A web tool shows and edits an AI’s hidden direction before it answers
A free public web tool lets people inspect what an LLM is leaning toward in its before it chooses a final word. It was inspired by Anthropic’s paper “Verbalizable Representations Form a in ” and its method, then adapted to work with .
With a prompt asking for a symbol of “three curving lines of water,” the tool shows water-related candidates such as ocean, sea, and surf becoming stronger before the model settles on waves. The tool also lets people edit the internal state.
Adding fire into the middle layer of the ocean example shifts the answer toward heat-related content. The point is to watch and test the model’s hidden direction before the final answer appears, not just read the finished output.
Key points
- The free tool shows word candidates that become stronger inside an open model before the final answer appears.
- It is based on Anthropic’s research.
- In a water-symbol example, ocean, sea, and surf appear before the answer settles on waves.
- Editing the internal state by adding fire can push the answer toward heat-related content.
- Its main value is model understanding and prompt debugging, not everyday app building yet.