Name: Yann LeCun 準備離開 Meta，創辦新 AI 新創公司！ (Yann LeCun prepares to leave Meta for a new AI startup)
Uploaded: 2025-12-16T08:58:18.000Z
Duration: 12 min 39 s

Why don't we talk about friend of the show, Yann Lecun?

我們來聊聊 friend of the show Yann Lecun 吧？

過去簡單式

I say friend of the show because Yann Lecun was a guest a number of months ago on this very show.

我說 friend of the show，是因為 Yann Lecun 幾個月前也上過我們這個節目。

祈使句型

If you have not seen the episode, definitely check out the interview that Jeff and I had.

如果你還沒看過那一集，一定要去看看我和 Jeff 的訪談。

We'd love to have him back sometime and maybe we'll have reason to have him back because.

我們很希望他能再來，也許我們會有理由讓他再來，因為。

Not that we need a reason, we'd have him anytime.

並不是說我們需要理由，我們隨時都歡迎他。

But apparently he is reportedly getting ready to leave Meta where he has worked and where he was working when we had him on for our interview with him preparing to launch his own AI startup.

但據說他正準備離開 Meta，他之前在那裡工作，也是我們採訪他時在那裡工作，準備推出自己的 AI startup。

Honestly, I'm kind of surprised that we hadn't heard a little about this sooner because he's such a, you know, such a predominant mind in the world of AI and, you know, his focus and much of the conversation that we had with him was about world models and the need for world models in AI so that those systems can simulate real world environments and predict real world outcomes, that sort of stuff.

老實說，我有點驚訝我們沒有早點聽到這個消息，因為他是 AI 領域裡一個非常、你知道，非常傑出的思想家，而且，你知道，他的重點以及我們和他談論的很多內容都圍繞著 world models 和 AI 中對 world models 的需求，這樣系統才能模擬真實世界的環境並預測真實世界的結果，諸如此類。

And reportedly he's been fundraising, at least this is according to anonymous sources to the Financial Times.

而且據說他一直在進行 fundraising，至少這是根據《Financial Times》的匿名消息來源。

And we could see an exit coming from Yann Lecun out of Meta and into his own thing sometime in the coming months.

而且我們可能會在未來幾個月內看到 Yann Lecun 離開 Meta，創辦自己的公司。

誰知道那到底什麼時候會發生？

So I like you, I just jumped on this news and it's a Financial Times report and it doesn't say much except that he might leave to start something new.

所以我跟你一樣，我剛看到這則新聞，這是《Financial Times》的報導，報導內容不多，只說他可能會離開去創辦新的事業。

I would kill to know what it is that he's working on.

我非常想知道他到底在研究什麼。

我們可以在幾分鐘內猜測一下。

But I often quote Yann Lacone on this show.

但我經常在這節目上引用 Yann Lecun 的話。

I think that he is the smartest and sanest person I know leading AI.

我認為他是領導 AI 領域裡我認識最聰明、最理性的人。

And not that Yann lecone doesn't believe that AGI is possible, but

並不是說 Yann Lecun 不相信 AGI 是可能的，但是

he's just, he's a lot more metered and sane about it and tied to reality.

他對這件事的態度比較謹慎、理性，而且腳踏實地。

And unlike somebody like Sam Altman, is not an engineer.

而且不像 Sam Altman 那樣，他不是工程師。

Yann Lacan is a professor and an expert in computer science and understands this deeply.

Yann Lecun 是一位教授，也是電腦科學的專家，對此有深入的瞭解。

And because he's an educator, I think that he can express it more clearly than some of the others who were involved in their companies.

而且因為他是一位教育家，我認為他能比那些參與公司事務的人更清楚地表達。

So I just can't wait to hear what he's working on and tie into this.

所以我迫不及待想聽聽他在研究什麼，並與此連結。

Surely would have been the politics at Meta.

肯定會是 Meta 裡面的政治角力。

Because Zuckerberg hired the boy genius to head AI.

因為 Zuckerberg 找來了那個天才少年來領導 AI。

Jan was always off in a strategic laboratory for AI.

Jan 一直在 AI 的策略實驗室裡。

He was looking at far end strategy about AI and he wasn't doing, I think, so much the, the close end product work.

他關注的是 AI 的長遠策略，我認為他並沒有做太多近期的產品開發工作。

So people thought, well, what happened to Yann Lecun when hired?

所以大家就想，Yann Lecun 被聘用後發生了什麼事？

I'm sure some people will get this and say, well, he had to leave because other people were hired and maybe there were internal politics and maybe all that.

我確定有些人會明白並說，嗯，他必須離開，因為其他人被聘用了，而且可能存在內部政治鬥爭，也許還有其他原因。

But I've anticipated for some time that he was going to go off on his own.

但我預期他會自己出來發展已經有一段時間了。

It's kind of felt that way because it has seemed, at least from the outside looking in and different reports here and there, that there have been some disagreements within Meta about it, you know, the direction that it has really been investing in and focusing in and increased, you know, focus on its, on how, on how it's developing its LLM products and kind of has seemed like Jan has gone public with, I don't know, perspectives that seem to go a little counter to the company line within Meta, which maybe isn't always necessarily a bad thing.

感覺上是這樣，因為從外面看，以及這裡那裡的各種報導，都顯示 Meta 在這方面有一些意見分歧，你知道，關於它真正投資和專注的方向，以及對其 LLM 產品開發的關注度增加，感覺上 Jan 公開表達的觀點，我不知道，似乎有點與 Meta 內部的公司立場背道而馳，這也許不一定是壞事。

You want someone, hopefully you want someone on the inside challenging your assumptions so that you get to a strong, you know, grounded place on the other end of that.

你希望，希望你能在內部找到一個能挑戰你假設的人，這樣你最終才能達到一個堅實、穩固的立足點。

But it has felt at times that he's, his perspective and his opinion is maybe going in a different direction than what Meta seems to be developing towards.

但感覺上，有時候他的觀點和意見，可能與 Meta 似乎正在發展的方向不太一樣。

Yeah, a few important things about Yann Lecun's worldview.

關於 Yann Lecun 的世界觀，有幾件重要的事。

Number one, he's been a very strong backer of open source and I think that's why Llama had open source models.

第一，他一直非常支持開源，我認為這也是 Llama 推出開源模型的原因。

但是，但是我們在節目中談過。

Seems to wobble a bit on that, so we'll see.

感覺有點動搖，所以我們拭目以待。

But I think open source is critical because it enables development entrepreneurially and educationally.

但我認為開源至關重要，因為它能促進創業和教育上的發展。

Second, he has argued, I, I, I shared, I think probably about two months ago, a very long PowerPoint he gave to students and alums at NYU where he outright said that LLMs have hit the wall, that if you really want to work on the future of AGI, LLMs are not the place they've gotten about as far as they're going to get.

第二，他曾說過，我，我，我分享過，我記得大概兩個月前，他給紐約大學的學生和校友做了一個很長的 PowerPoint 演講，他直言 LLMs 已經遇到瓶頸了，如果你真的想研究 AGI 的未來，LLMs 就不是方向，它們已經發展到極限了。

And I think that's really important as others are throwing just in huge capital expenditure trying to just throw more power at LLMs as they stand.

我認為這點非常重要，因為其他人正投入巨額資本，試圖為現有的 LLMs 投入更多算力。

He's looking at the next level and he has argued strenuously, as he did on in our interview with him, that that next level is real world.

他正在看向下一階段，他強烈主張，就像我們對他的採訪中一樣，下一階段是真實世界的。

And in that he is in sync, in harmony with Jensen Huang and others versus the scale for scale's sake argument.

這點他和 Jensen Huang 以及其他人一樣，都與「單純追求規模」的論點不謀而合。

So I've got to believe that whatever he's going to do is probably going to be some paradigm well past LLMs potentially into real world models.

所以，我相信他接下來要做的事，很可能是一種超越 LLMs 的範式，甚至可能進入真實世界模型。

That's just a guess of mine and there's maybe a few clues here.

這只是我的猜測，也許這裡有一些線索。

Jason generous papers in later, but I'm sneaking one in now.

Jason 稍後會分享很多論文，但我現在要偷偷放一篇進來。

I love a paper by a a dozen authors including Yann Lecun and Lee Fei Fei Cambrian s towards spatial super sensing in video.

我喜歡一篇由 Yann Lecun 和李飛飛等人撰寫的論文，標題是「Cambrian s towards spatial super sensing in video」。

And it's a long detailed paper that's trying to use video to see what AI can sense within that, can understand within that.

這是一篇冗長詳細的論文，試圖利用影片來觀察 AI 在其中能感知、能理解什麼。

So one video, just for example, I didn't get into great depth in this, but

所以，舉個例子，其中一個影片，我沒有深入探討，但是

they had a video going around a room and it's how many chairs are there and as it goes or circles around, is it going to recognize when the video circles around and seeing the same thing again, is it going to know what a chair is?

他們有一個影片在房間裡繞，然後問裡面有多少椅子，當它繞著轉，看到相同東西時，它會認出那是什麼嗎？

Is it going to be able to count them and separate them?

Those are the kind of real world things that he's working on.

這些都是他正在研究的真實世界問題。

And note again that Li Fei Fei is co author of this paper along with Rob Fergus also at NYU and others.

再說一次，李飛飛是這篇論文的共同作者，還有紐約大學的 Rob Fergus 等人。

And then there's another announcement that just came out that Lee Feife is doing often in American media is called Fei Fei Li is doing another her own startup which is which is related Marble or she's already done the startup.

另外還有一個剛出來的消息，李飛飛在美國媒體上經常被提及，她正在創辦一家新的 startup，叫做 Marble，或者她已經創辦過 startup 了。

It's really the first product, the first model race with Marble.

這真的是第一個產品，第一個與 Marble 相關的模型競賽。

It's a commercial product that where she says that you're going to create, you're going to take a photo, a video or anything else and then create a world around that, a 3D world from a simple prompt.

這是一個商業產品，她說你可以創建，你可以拍攝照片、影片或任何東西，然後從一個簡單的提示創建一個圍繞它的世界，一個 3D 世界。

So this all goes to the real world model question.

這一切都關乎到 real world model 的問題。

Can you figure out what's actually happening in real life?

你能弄清楚現實生活中到底發生了什麼嗎？

Because we live in 3D, thank goodness it ain't a matrix and how to reason it is a form of reasoning.

因為我們活在 3D 世界，謝天謝地這不是電影《駭客任務》裡的矩陣，而且如何推理，這就是一種推理。

I think what's going on in that world to understand the rules of that, to apply those.

我認為那裡發生了什麼事，是為了理解規則，然後套用。

So that's where the two of them have gone in one paper together and in one separate announcement.

所以這就是他們兩個人在一篇論文裡，以及在一個獨立的公告裡所做的。

And so I'm just looking for tea leaves here because I'm really interested in what Yann Lecun is going to do next.

我只是在這裡尋找一些跡象，因為我真的很想知道 Yann Lecun 下一步會做什麼。

But I've got to believe it's something in a 3D real world.

但我相信這一定跟 3D 現實世界有關。

It's so easy to see things like this and I haven't read through the Fei Fei Li World Labs article that you put in here, but

看到這種東西很容易，我還沒仔細看你放進來的 Fei Fei Li World Labs 的文章，但是

it's so easy to look at this and just immediately think video game.

看到這個很容易就立刻想到電玩遊戲。

It's certainly like, that's certainly an application.

這絕對是，這絕對是一個應用。

You can see how this would be really useful in that.

你可以看到這在其中有多麼有用。

But that's, I feel like in the long term that's really scratching the surface.

但我覺得長期來看，這真的只是冰山一角。

Well, this is where I always also call up my obsession with Hanson Wong's keynotes because he goes on at great length about digital twins and to be able to recreate a real world environment, factory, warehouse, car, on street and so on, and

嗯，這就是我總是想起 Hanson Wong 的演講，因為他會詳細談論數位分身，以及能夠重現真實世界的環境、工廠、倉庫、汽車、街道等等，

understand the laws of that real world so that you can then act on it appropriately.

並理解那個真實世界的法則，這樣你才能適當地採取行動。

That's where there's going to be a tremendous amount of work in AI and I think that is the next level.

這將會有大量的 AI 工作，我認為這就是下一個層次。

And if such a thing as AGI or something approximating it comes, it's going to require that kind of diligent work rather than the kind of one to two dimensional world of text and idea.

如果有 AGI 或類似的東西出現，它將需要那種勤奮的工作，而不是那種一維或二維的文字和想法世界。

So does, does a story like this about Yann Lecun being made public through, you know, anonymous sources or whatever, does that move the transition?

那麼，像 Yann Lecun 的這種故事，透過匿名消息來源公開，這會推動轉變嗎？

Does something like move the transition faster or do you think this is something that Meta probably already knew about Yann Lecun, that he was probably.

這會加速轉變嗎？還是你認為 Meta 早已知道 Yann Lecun 的事，他可能。

I can't imagine he's been such a good executive at Meta.

我無法想像他在 Meta 擔任過這麼好的主管。

And of course, don't forget he also has another life as an NYU professor.

當然，別忘了他在 NYU 也是位教授。

He just got awarded by the King of England, he's been awarded by the President of France.

他剛被英王頒獎，也被法國總統頒獎。

He's a world renowned AI expert and he's not going to screw with his reputation in those worlds.

他是世界頂尖的 AI 專家，不會毀損他在這些領域的名聲。

He's a very mature academic and executive.

他是一位非常成熟的學者兼主管。

So no, this had to been known and had to be negotiated and I think it's a little bit of a leak, you know, and

所以不，這一定是早就知道並協商好的，我認為這有點像是洩漏，你知道的，而且

Meta has been, let me just see what their stock is doing lately.

Meta 最近的股價如何，讓我看一下。

It's down 2% today, looking over a month.

今天下跌了 2%，看一個月來的表現。