Ali Ghodsi：别再吓唬人，真正的风险是网络攻击

Ali Ghodsi

作为企业领导者，这里有公地悲剧。如果你想停、想放慢——你自己为什么不先放慢？我在竞争，我想赢。

As a business leader, there’s a tragedy of the commons. If you want to stop, if you want to go slower—why don’t you go slower? I’m competing. I want to win.

Martin Casado

几乎分成两派：一派相信这其实是工程问题；另一派认为必须把速度降下来。

There’s almost two camps. There’s one camp which believes that this actually is an engineering problem, and there’s others which actually believe you have to slow it down.

Ali Ghodsi

人类对正在发生的攻击反应不够快，你必须把这一切自动化——大多数组织其实离做到还很远。实验室在做的 RSI、递归自我改进，会不会把我们带到那里？这是大问题。

Humans don’t respond fast enough to the attacks that are happening. You need to automate all of those. And most organizations are actually not close to doing that. Is RSI and recursive self-improvement that the labs are doing leading us there? That’s the big question.

Sarah Wang

Elon 说过一类话：这简直是某种精巧的四维象棋——一边说全人类都会死，另一边问你想要多少 IPO 配额。

Something Elon said—this is some elaborate 4D chess, because on the one hand you’re saying all of humanity will die. On the other hand, you’re saying, “Hey, what do you want for your IPO allocation?”

Martin Casado

我第一次在规模化公司里听到：上周有人说他们要从前沿模型迁到 GLM。你觉得这是趋势，还是一次性轶事？

For the first time ever, a company at scale last week said that they’re moving from the frontier models to GLM. Do you think that that’s a trend or do you think that’s just like a one-off anecdote?

Martin Casado

谢谢你来，Ali。我们当然想谈 Databricks，但眼下有一场更广的对话：Dario 发了声，其他人发了声，Elon 也发了声。我们想听 Ali Ghodsi 怎么看——若把话题粗称为「前沿节奏」（pacing the frontier）——你最同意什么、不同意什么，哪里有被漏掉的 nuance？

Thank you for being here, Ali. We obviously want to get to Databricks, but there is a broader conversation going on right now about AI—Dario’s weighed in, others have weighed in, Elon’s weighed in—but we want to hear what Ali Ghodsi thinks. If you called the topic broadly speaking pacing the frontier: what is your strongest agreement with what’s being out there? Where do you disagree? And maybe where is there nuance that’s not being captured?

Ali Ghodsi

乐意谈。我和 Martin 经常吵，肯定很快就会吵起来。

Yeah, happy to cover it. Me and Martin argue a lot. So I’m sure that’s not gonna take long.

Martin Casado

这次我们尽量收一收。

We’ll try and rein it in this time.

Ali Ghodsi

尽量冷静。首先——也许我们同意——领导者有责任：除非有非常非常充分的理由，否则不要无必要地吓人。社会上不同人处在不同心理位置；拿存在性风险、全人类被抹掉这类场景出来谈，我认为是不负责任的——可能把很多人推过边缘，造成大量心理健康问题。

Try to stay calm. But I do think first and foremost—maybe we agree on this—that leaders have responsibility to not freak people out unnecessarily unless there’s really, really good reason. There’s always different people in society that are at different places in their mind space. So talking about these kind of existential risks and scenarios where all of humanity is going to be wiped out, I think is irresponsible—it can tip a lot of people over and it can cause a lot of mental health issues.

Martin Casado

——除非你真有会把人抹掉的东西。

Unless you have something that’s going to wipe people out.

Ali Ghodsi

对，如果真有实际理由，另说。但我认为现在存在性风险接近零。为什么要让所有人恐慌？没必要。风险是有的，我们会谈到——那大概是我们分歧所在。但首先，领导者不该把所有人吓疯。技术研究上的细微差别，研究者可以内部讨论；不必每次上电视或对着 Twitter 上百万人喊：嘿，我认为有比如 10% 的风险全人类会被抹掉。这对很多人没有帮助，反而伤害很多并不处在这些 nuance 里的人。我不认为该这么做，没有建设性，帮不了任何人。

Yeah, as I said—if there is an actual reason for it, then that’s a different story. But I think that right now the existential risk is close to zero. So why freak everybody out? It’s not actually needed. There are risks. We’ll get into it. That’s probably where we disagree. But first and foremost, I think that leaders should not freak everyone out. If there’s technical nuances in how we’re doing AI research—researchers can discuss that. You don’t need to every time go on TV or blast on Twitter to millions of people that, hey, I think there’s like this percentage—10% risk that all humanity is going to be wiped out. I don’t think that’s helpful for a lot of people. Actually I think it causes a lot of harm for a lot of folks who get stressed out and are not in the nuances of all of this stuff. So that I don’t think we should do. I don’t think it’s fruitful. It doesn’t really help anyone.

Martin Casado

对一般公众，这非常非常对。我妹妹——很棒的人，亚利桑那州乡下当老师——周日给我发短信：要不要给你准备好小屋？她本来就有点 prepper：水都备好了，你什么时候现身？我说：等等。我们还没到那一步。显然这已经溢到大众，我认为没必要，也有反噬。

I think this is very, very true for the general public. Like my sister—who’s great, who’s a school teacher in rural Arizona—on Sunday texted me and said, should I prepare the cabin for you? She’s kind of a prepper anyways, but should I prepare the cabin for the AI apocalypse? I’ve got water set up. Like, when are you showing up? I’m like, hold on. Like, we’re not there yet. So clearly this has kind of spilled over the populace, which I agree is unnecessary and has blowback.

Martin Casado

第二点：我刚进来时刷 X，Elizabeth Warren 刚谈要暂停所有 AI 开发——当然是踩着 Bernie 的尾巴，Bernie 还在跟 Steve Bannon 之类合作。所以除了吓人，联邦机器也开始转起来了；我认为这可能与信息本身的目标背道而驰。利害不只是公众歇斯底里。

There’s a second one which is—I don’t know if you saw, walking in here I was checking X and Elizabeth Warren just talked about pausing all of AI development. That of course is on the coattails of Bernie who is also working with Steve Bannon. So in addition to just scaring people, the federal complex is now spinning up and I think that could be actually quite contrary to the actual goals of the message. There’s more than just public hysteria at stake here.

Ali Ghodsi

政治很多，但我两边的圈子都在——两边都有很重的政治。

Yeah, there’s a lot of politics going on, but I’m in all of these groups and I see both sides. There’s heavy politics happening on both sides, we should say.

Martin Casado

对，两边都在发生。我认为不包括 Trump 本人的两党，某种程度上都同意 AI 应被约束。

Oh yeah, this is happening both sides. I think both parties that do not include Trump himself agree that AI should be constrained at some level.

Ali Ghodsi

我说的也包括争论的另一边。举例：甚至 Greg Abbott 也会说德州不能有数据中心——但我不是在谈政客。商业这边也有政治：想要漂亮 IPO、想收回投资的人，会说别搞砸我的 IPO，大家闭嘴让我们把钱拿回来。他们有资源、会拉线。另一边则是：怎么把这条推文武器化、怎么带节奏。

I’m talking about the other side of this argument as well. Even Greg Abbott was like, you can’t have data centers in Texas—well, I’m not talking about politicians. There’s politics going on on both sides. There’s politics on the business side. People who want to see great IPOs and they want to get returns on their investments—and they’re like, don’t mess up my IPO—and they want everybody just shut up so that we can get our money back. They have resources and they’re using them. On the other side, there’s all the people that like, okay, how do we weaponize this? This guy tweeted this—let’s weaponize this one, let’s plant this, let’s pump these threads.

Martin Casado

说说大家共用的那个杠杆点——我认为这是经典公关失误，不只是末日阴郁那一类。行业自我监管并不罕见，也说得通：安全与安保重要，每个 tech epoch 都重要，我们要有某种监督——非常sensible。问题是它被包进「节奏／pacing」这个词。Pacing 有一堆问题：首先它与安全安保正交——你可以慢慢造武器，那还是造武器。人们不觉得真诚，因为这些公司一直在死跑，还在买更多算力跑得更快。它感觉像对「暂停派」有点无趣的妥协：你说 pause，我说 pacing，几乎像 pause 又不是 pause。他们挑了这面旗。但你读没读 Dario 那份文件？内容其实完全sensible——只是跟 pacing 几乎没关系。

Let’s talk to the specific leverage point that everybody is using, because I actually think that this is like a classic case of a PR misstep and it’s not just the doomy gloomy type stuff. It is not unusual for industries to try and regulate themselves. Saying security and safety is important—it is with every tech epoch—and we want to have some oversight. That was very, very sensible. The problem is it’s just couched in this notion of pacing, and there’s a number of issues with pacing. First off, it’s orthogonal to safety and security. Like you can slowly build a weapon. That’s not different than building a weapon. People don’t feel it’s genuine because these companies have been at a dead run—they’re still buying more compute to be even faster. It kind of feels like this kind of almost milquetoast capitulation to the pause people. You say pause, well I say pacing, which is almost like pause, but it’s not like pause. They chose this kind of flag to follow around pacing. But if you actually read the document that Dario wrote—it’s a totally sensible doc. It just has nothing to do with pacing.

Ali Ghodsi

我读了。他确实提到。我有点不同意。这里有公地悲剧：如果你想停、想放慢，为什么不自己放慢？为什么写文章？很多人这么抬杠。但作为企业领导者我理解公地悲剧——我在竞争，我想赢；市场均衡也暗示 pacing 多半不切实际。所以有人说「如果你们不自己停，公地悲剧会继续；我不会单边停，那我是傻瓜，你先停」——然后他们说：能不能进来阻止我们？另一面，看 Hugging Face／OpenAI 事件——顺便说这些公司很棒，大概也投了很多资源——但读发生了什么就很清楚：他们并没有监控每一个吐出的 token；他们是在跑这些 RL 实验，事后再来查发生了什么。在那件事上，他们本该更慢、更有节奏。

Yeah, I read it. No, he does mention it. I kind of a little bit disagree. There’s a tragedy of the commons. If you want to stop, if you want to go slower, why don’t you go slower? Why do you write articles? There’s a lot of people making that argument. But as a business leader, I understand that there’s a tragedy of the commons. I’m competing. I want to win. And you’re also at the market equilibrium, which suggests that pacing is probably impractical anyways. So it makes kind of sense for people to say, hey, if you guys don’t stop, this tragedy of the commons is going to continue. I’m not going to stop racing because there’s IPOs at stake, there is competition at stake, there’s also some animosity between the people. So I’m not going to stop unilaterally—I’ll be a sucker. Why don’t you stop first? So then they’re saying, hey, can you come in and stop us? But you could also make the argument that if you look at the Hugging Face–OpenAI incident—by the way I think these companies are great and they are probably investing a lot of resources—but it’s very clear from if you read what happened that they weren’t monitoring every token coming out. They were just running these RL experiments and then after the fact coming in and checking what happened. So they should have paced—they should have been much slower in that particular incident.

Martin Casado

我不想只抠句法，但公关里用词很重要。读 Hugging Face 时我的反应不是「OpenAI 该 pacing」，而是：把你的东西做安全控制——我们一直在做的那种。

I just don’t want to quibble on syntax, but words matter with PR. When I read the Hugging Face incident, my reaction was not, oh OpenAI should pace—it’s like, dude, secure your thing. Do security controls like we always have done.

Ali Ghodsi

那就是 pacing。你看：我在 Databricks 有法务、有安全，他们整天说什么都放慢——不是 AI，是每一件小事：你要上播客？脚本是什么？你能说什么不能说？一切必须实质真实。跑 RL、训下一模型时，安全团队要不要坐着开所有监控、看数百万 GPU 小时吐出的 token、沙箱里 mock 跑的智能体？若安全团队盯着看，会显著拖慢。他们会说：我们若这么做会变慢，对面未必这么做——能不能请你们进来放慢我们，设护栏，我们乐意按规则做安全的事；否则没意义，我们会被甩开。

It is pacing though. Look, I face this all the time. I have a legal department at Databricks. I have a security department at Databricks and they’re always like, hey, slow everything down—for everything, not AI—literally every little thing. You’re going to go on a podcast—well what’s the script for it, what are you gonna say, let’s review that, you cannot say this, you can say that. Everything you say has to be materially true. So you’re running an RL experiment, you’re training the next model—should the security team be there and run all their monitors and look at everything? Millions of hours of GPU hours of tokens were produced and these agents were running a mock in the sandboxes. It would have slowed them down significantly if we had security team sit there and look at all the stuff. And they’re saying, hey, if we do that it’ll slow us down and I’m not sure the other side is doing that—so can you guys come in and slow us down, put some guardrails around us, we’ll happily then follow the rules and do the secure thing. Otherwise it doesn’t make sense because we’ll get outrun.

Martin Casado

人们真害怕时，二阶 nuance 词不好使。你说「我要 pace」因此「这类事不会发生」——我真觉得本该直说：安全与安保至上，我们要上这些控制。那才重要。看 Zuck 怎么说——

I just think nuance second-order words don’t work when people are really afraid. You’re like, I’m going to pace and therefore these don’t happen. I literally think we should have just been like: safety and security is paramount. We’re going to put in these controls. That’s the important thing. And I do think that nuance actually got lost if you look at what Zuck said.

Ali Ghodsi

我觉得 Zuck 像：我们会自我节奏，我们会加安全——我们之所以晚发布是因为安全。

I thought Zuck did like, hey we’re going to pace ourselves. We’re going to put in security—like the reason we released this later is because of security.

Martin Casado

Zuck 好就好在他很聚焦安全、安保与自我监管。Dario 开篇差不多就是「我们需要 pace the frontier」——把你放进完全不同的心态，而他本可以说「我们需要 secure the frontier」。

What I thought was so great about Zuck is he was very focused on security, safety and self-regulation. Dario’s first five words or whatever are like we need to pace the frontier. It just puts you in a very different mindset than what he could have said is we need to secure the frontier.

Ali Ghodsi

好，我们需要安全的前沿。某种程度上他们想同时讨好喊暂停的 doomer 和政客，结果两边都不满意。但实验室里真有人在慌，很多安全的人是真心慌——不全是 EA 那一路。人们是真的吃惊。

Fine—we need safety at the frontier. At some level I think they were trying to optimize both for the doomers which call for pause and for politicians and they kind of didn’t satisfy either. But those people are actually freaking out inside the labs and there are a lot of safety people that are freaking out genuinely by the way—and not all of them are EA people and so on. People are like, hey, they’re surprised.

Martin Casado

但用「pace」这个词帮不了任何一边。Pacing 像 uncanny valley：doomer 不高兴——那不是 pause，只是 pacing；其他人觉得你反正不会真做、也不聚焦安全。所以不论我们该做什么，呈现方式就是糟，所以才有反噬。

But using the word pace doesn’t help either of them. Pace is like the uncanny valley of making the doomer people unhappy and the policy people unhappy—because the doomer people are like that’s not a pause, this is pacing. And everybody else is like this isn’t real, you’re not going to do it anyways and you’re not focused on security. So again independent of what we should do—which we should talk about—I just think that the way it was presented was just bad and it just didn’t work and that’s why we’re having the blowback.

Ali Ghodsi

这些人不是训练有素的公关。有些我同意。我同意核心前提：不该惊吓公众；我认为现在存在性风险接近零。但核心是：任何人做大规模强化学习、给奖励函数——比如一万个智能体、一亿美元，并行丢上巨型集群跑一两个月，试图解决任何东西——不必是安全题，也可以是解数学谜题——真的可能出很坏的事，意思是东西被黑掉，网络是首要的。那是真实的。把它做成「这是存在性风险」我认为是错的，不该那样吓公众。它已经变成不只你妹妹、全球都在谈的话题；从不关心这事、觉得极无聊的人也会来问我到底怎么想、开始担心。然后变政治问题——美国有选举，全世界也有选举，别处也不会坐着不动。我们有责任用平衡方式谈、真正暴露风险。

These guys are not trained PR people. And yes, some of the stuff I agree with. I agree with the core premise that we shouldn’t freak the public out. I think that existential risk right now is close to zero. But let’s talk about the core thing: anyone who’s doing big reinforcement learning runs and they’re giving it a reward function—unleashing, saying hey here’s like 10,000 agents and here’s $100 million, let’s put them in parallel and let them run on a gigantic cluster for a month or two, try to solve anything—and it doesn’t need to be a security thing, it could be like solve this math puzzle—really bad things can happen. Really bad things meaning things get hacked, and cyber is the primary one. That is real. Making it, hey this is an existential risk and so on, which I think was a mistake—I think it’s not good to scare the public that way. It’s become something that everyone—not just your sister—everybody around the planet is now talking about. I’ve had all kinds of people that never care about this stuff and they find this extremely boring ping me and say what do you really actually think about this—is it really important—now I’m starting to worry about it. Then it becomes a political issue and we have elections here coming up but there’s elections all around the world. You’re going to see they’re not going to sit still in other parts of the world either. But I think that’s our responsibility to talk about this in a balanced way and actually expose the risks.

Ali Ghodsi

我认为那本书里的「超级智能」想法非常非常远。我看不到我们在朝那走、或那会发生的证据。据说实验室里有人慌，觉得也许在朝那进步——我认为来自 RSI：模型在自我改进。我想弄清他们看到了什么我们不知道的。有四个判据：若这四件事同时发生，我很想理解。一、下一模型训练需要更少资源、更少 GPU，而且是超线性地少，不是一点点；二、下一模型训练时间也更短；三、模型准确度／智能在上升；四、前三项可以一遍又一遍再来。四件同时，不是任选。若资源恒定，那还好——硬件会不够，它会自己 pace。若只是「软件在写自己」——我们已经在那了：Databricks 大约百分之九十多的软件由 AI 写。最后那几个百分点也由 AI 写重要吗？没那么重要。但若四条件成立，你可能得到加速：下一模型时间减半、资源减半、更聪明，反复——或许更危险。他们若能分享数据、增加透明，会很好。我甚至不确定那是否必然通向超级智能——也许仍会收敛——但可能。

I think that superintelligence—that idea from that book—is very, very far away. I don’t see any evidence that we’re actually marching towards that or that’s going to happen. Apparently some people at the labs are freaked out that maybe there’s progress towards that and I think it comes from RSI—recursive self-improvement, the model is improving themselves. I would love to understand what have they seen—something we don’t know. There’s kind of four criteria—if those four things are happening I would love to understand them. One is: the next model requires less resources, less GPUs to train, and super-linearly, not just a tiny little bit. Second: the next model takes less time to train as well. Third: accuracy of the model—the intelligence—is increasing. And fourth: we can do the former three again and again and again. All of those at the same time—not just any of them. If all four are happening, then you can imagine—because if any of them does not happen, for instance if resources is constant, then that’s okay because we’re going to run out of hardware so then it’ll pace itself. We will not have enough hardware, not enough GPUs. Time the same. So it needs to be that you end up in this situation. If you just mean that the software is writing itself—we’re already there today. Like 90-some percent of the software in Databricks is written by AI. Does it matter if the last few percent is also written by AI? No, it doesn’t matter really that much. But if you’re getting these four conditions then you might get a speed up where the next model takes half amount of time and half the resources and it is more intelligent and you keep doing that—then you might end up in a situation—I don’t even know if that necessarily leads you to superintelligence per se—still could converge—but it could. So then that would be more risky. So it would be nice if they can share all that data and we can shine some light and transparency on that.

Martin Casado

这个拆解很好。我不认为有人真用那当定义。很多人慌的是涌现行为、「它在创造自己」；很多人的定义只是：我都不写代码了、它在写自己。但他们把「我的价值还在吗、对我可怕吗」和「因此就会得到 2014 年 Bostrom 假设的那种超级智能」混在一起。

I actually think it’s a great breakdown that you have. I don’t think anyone is using that as a definition actually. There’s a little bit of people freaking out about emergent behavior—now it’s creating itself—but a lot of people’s definition is just, hey, if I’m not even coding anymore and it’s coding itself. But I think they’re conflating what’s my value and is it scary for me versus hey that then means we’ll get that superintelligence that 2014 theoretically was hypothesized by Bostrom.

Sarah Wang

你关于算力的点，很多 RSI 争论里被漏掉了。就我们能看到的，训出好模型所需的最低算力门槛一直在升——以前大概一亿、十亿，现在大概五十亿量级。训前沿模型大概五十到一百亿。

Your point on compute though is a really good one that’s missed in a lot of arguments on RSI. Because as far as we can tell, the minimum threshold for compute needed to train a good model just keeps going up. Like it was 100 million, a billion, now it’s probably like 5 billion. To train like a frontier model—5 to 10 billion.

Ali Ghodsi

对。Sarah 这点是反对整套叙事的好论据：下一模型——首先这些实验室一年大概只做一两轮这种 run——而且正好是我那四条件的反面：要更多资源、更多人、更脆，还要建巨型数据中心、拿 GPU、搞对网络、工程上容忍每一数量级多塞 GPU 带来的以前不用操心的错误、提高鲁棒性。过程很脆，失败就浪费巨资，所以他们对那轮 run 极其小心，也有多轮搞砸过。所以不是「下一模型更快、更便宜、更聪明、递归改进」——相反：更久、更脆、更多人、更难做成。就 RSI 而言如是；但就网络风险、东西被黑而言，我们必须极其严肃对待。

Right. Exactly. Yeah. Frontier is very expensive. I think Sarah has a great point which is this is a good argument against this whole thing: the next model—first of all there’s only one or two such runs a year that each of these labs do—and they take—it’s the opposite of the four criteria that I mentioned—which is it’s going to take more resources, more humans involved, and it’s even more brittle and they have to build out the data centers. The labs are not necessarily doing that but others have to build the data centers and they have to be gigantic and they have to get the GPUs and they have to get the networking right, they have to do the engineering to make sure that they can tolerate—because every order of magnitude more GPUs you cram in there, you have to now worry about errors that before you didn’t have to worry about. So you have to increase robustness. It’s like a very brittle process and if it fails, you’ve squandered so much money. So they’re very, very careful with that run and there’s been multiple runs that have been botched. So it’s the opposite of that—hey the next model is faster, cheaper, smarter and recursive improve. It’s the opposite. It’s like it’s taking longer and it’s more brittle and it’s more people and it’s harder to pull off. So I do think that that is true with respect to RSI. With respect to actually cyber risks and things getting hacked—we need to take it super seriously.

Martin Casado

我本来等着跟你四条件抬杠，结果觉得很好。给你一个黑盒版：对付这类动态自适应系统，你信数字还是信「自己的眼睛」？我觉得得看数字。若这些公司一边继续增长、一边减少投入的人数和资金，那肯定有事在发生。你可以黑盒观察。但现状都不指向那——他们疯招人。

So I’m going to have like another miss here—I actually love your four criteria. I was like literally just waiting to argue with it, but I actually think this is very good. Let me give you like a black-box—when you’re dealing with these like dynamic adaptive systems, what are you going to believe? Are you going to believe like the numbers or your lying eyes? I think you kind of have to go to the numbers on these ones. What are the numbers to look at? If these companies continue to grow, reduce the number of people and the amount of money that goes into them, then I would say something is definitely happening here. You can actually black-box this and take a look. But none of those indicate—like they’re hiring like crazy.

Ali Ghodsi

那不公平，因为公司未必高效。OpenAI 自己以前做一百万件不同活动，很小的大概十人团队在做 LLM，而 LLM 那块有用。Twitter 以前很多人，现在少很多。

That’s not fair because companies are not necessarily efficient. What if you have—OpenAI itself was doing like a million different activities. A very small team of like 10 people were doing LLMs and the LLM stuff was useful. Twitter—there was a lot of people, now it’s much less people.

Martin Casado

同意，只是另一条试金石。我们可以有两条：你的很好，但要能仪器化；再加黑盒试金石——若 Anthropic 两周后剩 12 人还继续增长、模型发布越来越快，我们大概该注意。那是充分条件，不是必要条件。

I agree it’s just another litmus test. We can have two litmus tests. We have your litmus test which I think is great but then you would actually have to have a way to instrument it. And then we should have the black-box litmus test. If Anthropic in two weeks is 12 people and they continue to grow and they’re putting out models at an increasing rate I think we should probably take notice of that. That’s sufficient criteria but it’s not necessary condition.

Ali Ghodsi

对。也可能他们资源太多，做很多不必做的事，因为有无限钱就招人。真正该看的是：训下一模型的那支团队是否越来越小、做得越来越少、AI 在做、后训练也一样、GPU 越用越少——现状不是那样。

Right. Exactly. But I’m just saying it could be that they have so much resources that they might be doing a lot of other stuff they don’t need to do but they’re doing it just—and they can just hire the people because they have infinite money. So really the people that are training the next model—is that team tiny tiny and it’s actually getting reduced and they’re doing less and less work and just the AI is doing it and the post-training—and then they’re all just using less GPUs. That’s not the case.

Martin Casado

行业以前吵过很多次类似的。记得我们学会真正集群计算机——大型机受内存一致性之类限制，只能做这么大；然后客户机／服务器；然后超级计算机基本就是集群。大约那时互联网起来，GPU 也开始变好。还记得我们曾对 PlayStation 做出口管制，担心萨达姆用它们做仿真？论证很像：这些东西无限强大，我们用来仿真核武器，有存在性风险，该停——结果都没成真。所以很合理的讨论是：这次是否不同？我没有答案。

We’ve had this argument many times as an industry before. I remember when we learned how to really cluster computers because the mainframe was actually kind of limited by things like memory coherence—you can only make it so big—and then we went to client-server and then we started creating supercomputers which were basically just clustered computers. And at some point internet happened—and that was kind of also roughly when GPUs started getting good. And do you remember that we would actually export-control PlayStations because we were worried that Saddam Hussein would use them to do simulation. And the arguments were very similar: these things are getting infinitely powerful, we’re using them to simulate nuclear weapons, this has existential risk—actually they didn’t use those words—but like potential for nuclear weapons, and we should stop it. And none of that came to pass. So I think a very reasonable discussion is: is this time different? Yes or no? I don’t have an answer to that.

Ali Ghodsi

我不够老，记不清——无知是福。我不记得 PlayStation 非法、萨达姆什么的，也许瑞典不在乎美国出口管制。无论怎样，我认为现在是不同量级——AI、开发节奏等。他们在前沿慌。我确实认为网络是我们会看到的最大之一：全球基础设施比你说的萨达姆／Xbox 时代多太多，互联依赖完全不同。有大量不安全的基础设施；若放这些智能体出去，它们会找漏洞、找 exploit、东破西闯。这是真风险。这次没做成那样，但你可以想象它开始 hopping：拿走资源、到别处执行自己，有点像病毒扩散。那是真风险。

Yeah. I mean, look, I think I’m not old enough to remember. So ignorance is bliss. I don’t recall PlayStations being illegal and Saddam Hussein—maybe they didn’t care in Sweden. Sweden doesn’t care about the export controls in the US. Whatever it is, I think it’s at a different scale now with the AI and the pace of development and so on. They are freaking out at the frontier. I do think cyber is actually one of the biggest ones that we’re going to see. There’s just so much infrastructure on the planet—way more than whenever whatever Saddam or Xbox. We’ve just interconnected way more things and they’re dependent and the planet just looks different today from internet tech dependency interconnection than 30 years ago. There’s so much infrastructure that’s insecure. And if you’re going to unleash these agents they’re going to find loopholes, they’re going to find exploits, they’re going to break in here and there. So this is a real risk. And by the way this time it didn’t do that—but you could imagine a scenario also where it starts hopping—like it takes resources and it starts executing itself elsewhere so it kind of spreads like a virus a little bit. That’s a real risk.

Martin Casado

纯好奇，不是抬杠：那你觉得我们为什么还没看到很多？互联网刚出来时，到这个阶段我们已经搞掉大约 10%、瘫痪医院、拿下关键基础设施、蠕虫造成数百亿美元经济损失——而那时建设更少、经济上线更少。AI 这边那么多人想找风险威胁、跑得那么快、砸了那么多钱，我却没看到与早期蠕虫相称的东西。落差从哪来？

So this is pure curiosity. I promise I’m not trying to be a foil here—but why do you think we just haven’t seen very much then? When the internet came out, by this point we had literally taken out 10%, we’d disabled hospitals, we’d taken out critical infrastructure, we’d caused tens of billions of dollars in economic damages from worms. And to your point we had much less buildout—less of the economy was on it. AI has so many people that want to find risks and threats, we’re running so fast, so much money has been poured into it—and I haven’t seen anything commensurate with the early days of worms. What is that disconnect?

Ali Ghodsi

我记得那些日子。我睡得着，不认为现在有存在性风险。但有大量基础设施需要加固。我们有检测市场上的产品 Lakewatch 帮你做检测，这片空间动得极快。以前有 SOC——安全运营中心——的人看入侵、攻击；现在人类跟不上。整个网络安全空间在转向用智能体做全自动检测。若不做，我认为你会开始看到站点挂掉、整系统停一阵，后果不是存在性的，但有经济损害、人受伤等。所以我们必须极快地赛跑去做。人类对攻击反应不够快，必须自动化；大多数组织离做到还很远。银行在做，特别有安全意识的在做，但今天多数行业仍是老派 SOC，每天醒来几百封告警邮件，许多是误报可忽略，有些不是——他们没时间逐条看。你需要自动化威胁狩猎：用智能体自动攻击自己的系统等。还没发生。若只说「这就像早期互联网」，坏事会发生。所以有一场赛跑。

Yep. I do remember those days. Look, I would just say that I am sleeping well at night and I don’t think there’s existential risk right now. I do think there’s a lot of infrastructure that needs to be secured. We have a product in the market in the detection market—Lakewatch—that helps you do detections and the space is just moving so fast. You used to have these SOC teams—security operations center people—that would look at what intrusions are happening, how are we being attacked. And now the humans just can’t keep up. So this whole space—security cyber space—is being transitioned into fully automated using agents for detection. If we don’t do that—I mean now we’re rushing, the industry is rushing to do that super super fast—if we don’t do that, I do think you will start seeing those kind of things like sites going down, whole systems that stop working for a while and there will be consequences—not existential—but economic damage and people getting hurt and so on could happen. So we just have to race very very fast to do all of those things. Humans don’t respond fast enough to the attacks that are happening. You need to automate all of those and most organizations are actually not close to doing that. The banks are doing it. Some of the people that are super security conscious are doing it but most of the industry today is running with old-school security operation centers and people that are waking up every day and there’s like hundreds of emails of detections that have fired. Many of them are just false positives so you can ignore them but some of them are not—they just don’t have time to go through those. You need to identify that. You need to have threat hunting that’s automated where you’re actually attacking your own systems automatically with agents and so on. It hasn’t happened. So I do think if we just say hey this is just like the internet in the early days—bad things are going to happen. So there is a race going on.

Martin Casado

今早一个电话让我很惊讶——我觉我和你挺熟、对 Databricks 也了解不少。一位创始人说他们在做可观测性、智能体威胁检测，用的是 Databricks。我居然不知道你们有这档产品。智能体／AI 可观测／安全这块你们铺到多深？

I was actually very surprised. Earlier this morning I was on a call—I feel like you and I are pretty close, we talk periodically, I feel like I know a fair bit about Databricks. I was on a call this morning where a founder was basically like, yeah listen, we’re doing all of this observability, agent threat detection and we’re using Databricks. I didn’t even know that you had this offering, quite frankly. From an education standpoint—how extensive have you gotten in the agent AI observability security safety thing?

Ali Ghodsi

今年我们在 RSA 跟 Ben Horowitz 还做过分享。问题是：数据与 AI 正在和网络融合，两个市场在坍缩。以前数据／AI——Databricks 这类公司做的——和网络世界分开：一边是数据上跑 AI／机器学习，一边是检测坏人黑你。现在公司内部跑智能体，智能体之间互动，产生大量数据、日志、轨迹、指纹；内部智能体也在做这些，两个世界越来越合并——所有产生的数据都要分析，规模比一两年前高很多个数量级。变化剧烈：2018、2019 年从 CVE 漏洞公开到被武器化大概两三年；到 2022 显著下降，但仍约八九个月——那还行。从 2022 到现在，曲线压到基本以小时计，几乎立刻被武器化。所以必须用数据与 AI 平台方式自动化。我会说这些市场就是会坍到一起。

Yeah, I mean we gave a talk this year at RSA actually with Ben Horowitz. The issue is that data and AI is blending with cyber. These two markets are collapsing. It used to be like okay we have data and AI—the kind of stuff Databricks and these kind of companies used to do—which is you have a bunch of data and you run AI and machine learning and that lives separately—and then you have the cyber world: we want to detect if bad people are trying to hack us. But now on the data and AI side we have agents running internally in the company—people are having agents running and the agents are also doing things with other people’s agents and they’re producing a lot of data, logs, trails, fingerprints that are being left. So now you have internally these agents that are doing that—so these worlds start merging more and more: all the data that’s being produced needs to be analyzed and the scale at which you need to do that is just many many orders of magnitude more than just one or two years ago. Things have changed dramatically. Like 2018–19 the time it would take from a CVE vulnerability being published until you see it actually be weaponized in the industry would be like two, three years. That went down to 2022 significantly but it was still like eight, nine months. So that’s kind of fine—you have eight, nine months from a vulnerability—that was 2022. Now if you look at the curve from 2022 until now, now it’s down to like basically hours. So it’s like down to basically no time—things get immediately weaponized. So you need to just do it in an automated with the data and AI sort of platform approach. So these markets I’m going to argue are just going to collapse actually.

Martin Casado

同意。拉回存在性风险讨论：几乎两派——一派相信这是工程问题，像 Databricks 这类公司能用产品、工程、服务解决，行业把问题解决就行；另一派似乎相信没有工程解，必须减速、用监管，更像核武器。你认为这是工程问题吗？还是还没到能舒服说这句话？不然干嘛买 Databricks，直接丢进国家实验室？大家自己 pace 一点就好——开玩笑。

I agree. To pull back on the existential x-risk discussion—it feels like there’s almost two camps. There’s one camp which believes that this actually is an engineering problem and companies like Databricks can solve it through product and through engineering solutions and through services and so like we just as an industry need to solve that problem. And there’s others which actually believe it seems to me that there is no engineering solution. You have to slow it down, you have to use regulation. It’s more like a nuclear weapon etc. So does this mean you believe it is an engineering problem or are you not quite comfortable saying that yet? Because if it’s not—why would you even buy Databricks, man, let’s just put the stuff in a national lab. Just pace it is the only solution. Just kidding. Everybody just pace themselves a little bit then it’ll be fine.

Sarah Wang

我们学到的是：Martin 真的讨厌「pacing」这个词。我以后再也不跟你用了。

What we’ve learned here is that Martin really hates the word pacing. I will never use that word with you ever again. Okay. Clearly duly noted.

Ali Ghodsi

我们在谈哪个问题？有两个常被混在一起。一是超级智能问题——很多来自 Bostrom 2014《超级智能》。若你读他的定义，有点疯狂：AI 几秒写出有新颖同行评审内容的整篇博士论文、瞬间完成人类千年量级的思考——速度与智能是许多许多数量级之差，问题尺度完全不同。若真有那种东西，我认为那不只是工程问题能解的——会非常存在性。大家对此同意。但那被混进：我们现在有的智能体离那十万八千里，甚至没有通往那条路的东西。可这些智能体有能力，能做人类史上从未做过的事。我认为拐点已经发生：Databricks 有很好的安全研究员，但我以前从不能说「让我们在沙箱里放一万个、干一个月、做一亿美元工资量级的活」——现在可以，按个按钮就能十万个。数学也一样：要解猜想，找相当不错的数学家，但让一万个协作，进展可以极快。这通向网络风险——我认为网络是主要问题。这我认为可用工程解决；我们在做，很多人也在做；仍有风险，但非存在性。超级智能那种——能写新颖博士论文、在十一维物理空间里瞬时直观推理、人类做不到的那种——大问题是：实验室做的 RSI 是否通向那里、多快。他们建议有检查员进来看——我认为是好主意。问题是检查员是谁，因为你可以叠牌。

Is it an engineering problem that can be solved by engineers or is there more to it? Which problem are we talking about? There’s two separate problems that I think are being conflated. There is the superintelligence problem. I think a lot of this comes from like Bostrom 2014 Superintelligence book. And if you look at the definitions—people don’t have clear definitions of what super—if you read his book those definitions are kind of crazy. What he had in mind when he said superintelligence—AIs that write a whole PhD thesis with novel peer-reviewed stuff in a couple seconds and they can do like millennia worth of thought instantaneously—and it’s just many many many orders of magnitude. The scale of the problem is just completely different. If such a thing exists, do I think that’s just an engineering problem to solve? No. I think if such a thing would happen that would be very existential. And that’s what everybody agrees on. So I think that’s being mixed with: now we have agents that are nowhere near that—it’s not even like there’s nothing like that and we don’t have anything towards that path right now. But these agents are capable and you can do something with them that you could never do before in the history of mankind. So I do think an inflection point has happened. Something has changed which is: we have good security researchers at Databricks but I could never say let’s get 10,000 of them in a sandbox for a month and have them do hundred million dollar worth of salary wage work. We can do that now. We just turn on a button and we can get 100,000 of them. Or mathematics—we can say hey we want to solve a conjecture. Okay, let’s get pretty good mathematicians, but let’s have 10,000 of them collaborate, and then you can make very fast progress. So this I think leads to all these cyber risks. I think cyber is the major problem here. This I think you can solve with engineering. And I think we are working on it. Many others are working on it. There’s still risks—they’re not existential. I think we should do it. There is the superintelligence thing—that’s the thing that could write a novel PhD thesis or reason intuitively in 11-dimensional space physics instantaneously without writing anything down—something humans can’t do. Like that kind of superintelligence. The question is: is RSI and recursive self-improvement that the labs are doing leading us there—are we going to get there—and how fast is that going to happen? That’s the big question. And they’ve suggested that hey we should have inspectors that come in and look at what we’re doing. I think it’s a good idea. Have them go in there and get the data. I would love to. The question is who are the inspectors—because you can stack that.

Martin Casado

你觉得谁合适？

Who do you think?

Ali Ghodsi

两边阵营都有人。若检查员是已经站定一边的人，他们说什么我不会太服——进门前就定了结论。但举例：若 Yann LeCun——深度神经网络技术的发明人之一、先驱——进去看完出来说：没什么可看的，没有风险，超级智能那套是胡说，继续快、快、快——我是在转述他的立场——那我会感觉很好。或者他出来说天哪、立场晃了一下，也是很有意思的信号。关键是我们挑谁当检查员；我认为是好主意，挑一群多元的人，拿到不同 nuance 的观点。

There’s a bunch of people that like on either camp. Actually I wouldn’t care if they’re the inspectors—I would not be very impressed by what they say because they’ve already made up their minds even before they would go in there. But let’s say as an example, if Yann LeCun—who was one of the inventors of this deep neural network technology, one of the pioneers—if he said, hey, there’s nothing to see here, there’s no risk—I’m paraphrasing him—this superintelligence, this is just nonsense, keep on going, go fast, fast, fast—none of us would believe it if he hadn’t looked. If he was one of the inspectors and he went in there and he had a look and he came out and he said, hey, I’ve looked and it’s just what I said, there’s nothing to see here, just keep going—I would feel very good about that. Or if he comes out and says oh my god—he’s wobbling and he would change his mind a little bit—that would also have a lot of interesting signals. So I think it comes down to who we pick as inspectors and I think it’s a good idea—let’s have some of them and pick a diverse set of people so that we can get different nuanced points of view.

Martin Casado

Elon 那套呢？感觉有三种方案：OpenAI／Anthropic 是第三方；Elon 据我理解是实验室互相交叉检查，像科学同行评审；Zuck 是自我监管。中间那种你怎么看？

What do you think about this kind of Elon Musk view? It’s less third party. It feels like there’s kind of three proposals: the OpenAI–Anthropic one is a third party. The Elon Musk one, as far as I can tell, is the labs cross-check each other like peer review like you do in science. And then the Mark Zuckerberg one is police yourself. What do you think about this middle one?

Ali Ghodsi

他们互相评价？我觉得像拳台上拳手互相当裁判。会管用吗？不会。他们会一直喊犯规——对方一出厉害模型就喊：啊，超级智能风险冒出来了。有 IPO 计划、两家竞争激烈、还有历史恩怨时，既得利益在场——他们会对彼此非常「公平」？所以才需要第三方。世上为什么有法官？为什么有第三方？不能大家都自己谈拢吗？他们若想试可以试，但我怀疑他们不会不以多种方式自我偏袒地评判对方。

That they should evaluate each other? I think like if we have boxing matches in the ring, the boxers should just be the judges of each other. Would that work? No. They would scream foul all the time. Foul, foul, foul. The moment the other guy puts out the great model and it’s like—ah, big superintelligence risk surfacing. Absolutely—when vested interests are at play and there’s like IPO plans and these two companies are so competitive and they have like this history also between them—yeah they’ll be very fair to each other I’m sure. That’s why you need a third party. Why do we have judges in the world at all? Why do we have third parties at all? Why can’t just people figure things out between themselves? But they should try. If they want to do it, they should try. But I’m skeptical that they wouldn’t just be biased in multiple ways to self-judge each other.

Sarah Wang

必须问：Elon 好像在 All-In 峰会上说过——这是精巧的四维象棋：一边说全人类会死，另一边问你要多少 IPO 配额。更犬儒，但那种失调很多人有感觉。你怎么调和？

So I have to ask, Ali—do you think something Elon also said, I think it was on the All-In summit—he was like, this is some elaborate 4D chess because on the one hand you’re saying all of humanity will die; on the other hand you’re saying hey what do you want for your IPO allocation? That is probably a more cynical view but how do you reconcile that? The dissonance I think gets a lot of people—how do you think that gets reconciled?

Ali Ghodsi

这些东西会搅在一起。有人是真慌；也有人觉得若有监管来「给我们节奏」——抱歉又用这词——对他们有好处。人也有既得利益。人通常能在脑子里把这些调成和谐。过去确实有用营销噱头的倾向：天哪我训的最新模型好到难以置信、几乎吓到我——全世界就开始盯着。有那种营销。但同时，CVE 到武器化 exploit 从数年压到几分钟，就三四年——网络攻击是真的。每次训新模型，围绕「对世界有多疯狂风险」大造声势，也是很好的营销，对你有帮助。所以这些未必矛盾。

Look I think all of these things get mixed. I think there are people that are freaked out and I do think that there are people that saying like hey if there was regulation that would pace us—sorry to use the word—that would be good for us. But I also think that people have vested interests. Usually people figure out a way to always get all of these things to align harmonically in their head. Do I think that there has been a tendency in the past of in general using also marketing stunts by saying oh my god this latest model is so good that I train—it’s like unbelievable, it’s like almost scaring me—and then the whole world kind of starts focusing on it. Yeah there’s been that kind of marketing going on. But at the same time also as I said the time from CVE to actually weaponized exploit has been going down from years down to like minutes now just in like three, four years. So it’s real—the cyber attacks are real. But there’s also a great marketing ploy to whenever you train a new model make lots of noise around how much of a crazy risk it is to the world. It helps you. So maybe they’re not in contradiction these things.

Martin Casado

你我都是网络人，历史上有很多第三方来仲裁——IETF、IEEE 等。我认为他们提第三方其实很sensible，我也同意你说的要尽量独立——现在未必独立，人选会吵个不停。但你说为什么有法官——国家介入和行业自我监管是很不同的事。什么时候该认真考虑联邦介入？还是现在就该考虑真联邦介入，而不是更多行业自律？

You and I are networking folks and there’s a long history of forming third parties to help arbitrate things—like IETF or IEEE. I think it’s actually a very sensible proposal that they actually have. I actually agree with you—you probably want to make sure it’s independent which is not fair right now whatever—and there’s going to be a lot of arguments so who you put and everybody’s going to disagree. But you said why do we have judges—so that’s actually like the state stepping in is actually quite a different thing than basically industry self-policing. So at what point in time do you think it makes sense to actually consider federal involvement? Or do you think now is the time to actually consider actual federal involvement as opposed to more industry self-policing?

Ali Ghodsi

这些很不同，但会渗进彼此。比如 FINRA 不是完全独立的自我监管，它连着政府。历史上行业先自律，再演化成监管。若他们一边说有存在性风险、一边说来管我们、监管我们——监管者很难说不。目前他们说了不，但我认为撑不了太久。

Well, these are very different. They are different, but they kind of bleed into each other. For instance FINRA is not like a completely independent self—it is but it’s linked to the government—so these things will bleed over. Historically they’ve industry self-police and then it evolves into regulation. If they are saying there is existential risk—which they’re saying—and then saying come police us and regulate us, I think it’s very hard for regulators to say no we’re not going to do that. So far they’ve said that but I think that’s not going to last very long.

Martin Casado

David Sacks 好像说过：我从没遇过 CEO 求我们监管他们。我最爱的是——我也从没遇过监管者对那种请求说不。元政治机器其实已经在动：人人有话术，Obama 也出来说话，这会是中期选举重大议题。你觉得是不是太晚了——我们会走向重手联邦监管，前沿进 DOE、整个暂停——还是还能落到sensible的自律式监管？头条已经收不住。

But wasn’t David Sacks like, I’ve never had a CEO ask us to regulate them? And my favorite thing—and the CEO—I’ve never had a regulator that says no to that. The reality is the actual metapolitical machinery is actually in motion already. Everyone has a talking point. Obama has come out. It is a major issue. Do you think that there’s a reality that it’s too late—this will be a major issue in the midterms and we’re actually going to like heavy-handed federal regulation and this is all going to be paused—goes into Anthropic, goes into the DOE—and we’re past that point—or do you think we can actually end up with like a sensible self-policing regulation? Because the headlines you cannot—

Ali Ghodsi

我们仍应争取做对的事。事情怎么演化还有一些自由度，还有时间。你说得对：这些公司我们灌进那么多亿美元；强化学习的方式是给可验证奖励——解这道数学题、这一窄条编程——砸很多钱，能在那条窄带上拿到很好结果，并不等于你在得到超级智能。

We should strive towards doing the right thing. I think there’s still some degrees of freedom of how things evolve and there’s still time. And yeah you’re right that largely you have these companies where we’re pumping in so many billions of dollars. And the way reinforcement learning works is that you give it the reward function that’s verifiable—like we’re going to solve this math problem or this kind of narrow area of programming—and we pour in so much money into that. You can get quite good results in that narrow kind—that doesn’t mean that you’re getting that superintelligence.

Martin Casado

你甚至可能骗自己：更少输入给更好结果——只因为你跑了那么多实验、想了那么多。但投入这么多资源时，很难做封闭实验。融资额天文数字上涨——所以他们要上市。

No, but you can even trick yourself into thinking that like less inputs are giving you a better outcome just because you’re running so many experiments and thought about it so much. But it’s actually very hard to do a closed experiment this way given how many resources are going in. Well, the fundraises are going up astronomically to your point. That’s why these companies are going public, right?

Ali Ghodsi

对。作为跑大规模私有公司的人，我认为他们本来更愿保持私有。为什么上市？因为需要资本，他们把扩展定律与资本视为战略优势。但回到我列的四点：若那四点为真，你想不想知道？那会不会失控？现在没有证据四点在发生；但若真朝那去——

Yes. And that’s why these companies are going public. I think otherwise they would prefer to stay private. Why are they going public? Because they need the capital and they consider the scaling laws and the capital to be a strategic advantage. So that’s why they’re going public. But I would say let’s go back to the four things that I listed. If those are true—if those four are true—would you want to know about it and would that be worrisome that that could get out of hand? Now there’s no evidence that those four are happening—but if there was—like that is actually where it’s headed—

Martin Casado

理解任何系统的自推进属性都很重要——我们过去对动态系统、编译器、纳米技术都这么做。兴趣并不新。怕的是这些系统是元经济、太复杂——风险是你「以为看见了」其实没看见；我认为现在很多就在发生。当然真看见了你也想知道。

Yeah I actually think understanding for any system like any sort of self-propelling property is important and we’ve done this in the past with dynamic systems—compilers, all the research on nanotechnology. I don’t think that that’s new that it’s an interest. The fear is that these particular systems are meta-economic systems that are so complex that the risk is crying that you’re seeing it when you’re not seeing it and I think a lot of that’s happening right now. But of course if you see it—you want to know.

Ali Ghodsi

但公平说，实验室现在大量聚焦 RSI，那是下一步方向。也许他们只是没道理地自己吓自己，就像当年吓 GPT-2——说 GPT-2 会终结世界，结果没有，GPT-3、4 出来了。

But it is fair to say that the labs are now focusing a lot on RSI and that’s where they’re headed next. And maybe they’re just unjustifiably worried themselves just like they were worried about GPT-2—they were like GPT-2 is world-ending and then it wasn’t and GPT-3 and 4 came out.

Martin Casado

很多时候他们说 RSI，其实在说自催化效应——我们行业很久就有：没有芯片造不出芯片；有 CS 学位的人都知道编译器写自己的编译器；蒸汽机也是自催化。我全职工作就是实验室出来的人来创业，人人说 RSI，大概只有 1% 真是 RSI——更多是「我们用 AI 做数据清洗、用 AI 做某某」。

So I don’t want to quibble. A lot of the times when they say RSI, they’re actually talking about autocatalytic effects and autocatalytic effects have been our industry for a very very long time. There’s no way you can create a computer chip without a computer chip. Anyone with a computer science degree—a compiler writes its own compiler. The steam engine was autocatalytic. My full-time job is people coming out of labs and starting companies and they all say RSI because everybody says RSI—and maybe 1% of those are actually RSI. They’re more like we use AI for data cleaning, we use AI for making—

Ali Ghodsi

区分一下：你是说催化剂意义上用 AI 加速？

Let’s make the distinction. So you’re saying it’s basically catalyst in a sense that they’re using AI to speed things up.

Martin Casado

自催化。用模型写 GPU kernel、做数据清洗，就像我用计算机设计计算机。互联网也是自催化，因为它让人远程协作。轶事上大概 90% 的「卡路里」是自催化——完全符合预期，也不是新闻。但现在也确有聚焦：我们能不能让模型训练自己——有点像 Karpathy 做的 auto research，但他们想做那个。有团队在做，远没有你以为的那么多卡路里；我采样还行，因为他们都会来找我们谈。

It’s autocatalytic. Yes. You’re using a model to build a GPU kernel. You’re using a model to do data cleaning just like I use a computer to design a computer. It’s autocatalytic—which is every tech—the internet was autocatalytic because it allowed people to collaborate remotely. This is anecdotal—90% of the calories are autocatalytic which is 100% what you would expect—and that’s been going on for a while though—that’s not even new. But also now there is a focus on let’s move towards actually can we get the model to train itself—this kind of like the auto research that Karpathy did but now they want to do that. There are teams that do that. It is not nearly as many calories as you would expect and I just feel like I actually have a good sampling of this because they all come and talk to us.

Ali Ghodsi

那也许你可以当检查员之一——我们拿到数据大家一起看，也许没什么可看。我个人不认为那四条件发生的概率很高。Brockman 这周上播客说我们已进入 AGI 时代——然后人人都说我们有 AGI。人们会跟风。但我问很久：AI 是否多数时候比你身边多数人更聪明——大概从去年三四季度起他们就这么说。再问：有多少人在管成百上千互相协调、谈判、自动化生活的智能体蜂群？举手几乎没有——当然 Martin 在家大概干过。多数企业在用 Microsoft Copilot，AI 程度就那样。我问的多数企业说没有那些；那你们在干嘛？用聊天机器人问问题——基本是以前 Google 搜索结果的更快、更光荣版。编码在发生，ROI 可以另议，但没有把整个企业自动化的智能体工作——就是没发生。为什么？我认为真实原因是：模型够聪明了，只是没有组织内部存在的上下文——没参加过每次会议、不知道大家脑子里有什么、不知道流程。每个组织总有那么几个什么都懂的员工，你拍他们肩，所有人想：他／她要是走了怎么办。他们没有那种上下文。若把那上下文融进今天的前沿模型，我认为地球上任何组织都能拿到巨大生产力；为此我们其实不需要更聪明的模型——不需要能解 Navier–Stokes 或猜想、或把 Humanity’s Last Exam 从 60 拉到 70 的模型。这些都不需要。有人若「pace the frontier」会很不爽——但前沿不推进，对绝大多数组织其实无所谓：他们在自动化与拿价值的采用曲线上落后太多。但对实验室会是灾难，因为智能价格在渐近下降——大概每六个月降到约十分之一——若不推前沿会剧烈改变他们的生意。

Right—so maybe you can be one of the inspectors—can we get all that data and all of us look at that data—maybe there’s nothing to see here. I personally don’t think it’s like very high probability that those four criteria are happening. Brockman went on the pod this week and said we’re at AGI era—and now everybody’s saying we have AGI. People follow. But for a very long time when I asked this question—said that AI is smarter than most other people around me most of the time—that’s been almost like since Q3 Q4 last year they’ve been saying that. Then I asked them how many of you have hundreds or thousands of agents that you are managing that are coordinating with each other in swarms and negotiating and automating your life and everything around you—and if so raise your hand—it’s like almost nobody raises their hand. Of course Martin has done that at home. Most enterprises are on Microsoft Copilot—like that’s the extent of their AI. Most enterprises I talked to when I asked this question they’re like no we don’t have any of that. So what are you doing then? They’re using a chatbot—asking questions from a chatbot that’s basically very very glorified efficient Google search of the old day results—so it’s just faster Google search. And then coding is happening—people are using it for coding—though the ROI we can discuss—but there’s no agentic work that’s automated the whole enterprise—that has just not happened. So then why is that? I think that the real reason if you actually look at it is the models are smart enough—but they just don’t have the context that exists inside of any organization. They have not been in every meeting. They don’t know what’s in everybody’s heads. They don’t know all the processes. There’s always like a couple of employees who know everything in every organization—you go tap on their shoulder and everybody’s like oh my god what would happen if he or she quits. They don’t have that context. And if you just fused that and gave that context into AI models—just a frontier today—I think there’s so much productivity gains you could get for any organization on the planet. For that, we actually don’t need smarter models. We don’t need a smarter model that can actually solve Navier–Stokes or conjectures or do better on Humanity’s Last Exam—like we needed to just go from 60 to 70%. None of that is needed. So I think actually people are very upset on some like oh if we pace the frontier—but actually if the frontier doesn’t advance it doesn’t actually matter. I think for vast majority of organizations on the planet they’re just so far behind in the adoption curve of actually automating things and getting value out of this stuff. But it would be disastrous to the labs because the price of intelligence is dropping asymptotically. I think it’s going down by one-tenth every six months or something like that. So that would dramatically change their businesses if you weren’t pushing the frontier.

Ali Ghodsi

但我们该聚焦的是：成本收益分析两边。我们谈了很多成本——存在性威胁、网络风险等。收益呢？现在这成了公众话题，公众在问：我得到了什么？好像什么都没有。

Yeah. But this is what we should focus on. We should focus on—there’s like two sides—we discuss here a lot the costs. There’s cost-benefit analysis that we should do on everything. We’ve discussed the costs a lot here—oh is there like existential threat, is there cyber risk, are there things we should be worried about. That’s like the cost side. What’s the benefit? And I think now that this has become like a public thing and the whole public cares about AI, they’re asking, hey, what’s in it for me? What am I getting out of it? Seems nothing.

Sarah Wang

那他们怎么到那一步？你至今见过哪些用例，上行惊喜超过你预期？

So how do they get there? What are some of the use cases you’ve seen to date that have maybe surprised you to the upside?

Ali Ghodsi

大家太担心存在性风险，很多人不知道有多酷的用例。Crisis Text Line：他们和我们一起用大语言模型检测青少年是否可能自我伤害／自杀——真棒，真的在救命。Omnipod：糖尿病患者戴上，用 AI 学习胰岛素释放与血糖，精确释放——以前要自己扎，现在自动，像为身体自学的 AI。Zipline：无人机全自动、AI 驱动，从电池优化到航线，给有需要地区送物资——开始是给难民送血，从非洲到别处——建在 Databricks 上。更前沿的：我们和 Merck 建的基于 transformer 的模型 TEDDY——Transformer Enhanced Drug Discovery——他们发了研究；不是预测英文下一 token，而是预测基因调控网络（GRN）会如何响应，能区分因果细胞与只是反应的细胞，从而用于药物发现、显著压低针对特定疾病的研发成本。Genie：有本体，可问任何问题。Novo Nordisk 在用——他们做了 GLP-1 药，现在用它跑试验：肥胖研究之类拿洞察的时间，从数周压到数分钟。有很多惊人的 AI 用例。我们不该忘这些上行；我们想要这些，也不想「粘贴／放慢」这些。

Yeah. First of all, there’s like so much worry about existential risk and so on. So I think a lot of people just don’t know what cool use cases where people are actually doing interesting things. We have a lot of use cases that are fascinating. One that I like is Crisis Text Line. They actually use large language models with us to detect if teenagers want to do self-harm, suicide. That’s an awesome use case and it actually saves lives. Another one that’s kind of interesting is the Omnipod, which is for diabetes patients. They can put the Omnipod and it uses AI to really learn your insulin release and your glucose levels and actually exactly release. People used to stick themselves—but this now happens automatically and it’s like self-learned AI for your body. Zipline is another one. They’re doing awesome—when they started it was like these drones that were completely automated, all AI-driven everything from the battery optimization to the routes and everything and they were delivering food in areas of need—blood to refugees. Started in Africa and then elsewhere in the world. That’s all AI use case built on Databricks. There’s more advanced ones also—one that I kind of like but it’s harder to maybe explain is this model that we built—transformer-based model that we built with Merck—it’s called TEDDY, Transformer Enhanced Drug Discovery. They published actually the research. It basically—instead of predicting the next token in English—it predicts what the gene regulatory network, the GRN, is going to respond and it can really detect which cells are causal and which ones are just reactive. And therefore they can start using this in drug discovery and get costs down significantly for developing drugs that are targeting specific diseases. There are lots of these. Genie I mentioned—you have this ontology and you can ask any questions. Novo Nordisk is using this. They built this GLP-1 drug—but what Novo is doing is now they’re using it for all of their trials that they’re running. And it can compress the time it takes to get insights versus if you’re doing obesity study or something from weeks down to minutes. So there are a lot of amazing use cases of AI. We should not forget these upsides also—like we want all of these and we do not want to pace these.

Sarah Wang

完全同意。若映射未来 12 个月，企业怎么真正拿到价值？你提了上下文——怎么运营化？

Yeah, exactly. You’re totally right. So how do they—let’s say if you map out the next 12 months—how do the enterprises actually get value? You dropped the word context, but how do they operationalize that?

Ali Ghodsi

比多数人想的更难。首先必须把组织里发生的一切数字化——没有魔杖。每次会议要转录，所有会议上下文、所有数字内容要喂给 AI。我们称之为建本体，但首先得收集。许多组织这本身就是问题：法务会说别录每个电话、别录每件事。你得用某种方式去做——

It’s actually harder than most people believe. First and foremost we have to make sure that we have digitized everything that’s happening in an organization—you cannot just have a magic wand and make that happen. Every meeting has to be transcribed. You have to be able to get all the context of all the meetings and everything that’s happening—all the digital content has to be fed to the AI. So you have to build—we call it an ontology. We build that but first and foremost you have to collect that. That itself is a problem in many organizations because legal teams will say don’t record every call, don’t record every thing. So you have to do that in a way where—

Sarah Wang

能给大家定义一下 ontology 吗？我知道 Palantir 老说这个词，但他们并不拥有这个词。对听众意味着什么？

Can you define ontology for everyone—because I know Palantir says the word a lot but it’s not like they own the word ontology. What does that mean and for the people listening how should—

Ali Ghodsi

Ontology 就是组织里所有抽象概念——目标、部门、人、项目——它们到底是什么意思、彼此什么关系，以及人和资源与公司做什么的关系。好比同样技能、同样教育背景、同样聪明勤奋的两个人：一个今天第一天上班，一个干了五年。差别是什么？后者有「这组织怎么转」的本体：人是谁、事怎么做成——别看组织架构图，别去问那个人，他办不成；去问这个人，他能帮你办成；文件其实不用那样交；这个项目在发生、这个才是关键。大量印在每个人脑子里的知识，知道组织怎么运作。所以创业圈说：若你失去大多数人，公司恢复不了——不是补招就能补上，人太关键。我们怎么拿到那上下文——那就是本体——给 AI？一部分是录音等；第二部分是如何蒸馏成数字图，再喂给 AI。

Yeah. Ontology just means that in an organization the relationship between all the abstract concepts of all the goals and all the departments and all the people and all the projects that are going on—what do they exactly mean and what’s the relationship between them—the people, the resources and what that company does. So it’s the difference between a person who is a new employee in the company and just started today and a person that has worked there five years. Let’s say they’re equally skilled. They have the same educational background. They’re equally smart and hardworking and all of that. But one—her first day today at work, the other one she’s been there 5 years. What’s the difference between these two people? The one has an ontology of how that organization works, who the people are, how you get stuff done. Don’t look at the org chart. Don’t go ask that person—he will not get anything done. You go ask this person—he’ll get it done for you. And that’s not how it works—you don’t need to file that paperwork here. And this is this project, this is what’s going on, this is essential. So there’s just a lot of ingrained knowledge that’s sitting in everybody’s heads who knows how an organization works. That’s why people say in startup land: if you lose most of your people, that company can’t recover from it. You can’t just replenish and hire new people—the people are so essential. How do we get that context—that’s the ontology—and give it to the AI? Part of that is we just have to have the recording and all of that—but the second part is how do you actually distill it down into a graph—actually a digital graph—that you can then feed to the AI.

Ali Ghodsi

今天很多智能体——Claude Code、Codex、pi、OpenCode 这一串——有智能体循环：能推理，但一次查一个资源，去这个 MCP server 看答案在不在、再合成。偏慢。我打比方：若 Google 二十五年前这样建搜索——搜关键词，但不给你十条蓝链，而是去一个网站用 LLM 摘要、找几个枢纽超链、并行跳几个站读几页，干十分钟再给你「最好的十条蓝链」——一则每次上网很贵，二则等十分钟，三则质量差，因为只看了极小子集。他们怎么做？有索引。你从不离开 Google 服务器，打到反向索引，不到 100 毫秒出十条蓝链。AI 也需要同样的事：本体就是我们要离线持续算那个索引——有点像当年 Google 的 PageRank，但更复杂：Google 看的是人人可上的网、链接已在；这里有权限——我能访问的你未必能——有隐私、访问控制，对象类型也不只是网站。问题更难点，但可管理。我确信能做，能拿到巨大生产力，因为我们在 Databricks 自己做了。

So the way a lot of the agents work today—like a Claude Code or any of them, Codex or pi or OpenCode or you can go through the whole slew of them—they have this loop, agentic loop—it can reason but then it goes and checks every resource one at a time. So it’ll go to this MCP server for your question and try to see is the answer here, is there another one—it synthesizes it and gives you an answer—but it’s kind of slow. I liken this to if Google would have built Google search this way 25 years ago—we would have said okay we’re going to get 10 blue links, we search for key terms here, but instead of giving you 10 blue links, it would have gone to one website, summarized with an LLM what it does, found a few hub hyperlinks, jumped in parallel to a few of them, read a few websites, done that for 10 minutes, and then given you like its best 10 blue links it would find. Well that would be very expensive, cost a lot of money to do that every time go on the web. Two, it would have taken a long time—you got to wait 10 minutes. And three, the quality would be bad because you’re actually only looking at a very small subset of everything that exists out there. So how do they do it? They have an index. You never leave Google servers. You search for it—hits the index, the reverse index immediately gets you the 10 blue links within less than 100 milliseconds. We need to do the same thing for the AI. So the ontology is that we need to compute that index offline all the time. It’s almost like the PageRank algorithm that Google had invented back in the day, but it’s more complicated because Google was just looking at a web where everybody can go on the web. Here there are permissions. And the links existed—yeah, here there’s permissions involved. The data I’m allowed to access might not be the data that you’re allowed to access. So there’s privacy, there’s access control. Also there’s many different types of objects here that we’re dealing with, not just websites. So the problem is a little bit harder. But it’s manageable—you can actually do it. I’m convinced you can do this and you can get massive productivity gains out of it because we did it for Databricks. We did it for ourselves.

Sarah Wang

你们一直在自用 Databricks，能多说说组织层面看到的影响吗？

I mean you’ve been dogfooding Databricks for Databricks forever—but maybe say more about the impact you’ve seen as an organization.

Ali Ghodsi

一旦有了本体、开始用——我们大概是所有客户里最大的本体，比任何客户用我们建的都大，因为 Databricks 用 Databricks 比任何人都多——图里有数百万个节点。组织里发生什么？树状结构，信息上下流：做不了决定就升级，老板或许打破平局，再往上需要补上下文再决策；决定做出后要向下渗透。有了本体，很多可由 AI 做。会上有人做分析、PowerPoint、漂亮图；分析是聪明人用 Excel、模型、数字——很多现在 AI 就能做：有全上下文，按你想要的方式呈现，你可直接问 AI 而不开后续会。有点像 Jack Dorsey 说的对组织能做的事，只是一种具体实现。对我们是游戏规则改变者：会上人人盯手机问 Genie；有人一说复杂的东西，你就看见大家拿起手机。

Yeah. Once we got this ontology and we started working on it—and we actually have probably the largest of all of our customers, we have the largest ontology—our ontology is bigger on us than any of our customers when they use us to build their ontology—because Databricks uses Databricks more than anyone else uses Databricks—and so it’s like millions of millions of nodes in the graph in the ontology graph that we have. What happens in an organization: you have a tree structure organization and information flows up and down the tree structure. If you can’t make a decision you escalate—your boss maybe they can tie-break it—escalates up. They need to get up to speed on what’s happening and they need to get all the context and then they make decisions. Once decisions get made, you have to percolate them down in the organization. A lot of this can now be done by AI if you have an ontology. Why? Because what happens in a meeting? In a meeting you go through—someone has done the analysis. They probably have a PowerPoint deck with some pretty graphs in it. That person that did the analysis is some smart person that used Excel, made some models, there’s some numericals. So a lot of that you can now just do with AI. So the AI can do the analysis for you. It has all the context. It can present it in a way that you want. You can ask questions about it instead of having follow-up meetings. You can directly ask questions directly from the AI. So it’s very similar—it’s along the lines of what Jack Dorsey has said that you can do to the organization. It’s just a concrete way of implementing it. So it’s a game changer for us. Everybody’s on their phones now in the meetings on Genie and they’re like asking Genie questions. You can see—soon as someone says something complicated—you see everybody go to the phone.

Sarah Wang

能分享你提过的那个财务轶事吗——董事会上那次？

Can you share that finance—the finance anecdote you mentioned once in a board meeting.

Ali Ghodsi

可以，内部董事会相关。我要做一个 presentation，需要知道我们有多少 Fortune 500 客户、渗透率多少。我问销售运营的人，以为她有数；她回短信说抱歉上不了 Genie，在飞机上。我说如果你只是要登 Genie，我自己也能登——我问你是以为你有我没有的别的东西。我有点生气，就改问 CFO Dave：Fortune 500 渗透率多少？他直接复制粘贴一张 Genie 截图回来。他也是问的。我说：还有人做任何新颖的事吗？还是人人都去 Genie、问本体？现在大家就说：谁来 genie 一下、从本体里拿。我认为是游戏规则改变者。但不是按个按钮组织就有本体。Palantir 其实做得很好：走进组织，把大量默会知识写下来放进组织。我们自动拿那些建图，再把图喂给智能体，用业务领导想看的方式回答——图、分析、可追问——然后把信息散到组织里。

Yeah sure—internal board meeting. Only if kosher. Yeah exactly. No—it’s actually needed for one of our presentations. I need to know how many customers do we have in Fortune 500—what’s our penetration of Fortune 500—and I asked one of the people in sales ops because I thought she would have it and she texted me back and said oh sorry I can’t log in to Genie right now I’m on a flight. And I said if you’re just going to log into Genie I can do that myself—I asked you because I thought you had like something alternative that I don’t have access to. So then I was kind of a little bit angry so I texted the CFO instead—Dave—and I said hey do you know what our Fortune 500 penetration—and he just copy-pasted a screenshot of Genie back. So he also asked that. So I said does anyone do anything novel here or there? Just everybody just going to Genie and asking the ontology for questions. It’s like let me Genie that for you instead of let me Google. That’s what everybody does now. We just say it. We say hey can someone just Genie this—can you just get it from the ontology. So I do think it’s a game changer. But it’s not just you press a button you have an ontology in an organization. And I think Palantir actually has done a great job of going to organizations and getting a lot of that tacit knowledge written down and getting it into the organizations. We automatically take that and build the graph and then we feed that graph into the agents so that we can answer the question and answer it in a way that business leaders would like to see it—which is in graphs, analytical way and a way where you can interrogate that question and continue asking questions and getting answers to those so you can make decisions and then disseminating that information in the organization.

Sarah Wang

很惊人。你标过：开发者显然在用 AI，价值有争议。你们是最早一批——我不想说 token maxing，负面意味太重——但就鼓励人用 AI 提高生产力，你们走在前面。然后循环变成：糟，人在浪费，我们要 value max。你们自己的旅程是怎样的？怎么想 value maxing 而不是 token maxing？再把 Unity Gateway 扯进来——管成本越来越重要，你们也在帮人做。

Yeah, it’s pretty amazing. You sort of bookmarked—developers are obviously using AI, questionable value. I want to follow up with you on that because I feel like you guys were one of the earliest. I don’t want to use the word token maxing because it has such a negative connotation, but I think in terms of applauding people who can use AI to become more productive, you guys were at the forefront of that. And then of course there’s this cycle of oh shoot, people are being wasteful, now we need a value max. What was your own journey on that and how do you guys think about value maxing not token maxing—and then I’m going to throw in Unity Gateway in this because I think the managing of cost piece is actually getting more important and you guys are helping people do that.

Ali Ghodsi

大概去年四季度模型变得真的很好，我们开始注意到生产力明显更好。我自己开始用模型把代码提交进 Databricks 生产——要一路走到生产。然后我推动组织：人人都要做；CEO 都能往有全部安全要求的敏感数据平台提交生产代码，你——任何经理、任何人——也应该能。四季度我们做排行榜；一二月开年时已经全开，人人在用、我们在管。Token maxing 大概二三月已经在发生。我们运气好，大概领先几个季度看到苗头，开始失控。我们已有网关叫 Unity Gateway，用来提供 token 容量：客户可拿 OpenAI、Anthropic、Gemini、Grok 容量——我们有关系——以及任何开源模型。我们开始设预算约束、警告：你预算还剩多少、接近天花板——按人、按组；再做很好的分析，精确预测成本走向。然后加智能路由：接近预算或问题简单时选更便宜模型。我们还建了叫 Omnient 的 harness，可在不同 harness 间复用。事实证明 harness 本身很重要：同一模型、不同 harness，成本差几乎有 2 倍。换 harness 就能在成本上拿到很多杠杆。用这些我们弯折了曲线：token 继续涨，成本基本持平——对我们极重要，需求也极大，每个组织现在都在经历这个。

Yeah. Around Q4 last year was when the models got really really good and we started noticing that okay it’s actually starting to give much better productivity. So I actually started using the models myself to start commit code into production for Databricks—the actual—I want to take it all the way to production. So I did that and started pushing the organization that hey everyone needs to do that. I have done it. Why are you not? Like if the CEO can commit code to production on a very sensitive data platform that has all these security requirements, you should be able to do that too—you being any manager, anyone in the organization. So started pushing everyone very hard and we started making leaderboards in Q4. And at the beginning of January, February when we kicked off the year, we were already full swing. Everybody was using the stuff and we’re pushing and we’re managing this. But the whole token maxing thing was happening around February–March period already. So yeah we just had the luck of being like maybe a few quarters ahead of folks to see what was happening here and it was getting out of hand. So we already had a gateway—it’s called Unity Gateway—where we were already—this gateway was being used to provide token capacity. So you can get OpenAI, Anthropic, Gemini, Grok capacity—like any customer can come to us and we’ll just provide them that capacity because we have relationship with those—and any open source model. So we started putting in budget constraints in place and giving people warnings like okay you have this much of your budget left, you’re getting close to your kind of ceiling. So we started doing that per person and for group and then we started doing great analytics so we could predict exactly where the costs were going. And then we added smart routers that could actually pick cheaper models if you’re getting close to your budget or if you have simple questions. We started doing that. We also built a harness called Omnient which can multiplex between the different harnesses. Turns out actually the harness itself matters. Like if you use the same model but different harnesses, there’s almost 2x different cost difference. Even exactly same model—same version but different harness you get 2x difference in actual cost. So if you can change harness you can get a lot of leverage in the cost. So we started using all of this that we were able to actually bend the curve and actually our cost for AI has been basically—the tokens continue to go up but the costs have been sort of stagnant. So that’s been actually super super important for us and there’s a huge demand for this. I think every organization is going through this now.

Martin Casado

我在二十来个董事会。第一次有规模化公司上周说要从前沿模型迁到 GLM——大型工程组织。你觉得是趋势还是一次性？我记得第一次 DeepSeek 时刻、英伟达股价，结果没成真；然后 Kimi 时刻、又一次 DeepSeek——好像都没对市场有可感影响。但现在我听到的轶事很真，似乎在发生。想听你的看法。

Yeah. For the first time—I do a lot of board meetings, I’m on 20-some boards—for the first time ever, a company at scale last week said that they’re moving from the frontier models to GLM. This is a large engineering organization. Do you think that that’s a trend or do you think that’s just like a one-off anecdote? Because I’ve been hearing about—I remember the first DeepSeek moment and like Nvidia stock and then that turned out to not be real. Then the Kimi moment, then the next DeepSeek moment. None of it seems to have actually had an appreciable impact on the market. But now then the amount of anecdotes that I have are pretty real and it seems to be happening. Love your view.

Ali Ghodsi

人两样都要：难任务、有 ROI 时要最新超聪明模型；大量琐碎事——人真的用 harness 去重命名文件之类——你为此多付数量级的钱，至少自己打字，别让模型转五分钟再帮你改名、花掉几美分。人在动：模式是专家模式——小而便宜的开源模型调用大专家模型，或反过来，来回乒乓；也复用／切换 harness 控成本。比如发现 pi 作为 harness 很高效。会有很多种。模型本身随机，每次答案不同，又变得飞快，实验很多。我们会进入：不再事事都用最聪明模型——过去几年的范式是新模型一出，连最简单琐事也用它。

I mean I think people want both. They want the latest model that’s super intelligent for the difficult task where they get ROI. But then there’s a lot of mundane dumb things—people literally use their harness to rename files and whatnot. You’re paying orders of magnitude more for that—at least type that in yourself. Don’t have the model do that. It’s going to spin for 5 minutes and then it’s going to rename the file for you and cost you cents. People are moving on it. But what they’re doing is that the pattern is either use this expert pattern where you have small cheaper open source model that uses expert model—the big ones—or vice versa or a way in which they can sort of ping-pong them to each other—but also multiplexing harnesses and just changing harnesses so that you can control the costs is also what people are doing. People have found for instance—there’s pi is very efficient when it comes to as a harness. So yeah I think there’s going to be a multitude of these. It’s easy—the models themselves are stochastic as you said—every time they give a different answer and they’re changing so much. So there’s just a lot of experimentation happening. So I think we’re going to get to a world where you’re not always using the smartest model for everything which is kind of the been the paradigm for the last couple years. Like new model comes out, it’s super smart. They use it for everything even really really simple mundane tasks.

Martin Casado

我看到有人用 Fable 和 Astra 做架构，便宜模型做实现，再用 Fable 或 Astra 审计——像在涌现的模式。

Yeah. I’ll tell you what I see—people using Fable and Astra for like architecture, a cheap model for implementation and then Fable or Astra for audit. Like that seems to be like this emerging.

Sarah Wang

你们在创业公司看到多少开源？按 token 还是按美元？

What are you guys seeing in the startups? How much open source—by token or by dollar—either.

Martin Casado

按美元开源大概 5%，很少；按 token 计数超过 60%。

So by dollar open source is like 5%. It’s very little, but by token count it’s over 60%.

Sarah Wang

我们跟 Decagon 这类聊——内部用 vs 外部产品不同。产品外部他们开源几乎到 90%；内部很多说不在乎、用前沿、不想成本控制——变大了才会。你们另一个董事会里他们也确实从浪费角度压下来了。产品侧我确实看到更多走向开源。相关的还有后训练：我记得 2023 跟你聊——你们什么时候买的 Mosaic？

Yeah I was going to say—we talked to let’s say a Decagon or something like that. It’s different internal use versus external for product. On the external for product I think they’re almost up to 90% open source. On the internal—and I don’t want to say for Decagon in particular—but a lot of them are like we don’t care we’ll just use frontier we’re not thinking about cost control—but as it gets bigger yeah. Right you and I were another board meeting where they actually did bring that down just from a waste perspective. So I definitely see that moving more toward open source on the product side. And actually that’s related to another question maybe around open source but post-training specifically. I feel like you were kind of early. I remember talking to you in 2023—when did you buy Mosaic?

Ali Ghodsi

2023。

2023.

Sarah Wang

你 23 年的愿景到 2026 有点成真了——我们到处听到：你会拥有自己的智能，会对开源模型做后训练等。创业公司肯定在做；企业还不确定。你觉得自己早了吗？

2023. Okay. So this vision that you had in 23 kind of came true in 2026. I don’t know if you guys would agree—that’s sort of what we’re hearing across—oh just sort of like hey we’re going to actually—you’re going to own your own intelligence, you’re going to be post-training your open source models, etc. And that’s definitely what the startups are doing. I don’t know if that’s what the enterprises are doing yet. Do you feel like you were early to that?

Ali Ghodsi

一开始还有「我们也会帮你预训练」——那现在没意义了，有很好的预训练模型可用。但你可以做后训练、做强化学习。我们其实在规模化做，很多这类创业公司是客户——我们帮他们用强化学习环境，让模型在特定任务上变得非常好。对重复性任务很合理：创业公司产品做特定的事，不是一般智能——拿很好的开源模型用 RL 把它练到那任务上极好，成本可砍、速度可快、IP 自己控。大型企业只需要基本自动化，现在做这个对他们太重了。

Yeah. First of all, when we started it was also, hey, we’ll also pre-train it for you—which that’s that doesn’t make any sense. You can—there’s so very good pre-trained models now that you can use. But that you can do actually post-training on the model and you can do reinforcement learning. Yeah, we’re actually doing it at scale and many of those startups are actually customers. So we actually help them using reinforcement learning environments where we can make the models very very good at the specific task that they are doing. It makes a lot of sense for them to do that. If you have a repetitive task—so if you have a startup and it’s offering a product and a product does something specific. It’s not just a general intelligence. It does something specific for you. It makes just a lot of sense to take a really good open source model and use reinforcement learning and make it really good at that specific task. You can cut the cost down. They can make it really fast. They control their own IP. So in that sense that is possible—but large enterprises they just need basic automation and it’s just too much for them to do this right now.

Sarah Wang

挑战之一是你需要好 eval——做好 eval 很难。创业公司能做也有动力；其他组织更容易的按钮可能是直接用前沿模型，而不是自己建 eval。

I think one of the challenges is you need good evals—and making good evals is hard. So while the startups can do that and they’re motivated to do that, other organizations—the easy button might be just to use a frontier model than having to create their own evals.

Ali Ghodsi

我们甚至在产品里自动给客户生成 eval，放在很显眼处——结果人不想用，我们就挪到后端变成可选，然后他们再也不去。总体像软件工程里的 TDD：人人说该做测试驱动开发，真正做的极少。训自己模型的诅咒一样：难的是 eval。

We actually generated even evals for the customer automatically in the product and we had it front and center—but then people didn’t want to use it so we said okay let’s move it to the back end so that it’s optional and then they would never go to it. So I would say in general—why—they just don’t want to get into it, it’s too complicated—or I think you want quick reinforcement of like hey there’s a new model, I want to try this out, I want to get this problem solved. You don’t have time to go do this the scientific method of let’s make an eval, let’s have a great baseline. And it’s sort of like TDD, test-driven development. In software engineering, did people actually do test-driven development? Very few did. Everyone said it’s the right way to do it, but nobody actually in practice did it. So that’s the same—that’s kind of a little bit of the curse of doing training your own model—is the eval is the hard part.

Sarah Wang

你们有 FD 模式——现在很热的缩写。要把这些带进那么大的企业，需要全职 FTE 模式吗？怎么演化的？

I know you have an FD model at Databricks that’s a very popular word right now or acronym—but does it—to get these at enterprises that large—is it a full FTE model that’s required or how do you and how’s that evolved maybe?

Ali Ghodsi

我们有这些 FD，需求显著上升。很多是：怎么建本体？本体可自动，但你若不收集任何信息、不录任何东西——对吧。那是我们做的关键之一。还有：我想建智能体，要面向客户、低延迟、有护栏防滥用或问我们不想答的东西——我们能建，比如 Fox 的 Sports AI，你可以聊体育赛事；试着问政治，它很会拒答并转回体育。所以 FD 建这些，帮组织真正开用 AI。重要，因为很多组织没有内部专长建这套，需要边上一点帮助就能起步。

Yeah. We’ve had these FDs and the demand for it has gone up significantly. A lot of it is how do we build that ontology? The ontology is automatic but if you’re not collecting any information—like you’re not recording anything right. So that’s one of the key things that we do. But also things like I want to build an agent, I want it to be customer-facing and have really low latency and I want it to have guardrails to not—people coming to abuse it or ask it things that we don’t want it to answer and so on. So we can build that—like Sports AI that Fox has—you can go chat with it about sport events. You can try to ask it actually about politics and it’s very good at rejecting you and moving and talking about sports instead. So the FDs built that—so we’ll help the organizations actually get started with AI. It is important because just many organizations do not have the in-house expertise to build this stuff. So they need just a little bit of help on the side and then they get started.

Sarah Wang

说得通。偏智能体这边：我最近看到第三方中立测试里，Lakebase 或 Neon 竟是智能体首选的 Postgres 数据库。一是为 Databricks 兴奋，二是一年前我大概猜不到。你们怎么拿下的？若现在赢了智能体，就赢了市场。

Yeah, makes sense. So this is more related on the agent side, but I saw recently that I think a third party—neutral third party—did some tests that Lakebase or Neon was actually the Postgres database of choice for agents. And I thought that was interesting. One because exciting for Databricks—but two I probably wouldn’t have guessed that maybe a year ago. Surprise for sure. It was pretty clearly number one. So I’m curious, how did you guys crack this? And what makes you win across the agents? Because if you win the agents now, you win the market.

Ali Ghodsi

很多功劳该给 Neon、Nikita 和团队。他们痴迷于：怎么让模型／智能体偏爱 Lakebase 或 Neon。智能体要实验、去建一小块软件，需要数据库——数据库必须很快起来。他们的执念是一切远少于一秒：库在远少于一秒起来；你可以克隆巨型库——petabyte 级——也少于一秒。高度弹性、高度响应。然后他们做了杀手功能：branching——同一库上很多轻量分支。我们在别的智能体工具上也看到这点：像 UV、ripgrep——把很多 Unix 工具重做成极快、极轻，并对智能体某种程度 fail-safe。他们把这套做到更难的问题上：数据库。于是你有 Postgres，又快又灵又 fail-safe，可回快照。智能体更好用。定价也想到：别因为智能体在实验就让费用飙——生产多人用你愿意付，纯实验不行。他们不是要赢数据库战争或压过某厂商，而是痴迷：我们怎么对智能体最好？这是新 persona——数据库 historically 痴迷帮 DBA、帮应用开发；他们改规则：聚焦智能体。现在 Neon／Lakebase 上创建的库超过 90% 是智能体建的，甚至不是人——数字自己说话。

Yeah. I mean I think a lot of credit should go to Neon and Nikita and team. What they’ve done is they’ve just been obsessive about how do you make the models—how do you make the models pick—and agents favor Lakebase or Neon as a database. So what do they do? The agents want to experiment. They’re going off. They’re trying to build a little bit of software. They need the database. So you need the database to come up quickly. So they had this obsession that everything should take far less than a second. So database comes up in far less than a second. You can clone gigantic database—kind of petabyte database. You can clone it in less than a second. So it’s like highly elastic, highly responsive. And then they built this killer feature called branching. Branching just lets you branch the database and you can have many many branches over the same database. And they just made this very very lightweight. We saw this with other things with agents—like UV, ripgrep—basically reimplementation of a lot of the tools on Unix making them really really blazing fast and lightweight—and also sort of failsafe for agents. They’ve just did this to a harder problem which is database. So now you have a Postgres database and the Postgres database has all these advantages that it’s really fast, it’s nimble, it’s fail-safe, you can go back to snapshots. So I think that’s why it’s just easier for the agents to use this. They also made sure that they had a pricing model that was like—you don’t want just because the agents are building some software and experiment, you don’t want the cost to run up. You’re okay paying for your database if it’s like production use and lots of people are using it, but just to experiment. So I think they were just obsessed. They were not trying to win the database war or trying to be better than some other vendor. They were obsessed with how are we the best for the agents? And that’s a new persona because in databases the obsession has been how do we help DBAs? How do we help app devs? How do we help the people that are using the database? Exactly. They changed the game and said hey how do we focus on agents and help agents get the best database they want? And now over 90% of their—the databases that are created on Neon and Lakebase are actually created by agents. So it’s not even humans. So numbers speak for themselves.

Martin Casado

很惊人。我开始把 Neon 当标准库用——进大公司通常会变慢，结果产品实质变好了。他们完全独立吗？跟其余部分怎么协作？

By the way, it’s remarkable. So I’ve started to use Neon as like my standard database and it was bizarre to me because normally when you enter a large company things slow down. It’s actually like the products got materially better. Are they totally independent? Do they work with the rest of—like how—

Ali Ghodsi

很棒的团队，合作很紧。我们爱数据库和数据——我们以此为生。团队把东西做得超快、干脆，对智能体很友好。

No, it’s a great team. I mean, and work very closely together. We love databases and data. So—it’s you know, we live that. But no, the team does a great job of just making super fast, snappy, and great for agents.

Martin Casado

好。Hey Ali，你的 P(doom) 是多少？

All right. Hey Ali, what is your P(doom)?

Ali Ghodsi

不到 10%。不——接近零。你呢？

Less than 10%. Uh, no. Close to zero. What about yours?

Martin Casado

我不知道。我唯一的答案是：没有 AI 时我的 P(doom) 远高于有 AI 时的 P(doom)。怎么样？

I don’t know. I say my only answer is my P(doom) without AI is much higher than my P(doom) with AI. How’s that?

Sarah Wang

哇。我也这么说。技术上这次我同意 Ali。

Wow. That’s what I say. I went on a—technical—I would agree with Ali on this one. Yeah.