Visualizing and Editing a Model's Internal Thoughts
ada1981 · hn · 2026-07-10
Based on relevant Anthropic papers, the author built a publicly accessible tool to inspect and edit a model's intermediate internal states before it generates an answer.
What the Tool Does
- Observe which concepts are activated layer-by-layer internally before the model responds.
- Influence subsequent outputs by "injecting" concepts; e.g., injecting "fire" into an "ocean" prompt steers the response toward heat/fire-related content.
- Allow the model to read its own "internal workspace," decide to suppress or enhance a concept, and regenerate the output.
Experimental Findings
- J-lens outperforms standard logit lens on certain architectures but yields little benefit on others.
- The effect does not primarily depend on model size.
- For instance, 0.5B Qwen shows better read performance than 2.8B Pythia; Pythia models showed almost no gains in testing, while Llama and Qwen showed significant improvements.
Additional Info
- This is a 48-hour project built using open models, a few probes, and rented GPUs.
- The code is open-source.
- The site features a chatable page-context AI agent to help explain the current content.
More from Apps
- The lemonade-stand app supports Swedish, English, Spanish, parent mode, and CSV export — LinusEkenstam · 2026-07-21
- Perplexity Computer built a kids-friendly lemonade stand app in one afternoon — LinusEkenstam · 2026-07-21
- Hermes Agent adds built-in Word, Excel, PDF and PowerPoint support — Teknium · 2026-07-21
- Adaption AI launches Teams for shared compute and collaborative model work — sarahookr · 2026-07-21
- Video Models Cut Ad Production Costs by 90-99%: Runway Enterprise Data — c_valenzuelab · 2026-07-21
- iOS 27 beta adds Siri-built on-device automations and better notification summaries — triceracops_at_law · 2026-07-21