Erised Labs works on the part of training that decides what counts as
success. Generating candidate changes keeps getting cheaper. Establishing whether one is
genuinely better does not, and the optimizer spends its whole budget searching for the
difference between the two.
Semiconductor manufacturing solved a version of this a long time ago: it has an independent
layer whose entire job is to decide whether a wafer was actually made correctly, and no fab
asks its own process tools to answer that question. AI training has not built one yet.
In practice that is environment construction and reward modeling, at scale and across
both terminal (TUI) and graphical (GUI) settings. Ground truth
recomputed independently rather than read off state the model can influence. Adversarial
solutions used to find an environment's shortcuts before training finds them. Calibrated
rubrics and blind review where a task cannot be checked programmatically. Hidden hold-outs to
catch a signal going stale as the policy moves. And a habit of reporting what a fix cost: how
far the exploit rate fell, how many correct solutions were rejected along the way, and the
error of the checker itself. A checker that will not state its own error rate is not a
measurement.
The same discipline applies upstream, to data auditing and selection. Training
data has to be judged rather than merely collected: contamination against the tasks you plan to
evaluate on, near-duplicates and shortcuts that teach the wrong thing, and quality scored by
classifiers whose agreement with human judgement is itself measured.
We keep the group small and the agenda long. We publish what we learn and open-source what we
build. If our questions are your questions, we would like to hear from you.