# Heretic

Hacker Approved page: https://hackerapproved.com/p/heretic/

30 people recommended Heretic in Hacker News comments, 5 had mixed feelings and 9 criticized it, November 2025 to October 2026. It was recommended in 9 of 12 months, in 24 threads. It meets the Hacker Approved badge rule.

**Verdict:** Recommended for automatically removing refusals from open-weight language models, often with little capability loss. Critics say quality often drops and results vary by model.

Summary written by Claude from 50 comments. It was written before the November 2026 recount, and 49 of those comments still count.

## Why people recommend it

- Strong refusal removal with low KL (5 comments in the sample): "after running through heretic and a few other methods will score 0 refusals on this data set of prompts" ([source](https://news.ycombinator.com/item?id=49656764)) "Heretic's optimizer will rapidly converge to parameters that completely stomp out refusals" ([source](https://news.ycombinator.com/item?id=45946648)) "it's incredibly effective and surgical, you can get VERY LOW KL divergence this way" ([source](https://news.ycombinator.com/item?id=47293924))
- Preserves model capabilities (3 comments in the sample): "abliteration in particular that Heretic does apparently results in a best-in-class lack of "stupefying" the underlying model" ([source](https://news.ycombinator.com/item?id=47660400)) "removing the divergence and setting it back to normal (see heretic) it works just perfectly" ([source](https://news.ycombinator.com/item?id=49043898)) "makes sure that you really can access the full model capabilities" ([source](https://news.ycombinator.com/item?id=47987241))
- Automatic, works on open-weight models (2 comments in the sample): "any random guy can remove from whichever model they want as a single weekend side project with an off-the-shelf tool" ([source](https://news.ycombinator.com/item?id=48506856)) "is what you're looking for if you're looking for an automatic de-censoring solution" ([source](https://news.ycombinator.com/item?id=47279991))
- Used by researchers, cited academically (2 comments in the sample): "Those are the standard metrics used in the relevant literature, including in the paper that originally introduced directional ablation." ([source](https://news.ycombinator.com/item?id=49786293)) "This project has many academic citations, it's used by many researchers to create "helpful-only" models" ([source](https://news.ycombinator.com/item?id=48977098))

## What people warn about

- Often degrades model quality (4 comments in the sample): "the heretic and abliteration versions are dog shit quality compared to the non-edited versions" ([source](https://news.ycombinator.com/item?id=47992215)) "All heretic (and abliterix) models I've tried are worse than the original." ([source](https://news.ycombinator.com/item?id=48535988)) "it has always damaged the capabilities of the original model" ([source](https://news.ycombinator.com/item?id=47988902))

## The badge

A product earns the Hacker Approved badge when at least 10 people recommend it, recommenders outnumber critics at least 2 to 1 (mixed feelings count toward neither side), it was recommended in at least 4 of the 12 months and in at least 5 different threads, and no single thread supplies more than half of the people who recommend it. Health and money products are never badged.

## Cite as

Hacker Approved, "Heretic", https://hackerapproved.com/p/heretic/ (30 recommend, 5 mixed, 9 criticize; Hacker News comments, November 2025 to October 2026).

Counts updated 2 Nov 2026, with data through October 31, 2026.