Steering, a technique to guide LLM outputs by manipulating model activations, is gaining traction with the release of DwarfStar 4, a local model running DeepSeek-V4-Flash. While promising, steering faces challenges like limited practical applications compared to prompting and the complexity of extracting meaningful steering vectors for complex concepts like “intelligence.” Despite scepticism, the open-source community’s growing interest in steering could lead to practical applications within the next six months.