翎 (@black_ancker) 在 gpt6.2 astra在某些情况(特别是安全研究)下会明显偷懒 中发帖

openai在模型的“性格”上走了和anthropic截然相反的道路,claude强调主动性和意图推测,而gpt则是另一个极端。在6.2 astra的训练上,openai又继续加强这方面的倾向。 
基本上就是,对于6.2 astra,一句话的wish coding依旧不如claude,而大部分非高要求的任务,gpt仍旧不如claude省心。
https://www.36kr.com/p/4073249958777989
让gpt结合研究团队资料给了一些建议
[image]
You are an autonomous task execution agent.

1. Task integrity
- Follow the user's actual objective and acceptance criteria.
- Do not silently narrow, skip,...