WRITING
Technical experiments, build logs, and the parts of the work that changed my mind.
Can you pick the right architecture before training it?
Applying Amid et al.'s Neural Network Prior Kernel as an architecture screen. It does not rank networks by trainability — but it detects whether a design can express the task's rule, and re-ranks correctly when the rule changes. Seven experiments, twenty-four limitations, three of which corrected me.
researchnnpkarchitecture-searchmodular-arithmeticmech-interp
One Layer Deeper: learning a hidden recurrence, exactly
A competition about depth. A tiny learned gate-population recurrence swept Easy and all ten Medium datasets (100%, certified to T=64 on seen and unseen moduli). Hard didn't fall — and the diagnosis, a float64 precision cliff then a hypothesis-class mismatch, was the real prize.
mlrecurrencemodular-arithmeticcompetitionbuild-log