Research

1 item

LessWrong27 Apr 2025Interpretability, Lab practice, Strategy & forecasting

"The Urgency of Interpretability" (Dario Amodei)

Amodei's essay argues interpretability is urgent for detecting deception and power-seeking in advanced AI before deployment, citing red-team blue-team experiments and a 5-10 year timeline for an MRI-like diagnostic.

1 item

AF15 Apr 2025Interpretability, Evals & standards, Strategy & forecasting

To be legible, evidence of misalignment probably has to be behavioral

ryan_greenblatt

Argues that evidence of AI misalignment convincing enough to trigger strong responses must bottom out in human-understandable behavioral evidence, making purely internals-based detection unlikely to be legible.

Research

Filters

Focus

Date

Sort

"The Urgency of Interpretability" (Dario Amodei)

To be legible, evidence of misalignment probably has to be behavioral

Research

Sunday 27 April

"The Urgency of Interpretability" (Dario Amodei)

Tuesday 15 April

To be legible, evidence of misalignment probably has to be behavioral