DeepSeek-V4-Flash Makes LLM Steering Vectors Practical for Individual Developers — 249 Points on HN

Sean Goedecke / Hacker News·medium signal

Sean Goedecke's blog post argues that DeepSeek-V4-Flash (284B MoE, few experts active) has made activation steering practical for individual engineers because it runs locally on consumer hardware. Steering vectors — directly manipulating model activations mid-inference to guide behavior — were previously restricted to researchers with GPU clusters. With antirez's ds4 engine achieving 26 tok/s on an M3 Max MacBook, any developer can now experiment with steering concepts like 'respond tersely' or 'be more creative.'