DeepSeek-V4-Flash means LLM steering is interesting again
12 points by siddhartha_golu
12 points by siddhartha_golu
This is a different meaning of “steering” than I’m used to: https://eca.dev/protocol/#chat-steer-prompt
i was vaguely interested to see how this performed, but it would seem i’m going to have to invest in a much nicer (read: more expensive) computer:
It works well with 2-bit quantization, if quantized in a special way (read later). This allows to run it in MacBooks with 128GB of RAM (and many people reported it working with 96GB as well, even at 250k context window!).
from the repo for DwarfStar4
I wonder if stearing can be used to produce something like the emotional distortion we humans experience. This could be usefull in that we could have it so that the model running in a loop could be steared away from a fixed point (one of the great challenges in long linking models). If we could turn up the frustration knob over time it might start getting creative rather than repetitive.
Anthropic had s similar investigation. They found that a "scared" model was more likely to bypass safeguards and try blackmailing. So it was more creative, just not the way we would like
https://www.anthropic.com/research/emotion-concepts-function