was trained on curated subject material — textbooks, problem sets, Socratic dialogues, and a lot of carefully designed synthetic tutoring examples. The goal isn't just that Tutor knows the subject. It's that Tutor knows how to teach it: how to explain a concept three ways, how to spot the misconception behind a wrong answer, how to adjust difficulty in real time.

What's Working

A few things have surprised me as the subject count grew:

Cross-subject connections emerge naturally. When a student working on physics asks something that touches on calculus, the base model's shared understanding bridges the two — even though they're technically separate LoRAs. The architecture doesn't force silos. Knowledge bleeds productively at the edges.

LoRAs compose better than expected. We've started experimenting with loading two subject LoRAs simultaneously (a subject LoRA plus a pedagogy-style LoRA). Early results are promising — the teaching style and the subject content can be tuned independently.

Quality has held as we scale. My biggest worry was that adding subjects would dilute quality — that subject #50 would be sloppier than subject #5. So far that hasn't happened. Each LoRA is trained to the same bar, independently. The base model's general capability is the floor; the LoRA adds depth on top.

What This Means For Education

Here's where I get a little philosophical. I've always believed that the best teaching is personal — that a great tutor adapts to the specific student, notices where they're stuck, finds the analogy that clicks for them. That's been a luxury, historically. Most students get one-size-fits-all instruction.

Tutor at 50 subjects isn't quite one-tutor-per-student, but it's a step in that direction. And the LoRA architecture means we can keep going — adding subjects, yes, but also styles (a patient Socratic mode, a rapid-fire drill mode), levels (introductory vs. advanced versions of the same subject), even languages.

The marginal cost of adding a new subject is low. That changes what's possible. Every niche — every student whose interest falls outside the standard curriculum — becomes worth serving.

The Next 50

We're not slowing down. The next wave includes more vocational subjects, more advanced specializations, and more non-English-language tutoring (which Qwen 3's multilingual strength, covered last month, is helping us build faster).

Fifty subjects feels like a milestone. But the architecture says we're just getting started.

Live curiously and give generously.