杨立昆抨击Meta AI及大语言模型为“死胡同”

杨立昆（Yan LeCun），Meta前人工智能负责人已离职，他离开时反向射出毒箭，把公司搅得天翻地覆。

Yan LeCun, the former head of AI over at Meta, has left Meta, and he is shooting darts, in reverse as he's leaving, lighting the place on fire.

他即将创办自己的新公司，但在离职过程中对Meta及其团队的批评丝毫未减，这预示着事态发展不容乐观，而扎克伯格对此显然非常不满。

He's going off to his own new startup, but he has not slowed down his criticism of Meta and the team over there on his way out, which doesn't bode well for basically what's coming out, and Zuckerberg does not appear to be very happy about this.

所以在今天的播客中，我将剖析他的言论、他接下来的动向，以及Meta AI在2026年因这场风波可能面临的批评与困境。

So today on the podcast, I'm gonna be breaking what he has said, what he's off to next, and some of the criticism and some of the struggles that Meta AI is gonna have going into 2026 with this whole situation.

在深入探讨之前，如果你想试用节目中提到的任何AI模型——包括所有MetaLlama模型，以及OpenAI、Anthropic、Google、Grok等超过40种不同模型，请访问aibox.ai。

Before we get into that, if you want to try any of the, AI models that I talk about on the show, including all of the MetaLama models, including everything from OpenAI, Anthropic, Google, Grok, you know, over 40 different models, go check out aibox.ai.

你可以构建无代码工具。

You can build no code tools.

我们提供了一个游乐场，你可以直接使用所有不同的AI模型。

We have a playground where you can just use all of the different AI models.

你无需订阅所有不同的平台。

You don't have to have subscriptions to all of the different platforms.

快去看看吧，aibox.ai。

Go check it out, aibox.ai.

我会在描述里留下链接。

I'll leave a link in the description.

好的。

Alright.

我们来聊聊Meta AI目前的状况。

Let's talk about what's going on with Meta AI right now.

《金融时报》刚做了一个重磅采访。

There was a big, interview basically done by the Financial Times.

他们采访了Meta前首席AI科学家Yan LeCun。

They were talking to Yan LeCun, who is Meta's former chief AI scientist.

用你们的话说，他对过去几个月在公司的经历进行了极其尖锐的控诉。

He had a very, as you could say, blistering account of his last few months over there at the company.

他提到最令人震惊的一点是Meta操纵了他们旗舰模型LAMA four的基准测试，并批评了公司的新领导层。

He said one of the most shocking things is that Meta manipulated the benchmarks for their top flagship model, LAMA four, and he criticized their new leadership.

他说大语言模型这项技术已经走到死胡同了。

He said that large language models as a technology are a dead end.

这里有很多内容需要分析。

There is a lot to unpack here.

那么他是谁？

So who is he?

我想，简单介绍一下背景，他无疑是人工智能领域最具影响力的研究者之一。

I guess, just for like a little bit of context, he's definitely one of the most influential researchers of AI.

他被称为深度学习之父。

He's kind of known as the godfather of deep learning.

他于2025年底离开了Meta。

He left Meta late in 2025.

他在那里工作了十多年，现在正在创办一家名为'高级机器智能实验室'的新公司，而且显然他很乐意谈论离开Meta的原因。

He was there for over a decade, and he's now launching a new startup, which is called Advanced Machine Intelligence Labs, and he's super excited apparently to talk about why he left Meta.

我认为所有后续讨论的核心焦点之一就是LAMA四，这是Meta的旗舰语言模型。

So one of the big things, I think, at the center of all of the follow is LAMA four, which is Meta's flagship language model.

它是在四月份发布的。

It was released in April.

所以，理论上他们应该很快会有新动作，但Lacan承认模型的基准测试结果实际上是，他们基本上在数据上做了手脚。

So, I mean, theoretically, they should have something new coming soon, But Lacan acknowledged that the model's benchmarks were actually, like, they basically fudged the numbers on it.

他原话说'结果稍微做了点手脚'。

He said, quote, results were fudged a little bit.

他接着说Meta在不同基准测试中使用不同的模型变体来夸大性能，基本上是选用最好的数字作为最终结果。

He then said that Meta used different model variants across benchmarks to inflate the performance, basically using whatever was the best number and and calling that the the final number.

当这件事被发现时，据说马克·扎克伯格非常不高兴。

When this was discovered, apparently, Mark Zuckerberg was not very happy.

据Yan所说，马克当时非常愤怒，基本上对参与此事的所有人都失去了信任。

According to Yan, he said that Mark was really upset and basically lost confidence in everyone who was involved in this.

据他透露，马克·扎克伯格随后决定边缘化Meta整个生成式AI部门。

According to him, Mark Zuckerberg then went and decided to sideline Meta's entire generative AI organization.

他表示后果仍在持续发酵。

He said that the consequences are still playing out.

Lacan Lacan说已经有很多人离职了。

Lacan Lacan said that a lot of people have left.

许多尚未离开的人将会离开。

A lot of people who haven't left will leave.

他还表示，自从最初的LAMA模型发布后，他就基本没有参与Meta的大语言模型工作，尽管公司正在推进代号为'Mango'和'Avocado'的新生成模型，预计今年推出。

He also said that he had been largely uninvolved in Meta's LLM efforts since the original LAMA model shipped, even though the company was pushing ahead with new generative models, which they codenamed Mango and Avocado, which are expected to come out this year.

他说这次采访中最有趣的一句话是：'你不能告诉一个研究人员该做什么，尤其是我这样的研究人员。'

He said one of the funniest lines from this interview though is he said, you don't tell a researcher what to do, especially not a researcher like me.

显然他很不喜欢被人指手画脚。

So he obviously did not appreciate being told what to do.

所有紧张关系的部分原因，也是他与Meta产生矛盾的核心在于：在Meta收购Scale AI后，他们引进了该公司的CEO——Alexander Wang。

And part of the the tension with all of that is that and part of the reason why he really was struggling with Meta is because in the wake of Meta acquiring Scale AI, they brought in their CEO, which is Alexander Wang.

他是Scale AI的亿万富翁联合创始人。

He is a billionaire co founder of Scale AI.

在收购其公司49%股份后，Meta聘用了他。

They hired him after acquiring a 49% stake in his company.

他们让他负责名为'TBD Labs'的新前沿AI部门，这实际上让他成为了Yan LeCun的上司，而LeCun对此非常不满。

They put him in charge of the New Frontier AI unit, was known as the TBD Labs, And this essentially made him Yan LeCun's boss, and and LeCun was not impressed by that.

他表示自己年轻且经验不足，并说虽然王学得很快，但缺乏基础研究经验，也不理解什么能吸引或排斥顶尖科学家。

He said that he was young and inexperienced, and he said that while Wang learns quickly, he lacks experience in fundamental research and in understanding what attracts or repels top scientists.

而且，他就是那种，你知道的，他就像在说‘你绝对不能对研究人员指手画脚’。

And, he's just, you know, he's he's like, you certainly don't tell a researcher what to do.

你绝对不能对我这样的研究人员发号施令。

You certainly don't tell a researcher like me what to do.

所以我觉得他就是因为有个上司、自己不在组织顶层而感到不爽。

So I think you're just rubbed the wrong way by having a boss and not being at the top of the the organization.

Meta更广泛的策略是提供巨额薪酬方案——据说他们用1亿美元的签约奖金从OpenAI等公司挖角研究人员——很多人对此持高度怀疑态度。

Meta's broader strategy of basically giving these really huge compensation packages, you know, they reportedly had signing bonuses that were $100,000,000 to poach researchers from OpenAI and others have been, you know, a lot of people have been very skeptical of them.

勒昆说：‘未来会证明这是否是个好主意。’

Lacan said, the future will say whether that was a good idea or not.

所以，他基本上证实了这种情况确实存在。

So, I mean, he basically confirms that this was happening.

不过我认为，这种冲突远比组织架构图或勒昆受伤的自尊心要深刻得多。

I think the conflict is a lot deeper than just the org charts and maybe Lacan's bruised ego, however.

他认为大型语言模型存在根本性缺陷，无法实现真正的超级智能。

He believes that large language models are fundamentally flawed and that they're incapable of delivering true super intelligence.

他在Meta内部表示，这种观点使他与领导层以及新入职员工（在他看来完全沉迷于LLM的人）越来越对立。

Inside Meta, he said that, that view he had put put him increasingly at odds with the leadership and with new hires who are, you know, completely LLM pilled according to him.

他说LLM在实现超级智能方面基本是条死路。

He said LLMs basically are a dead end when it comes to superintelligence.

他声称Meta内部曾有人试图施压让他弱化这一观点，但他拒绝了。

Essentially, his claim is that there were efforts inside of Meta to pressure him into softening that message, but he refused.

他说：'我不会因为有人认为我错了就改变想法。'

He said, I'm not gonna change my mind because some dude thinks I'm wrong.

'我没有错。'

I'm not wrong.

'作为科学家的操守不允许我这么做。'

My integrity as a scientist cannot allow me to do this.

他还描述了Meta研究团队与产品团队之间日益扩大的分歧。

He also described this kind of growing disconnect between Meta's research and product team.

所以当研究人员提出更冒险、更具野心的想法时，产品负责人总是倾向于更安全的渐进式方案。

So while the researchers are, you know, proposing riskier, more ambitious ideas, product leaders were always favoring the safer incremental bets.

我认为这在很多组织中都存在这种情况。

I mean, I think this is kind of true in a lot of organizations.

他说，这样做会导致落后。

He said, when you do this, you fall behind.

他还提到Meta内部的政治氛围变得复杂。

And then he also said that meta became politically difficult.

我认为这某种程度上是因为杨立昆一直以自由主义立场著称，而马克·扎克伯格虽然过去支持很多自由主义政策，现在却开始与特朗普会面讨论在美国的投资计划。

I think this kind of is because Yan Lecun has been very famously quite liberal, and Mark Zuckerberg, while he's basically supported a lot of liberal politics in the past, is now, you know, going to meetings and talking to Trump about how he's gonna invest in The United States.

所以我觉得勒昆并不喜欢这种他感受到的权力格局变化。

And so I think Lacan didn't really like that shifting dynamic that he felt.

勒昆的下一站是名为'高级机器智能实验室'的新公司，他将担任执行主席。

Lacan's next act, the next company that he's going to be working on, is called Advanced Machine Intelligence Labs, he's gonna serve as the executive chairman.

日常运营由CEO亚历克斯·勒布朗负责，勒昆表示自己不想当CEO，但希望参与组织内部工作。

It is led day to day by their CEO, Alex Le Brun, and then Lacan says that, you know, he doesn't really wanna be the CEO, but he does want to work inside the organization.

他谈到自己在公司的新角色以及为何不当CEO时还挺有趣的

It's kinda funny when he was talking about his new role at the company, and why he wasn't the CEO.

他说：'我这人太散漫了，而且年纪也大了'

He said, I'm too disorganized for this and also too old.

虽然是开玩笑，但他表示现在的工作是激发灵感

He's kinda joking, but he said his job now is to inspire.

这家初创公司专注于世界模型研究

So, the startup is focused on world models.

这本质上是另一种AI研究方法，不再局限于纯文本训练

This is basically an alternative approach to AI that moves It's not just like text only training.

该技术基于Lacan的VJPA架构构建

It's built around Lacan's VJPA architecture.

这些系统通过视频、空间数据和物理交互进行学习

So, these systems learn from video and then spatial data, and physical interactions.

本质上他们试图理解世界运作规律，而非仅仅预测模型中的单词——这毫无疑问是AI的未来方向

And essentially, they're trying to understand how the world works rather than just predict a word in a model, which honestly, this is definitely the future of AI.

这是一种更为复杂的方法。

It's a much more complex approach.

他认为他的方法避免了大语言模型的许多结构性问题，比如幻觉、非确定性推理，以及对连续多模态数据的处理不佳。

He argues that his approach avoids a lot of these structural problems of LLMs, which is, you know, hallucinations or nondeterministic reasoning, poor handling of continuous multimodal data.

早期原型系统可能在一年内问世，所以不会等太久。

Early prototype systems could arrive within a year, so this isn't gonna take too long.

当然，更强大的版本将在明年推出。

Of course, more capable versions are gonna come next year.

但他说的一个有趣观点是，他说也许存在我们尚未察觉的障碍，但至少还有希望。

But something that's interesting he said is, he said, maybe there is an obstacle we're not seeing yet, but at least there is hope.

所以这家公司有着非常大的影响力。

So the company has a really big footprint.

它与法国有着非常紧密的联系。

It has a lot of strong ties to France.

法国总统埃马纽埃尔·马克龙在听说拉康离开Meta后，显然给他发了短信。

French president Emmanuel Macron, apparently texted Lacan after he heard that he was leaving Meta.

目前我认为有一点很明确，那就是拉康不再受公司政治的束缚。

For now, I think one thing's pretty clear, that is that Lacan is no longer getting kind of constrained by corporate politics.

他无意低调离开，所以会直言不讳地批评他对Meta的所有不满。

He has no intention of going quietly, so he's gonna say everything that he disliked about Meta.

可以说马克·扎克伯格现在手上确实有个公关难题。

And Mark Zuckerberg definitely has a bit of a PR nightmare, I would say, on his hands.

除此之外，我认为还有组织架构的问题。

And beyond that, I think just organizational problem.

Meta在Meta AI方面做过一些有趣尝试，但未来显然要面临诸多挑战，毕竟它已经历过不少困境。

There's a you know, Meta has done some interesting things with Meta AI, but it definitely has a lot of fighting ahead as it's definitely had a lot of struggles.

如果你能帮个大忙——如果还没评价过的话——我会非常感激你能给这个播客留下评论。

If you could do me a huge favor, if you haven't already, I would really appreciate if you could leave a review for the podcast.

这对节目帮助极大，对我而言意义非凡。

It helps the show out a ton and it would mean the world to me.

非常感谢你的收听，祝你今天接下来的时光愉快。

Thanks so much for tuning in, and I hope you have an incredible rest of your day.

一如既往，记得访问aibox.ai。

As always, make sure to check out aibox.ai.

链接就在描述里。

It is linked in the description.

这是我自己的创业项目，如果你想不写代码就用AI构建AI工具的话。

It's my own startup if you want to use AI to build AI tools without knowing how to code.

好的。

Alright.

下期节目再见。

Catch you in next episode.