DeepSeek-V4-Flash means LLM steering is interesting again
Summary
The post dives into LLM steering and how local models like DeepSeek-V4-Flash and DwarfStar 4 enable experimenting with activations to influence outputs. It explains steering concepts, potential approaches (activation-level vectors and feature mapping) and contrasts them with prompting and fine-tuning. It also discusses practicality, limitations, and the open-source community's potential impact on future tooling.