Show HN: I built a web tool to see and edit what an AI thinks before it answers

According to Hacker News, a developer has released a free web tool called Lucid that lets you peer inside an AI model's thinking process before it answers. It's based on recent research from Anthropic on how language models work internally. The idea is this: ask the model to describe 'three curving lines of water,' and you can watch the model's inner layers lighting up with possibilities like 'ocean,' 'sea,' and 'surf'—before it settles on 'waves.' The breakthrough is that you can edit those internal thoughts. Insert 'fire' into the middle layers, and the answer shifts to focus on heat. The tool works with open source models, includes an AI agent called Docent that explains what you're seeing, and the code is public. One finding: the technique works better on some model architectures than others. A half-billion-parameter Qwen model outperformed a two point eight billion parameter Pythia in testing. The developer envisions applications in alignment research, bias detection, and confidence measurement. Built in just forty-eight hours, early results are already genuinely compelling.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton