Research
An open lab does not win by spending more. It wins by choosing the problems money alone does not solve: reproducibility, access, accountability, and the applications nobody is paid to work on.
Frontier-scale training is out of reach for us, and that is not the interesting constraint. What is missing is work that others can actually run: models, datasets and training recipes that reproduce on hardware a graduate student or a hobbyist already has.
Questions we care about: how small can a model be and still be useful for a given task; which training recipes survive contact with a different cluster; what a genuinely reusable dataset release looks like.
Plenty of applied AI never gets done because there is no business case for it. Health, education, climate and public services are full of problems that matter to many people and pay no one. We work on them with the communities who would use the result, not at them.
Questions we care about: what makes a deployed model trustworthy to the people it affects; how to evaluate a system where the metric that matters is not accuracy; how to hand a working tool to an organisation that has no ML team.
As systems get more capable, the gap grows between what they can do and what the public can check. Independent evaluation is one of the few jobs that is better done outside the organisations being evaluated.
Questions we care about: evaluations that resist being gamed; transparency that is useful to a regulator or a journalist rather than only to a researcher; what accountability means for a system no single person understands.
Most robotics results are reported in simulation or on hardware few labs own. We are interested in agents that work in the open world, and in the training methods that make that affordable.
Questions we care about: sample-efficient learning on real hardware; what transfers from simulation and what quietly does not; perception and control that degrade gracefully instead of failing silently.
How we choose
A specific answer, naming real people or communities. “The research community” is not an answer.
If yes, and they are already doing it, our time is better spent elsewhere. We look for the work that gets skipped.
If the result cannot be published and reproduced, it is not a Forall Labs project. The exceptions are rare and argued in the open.
Output
Papers, code and datasets are listed here and on GitHub as they are released, with every contributor named.