Spokes.wiki Search About
Created
Updated

Philipp Emanuel Weidmann

Independent developer, GitHub p-e-w, site worldwidemann.com, location given as "Anywhere the Internet is." Author of heretic, the tool that made abliteration a one-command operation. Profile read 2026-08-22: 37 public repos, 1,185 followers, account opened 2012-11-01.

Not an ML researcher by track record. His repos before Heretic are desktop and terminal software across four languages — maybe (6.3k, Python: show what a program would do to your filesystem before letting it), finalterm (3.8k, Vala, unmaintained), argos (1.8k, GNOME Shell extensions), ternimal (1.1k, Rust), plotinus (1.0k, Vala, a command palette for GTK+ apps). Heretic, at 28.0k stars, is larger than all of them together and is his first entry in this spoke's subject.

That matters for reading the work. Heretic's contribution is not a new result about model internals — the result is Arditi et al. 2024 — it is engineering: wrap a known edit in tunable parameters, point an off-the-shelf optimizer at it, ship it as a pip package that anyone can run. A developer with a decade of "make this usable from a command line" projects behind him is exactly the kind of person who would do that, and the reason the technique crossed from research to commodity is closer to packaging than to interpretability.

He licenses Heretic AGPL-3.0 and asks for a BibTeX citation, which is an unusual pair: copyleft strong enough to reach network use, plus academic-style attribution, on a tool whose stated purpose is removing safety alignment.

heretic · abliteration · arditi-refusal-direction