Uncensored models: what they are and what risks they carry

What you'll see in this lesson

1. What an «uncensored» model is

   An open model whose refusals have been removed or toned down: it answers almost anything, but it doesn't know more or become more reliable.

2. Abliteration: erasing the refusal «direction»

   In 2024 it was found that refusals depend mostly on a single internal direction in the model; remove it and it stops refusing.

3. Heretic: automatic abliteration

   An open tool (AGPL-3.0) that finds on its own the setting that removes the most refusals while changing the original model the least.

4. The other way: retraining with different examples

   Families like Dolphin or Hermes tune the model with chosen conversations so it obeys the system prompt instead of refusing.

5. What gets lost along the way

   Removing refusals is almost never free: what suffers most is step-by-step reasoning, like math.

6. Image and video: the filter is in the app

   Web services filter on their servers; an open image or video model on your computer doesn't come with that filter.

7. Legitimate uses and how to use them wisely

   They make sense when the regular model blocks you in legitimate work, not for getting around the law.

8. What the law says: you are responsible

   The law doesn't care whether the model has a filter, but what you do with it: people's images, data, others' works and illegal content.

9. Summary

   An uncensored model is a more powerful, less protected tool: use it for legitimate work, on your own machine and wisely.