Goodfire 联合创始人谈只读 AI 推理文本为何不够
Goodfire 首席科学家上节目聊个有意思的问题:模型写出来的推理和它脑子里想的不一定一致,得看内部才能发现。
Goodfire 联合创始人兼首席科学家 Tom McGrath 在 AI Deep Dive 第 5 期节目中讨论模型可解释性问题。他指出模型的书面推理文本可能遗漏真实内部过程,仅靠阅读推理不足以理解模型。节目探讨了直接查看模型内部机制的方法,可作为额外的安全防线。
Why isn’t reading AI’s reasoning enough?
For Episode 5 of AI Deep Dive, Goodfire co-founder and chief scientist Tom McGrath joins Rocket to explore what written reasoning can miss—and how looking inside the model could add another line of defense.
🚀 Watch today at 12 pm PT / 3 pm ET: https://t.co/kYW16CDQIa