Ameya Panchal

This is a linkpost for the write-up on my site; the full body is below, and the code, decisions ledger and devlog are in the repo. Base-model z-score calibration of the J-lens helps elicit hidden secret words from Cywiński et al.'s taboo organisms: 0.805 leave-one-out accuracy against 0.665 for their...

Confident in the measurements — copying score, OV logits, attention patterns. More exploratory on the interpretations of what the head "knows," two of which I ended up disconfirming myself. My first mechanistic interpretability post; corrections welcome. Full analysis and code: github/Ameya-bit/QuotesByNiche I trained a small character level transformer on Nietzsche...