Re:Show HN: Near-Realtime LLM Steering
by osmoscraft on Monday November 16, 2026 @11:51AM
(#50519717)
This is a conceptual prototype for working with near-realtime LLM models. I saw the recent rise of decision model (Jev) and specialized hardware (Cerebras, Groq, ChatJimmy) and thought it would be really cool if we can put LLM into this realtime feedback loop for shaping ideas. Here the questions I explored:
Focus monitor: can we visualize what LLM is paying attention to, and use that to guide human's focus towards what's important?
Attention lock: can we force LLM to pay more attention to specific phrases so as to bias its future behavior?
Future Vector Steering: can we decompose the task into eigenvectors and independently tune each to control steer the text generation?
Realtime re-generation: when text generation is cheap and fast, can we explore alternative futures by scrubbing?
I used Cerebras GPT OSS 120B for text generation (~200ms latency) and Gemini Decision API for detection and evaluation (~150ms latency).
Focus monitor: can we visualize what LLM is paying attention to, and use that to guide human's focus towards what's important?
Attention lock: can we force LLM to pay more attention to specific phrases so as to bias its future behavior?
Future Vector Steering: can we decompose the task into eigenvectors and independently tune each to control steer the text generation?
Realtime re-generation: when text generation is cheap and fast, can we explore alternative futures by scrubbing?
I used Cerebras GPT OSS 120B for text generation (~200ms latency) and Gemini Decision API for detection and evaluation (~150ms latency).