Research
An agenda built around consequences.
Our agenda starts from decisions people need to make, not from what current models happen to do well. Each area below is a place where better AI research could change an outcome that matters to people who will never use the model directly.
We are interested in work that can be evaluated in the setting where it is meant to be used. That usually means smaller projects, closer partnerships, and a longer view of what counts as success.
Open problems
Questions we want answered
01
Models that know when they are unsure
Scientific use requires calibrated uncertainty, not confident prose. The work we care about reports what a model does not know in terms a researcher can act on, and records which data and model version produced a result.
02
Evaluation for messy conditions
Most benchmarks assume clean data, stable power, and abundant staff. We need evaluation built for clinics, classrooms, and field offices where those assumptions fail, including the cases where current methods break.
03
Climate information at the scale of decisions
A regional projection is not a local plan. Research is needed to connect climate models to the choices a city, farm, or water authority actually makes, with uncertainty stated in useful terms.
04
Care with scarce resources
Decision support for settings with few clinicians and intermittent connectivity, designed around the constraints of the setting rather than around a demonstration.
05
Teaching in a crowded room
Tools that help a teacher with a full classroom, shared devices, and limited bandwidth, rather than tools that assume one learner and one laptop.
06
Accessibility by design
Models and interfaces built with disabled users from the beginning. Retrofitting accessibility is both more expensive and less useful than designing for it.
Methods
How the research is run
01
Start from a decision
Every project names the decision it is meant to improve and the evidence that would count as progress.
02
Fund small teams for long enough to learn
Grants are sized to complete a defined piece of work, with room to report what did not work.
03
Review with people outside the foundation
Proposals are reviewed by researchers and practitioners who know the setting, against criteria we publish.
04
Publish methods, code, and results
Outputs are released under open licenses unless privacy or safety requires otherwise, and the reason is stated.
Outputs
What gets published
Papers and preprints
Peer-reviewed work where the field allows it, preprints where speed matters. Each paper states its limitations plainly.
Open code and data
Released with enough documentation for another group to reproduce the result, subject to privacy and safety constraints.
Evaluation suites
Tests and datasets that other researchers can run on their own systems, including the conditions where current methods fail.
Field reports
Short, honest accounts of what happened when an approach met a clinic, a classroom, or a community.
Publications
No publications are listed yet. Results will be added here as they are completed and reviewed.