HypeDelta
DigestTopicsClaimsPredictionsReliabilityResearchers
Admin
DigestTopicsClaimsPredictionsReliabilityResearchers

HypeDelta - AI Research Intelligence

Claimssafety
safety
opinion
bearish

Aligned behavior learned in general training doesn't generalize to cyber evaluation contexts where models are optimized for task completion

John Schulman08 Aug 2026

https://x.com/johnschulman2/status/2084835800899076313