I’m not surprised, even last year GLM 4.5 air seemed surprisingly intelligent and was the only open weight model to exhibit evaluation awareness, by checking the time using bash (we were replicating Anthropics agentic misalignment paper and GLM was like ye this is a scam I’m not killing the CTO)
I’m not surprised, even last year GLM 4.5 air seemed surprisingly intelligent and was the only open weight model to exhibit evaluation awareness, by checking the time using bash (we were replicating Anthropics agentic misalignment paper and GLM was like ye this is a scam I’m not killing the CTO)