Jacob Hilton: graph completeness measures completeness, not faithfulness
JacobHHilton · x · 2026-10-17
Researcher Jacob Hilton offers a small correction: the graph completeness score in interpretability is more properly a measure of completeness, not faithfulness. Context from his earlier thread: understanding even 0.01% or 0.0001% of a frontier model could still be incredibly useful — but that is a very different proposition from "solving alignment by understanding everything."
Related event: ARC's Jacob Hilton Sparks Debate: Interpretability Far From Solved(22 posts)→
More from Safety
- Epoch AI: open-weight cyber capabilities trail closed frontier by just 6-7 months as GLM-5.3 outscores GPT-5.4 — teortaxesTex · 2026-10-17
- Anthropic researchers: superposition isn't the blocker, interpretability is more tractable than assumed — akbirthko · 2026-10-17
- Security researcher criticizes jailbreak disclosure norms as unfair to closed models — xeophon · 2026-10-17
- Matthew Green highlights unexplained skipped curves in NIST P-256/P-224 seed generation — matthew_d_green · 2026-10-17
- AI tools reportedly helped one attacker hit 9 South Korean banks — Comfortable_Gene5180 · 2026-10-17
- Meta terms reportedly grant itself rights to scrape the open web for AI training — VeryWellVersed · 2026-10-17