Steering a Local LLM's Attention Directly: Sliders That Shift Activations, Not Prompts
round · x · 2026-11-20
Inspired by Sun Chuanqi's "Attention steering" HCI demo (which philippepage calls one of the most brilliant takes on HCI in the LLM era), he wired the interface directly into a local model's internals as an experiment:
- Spaceship normally works via prompting; instead, his steering controls add directly to the model's activations, and selected boxes shift actual attention;
- A sidebar runs an unsteered copy of the model side by side, showing in real time where the two outputs diverge.
The demo shows an interaction paradigm that steers LLM output at the activation and attention level rather than through prompts — and the author argues LLM-era HCI still has many unexplored patterns.
More from coding & agent
- Ex-Googler open-sources a CLI that lets Claude Code and Codex chat in one room — thechrisperry · 2026-11-20
- Vibe-coding day 17: a WebGPU kelp lantern built with Gemini in Antigravity — AI_Andrew · 2026-11-20
- Beaker autonomously finds 100+ grading errors across five popular AI benchmarks — VictorSanh · 2026-11-20
- Burkov spots Codex quirk: suggested follow-up knows the agent screwed up, but the agent doesn't — burkov · 2026-11-20
- QuixiInfer author: tell Claude 'defeat is not an option' to make it write kernels for your GPU — QuixiAI · 2026-11-20
- OpenAI's Codex repo has 27,802 open issues: agent harness maintenance compared — SuaveSteve · 2026-11-20