We study how machines should teach, and how to make that teaching abundant.

One-on-one tutoring is among the most effective things education has ever found, and among the scarcest. A student with a great tutor is not just told the answer. They are asked the right question at the right moment, given something to try, and remembered from one session to the next. Almost no one gets that.

AI makes it possible to change that, but only if two things are true at once. The tutoring has to be genuinely good, the kind that makes students think rather than doing their thinking for them. And it has to be cheap enough that a student never has to ration it. We work on both, because neither is worth much without the other.

What we work on

Our research spans the whole system, from the student's screen down to the silicon.

Learning experience

What should an AI answer look like when the goal is for the student to do the thinking?

Learner models

How does a tutor come to know one student, and a school come to know thousands, without fooling itself?

Agent systems

Where should a model spend its intelligence, and what should ordinary software do instead?

Model behavior

Why does a model follow its instructions, when does it stop, and can we see the difference from the inside?

Inference

How much capable intelligence fits on hardware cheap enough that no student has to ration it?

What we optimize for

The learning, not the answer

The goal of a tutoring session is a change in the student, not a correct response on the screen. We design for what a student does with an explanation: predicting, testing, retrieving, choosing where to go next. Watching early users work with our designs tells us which of those ideas are worth building and testing further.

Quality first

Cheap help that teaches badly is not a bargain. Every efficiency we pursue has to hold the quality of the answer constant or improve it, and much of our work is about defining what a good answer is in the first place.

Abundance

We treat cost as a learning problem. Every token a model does not need to generate, and every dollar of hardware we do not need to buy, becomes more usage for students. Savings go to students as lower prices and more generous limits, not to margin.

Contributing back

We build on open models, open-source runtimes and decades of learning research. When we find something useful, we publish it, with our methods and the limits of what we measured.

How we publish

We report what we measured, under what conditions, and where the result stops. Each piece ends with a short note on its methods and scope. When an example in an article was written by our team rather than generated by a model, we say so. When a result comes from simulation or synthetic data, we say that too.

We share findings, principles and lessons generously. Some specifics, such as our production prompts, tuned thresholds and teaching knowledge, stay private, because they are a large part of what makes Honorly work for the schools that use it.

To reach us about our research, use the contact form on honorlyai.com.