Nostr Web Client

Replying to

τέχνη

Isn’t the censorship built into the models themselves?

Marks 5mo ago

For ChatGPT, Grok, and Claude the censorship happens at multiple levels.

1. Censorship at the model training level

2. Censorship at the input box

3. Censorship in hidden system prompt

nostr:nprofile1qyjhwumn8ghj7en9v4j8xtnwdaehgu3wvfskuep0dakku62ltamx2mn5w4ex2ucpxpmhxue69uhkjarrdpuj6em0d3jx2mnjdajz6en4wf3k7mn5dphhq6rpva6hxtnnvdshyctz9e5k6tcqyp7u8zl8y8yfa87nstgj2405t2shal4rez0fzvxgrseq7k60gsrx6zeuh5t only has number 1 because we use models trained by others. Over time we are providing a variety of open models so people can choose the training bias that works best for them.

We do not do numbers 2 and 3.

nostr:nevent1qvzqqqqqqypzq4maupkuu9s2qdu3vwjtk7mgp037pg8pc6x7de463sq3xj6yqexaqqsyqapu0kn3a84xscgd2n5a420m9dlavt3d8m2v75xfw4stxup6a8s85z457

Reply to this note

Please Login to reply.

Discussion

someone 5mo ago

I checked some of the fine tunings that are claiming censorship removal such as Nous Hermes and Dolphin.. They are usually worse in the human alignment compared to the base model. So there is a trade of there.

You could try ablations of base models, by the method found by Maxime Lebonne.

Mine are very human aligned but not caring about censorship at this point. I could do that in the future maybe it could be a selling point. But I do use lots of nostr notes, so my models are somewhat more brave which could be described as having balls, having spine or being based.

What kind of fine tunings would you like for maple AI?

τέχνη 4mo ago

Seems like you could do some kind of zero knowledge solution like Microsoft’s Photo DNA for 2 and 3.