AI should make you think.
Historia began as a research tool that cannot invent a source. Next is a place to learn, starting with maths, where the AI never does the work for you. This page says what is live today, what is being built, and what is still a plan.
The research tool is live. It searches public archives, libraries and scholarly indexes, not the open web. Titles, dates, links and quotations are copied from the archive’s own record, so none of them can be invented. A request to write an essay is refused on the server before any model runs. Start researching
Mastery practice for a student’s own maths class. It is being built and tested now. It is not available yet, and there is no release date.
Try the demo: A preview of the idea, not the finished product. It opens in a new tab.
Teachers can already send a whole class into the archives with one link. Next: a view of who is stuck, what they tried, and what they can do without help. The teacher decides the grade.
Nearly 1,000 high-school students in Turkey practised maths with plain GPT-4. Their practice scores rose by 48%. On the exam taken without AI, they scored 17% lower than students who had no AI.
In the same trial, a tutor that only gave hints removed the harm. It added no gain.
Khan Academy with its AI tutor, Khanmigo, over two years in 18 Tennessee schools: about 0.06 to 0.08 SD a year, the same as Khan Academy without the AI. Students messaged the tutor in 17% of the sessions where they made a mistake.
Help is limited in code. No hint before you try. One hint at a time. A worked solution only after an attempt, and that problem then earns no credit. The server enforces it; it is not a request to a model.
Working checked by code. Answers, and each line of working where a checker exists, are checked by code, not judged by a model. In one major review, software that checks each step came close to a human tutor: 0.76 SD against 0.79.
Proof without AI. Each cycle ends with a check taken with no help at all. Only that counts as knowing it, and it comes back days and weeks later.
You do the thinking. The AI points, asks and checks. It does not solve your homework or write your paper.
Our target is a gain of 0.2 to 0.4 SD on tests taken weeks later without AI, the range of good human tutoring. It is a target, not a result. No part of Historia has been tested for learning gains yet.
Teach maths, run a school, or think we have something wrong? thehistoria.company@gmail.com