
LeCun 公开否认自己拦过 Meta 的 LLM 研发,并列出四项证据;本期把这四项逐项核到原始文件,同时核对同日那条关于 JEPA 发明权、用感知机与直升机讲「落地才算数」的回复。

I have absolutely never ever "asked" to stop R&D on LLMs. In fact, I was very supportive of LLM efforts at Meta. Just look up my public advocacy for Galactica (November 2022), and the open sourcing of OPT-175B (May 2022), and of the Llama herd (mid 2023). I was actually influential in convincing the Meta leadership to start an AI-first product division built around Llama in January 2023.
| 他写的 | 可核到的 | 判定 |
|---|---|---|
| 「公开为 Galactica 说话」(2022 年 11 月) | 论文《Galactica》2022 年 11 月 16 日提交 arXiv,是 Meta AI 与 Papers with Code 的 125M–120B 模型家族,摘要写「我们开源这个模型以惠及科学界」;公开 demo 在 11 月 17 日下线。他在 11 月 17 日到 25 日之间发了至少八条帖为它辩护,其中一条是「别担心,Galactica 活得好好的」131415 | 成立。他在同一批帖里说过「我不是这项工作的参与者」16 |
| 「OPT-175B 的开源」(2022 年 5 月) | 官方博客日期 2022 年 5 月 3 日;放出来的是「预训练模型,以及训练和使用它们所需的代码」,许可写明是「非商业许可,聚焦研究用例」,175B 权重需申请,面向学术与公共利益机构17 | 成立,需要限定:不是无条件下载 |
| 「Llama 那一群」(2023 年中) | Llama 2 于 2023 年 7 月 18 日发布,「研究用与商业用都免费」,含权重与起始代码;Code Llama 于 2023 年 8 月 24 日发布1819 | 成立 |
| 「2023 年 1 月成立以 Llama 为核心的 AI 优先产品部门」 | 可查到的公开节点是 2023 年 2 月 27 日:Zuckerberg 当天宣布成立一个聚焦生成式 AI 的顶级产品组,由 Ahmad Al-Dahle 负责、向首席产品官 Chris Cox 汇报2021;而 LLaMA 1 的对外公告是 2 月 24 日22 | 月份查不到;「以 Llama 为核心」查不到 |
facebook/opt-175b 这个页面也已经打不开了。I have not been involved in any Llama, except very indirectly in Llama-1, and in pushing for the open sourcing of Llama-2. Since early 2023, and until recently, Llama 2, 3 and 4 have been in the hands of the GenAI organization.
What a misleading post. First of all, the so-called "JEPA" family of techniques (2022) is actually the 1992 Predictability Maximization family of techniques
JEPA is merely a name for a general concept. The question is, and has always been, how do you make it work (particularly how do you prevent it from collapsing), and how do you make it work at scale with SOTA results on non-toy problems. That's the hard part. Ideas are a dime a dozen. Making them work is what the community will give you credit for.
Conceptual ideas like this are a dime a dozen and have been around long before you and I started working on ML. The hard part is to reduce them to practice and to make them work.
Lots of people "invented" helicopters in the early 20th century. Some were built. Some took off. None were controllable enough to be practical.
reuters.com
venturebeat.com
drive.google.com
arxiv.org
the-decoder.com
ai.meta.com
axios.com
theverge.com
ai.meta.com
reuters.com
fortune.com
theverge.com
cnbc.com
theguardian.com
dl.acm.org
arxiv.org
arxiv.org
people.idsia.ch
threads.com
ai.williamtheisen.com
quoteinvestigator.com
bnaskrecki.faculty.wmi.amu.edu.pl
centennialofflight.net
nasledie-vvia.ru
aviastar.org
asme.org
airandspace.si.edu
knowledgeminer.eu
people.idsia.ch
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.