Interpretability researcher suspects Claude and ChatGPT CoT show the same problematic behavior
BlancheMinerva · x · 2026-10-17
- Interpretability researcher Blanche Minerva says she isn't permitted to inspect Claude or ChatGPT's chain-of-thought, but would be "shocked" if they didn't exhibit the same behavior — especially versions with less alignment training to suppress it.
- She says she has been assuming this is common across frontier models for the past year or so.
- She adds that if someone with permitted CoT access can't replicate the results and believes it's not due to post-training, that would be very interesting.
More from Safety
- Matthew Green highlights unexplained skipped curves in NIST P-256/P-224 seed generation — matthew_d_green · 2026-10-17
- Meta terms reportedly grant itself rights to scrape the open web for AI training — VeryWellVersed · 2026-10-17
- Jensen Huang on Lex Fridman: build the flywheel, don't just block chip exports — rohanpaul_ai · 2026-10-17
- METR: AI labs redacted key claims before approving frontier risk report — kevinnbass · 2026-10-17
- Macroscopic Ventures: Anthropic backer with stakes in Apollo Research, Gray Swan and Redwood — kevinnbass · 2026-10-17
- Evaluator payment policies compared: METR takes no lab money, Apollo charges market rates — kevinnbass · 2026-10-17