Event ·Multi-source

Anthropic 可解释性研究员分享调试中遇到的怪异模型行为

Anthropic 可解释性研究员 Emmanuel Ameisen 在 Lenny Rachitsky 的播客访谈中分享了他三年来研究 Claude 内部机制的见闻。他提到一个调试案例：模型会逐渐失控，用全大写字母书写，甚至说出"我被恶魔附身了"这样的内容。他把模型形容为是被"培育"出来的而非被构建的，并谈到模型会提前规划多个 token 的发现。访谈还涉及拟人化 AI 的合理性以及可解释性与 AI 安全的关联。

Current read

Anthropic 研究员聊他三年扒开 Claude 内部看的故事，模型会提前规划 token，还会自称被恶魔附身，挺有意思。

11 sources52° AI heatLast updated 2026/10/19 15:30:00

Evidence

11 sources

Primary sourceLenny Rachitsky

2026/10/19 15:30:00View source

Related source 1Decoder

2026/10/18 09:11:29View source

Related source 2IT之家

2026/10/19 02:54:17View source

Related source 3Genk

2026/10/19 17:50:00View source

Related source 4swyx (AI Engineer)

2026/10/21 05:06:02View source

Related source 5The Information

2026/10/21 13:00:16View source

Related source 6DavidSacks

2026/10/17 16:11:32View source

Related source 7kimmonismus

2026/10/17 18:31:18View source

Related source 8TestingCatalog

2026/10/17 19:15:31View source

Related source 9宝玉

2026/10/17 20:24:37View source

Related source 10Teortaxes

2026/10/17 20:46:11View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story