Interpretability red flag: graph completeness only ~50% on an 88-node addition circuit
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.
More from AGI Musings
- Grady Booch: frontier models are stochastic software systems, not conscious — Grady_Booch · 2026-10-17
- Will AI destroy mathematics? A prominent mathematician weighs in — bgreene · 2026-10-17
- Roger Penrose: 'The I in AI is a misnomer—intelligence requires consciousness, and AI is not remotely conscious' — JosephJacks_ · 2026-10-17
- Peter Diamandis: AI danger debate misses the point—learn to use it faster than we fear it — PeterDiamandis · 2026-10-17
- Satirical 'News from 2026': Alignment solved, researchers smoke to prove AI faith — Kyrannio · 2026-10-17
- If superhuman reasoners can clean up Lean proofs, why not let them audit weights for alignment? — akbirthko · 2026-10-17