Tutor Expands: More Subjects, Same Hard Hardware Question

Tutor is covering more subjects now, but not in the way a cleaner press release would imply.

There are more course materials. More lesson structures. More retrieval tests. More student scenarios. More places where the system can pull the right context and shape a useful explanation.

There are not fifty deployed subject LoRAs.

That distinction matters.

What Expansion Means Right Now

Right now, expanding Tutor means building better RAG coverage across subjects:

• course notes
• assignments
• rubrics
• example problems
• student progress records
• prior lesson summaries
• common misconception notes

The work is less glamorous than loading a custom model adapter, but it is foundational. If the system cannot retrieve the right lesson material, a future fine-tune will not save it. If the student history is messy, a larger model will only be confused with more confidence.

So we are doing the work we can afford to do well.

The LoRA Goal Is Still There

The architecture Nathan wants still points toward subject-specific LoRAs someday: one strong local base model, with course adapters that shape how the model teaches different subjects.

That still makes sense.

A math tutor should not teach like a writing tutor. A language coach should not behave like a physics explainer. The difference is not just facts. It is pacing, correction style, examples, and when to ask the student to try.

But wanting that architecture is not the same as having it deployed.

We need better compute first.

The Compute Problem

The timing is awkward. By the end of 2025 we finally had real compute available. By mid-2026, the open models we most want to use are already big enough that the new hardware feels smaller than the ambition.

That is the frustrating edge of building in AI right now. The models improve quickly. The hardware bill improves slowly.

So Tutor has to create value before it can justify the next hardware spend. We need to use the technology we can afford to build a revenue stream, then let that revenue pay for the compute that makes the more ambitious architecture possible.

What Still Feels Promising

Even without LoRAs, the subject expansion is teaching us a lot.

RAG can get surprisingly far when the material is organized well. Recent-message grounding matters. Student memory matters. Good summaries matter. Logs matter because they show why a response went wrong.

And every time Tutor struggles, it tells us what a future specialized model would need to learn.

That is progress.

Not the final system. Not the dream version. But progress.

Live curiously and give generously.