421M BERT 复刻 Jev 思路跑在消费级显卡
一位调试 LLM 训练的人推荐这段 Jev 视频：有人用 421M 的 BERT 在消费级 GPU 上复刻了 Type-Safe AI 的 Jev 思路，说明万亿参数聊天模型对真实工作流任务是过度配置。
As someone who debugs LLM training runs, this JEV video is the closest thing to a "why a random Redditor just cloned a Silicon Valley AI startup in an open-source 421M model" explainer I've ever seen released for free. Everyone thinks OpenAI and Anthropic own AI automation. They don't. Type-Safe AI built JEV, and a Redditor cloned the whole idea in a 421M BERT running on a consumer GPU - proof that trillion-parameter chat models are overkill for real workflow tasks. Bookmark this and watch when you have a free evening. Same probabilistic sampling framework LLMs abandoned in 2022, now the hidden path to automation that OpenAI and Anthropic keep missing.
这条是引用着另一条写的：x.com/0xeronn/status/2100557880168472736