On GEMS: training and evaluation as connected work

GEMS / Training Grounds is FDS’s flagship research initiative, and the idea underneath it is simple to state and hard to practice: training and evaluation are the same job.

A training run that produces no evidence about whether it worked is only half-finished. So the research direction pairs every training loop with the evaluation that tells you what actually happened — and treats that evidence, not the fact that a run completed, as the thing worth keeping.

Compute that adapts

The second thread is compute awareness. Available hardware differs from environment to environment, and the work of planning an execution — what runs where, and how — is part of the research rather than an afterthought. The public framing stays deliberately high-level here: the goal is a system that plans against the compute it actually has.

What this note is not

This is a note on research direction. It is not a product announcement, a benchmark claim, or a description of a finished autonomous pipeline. The point is to document the current hypothesis without presenting it as a completed result.