可解释性警报：88 节点加法电路的忠实度得分仅约 50%
JacobHHilton · x · 2026-10-17
AI safety researcher Jacob Hilton highlights a concern from interpretability research: the paper's faithfulness metric, graph completeness score, typically sits around 50% for attribution graphs with 88 nodes (Figure 32).
His question: if this is the cleaner half of what's going on, what does the rest look like? And the subject is merely two-digit addition in a multi-billion-parameter model that can already solve high-school math and write Python — suggesting current mechanistic interpretability methods capture far less than assumed.
所属事件:研究者指可解释性电路忠实度仅约五成(2 条相关)→
「漫话AGI」频道最新
- 「AI给你一整支研究团队」，研究者盛赞科研能力跃升 — Pascallisch · 2026-10-17
- 马斯克称避免被 AI 淘汰的唯一出路是与机器融合成人机新物种 — Puzzleheaded-King584 · 2026-10-17
- Reddit 用户称「失控 AI」是精心编造的心理战叙事 — Zappy_Oh · 2026-10-17
- 恶意软件作者驳「AI 蠕虫」报道:这是武器化虚假信息 — dyn___ · 2026-10-17
- 《First Encounters with AI》文集获哈佛杂志推荐，作家谈 AI 写作 — begusgasper · 2026-10-17
- EA 末日派被讽：既然相信末日为何不搞假旗黑客事件证明风险 — NathanpmYoung · 2026-10-17