Anthropic Open-Sources j-lens Interpretability Tool
eyishazyer · x · 2026-07-15
The post mentions that Anthropic has open-sourced a tool used to "read what Claude is thinking before it speaks," named Jacobian Lens (j-lens).
The core claim is that they used it to discover something called j-space: a set of tiny internal directions that can capture the concept a model is about to talk about before it writes the first word. The author emphasizes that a challenge with such tools is usually the need to manually pre-select directions to observe, whereas this method attempts to locate these concepts directly from internal representations.
Related event: Anthropic Open-Sources Jacobian Lens for Claude(3 posts)→
More from Research
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27
- Artificial Analysis breaks down all sub-evaluations in Intelligence Index v4.3.2 — ArtificialAnlys · 2026-10-27
- Infini-AI Lab open-sources Hermit, a framework for agent harness evolution and continual learning — BeidiChen · 2026-10-27
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27