कर्मसु कौशल्यम् · “skill in actions”
Uday Phalak
Hey — I am Uday Phalak. This space documents my applied research at the intersection of advanced AI, human-centric design, and my primary focus: AI safety. Building on early work in adversarial ML defense (2017–2019), the goal is the same: as the models get more capable, they should stay interpretable and actually safe.
Alignment · interpretability · neurosymbolic models · generative UX
Notes
-
ARC White-Box Estimation Challenge 2026 - Phase 1
The goal was last-layer ReLU means of a 256×32 random MLP, under a FLOP cap. I started with the closed form, then let each result decide the next experiment.