Philipp Emanuel Weidmann is a mathematician and software engineer doing independent research on language-model alignment and interpretability. He created Heretic, software that automatically adjusts refusal behavior in supported language models.[1][3]
Start with Heretic →
Links
GitHub Personal blog Background
Weidmann describes a background in mathematics and software engineering, alongside an interest in philosophy. His personal blog has covered experimental mathematics, information security and independent software projects. He has also designed language-model sampling algorithms, including DRY and XTC, which concern how a model chooses its next output token.[1][2]
Contributions
Heretic applies directional ablation, an intervention on internal model representations associated with refusal, and searches for parameters automatically. Its optimization balances reducing refusals against divergence from the original model’s behavior. This makes a manual model-modification workflow more systematic. The project evaluates changes on selected prompt sets; such measurements do not establish that a modified model preserves every capability or will behave reliably in every application.[3]
Tools, writing & talks
Tools
Start here
Hereticgithub.comAutomates directional-ablation parameter search to change refusal behavior while measuring divergence from the original model.
Read the README’s explanation of the optimization objectives and comparison metrics first. Understanding the tradeoff is more useful than starting with a modified model; running experiments requires the relevant model files and supported hardware.
Creator[1][3]Original source