冯·诺依曼也是“巨型记忆检索”：LLM 能否提出新解法
LeCun 说 LLM 只是记忆检索不能创造，这条帖子拿冯·诺依曼和权重不存事实来反驳，角度很别致。
一条 X 帖子回应 Yann LeCun 的观点——LLM 只是“拥有巨大记忆与检索能力的系统，无法为新问题发明解法”。
作者反驳称人类天才如冯·诺依曼同样以超大记忆检索著称，且 LLM 权重中并不存储事实字符串，而是以表征形状在推理时动态重算知识。作者认为这正是 LLM 不会被训练数据完全束缚的原因，它们确实能为新问题找到解法。
«“It’s a system with gigantic memory and retrieval ability, not one that can invent solutions to new problems.”» Von Neumann was a system with gigantic memory and retrieval. Somehow that's a frequent mark of creative genius in humans. I think the core intuition one needs to respect LLMs is very basic. They are shapes, not code or mechanics. Weights describe the shape of the representation of the mind. The data – rather, the training signal – defines this shape, but we explicitly don't have facts in there, they get recomputed on the fly. Even engram-style memory modules contain no strings. This is, in fact, regrettably wasteful, but it's also why the relation of LLMs to data is not quite as hopelessly slavish as Yann imagines. They can and do find solutions to new problems. It's a bit too late to play these games.
本条情报归属 AITOP「论文」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「中美大模型技术代差与商业化对照」「各国 AI 监管法案、合规治理与主权算力政策」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：LLM、Yann LeCun、推理模型、AI安全，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「Teortaxes」（@teortaxesTex），经 AITOP 多维评分管线结构化提炼。