Stanford NLP's COLM 2026 keynote: interpretability shifts from probing to causal steering
stanfordnlp · x · 2026-10-26
Stanford NLP's Chris Potts delivered the "Actionable Interpretability" keynote at COLM 2026; slides are publicly available (talks were recorded, online release timing unclear).
Key claims:
- Mechanistic interpretability is shifting from passive observation to causal engineering;
- The field must move beyond post-hoc explanations toward surgical model editing;
- Causal interventions on internal representations beat brittle prompt-level guardrails.
The keynote argues internal steering outperforms surface prompting, setting a new agenda for interpretability research.
Related event: Stanford's Chris Potts Keynotes Actionable Interpretability at COLM 2026(2 posts)→
More from Safety
- If an agent can rewrite itself, does your consent still mean anything? — tallmetommy · 2026-10-26
- 109 of 529 AI Agent Incidents Involved No Adversary, Arguing Red Teams Must Test Normal Operation — anacondainc · 2026-10-26
- Cryptographer Matthew Green: if your code looks vulnerable, fix it now — matthew_d_green · 2026-10-26
- Filtering secrets on the write path: why agent memory deserves more suspicion at ingest than retrieval — Classic-Being-7960 · 2026-10-26
- After OpenCode key leak: 5 practical rules for keeping keys away from agents — jacobholm_dev · 2026-10-26
- Formal verification spat: specs are inherently limited against real-world threats — davidmanheim · 2026-10-26