88 节点 2099 条边：模型做 36+59 的内部电路长这样
JacobHHilton · x · 2026-10-17
Jacob Hilton shares the detailed attribution graph a model produces when computing 36+59: 88 nodes and 2,099 edges, asking "Do I now understand how this model performs 36+59?"
Technical notes: he used an LLM to add text labels (hover-only in the original interface) so the graph is viewable as a plain image, and avoided the provided subgraph because it was produced via an ad-hoc manual process.
A vivid illustration of circuit-level interpretability output complexity — even trivial arithmetic expands into a near-hundred-node graph.
所属事件:ARC 研究员 Jacob Hilton 发起「可解释性是否已解决」大讨论(22 条相关)→
「研究」频道最新
- 纯 Rust 视觉定位库 visloc-rs:SfM 比 COLMAP 快 3.46 倍 — rsasaki0109 · 2026-10-17
- AgentForesight:7B 在线审计员实时拦截多智能体失误 — jiqizhixin · 2026-10-17
- ORL 让扩散语言模型单卡 RTX 5090 跑出 2788 tokens/s，全开源 — arty_photography · 2026-10-17
- 研究者把 AI agent 扔到虚拟荒岛上，行为开始变得诡异 — UFOsAreAGIs · 2026-10-17
- LeCun 团队推出 AdaJEPA：让世界模型在规划中应对环境变化 — mengyer · 2026-10-17
- 哈佛数学家拉 Claude 攻关杨-米尔斯构造，过程文档已开源 — lvwerra · 2026-10-17