Yann LeCun：仅靠扩展LLM规模无法实现AGI
Yann LeCun's (@ylecun ) latest talk at ETH Zürich
Yann LeCun对AGI路径的底层逻辑批判，直接挑战当前Scaling Law的主流叙事，值得所有关注模型本体的从业者深入思考其论点的有效性。
Yann LeCun's (@ylecun ) latest talk at ETH Zürich
Yann LeCun (@ylecun) 在苏黎世联邦理工学院的最新演讲
Scaling LLMs to reach AGI is "impossible"
将大语言模型扩展至实现 AGI 是“不可能的”
A large language model is trained on about 30 trillion tokens, which is roughly 10^14 bytes of text and would take a person about 400,000 years to read.
一个大语言模型在约 30 万亿个 token 上进行训练，这大约相当于 10^14 字节的文本，一个人阅读它需要约 40 万年。
A 4-year-old child receives about the same amount of data, 10^14 bytes, through vision alone in about 1 year and 10 months.
一个 4 岁的儿童仅通过视觉，在约 1 年 10 个月内接收到的数据量也约为相同数量级，即 10^14 字节。
In his view, intelligence is the ability to learn new tasks quickly or perform them without prior training, as a teenager learns to drive in about 20 hours. Scaling increases stored knowledge, but it does not produce this ability to adapt.
在他看来，智能是指快速学习新任务或在无需先前训练的情况下执行任务的能力，就像青少年在约 20 小时内学会开车一样。扩展规模会增加存储的知识，但并不能产生这种适应能力。
---- From "Perfology Clips" YouTube channel, (link in comment)
---- 来自 YouTube 频道 "Perfology Clips"（链接见评论）
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力