gpt-6POST
K3-trained SWE-2 had almost no safety alignment, user recalls, hoping K3.1 keeps it
The author recalls Cognition's K3-based SWE-2 having almost no safety alignment and hopes a K3.1 fine-tune that rivals GPT-6 Astra keeps that trait.
雨夹雪❄️
@mizorewwwArchived 4 days agoOriginally posted 2026.10.24
雨夹雪❄️ @mizorewwwOriginally posted 2026.10.24
昨天Kimi K3.1发布了，这让我想起来一件事情 之前 Cognition 基于 K3 后训练的 SWE-2道德感极低,似乎完全没做安全对齐 这模型基本上什么活都肯接,甚至直接让它利用漏洞去黑网站都行 希望他们会基于k3.1这个能打gpt 6 astra的模型后训练,并且保留低道德得特点🤣
x.com/mizorewww/status/2113831023691874475Archived 2026.10.25 10:05