Learning
Universal Tools for Learning
A tutoring system that works only in three languages and on fast connections is not an education technology. It is an amplifier for existing advantage.
Noor Haddad · Amara Osei

Every wave of information technology has arrived with the same promise — knowledge for everyone — and kept it selectively. Books required literacy; broadcast required infrastructure; the web required connectivity and, in practice, English. Each wave widened access and widened gaps at the same time, because the tools calibrated themselves to the users they already had.
Language models are the first learning technology that can, in principle, meet a learner where they are: in their language, at their level, on their schedule, in dialogue. Whether they do so in practice is an engineering and institutional choice, and the early defaults are not encouraging. Evaluation suites are overwhelmingly English. Interfaces assume fluent typing, stable connectivity, and prior schooling. The systems are universal in architecture and parochial in deployment.
Three design commitments
- The learner's language is the interface, not a translation layer. Quality must be evaluated natively per language, with local educators defining what good explanation looks like — explanatory conventions differ across cultures, and a model that explains in translated English idiom teaches less than it could.
- Degrade gracefully, not punitively. The same tutor should function in a text-only, intermittent, low-bandwidth mode without becoming a worse teacher — shorter exchanges, more durable artifacts the learner keeps offline.
- Teach toward independence. The tutor's success metric is the learner's performance without the tutor. This single choice — the same returned-capacity standard we apply everywhere — rules out engagement-maximizing designs from the start.
What we are building
We are developing an open evaluation suite for explanatory quality in eleven languages, built with teachers in each language community rather than translated from a master set; and a reference tutor architecture designed for the constraints above, which we intend to publish in full. The point of publishing is the point of the project: a learning tool is universal only if people we have never met can run it, audit it, and adapt it to learners we have never imagined.
The stakes are easy to state. Done carelessly, AI tutoring becomes one more amplifier for existing advantage. Done well, it is the closest this generation will come to the old promise — that the time and means to learn stop being an inheritance and start being a baseline.
More research
All research →Machines That Return Time
AI is evaluated by whether it completes tasks. We argue for a complementary standard: whether it expands the time and capacity people have for judgment, learning, care, and shared work.
Amara Osei, Daniel Reyes, June Park
Interfaces for Collaborative Intelligence
Most AI interfaces assume one person and one model. Most meaningful work happens in groups. Design patterns from a year of building shared-context systems for small teams.
June Park, Tomás Carvalho
Measuring Human Capacity, Not Just Model Capability
A benchmark tells you what the model can do. It tells you nothing about what its users can still do. First results from paired capability–capacity evaluations.
Daniel Reyes, Priya Raghavan