为什么基于能量的模型可能是人工智能的下一次重大转变

你再也不必猜测下一个词了，你不再需要玩猜词游戏。

You never have to guess the next word, or you never you you don't really play in a guessing game anymore.

你可以看到你的能量景观，知道正确答案在哪里，然后立刻知道该往哪里走。

You see your energy landscape, you know where the right answer is, and you just know where to go right away.

这正是节省你时间的关键。

And this is what's saving your time.

所以我们的模型非常小。

So our models are very small.

我们的模型参数规模从两千万到两亿不等，而且我们是在最便宜的H100显卡上运行的。

Like, we we scale in this model from, like, 20,000,000 parameter to 200,000,000 parameter, and there is a range And in we run it on, like, the cheapest h 100 GPU.

我当时就想，这太令人印象深刻了。

So I I was like That's impressive.

是的。

Yeah.

我深受你大脑的启发。

I was like a lot inspired by your brain.

对吧？

Right?

比如，你现在和我说话的时候，如果有人问：‘我能为你创建一个数字孪生体吗？’

Like, as you're speaking to me right now, if somebody says, hey, can I create a digital twin of yours?

而我正试图映射你的身体内部状态、你听我讲话时处理的信息、你的视觉感知等。

And I'm trying to map, like, your internal body state, the information you process as you're listening to me, your visuals.

这需要海量的GPU、能源和协调工作，但你的大脑却能自然地以不到20瓦的功耗完成这一切。

It's gonna take, like, enormous amount of GPUs and energy and orchestration, but your brain just naturally can do it, like, less than 20 watts.

所以，如果你的架构设计得当——我必须明确说明，我并不是在拿人和机器直接比较，因为我们已经进化了数百万年。

So if you make the architecture right well, I'm not comparing people, full disclosure, with the machines because we evolved, like, for many years.

是的。

Yeah.

对。

Right.

对。

Right.

是的。

Yeah.

所以背后确实有很深层的进化机制，但在这里，核心理念是：如果设计得当，根本不需要那么多GPU。

So there's some some very hardcore evolution behind, but here, the idea is there if you make it right, it's not supposed take you that many GPUs.

这很巧妙。

That's neat.

我在使用Kona时注意到一件事，特别吸引我：观看它让我想起了扩散模型。

I noticed something I noticed that really stood out to me in playing with Kona is that to watch it reminds me of, like, diffusion.

这种风格上是否存在相似性或关联呢？

Is is there a similarity or relationship there in that style?

比如，看着它运行时，那种感觉和扩散模型很像，仿佛在反复迭代同一个东西，不过我知道两者有巨大差异。

Like, watching it, it it's that same feel as it's going over, and it looks like it's iterating over the same thing, like but I understand there's there is a big difference.

对吧？

Correct?

是的。

Yeah.

所以这可能会是一个非常深入的讨论，因为扩散模型有太多不同的版本。

So it it can be quite deep discussion because there is so many, like, different versions of diffusion models.

是的。

Yeah.

对我来说，进行技术讨论时，我总是希望先明确定义概念。

And to me, to have a technical discussion, like, I always want to define things first.

我们不能含糊其辞。

We don't we don't go on ambiguous.

对吧？

Right?

但显然，扩散模型之间有一些相似之处。

But there's some similarities with diffusion models, obviously.

一般来说，基于能量的模型理念并不新鲜。

And the ideas of energy based models in general, they're not that new.

它们已经存在了大约二十年，甚至更早的科学研究可以追溯到四十年前。

They've been out there for, like, twenty years and some early, like, science even forty years ago.

问题是，没有人尝试过构建基于能量的推理模型。

The the problem was nobody tried to build energy based reasoning model.

人们尝试将基于能量的技术和建模应用于现有的大语言模型或图像识别。

People tried to apply, like, energy based techniques and modeling to, like, existing LLMs or image recognition.

但真正去设计推理部分本身。

But to actually design the reasoning part itself.

这才是真正新颖的地方，我们只是运气好，成功了。

This is where things are really new and we just got lucky that we made it.

我觉得是的。

I think Yeah.

这太棒了。

That's awesome.

我喜欢

I love

这个。

that.

我也欣赏你的坦诚。

I appreciate the honesty too.

我真的觉得这是一个令人兴奋的方向。

I I really feel like this is an exciting direction.

当我想到，你知道的？

When I think of you know?

但在思考这个问题时，很难摆脱大语言模型的思维定式。

And it's hard to get out of the LLM trap in my mind when thinking of it.

我也得说，因为我一直想到代理中的节点。

I've gotta say too because I keep thinking of, like, nodes in agents.

比如，速度简直快得离谱。

Like, speed is insane.

作为一个经常做数独的人，这真的让我印象深刻。

And as a guy who does a lot of sudoku, it's also really impressive.

嗯，数独只是其中一件事。

Well, sudoku is just one of the things.

我们只是想，有没有一种最简单的方式来说明，存在一些不依赖语言的任务，这些任务是人们熟知且喜爱的，并且能够立即理解，同时也能用大语言模型进行测试。

It's just we we thought of like, what's the simplest way to illustrate that there are different tasks which are not based on any language and people, like, know and love and can get immediately and something which can be tested with the LLMs.

因为大语言模型常被宣传为具备推理能力的模型。

Because LLMs, we compare it to advertised as LLM reasoning model.

所以它的目标是推断和扩展知识。

So it's meant to be extrapolating knowledge.

它从一些游戏中学习，然后应该能够推导出其他游戏的规则。

It learns from some game games, and then it's supposed to, like, sort of extrapolate the rules for other games.

但在这里，我们离这个目标还很远。

And here, we're not even close to this.

而能够推断和扩展知识，是自然智能最关键的能之一。

And being being able to extrapolate knowledge is one of the most crucial abilities for natural intelligence.

对吧？

Right?

所以这就是大语言模型所缺乏的。

So there's like this is what LLM doesn't have.

所以如果你拿一个大语言模型，教它做数学并赢得IMO和其他各种奥赛，我们自然会假设：哦，这个模型真聪明。

So if you take LLM and you teach it to do some math and win IMO and all of these fancy Olympiads, It's just we have a natural assumption, like, oh, this model is so smart.

那我给它一些代码，或者给它一些其他数学问题，它应该也能解决。

Let me give it some code or let me give it some other problems in mass and it's gonna solve it.

但现实并非如此。

And reality is not.

对吧？

Right?

并不是这样的。

It's not.

它只是非常擅长你训练它做的那一件事。

It's just really good at one thing you're trained for.

但如果你让一个孩子去学习数学，他们很可能会擅长数学建模。

But if you take a child and you force them to, like, learn some mathematics, they're probably gonna be good at, like, mathematical modeling.

他们可以尝试理论物理，甚至走得更远。

They can try theoretical physics or they can, like, go even further.

他们甚至可以学习法律。

They can even study the law.

对吧？

Right?

法律就像是语言中的逻辑，我个人做不到，但我有很多物理系的朋友去了哈佛法学院。

Which is like a logic in language, which I I I can't do personally, but I have a lot of friends from physics department went to law school in Harvard.

哦，哇。

Oh, wow.

真有趣。

Funny.

我之前没意识到那里有这么多交叉之处。

Didn't real didn't realize there was so much crossover there.

是的。

Yeah.

因为人们天生擅长在不同领域之间迁移知识，而创造力就来源于此。

Because, like, naturally, people are good at extrapolating knowledge across different domains, and that's where the creativity comes from.

对吧？

Right?

你有时会从一些你从未想过的地方获得灵感，然后突然就奏效了。

You sometimes, like, get idea from some other areas which you never dreamed of, and all of a sudden it works.

是的。

Yeah.

不幸的是，在这个阶段，它们还做不到，而且我认为它们永远也做不到这一点。

Unfortunately, at this stage, they can't do it, and I don't think they will ever be able to do this.

哇。

Wow.

这很有趣。

That's interesting.

接下来是什么

What's the next

在Kona的发展上，你现在处于什么阶段？

where would you say you are right now in terms of your development of Kona?

那你的路线图是什么样子的？你能分享到什么程度？

And and, like, what are your what's your road map look like to to what extent you can share?

我明白这可能问得太多了。

I understand that that might be asking too much.

不。

No.

当我们刚开始这家公司时，我只有些理论上的想法。

It's so the first, when we started this company, I just had, like, some theoretical idea.

对吧？

Right?

然后我身边有一群才华横溢的工程师，他们把我的想法变成了一个概念验证，那还是几个月前的事。

And then I was surrounded by talented engineers who just brought this idea into a form of proof of concept, and that was like few months ago.

自然而然的问题就是：这个概念验证能不能成为我们今天这个模型的原型？

And the natural question was like, oh, can the proof of concept be the actual, like, toy model for the model for the model we have today?

所以答案是肯定的，但要达到这一步，我们必须进行一系列实验，来评估哪些是对的、哪些有效、哪些无效。

So the answer was yes, but to get there, we had to perform like a series of experiments to evaluate, like, what's right, what works, what doesn't.

所以当架构完全设计好后，下一步就是：它是否兼容LLM或一般的Transformer？

So when the architecture was fully designed, the next step would be, oh, is it compatible with LLMs or with transformers in general?

因为它从根本上来说非常不同。

Because it's so fundamentally different.

我们甚至不确定这样做是否可能。

We didn't even know, like, it's it's possible to do.

所以第一步是连接Transformer并尝试稍微扩展它，然后再将其缩小回玩具模型的版本。

So the first step was to attach transformer and try to scale it a little bit, and then kinda shrink it back to the toy model version.

我们成功做到了。

So we successfully done so.

然后我们想，是否可以不使用LLM，而是连接一个最简单的、与LLM相关的模型，这个模型很小，而且我们能理解。

And then we're like, oh, can we, like, not even have an LLM, but the most simplest version of something related to LLM attached to transform it, which is small and we understand.

所以我们把它连接上了。

So we attach that.

我们还对其进行了扩展，然后再缩小，结果发现：嗯，这个方法是可行的。

We also scale it and then scale it back and like, okay, that works.

我们不如直接把真正的LLM连接到EBM上，看看它作为用户界面表现如何？

How about we just attach the real LLM to the EBM and see how it how it is as a user interface?

它能否以我们期望的方式提示EBM？

Can it prompt the EBM in the way we want?

答案是肯定的。

And the answer was yes.

于是我们再次进行扩展并测试，我们有一组与空间思维和层次规划相关的基准测试。

So we again scale it and then test it, like, we have a set of benchmarks, which is related to spatial thinking and hierarchical planning.

我们为模型的最小版本、概念验证版本和真实版本都设定了基线，然后相互对比，结果似乎都有效。

So we, like, had baselines for the smallest version smallest version of the model, then proof of concept, then the real version of the model, and kind of like compare it back and forth, and seems to be working.

于是我们想，好吧，让我们尽可能地扩大规模。

So then we're like, okay, let's actually try to scale it as much as we can.

我们进行了一系列实验，并且对它的运行机制有了相当不错的理论理解。

And we performed a bunch of experiments, and also we have pretty decent theoretical understanding how it works.

因此，我们没有看到任何障碍。

So we don't see any obstacles.

但你知道，工程工作可能会很棘手。

But, you know, engineering can be tricky.

所以有时候事情能运行，有时候你得调试。

So sometimes things work and sometimes things you need to debug.

对我个人来说，最大的部分是如何表达这一点？

So the biggest part for me personally was to how to say it?

所以EBM本身并不是自然的自回归模型，因为它不涉及逐字生成，而且也不是极度激进的。

So the the EBM is not naturally auto regressive, because there's no talking and it's also non ultra aggressive.

这意味着它同时在审视所有可能的情景。

So meaning it's overseeing all possible scenarios at the same time.

但当你试图将其与变换器连接时，变换器是非常激进的。

But when you try to attach it to transformers, transformers are very ultra aggressive.

所以你得把这种狂野的东西，连接到一个思维非常线性的东西上。

So you have to, like, take this wild thing and attach to something which is, like, thinking very sort of linear.

一步接一步。

One step after another.

是的

Yeah.

对

Yeah.

所以你在中间面临巨大的信息丢失。

So you you facing a huge information loss in the middle.

当你尝试用LLM来提示EBM时，情况也是一样。

And then the same the same thing when you try to prompt using LLM, the EBM.

所以这一层的信息也大幅减少了。

So there is also like a giant reduction of the information on that layer.

所以我们当时试图单独协调这一层，这花了一些时间，也尝试看看它的扩展性如何。

So we were trying to, like, orchestrate this layer alone, which took some time and try to see how it scales.

所以我认为这是我们遇到的最大困难。

So I think that was the biggest difficulty we faced.

但现在架构已经建立了。

But now the architecture is there.

它是可扩展的。

It's scalable.

它已经按照我们的预期在进展，甚至还有一些超出。

It's it's already, like, progressing the way we expected and a little bit even beyond.

是的。

Yeah.

对。

Yeah.

我们现在就处于这个阶段。

That's where we are.

太棒了。

That's awesome.

它是否永远都需要语言模型作为接口，还是这只是临时步骤？

Will it always need the the language model attachments to it as like the interface or is that just like an interim step?

明白。

Okay.

这只是因为，正如我所说，GI 是一个生态系统。

It's just for you know, like, I said to me, GI is the ecosystem.

你需要能够适应和规划。

You need to be able to adopt and plan.

有时候你需要语言，因为这是为了那些可能想和你交流的人，而你作为AI，我的意思是。

And sometimes you need language because it's for people who might want to speak to you and you as AI, I mean.

你需要具备说话的能力。

And you need to have an ability to speak.

你需要具备执行空间任务的能力，比如导航等等。

You need to have an ability to, like, perform spatial tasks like navigation and so on.

在这里，你不一定非得有这个能力，但拥有它会更好。

And here, like, you don't have to have it, but it's nice to have ability to have it.

对于机器人来说，你根本不需要它。

For robotics, you don't need to have it at all.

你只需要把EBM用来控制你的能量分配，比如城市中能量的分布。

You could just put the EBM to control your, like, energy grades, like how much energy distribute in the town.

而且你可以控制它，因为它会实时分析海量数据。

And you can control it, like, because there's it's gonna analyze a giant chunk of data in real time.

这就是为什么我们具备毫秒级响应能力在这里很重要，因为这对交易也同样重要。

And this is why the millisecond skill we have is important here because also important for trading.

这就像我所说的软性优势应用场景。

This is like my soft soft card use case.

说得通。

Fair enough.

我只是非常喜欢整个算法的理念，我个人也特别喜欢博弈论，而交易场景中就蕴含了大量相关原理。

Just I just love this whole idea of, like, algorithms, and it's it's like I love game theory, personally, and there's a lot of it in trading use case.

但确实如此。

But yeah.

所以你并不总是需要大语言模型，但拥有它确实很好。

So you don't have to have LLM all the time, but it's nice to have it.

你能

Could you

做同样的事情，比如你把这个用在机器人上。

do the same thing where you attach like, let's say you have this in a robot.

你能以同样的方式给它附加一个视觉动作模型吗？

Could you attach like a vision action model to this in the same way?

这是不是意味着

Is that how that

会这样运作？

would work?

或者

Or

是的。

Yeah.

我认为我们其实不需要视觉动作模型。

I don't think we actually need a vision action.

你只需要一个能获取视觉信息的传感器。

All you need is like a sensor which takes the visual information.

所以，它完全可以只是摄像头，用来采集大量图像，同时也包括温度传感器等，任何用于自动驾驶或一般自主导航系统的传感器都可以。

So it it can be just the cameras which take in, like, a bunch of pictures for you, but it also, like, a temperature sensors as well, whatever sensors people use for, like, self driving cars or, you know, self navigating systems in general.

这些全部都可以映射到能量空间中，并且因为我们在EBM中具备推断知识的能力，所有这些都可以整合成一个统一的空间。

And this alone can be all mapped in energy space, and it all can be like one space for all because we do have ability to extrapolate knowledge in in the EBMs.

这另一部分让我非常兴奋，因为现在你不再需要专门为生成视频或图像而设计独立的模型了。

That's another part I'm so excited about because now all of a sudden you don't need to have, like, a separate model just for generating your videos, generating your images.

你可以将所有功能整合在一个系统中，而且

You could just have it all at one and

哦，这太有道理了。

Oh, that makes so much sense.

我们一直认为，外部存在一个完整的世界，最终必须被它们所理解。

You know, we keep thinking about this idea that that there is this whole world out there that is eventually going to have to be understood by them.

LLM和我们之间的一个重大区别在于，我们拥有这种经验性知识，所有这些空间感知信息，而它们并不具备同等程度的这类能力。

That one of the big differences between, like, an LLM and us is that we have this experiential knowledge, all of this spatial stuff that that they do not have in the same way.

你听完这些之后，是否觉得EBM可能是其中的一环？也许正是那个最终通向AGI的认知缺口？也许这类系统需要多个组成部分共同构成？

Do you see after saying that, do you see that EBMs are a piece of this, maybe the cognitive hole that becomes AGI at some point, that maybe there are multiple elements to such a thing?

如果我说得有点离谱，抱歉。

I'm sorry if that's a little out there.

好的。

Okay.

这绝对是朝着更大目标迈出的一步。

I think it's definitely a step towards something bigger.

这确实是一步进展。

It's definitely a step.

令人兴奋的是，这一步能与我们已有的技术兼容。

And what's exciting, this step is compatible with what we already have.

所以我们并不是说，要完全抛弃LLM，因为EBMs的存在。

So we we we don't say, like, I am gonna kill the LLM entirely because of EBMs.

实际上，语言很重要，LLM也值得拥有自己的位置。

You actually like, language is important and, you know, LLM deserve the place to be.

是的。

Yeah.

所以它就会是那样。

So it's gonna be that.

我们会看看它如何扩展。

And we'll see how it scales.

对吧？

Right?

我们会看看它如何学习、如何适应。

We'll see how it learns, how it adopts.

将来会有一个世界，在那里我们会知道EBMs特别擅长某一项任务，但在另一项任务上却不擅长。

There's gonna be a world when we're gonna know that EBMs are specifically good at one one particular task, but not good in another particular task.

如果是这样，就会出现一些我们目前还不知道的EBMs版本。

And if that's the case, there's gonna be some versions of the EBMs, like, you know, we don't know of.

但人们很有创造力，我们会想出一些办法，这整个就是生态系统。

But people are creative, and we're gonna come up with something, and it's it's the ecosystem.

是的。

Yeah.

对。

Right.

我认为人也是如此。

Same's true of people, I would say.

我的意思是，无论是人还是元素，似乎一切都存在我们擅长的领域，也有一些我们特别不擅长的愚蠢方面。

I mean, whether it's people or elements, it seems like everything has here's what we're here here are areas where we're really good, and here's something silly that we're just awful at.

没错。

Exactly.

而这正是构成我们社会的原因。

And this is what makes us as a society.

对吧？

Right?

我们只是，比如说

We just, like

嗯。

Mhmm.

总的来说，你知道的，作为集体意识共同发挥作用，为所有人带来益处。我认为人工智能也可以成为其中的一部分，同样的道理。

Altogether, you know, function as collective consciousness and collectively contribute And to the benefits of all of I could see AI to be a part of it and the same the same thing.

这就像，循证医学只是故事的一部分。

Is it like, EBM is one part of the story.

对吧？

Right?

你是在问我接下来会怎样。

You're asking me what's next.

循证医学只是故事的一部分。

EBM is one part of the story.

将循证医学与大语言模型结合，是故事的另一部分。

EBM attached to the LLM is another part of the story.

但你还得在上面加上智能代理层。

But then you have agentic layer on top.

我们实际上已经拥有相当智能的代理层了。

We actually do quite quite intelligent agentic layer.

然后你可以在LLM之间、EBM之间协调智能代理层，或者直接克隆EBM和LLM的混合体。

And then you can orchestrate the agentic layers between LLMs, between EBMs, or you can just clone the hybrids of EBMs and LLMs.

接着你可以设置某种博弈论情境，包括传递性博弈和非传递性博弈。

And then you set up some sort of game theory situation, transitive games, non transitive games.

这样一来，你突然就拥有了一个完整的AI生态系统演化过程。

And this is you you have your full evolution of AI ecosystem all of a sudden.

因为它会自我训练、自我对齐，并创造出我们根本无法想象的东西。

Because it's gonna self train, it's gonna self align, it's gonna create something we can't even dream of.

所以这才是令人兴奋的部分。

So that's exciting part.

我不知道在哪个时刻你会称之为通用人工智能（AGI）。

And I don't know at what moment you call it AGI.

你是说，当你已经有代理能为你解决黎曼猜想时，这才算AGI吗？

Do you call it, like, when you already have agents, I don't know, bringing you solutions for Riemann hypothesis?

还是说，只要能控制能源电网和汽车，就算AGI了？

Or it's just ability to control the energy grid and the car?

那它到底是什么？

So what is it?

这是个好问题。

That's a great question.

好吧。

Alright.

是的。

Yeah.

我不知道。

I don't know.

经典的定义是它能够泛化。

The the the the classic is is it can generalize.

对吧？

Right?

所以，目前能量模型能够泛化吗？

So, I mean, at this point, can can the energy model generalize?

你是说你可以输入任何数据，它就能给出答案，通常都能提供可靠且准确的回答。

Can can you give it you're saying you can give it any data in, and it can give you an answer it can give you reliable answers, usually accurate.

是的。

Yeah.

我的意思是，这在我看来就是通用人工智能。

I mean, that sounds like AGI to me.

但现在我们看到了。

But now we see.

但你知道，十年前，如果人们知道现在的大型语言模型能有这样的表现，他们也会称它为通用人工智能。

But, you know, ten ten years ago, people would like, if they would know how LLMs perform right now, they would also call it AGI.

我认为通用人工智能的定义也会随着发展而演变，因为我们将会迎来新事物。

I think the definition of the AGI is gonna evolve as well because, like, we're gonna have a new thing.

我们会看到新事物的缺陷，然后说：‘哦，不。’

We're gonna see the flaws in the new thing and be like, oh, no.

不。

No.

不。

No.

这完全是胡扯，AGI应该是更高层次的东西。

This is bull Like, AGI is something higher.

改变标准。

Move the goalposts.

是的。

Yeah.

我可以想象，大学里会开设关于AI和AGI的课程。

I could see, like, textbooks, the course on AI and AGI in the universities.

他们会讨论不同领域中曾经被称为AGI的东西。

They're gonna talk about different areas what was called as an AGI.

所以

So

所有的

All of the

在不同的时间，我们不断把AGI的门槛推得更远。

different times we've pushed the bar further away on AGI.

没错。

Exactly.

AGI的百科全书。

The encyclopedia of AGI.

我有太多相关的问题要问，但我想回过头来谈一点：假设你在机器人例子中使用了能量模型，并且你想和机器人对话。

I have I have so many branching questions on this, but one of the things I wanted to go back and touch on is, let's say you were to have the energy model in the robot example, and you wanted to talk to the robot.

会不会也是类似的情况？比如，假设它有传感器，但它是否需要某种语言或语音模型才能真正和你交流？

Would it be the say would it be a similar thing where, okay, let's say it has sensors, but would it need some sort of like language speech model to to actually talk to you out of it?

然后，对于智能体，我也有同样的问题：你知道，智能体是否需要某种类似……

And then I guess the same question for agents, which is, you know, does does does the agent need some sort of like, I guess, like

比如为了人类互动？

Like for human interaction?

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

但这始终取决于人们的需求，对吧？

Well, it's it's always up to people, right, what we need.

这不取决于AI模型。

It's not up to AI model.

所以，如果有人告诉我——而我是一名工程师——如何设计这样的机器人，我会问：我们有哪些可用的数据？

So the way if if somebody would tell me and I were an engineer how to design such robot, I would say, okay, what what kind of data we have available?

我们可能有语言部分，也就是人们跟我交谈。

We have probably the language part, which is, you know, people talking to me.

我们还有视觉部分。

And we also have the visual part.

我们还有传感器测量距离，判断你离物体有多远。

We also have sensors measuring the distance, how far you are from things.

你可能还会从周围的整体环境中获取输入，从而调整你的行为和身体移动。

You probably have, like, an input from, like, overall environment around you, so you can adopt your behavior and your, like, moving your body around.

所以你可以在这里获得大量输入信息，同时你也有一个目标。

So you can have, like, a bunch of input information out here and you also, like, have a purpose.

那么，这是一种什么样的机器人呢？

So what kind of robot is it?

它是你的家用清洁机器人，还是别的什么？

Is it gonna be, I don't know, your house cleaner or something else?

因此，目标将决定你需要多少以及何种类型的互动。

So the purpose is gonna define how much and what kind of interaction you're gonna have.

如果这是一个照顾病人的机器人，而这个人独自一人，没有看护者。

So if it's a robot who is taking care of like a sick person and this person just leave themselves and they don't have a caregiver.

那么它们会有一套固定的日常流程，这些流程可能需要通过语言来传达，但同时也必须具备评估周围环境的能力。

So they they gonna have a set of routines and they probably should be communicated using language, but it's also they should be having ability to, like, evaluate what's around them.

这个人有危险吗？

Is this person in danger?

我该拨打911吗？

Should I call 911?

你明白我的意思吗？

You know what I mean?

这取决于机器人的用途。

It depends on the purpose.

因此，将所有这些要素整合在一起，并确保它们按预期方式协同工作，这一点至关重要。

So this is why it's important to have all of these ingredients in place and be combined in in the way and making sure it's functioning the way it's meant to be.

所以，仅靠大语言模型是无法做到这些的。

So LLMs alone cannot do this.

而且这也很昂贵。

And it's also, like, very expensive.

在计算方面，你得等上几分钟才能知道它是要往左还是往右。

And in terms of compute, you have to wait for, like, few minutes before it knows is it going left or right.

所以你需要在秒级、毫秒级的范围内理解周围的情况。

So you need to be able, like, on on the seconds, milliseconds to understand what's around you.

人类现在能做到这一点，真是令人惊叹。

Kind of amazing that humans could do that now.

我知道。

I know.

我知道。

I know.

也许我们本身就是通用人工智能。

It's like May maybe we are an AGI.

我们正在寻找它。

We're looking for it.

我们只是在尝试创造一个自己的版本，但我不确定。

We're just trying to create a version of ourselves, but I don't know.

是的。

Yeah.

更像是一种玩笑，说有一天，我们的孩子会回望过去，觉得你们竟然让人开车载你们？

More or joking that one of these days, our kids will will look back and think, you let people drive you in a car?

像这样，你们怎么会做这种事，人类？

Like, why would you ever do that, humans?

是的。

Yeah.

那么这里的扩展范式会是什么？

What would what would then be the the scaling paradigm here?

也许这更多是关于如何训练这些东西。

And perhaps this is more about, like, how you train one of these things.

比如，是不是只是尽可能多地使用GPU进行同样的训练运行？

Like, is is it is it just doing doing the same training run on as many GPUs as you can do?

它是以相同的方式运行的吗？

Is it is it work the same way?

或者怎么说呢。

Or how do yeah.

我们会找到答案的。

Well, we're gonna find out.

我只能猜测它会变成什么样。

I can only speculate how is it gonna be.

比如，目前我们所做的扩展，暂时还是只用一个GPU。

Like, right now, the scaling we're doing, it's all still one GPU for now.

但我们还没有让模型真正进入现实世界足够久，去真正理解它。

But we haven't, like, forced the model to be out there in real world enough yet to understand.

所以对我来说，当我从理论角度研究它时，最重要的问题是：架构是否可扩展？

So for me, like, when when I was working on it from the theoretical perspective, the most important question is, is architecture scalable?

而那些参数到底在哪里？

And is that what what parameters are out there?

因此，有办法在不使用太多GPU的情况下分解你架构中的复杂性。

So there are ways to break complexity in your architecture without having too many GPUs.

我来举个例子说明我的意思。

I'm gonna give you an example by what I mean.

所以，大语言模型具有一些性能水平。

So LLMs, they have some level of performance.

一旦你达到GPU的临界数量，拥有大量计算资源和数十亿参数时，就会出现某种基础性转变。

And then once you reach a critical mass of GPUs, so you have a lot of compute and a lot of, like, billions parameters, then you have some sort of base transition.

突然之间，你就会观察到不同的行为。

And all of a sudden you've seen a different behavior.

而正是在这个阶段，大语言模型才真正开始发挥作用。

And this is where, like, things really start working for the LLMs.

需要有足够的GPU，才能观察到这种复杂性带来的行为变化。

Like, there needs to be enough GPUs to see this kind of complexity change behavior.

在我们的情况中，当我们使用混合架构，将EBM与大语言模型结合时，就能观察到这种相变。

And for our case, we see the phase transitions when we work with the hybrid, when we have the EBM attached to the LLM.

存在一些阶段，LLM占主导，但随后EBM开始占主导。

There's there are regimes when the LLM dominating, but then the EBM starts dominating.

如果EBM阶段占主导，那你其实不需要GPU。

And we just like, if the EBM regime is dominating, then you you don't really need GPUs.

但如果LLM占主导，你仍然需要GPU。

But if there is a regime on LLM dominating, you still need the GPUs.

所以简短的回答是，这可能取决于具体的应用场景。

So it's the short answer here, it's probably gonna depend on the use case.

如果你处理的是需要实时处理大量语言的应用场景，那么在使用EBM与LLM的混合模型时，很可能有一部分时间LLM会占主导。

If you're working with a use case when you have to process a lot of language in real time, so most likely there's gonna be a portion when the LLM is gonna be dominating if you're using the hybrid of the EBMs.

但我们直到实际测试后才知道。

But we don't know until we, like, actually test test.

但当我看到这种相变时，相变是物理学中的一个非常标准的术语。

But when I when I saw this, like, phase transition and phase transition is, like, a very standard term in physics.

你通常在凝聚态物理中看到它，这些模型背后有着优美的数学原理，这也是我博士研究的一部分。

You typically see it in condensed matter, and there's beautiful mathematics behind of these models, and that was a part of my PhD studies.

当我看到这一点时，简直让我震惊。

And when I saw it, it's, like, blew my mind.

我当时就想，这太疯狂了，居然真的会发生。

I was like, it's just you it's it's crazy that it happens.

所以

So

好的。

Okay.

我有个问题想问你，但我相信科里也有同样的疑问。

I've I've got out there question for you, but I'm sure Corey has this one too.

你认为人类是否有点像在运行一个基于能量的模型？

Do you think that humans are kind of like a are we running an energy based model up here?

你认为这是不是

Do you think that's kind of

和我们正在做的事情很接近？

close to what we're doing?

当我学习的时候，好吧。

When I was when I was studying so okay.

坦白说，没人知道大脑是如何工作的。

Full disclosure, nobody knows how the brain works.

外面有一些理论。

There are theories out there.

有理论，也有假设。

There are theories, there are hypotheses.

一个好的理论的标准是，你建立一个模型，试图将其与真实数据对应，观察你在现实生活中实际看到的现象，同时还能做出一些预测，而这些预测在某种程度上会成真。

And what makes you a good theory is if you take a model, you're trying to map it in real data and you see what you actually see in real life, but you also can make a little bit of prediction and then prediction to some degree comes true.

所以，这就是一个好模型的标准。

So that's what makes it a good model.

对于视觉皮层，比如在人类和动物中是如何运作的，已经有一些很好的模型。

And there are good models for, like, how the visual cortex, for example, performs in humans, in animals.

其中一些模型已被应用于人工智能，尤其是在图像识别方面。

And some of it is adopted to AI, especially for the image recognition.

因此，有一些基于能量原理的技术。

So there are techniques for the using the energy based principles.

但这是描述大脑的模型吗？

But is this the model to describe the brain?

没人知道。

Nobody knows.

但这仍然是一个非常开放的科学问题。

But again, it's a very open scientific question.

如果是的话就太酷了。

Be cool if it was.

是的。

Yeah.

从我个人作为物理学家的职业经历来看，我发现这个世界上的一切都倾向于最小化能量，几乎任何事物都可以写出拉格朗日量。

I have From a my personal career, as a physicist, I see that everything in this world wants to minimize energy and pretty much for everything, you can write the Lagrangian.

至于这个拉格朗日量是否能给出正确的运动方程，我就不得而知了。

Whether this Lagrangian is gonna give you correct equations of motions, I don't know.

这正是实验物理学要告诉你的。

That's what the experimental part of physics is gonna tell you.

但事实上，你知道，这似乎走在正确的道路上。

But in reality, you know, this seems to be on the right track.

这对我来说。

This to me.

我有个问题。

It's a I have I have a question.

在考虑能量基模型不需要GPU负载以及其他相关开销时，当我想到像我们之前讨论过的与大语言模型结合的混合场景，这感觉像是一个能带来更高效率、更可持续的AI的解决方案。

In thinking about EBMs as being as not requiring the GPU load and all of the other things that are involved, when I think about that in a in a hybrid scenario like we were discussing with an LLM, that feels like a solution that could lead to much greater, you know, efficiency and more sustainable AI in the long run.

这是过度延伸了，还是我方向对了？

Is that a stretch, or or am I on the right track?

这取决于你怎么定义可持续。

Well, it depends how you define sustainable.

对吧？

Right?

那么，什么是可持续人工智能？

So what is sustainable AI?

这是个很好的问题。

I that's a great question.

我们假设它更节能，因此运行成本更低。

Let's say let's say it's more energy efficient, so therefore, it's not costing as much to run.

是的。

Yeah.

我就从这点开始。

I'll start there.

对。

Yeah.

我认为无论是从环境角度还是财务角度来说都是如此。

I I would say both environmentally and financially.

是的。

Yeah.

会很好。

Would be nice.

对吧？

Right?

不。

No.

但你知道吗，当我看到这个时，这叫什么？

But you know what's when I saw this what's this called?

Molten book on x。

Molten book on x.

哦，是的。

Oh, yeah.

当有不同的智能体在互动时，你知道，它们在彼此交谈，并且产生不同的主题。

When there's, like, different agents playing, you know, they're talking to each other and they're coming up with different topics.

我最初的反应是，哦，你其实可以做到同样的事，但为了节省你所使用的计算资源的成本。

My initial reaction was like, oh, you can actually do the same, but to preserve your money for how much compute you're using.

因此，你可以通过使用LLM或EBM来优化这种智能体系统，以优化智能体自身资源的使用。

So you can actually optimize this agentic game either using LLM or EBMs on optimizing for the agentic system its own resource.

对吧？

Right?

而且

And

我们是在讨论开放的Kona吗？

Are we talking about open open Kona?

我们是在讨论Kona四核吗？

Are we talking about Kona quads?

所以我说的是在Kona上以及可能在LLM上的智能体层。

So I'm I'm talking about agentic layer on Kona and maybe on the LLMs.

因此，我们实际上可以设置一个不变量，使系统能够以某种方式运行，从而保护其自身的资源，也就是金钱和时间。

So we can actually set up an invariant potentially so we can, like, function in a way to preserve its own resources, which is money and time in this case.

是的。

Yeah.

所以你可以相当有创意。

So you can be pretty creative.

你知道，关于MoltBook、Open Claw，不管这周是哪一个，我想说的是。

You know, the thing I'll say about that about about MoltBook, Open Claw, whichever one it is this week.

Moldt Book是社交平台。

Moldt Book is the social

AI的社交网络。

network of the AI of the AI.

稍微有点吧。

With it a little bit.

每个提示的令牌负载都多得离谱。

The token load on every single prompt is obscene.

比如，还没打一个字，就已经有16000个了，类似这样。

Like like, 16,000 before you ever type a word, something like that.

这非常耗资源，而且如果不在本地运行，成本会很高。

It's very, it it's intensive and expensive if you're not running something local.

所以我认为，这种技术在代理式应用场景中其实有很大的潜力。

So I think, you know, there's there could be a really cool opportunity for this type of technology in an agentic setting

就像那样。

like that.

当然了。

Oh, definitely.

但当我看到那些大科技公司如此慷慨时，我觉得简直不可思议。

But we are like, when I see how generous those big tech companies are, it's, like, amazing to me.

我现在不会只是为了让人玩玩就这么做。

I wouldn't do it for just people to play with it for now.

对吧？

Right?

我只会，嗯，做企业对企业（B2B）。

I would just, like, we b to b.

我们把它卖给企业。

We sell it to businesses.

我们与关键任务行业交流，他们说：我需要一个专门为我服务的模型，来控制我硬件的这部分功能。

We talk to mission critical industry and they're like, hey, I need a model just for me, which is gonna control like this part of my hardware.

我们会实际获取数据，然后训练模型，根据模型的规模，这个过程需要几天时间，之后就可以投入使用了。

And we actually take the data and then we train the model, which takes a couple of days given the size of the model, and then it's good to go.

在这种情况下，我们清楚地知道资源用在了哪里，能够掌控这些资源，人们也能掌控自己的安全数据等等。

So in this case, we, like, know exactly what the resources are spent for, and we sort of can control this resource, and people can control their security data and so on.

所以这使得整个模式既简洁又高效，而且由于模型体积小，我们的成本也不高。

So that's what makes it, like, nice, clean, but also, like, itself, it doesn't cost us much because the models are small.

所以，是的，这些商业模式非常出色。

So, yeah, the the business models are quite amazing.

我知道这些大科技公司是怎么做的，比如生成我的照片，或者生成视频供人娱乐。

And I know, like, how these big tech companies, like, hey, generate the picture of me and, you know, generate the video, play with it.

我就想，这有什么意义呢？

And I'm like, what's the point?

然后他们还亏了一大笔钱。

And then they, like, losing a bunch of money.

所以我不确定。

So I don't know.

我觉得这全都很好。

It's I think it's it's all great.

我们刚刚开始把大语言模型作为第一个历史性的AI。

We're just starting LLM as the first historical AI.

而且，你知道，这是第一个商业模式。

And, you know, this is the first business model.

人们最初只是想，好吧，把它交给人们，然后他们可以随意使用。

People just initially thought, okay, give it to people and then they can do with this whatever.

但现在我们进入了真正可以学习的阶段。

But now we are at the stage when we can actually learn.

我们从这些商业模式的优缺点中学习，尝试借鉴并改进它。

We learn from all the pros and cons of these business models and try to adopt it and make it something better.

目前这个领域的科学进展真的超乎想象，而且围绕已发生事情的研究数量也非常庞大。

The amount of science right now is really otherworldly in this space and and kind of the amount of research being done around what's already happening.

而最令人着迷的一点在于，事实上，如今在这个领域所进行的科学研究，与十年前相比，简直是天壤之别。

And, that's that's probably the most fascinating thing is the truth of it is is the amount of science being done now compared to ten years ago on this subject is is night and day.

那么，十年后或二十年后，这又意味着什么？

What does that mean for ten years from now or twenty?

是的。

Yeah.

对我来说，最令人兴奋的部分是，似乎人类经历了一个进化过程，接着是我们人类推动了人工智能的进化，而终将有一个时刻，我们会与之共同进化。

This exciting part to me is like it it seems that there was an evolution just for us for people, then there was us people pushing the evolution of AI, and there's gonna be a moment when we sort of co evolve together.

也许二十年后，我们会生活在一个完全无法想象的世界里，比如：‘没有AI的时候，我到底是怎么做到这些的？’

Like, it's gonna be a world when maybe in twenty years we couldn't even imagine, like, oh, how many did I do this without AI?

就像现在，孩子们写博士论文时，可能会想：‘以前的人是怎么靠手工完成的？’

Like, right now, kids probably writing the PhD thesis, and they're like, oh, how people do it by hand?

我以前还是靠手工做的，但现在你可以直接交给EPT之类的工具，它就能做得很好。

I was still doing by hand, but now it's like you could just put it in charge EPT or whatever, and it it did great.

你认为总体来看这是好事，还是认为这会损害人们的的学习过程？

Do you think that's a good thing on net, or do you think that that's gonna hurt people's, you know, learning process?

当然会损害学习过程。

Oh, definitely gonna hurt the learning process.

好的。

Okay.

我有孩子，我知道这种情况已经普遍存在了。

I have kids and, you know, I see it's already out there.

这很有趣。

It's it's interesting.

我很想知道整个教育体系将会发生什么变化，人们将如何应对这一点。

I'm curious what's gonna happen with educational system in general and how people are gonna navigate this.

我们现在在教育方面确实处于一个奇怪的境地，不是吗？

We're kind of in this weird spot right now with education, aren't we?

是的。

Yeah.

我想问一下，回到商业应用场景，因为我觉得你所提供的东西真的很棒。

I do wanna ask you going back to the business use case because I think what you're what you're offering is really cool.

如果有人想使用Kona模型或与您合作创建自己的模型，他们该如何操作呢？

How could someone work with a Kona model or or create work with you to create their own model if if they so wanted to do that?

你是说，如果有人想用Kona创建自己的模型？

You mean if somebody wants to create their own model with Kona?

对，就是这样。

You like Yeah.

要合作的话

To work

基本上就是与你们合作。

with basically, to work with you.

因为我知道你们有Kona，还有Alif，对吧？那是一个解决方案。

Because I know you have Kona, and then you also have Alif, I believe, the is the solution.

Alif只是专门用于需要修改部分的智能代理层。

Alif is just the agentic layer specifically for the parts where you need for modification.

所以我们用它来做cogen。

So we use it for cogen.

这原本是我们内部的一个工具，后来对外发布了，我们在铂金数据上做了测试，人们看到后惊呼：天哪，它真的能解决铂金问题，于是我们就稍微公开了一点。

Like, it was our internal tool which just got out and we tested on platinum and people saw, like, oh my god, it actually solves platinum and we, like, made it a little bit public.

是的。

Yeah.

起初，它根本没打算对外，因为还不足以胜任实际的代码生成场景。

At first, it was, like, never meant to be the public because it was not, like, good to go for the actual code generation use case.

明白了。

Got it.

是的。

Yeah.

在把它开放给不认识的用户之前，对我们来说这将是一段探索之旅。

So it's gonna be a journey for us to see, like, before we give it to just people we don't know.

我们想先了解它的边界在哪里，它是如何扩展的，有哪些问题，擅长什么，不擅长什么。

Let's say, we want to understand where the boundaries of this, like, how it's scaling, what's the issue is, what it's good at, what it's bad at.

这纯粹是为了安全考虑，你不想随便搞出个东西，然后放出去让别人胡乱使用。

And it it's just for the sake of safety because you don't wanna, like, create something wild and then put out there in public and people start doing crazy things with it.

所以我们只是想了解、熟悉它，并弄清楚什么是安全的。

So we just wanna, like, understand, know it, and figure out what's safe.

然后我们才能把它发布出去，也许可以附上API，让人们可以在其基础上进行开发。

And then we can put it out there and maybe attach API to it, and then people could start building on top of it.

但我们还不到这个阶段。

But we're not we're not at this stage.

是的。

Yes.

这很合理。

That's fair.

这完全合理。

That's totally fair.

是的。

Yeah.

确实如此。

It is.

确实是。

It is.

我想知道具体来说，你提到过能量模型已经存在很久了。

I guess I was wondering specifically, so you mentioned that there's energy models have been around for a while.

你说在某些形式上，可能已经有四十年了。

You said in some forms, maybe four four decades old in some way.

到底是什么让推理能量模型如今成为可能？

What what in particular made the reasoning energy model cape like, capable of happening now

嗯。

Mhmm.

而以前却不行呢？

Versus before?

看起来是你的想法把这一切串联起来了。

It seems like it's your ideas were what put it together.

但还有什么其他因素让你觉得，嘿。

But what's what what what else led you to say, like, hey.

我们现在实际上可以做到以前做不到的事情。

We could actually do this now when we couldn't have done it before.

是的。

Yeah.

我认为，对我个人而言，缺失的部分是人工智能的延迟空间概念。

I think that the to me, personally, the missing part was missing is the idea of the latency space for AI.

而延迟空间就像是你大脑中的延迟空间。

And latency space is something like, literally, like your brain latency space.

它是你大脑中某个部分，能让你把任务一直保留在潜意识里。

It's some some part of your brain which keeps the task on the on the back of your mind.

而且，至少据我所知，以及根据其他人的认知，因为直到我们出现之前，基于能量模型的推理并不存在。

And you just like, at least to my knowledge, and apparently to other people's knowledge since the EBMs for reasoning were not there until us.

至少我们之前没看到多少真正有效的成果。

At least we haven't seen much, like, which is actually working.

我觉得，延迟变量才是关键所在。

I felt like the latency variables is what was crucial.

而且，你们训练它的方式也不同。

And also, like, the way you train it was different.

而且，我们在训练过程中让整个过程变得透明了。

And also, we sort of the training process, we made it transparent.

模型具备自我对齐的能力。

It's there is ability to self align for the model.

因此，这在某些使用场景下类似于自我监督。

So almost like a self supervision for certain use cases.

所以，这就像多个部分需要以特定方式协同设计。

So it's like multiple pieces which need to be engineered together in a certain way.

此外，关于能量景观的导航算法，源自一个我本人在博士期间研究并热衷的完全意想不到的领域。

And also navigation algorithm on the energy landscape came from completely unexpected area I personally was working and passionate was passionate about during my PhD years.

所以

So

太棒了。

Excellent.

那么，Eve，人们可以去哪里了解更多关于你们在Logical Intelligence所做工作的事情呢？

Well, Eve, where could people go to learn more about what you all are doing over at Logical Intelligence?

这真是非常有趣的内容，我相信我们的观众会有很多问题。

This is really fascinating stuff, and I am certain our viewers will have questions.

是的。

Yeah.

所以我们差不多一周前刚成立了公司，我收到了超过2000封邮件和LinkedIn消息，这很好。

So we launched the company basically a week ago or something, and I received, like, over 2,000 emails and messages on LinkedIn, which is good.

我始终心怀感激。

Like, I'm always grateful.

看到人们充满好奇总是很有趣，这说明我们做对了些什么。

And it's always, like, fun to see that people are curious, so it means, like, we're doing something right.

我们正在与多位教授和领域专家合作，尝试撰写一些教育材料并发布到我们的网站上。

And we're partnering with a bunch of professors and fields in experts, so we can try to write some education materials and put it on our website.

我们还有一个小型科学团队，正在撰写关于LLM和各类EBM背后基础数学原理的内容，并试图发表论文。

We also have like a small science team who just writing about foundational, like mathematics behind LLMs, behind EBMs of different kind, and they're trying to publish papers.

所以你看，我觉得论文可能是最好的起点。

So see, I I feel like maybe papers would be the best place as a start.

但同样，这真的很棒。

But again, like, is amazing.

他正在尝试许多不同的技术，但还有更多即将推出。

He's doing, like, lots of different techniques, but there's also a lot more coming up.

所以，如果你持续关注的话。

So I guess if you just stay tuned.

当然了。

Heck yeah.

是的。

Yeah.

比如Amylabs在生产新材料，而我们也在生产新材料。

For, like, amylabs producing new materials and us producing new materials.

所以，我们实际上是在开创一个全新的领域。

So, like, we we build in a new field, like, literally.

不是EBM本身，而是EBM的推理部分是AI领域的一个新领域。

Not not the EBM part itself, but the EBM reasoning part is a new field in AI space.

我想我们会继续与像你这样了不起的人交流，向公众进一步解释，并在网上传播，希望有所帮助。

And I guess we're just gonna speak to amazing people like you and, you know, try to explain more for public and put it out there in the Internet and hopefully it helps.

这就是创造力如何融合在一起的例子。

Is the example of, like, how the creativity comes together.

你在学习一些完全不相关的东西，然后突然之间，啪的一下就明白了。

You're learning something which completely unrelated, and then all of a sudden it just like boom.

有道理。

Makes sense.

这里面当然有很多运气成分，因为人们知道这些碎片，但把它们组合在一起并不容易，没错。

So, obviously, a lot of luck in here because, you know, people know the pieces, but bringing pieces together is not easy and Right.

花了一些时间。

Took some time.

这太令人着迷了。

That's fascinating.

有时候，正是具备多元知识背景的合适人选才能出现，并提出那个关键的想法。

And, you know, sometimes it takes the right person with that diverse knowledge set to come through and have the idea that needed to happen.

当然。

Well, definitely.

没错。

Definitely.

我们非常期待见证你们的成长，并看看这一切将走向何方。

Well, we're super excited to watch watch you all grow and see see where this goes.

如果有机会，格兰特和我真想和你彻夜畅谈。

Grant and I would nerd out for days with you if given the opportunity.

简直有上百个问题想问你。

Sounds So like a 100 more questions.

是的。

Yeah.

是的。

Yeah.