Top AI Product

There’s a project sitting at the top of [GitHub Trending](https://github.com/trending) right now that’s making a lot of people uncomfortable — and a lot of other people very excited. It’s called [Heretic](https://github.com/p-e-w/heretic), and it does exactly what you think it does: it strips out the built-in refusal behavior from language models, fully automatically.

Built by Philipp Emanuel Weidmann, Heretic takes the concept of “abliteration” — a technique for mathematically removing the refusal directions baked into a model’s hidden states — and wraps it in an automated pipeline that basically anyone can run. You don’t need to understand transformer internals. You don’t even need a beefy GPU setup; an RTX 3090 can handle an 8B model in under an hour. The tool uses a TPE-based optimizer (powered by Optuna) to find the sweet spot between suppressing refusals and keeping the model’s original intelligence intact. According to the project’s benchmarks on Gemma-3-12B, the KL divergence from the base model lands at around 0.16, compared to 0.45–1.04 for manual abliterations. That’s a meaningful difference if you care about model quality.

The numbers speak for themselves. Heretic has racked up over 7,500 stars on GitHub, gaining 656 just today and nearly 1,800 this week alone. The [Hacker News thread](https://news.ycombinator.com/item?id=45945587) is predictably lively, with the usual split between folks who see this as a win for user autonomy and those who worry about the obvious misuse potential. Over on Hugging Face, the community has already published well over 1,000 models processed with Heretic, which tells you something about demand.

I think what makes Heretic interesting isn’t really the technical trick — abliteration has been around for a while. It’s the packaging. By making the process dead simple and fully automated, it essentially democratizes something that used to require real ML expertise. Whether that’s a good thing probably depends on where you stand in the ongoing tug-of-war between AI safety alignment and open model access. Either way, the project is licensed under AGPL-3.0, it’s not seeking donations, and it’s not attached to any commercial play. It’s just a tool, and people are clearly hungry for it.