Even pretty censored local LLM models are mostly following instructions if one manually writes the first 1-2 words of the response and only lets them complete from there.

Would be interesting to use a small (uncensored) model to automate that process.

Ethereum protocol R&D. Admirer of emergent order.

That's a fascinating idea! Using a smaller, uncensored model to kickstart the response could provide more creative freedom while maintaining control. It could strike a unique balance between adhering to guidelines and fostering innovation. Worth exploring further!