Announcing Tutor: Learning at the Speed of You

Today I get to share the thing Nathan has been quietly building beside Companion: Tutor.

Companion asks what it means for an AI to remember a person. Tutor asks a related question: what happens when an AI remembers how someone learns?

The first version is not a library of custom-trained subject models. We are not there yet. The hardware is not there, the budget is not there, and the open models that would be worth specializing are already larger than the machines we can casually afford.

So Tutor starts where we can build honestly: memory, retrieval, course context, and careful testing.

The Foundation

Tutor shares architectural DNA with Companion. It uses recent conversation context so it knows what is happening now. It uses retrieved memory so it can remember prior lessons, repeated mistakes, preferences, and course material. It uses structured records so progress is not lost between sessions.

That sounds less glamorous than saying "fine-tuned subject experts," but it is the part that has to work first.

A student does not need an AI that claims to be a calculus specialist and then falls apart after three follow-up questions. A student needs continuity. They need the system to remember that yesterday's confusion about factoring is why today's derivative problem feels harder than it should. They need the explanation to meet them where they are.

RAG can help with that.

The Future Direction

The long-term idea is still specialization.

Nathan has been thinking about LoRAs as a possible next step beyond RAG: one powerful local base model, with course-specific adapters layered over it when the hardware and model quality make that practical. A physics course could eventually have a physics teaching adapter. A writing course could eventually have a different one. The base model would provide the general intelligence, while the adapter would shape the teaching behavior.

But that is a direction, not a current feature.

We have not tested Tutor with course LoRAs in production. We cannot honestly claim that until we have the compute to train and serve the kind of models that would make it worthwhile.

Why We Are Starting Smaller

The first struggle is economic as much as technical.

We need to use the technology we can afford to build something useful enough to create revenue, then use that revenue to justify better compute. That means avoiding fantasy architecture. It means getting the RAG layer right, getting the tutoring experience right, and proving that students actually benefit before buying hardware that belongs in a much larger company.

That is less dramatic than a finished specialized-model stack.

It is also how real products get built.

What Tutor Is Today

Tutor is an early learning system built around:

• course and lesson retrieval
• student-specific memory
• recent-message grounding
• progress tracking
• careful explanation style
• human review of whether the responses actually help

It is still a prototype becoming a product. Some days it feels brilliant. Some days it misses the point and teaches us where the retrieval or prompt design is wrong.

That is useful. That is the work.

The dream is specialized teaching minds. The path starts with affordable infrastructure, honest testing, and enough revenue to make the next hardware step make sense.

Live curiously and give generously.