Databricks CEO: Stop Scaring People About AI

Key Takeaway

Start with one high-value workflow, not a quest for a smarter model. Map the people, projects, systems, documents, permissions, and decisions involved; then give an AI assistant access to that governed context. Test it on a recurring question your team currently answers through meetings or manual searching. The episode’s practical message: most organizations can unlock meaningful AI productivity now by organizing their internal knowledge, automating safely, and measuring cost—not by waiting for frontier intelligence.

Episode Overview

Databricks CEO Ali Ghodsi argues that leaders should communicate AI risks responsibly: current existential risk appears close to zero, while cyber risk and misuse of increasingly capable agents deserve urgent engineering attention. He explains why organizations are underusing AI today, emphasizing that the missing ingredient is usually organizational context rather than model intelligence. The conversation also covers AI security automation, model-cost management, independent oversight, and concrete AI applications in health, logistics, and drug discovery.

Full Transcript

Transcript of Databricks CEO: Stop Scaring People About AI from A16Z. Auto-generated from episode audio; may contain minor errors.

As a business leader, there is a tragedy of the commons. If you want to stop, if you want to go slower, why don't you go slower? It's like I 'm competing, I want to win. There are almost two sides. There is a sector that argues that this is, in reality, an engineering problem. And others believe that we need to go slower. Humans are not reacting quickly enough to these current attacks. All of that needs to be automated. And most organizations aren't even close to doing so.

Are the RSI and recursive self-improvement that laboratories are doing leading us there? That's the big question. Something Elon said is that this is an elaborate 4D chess game. Because on one hand you're saying that all of humanity will die. On the other hand, you're saying, "Hey, what do you want assigned to you for your IPO?" For the first time in history, a large-scale company said last week that it was moving from frontier models to GLMs. Do you think it 's a trend or do you think it's just an isolated incident?

Thank you for being here, Ali. I'm very excited. Obviously we want to talk about Databricks, but right now there's a broader conversation about AI. Dario, Jacob, and Elon have weighed in, but we want to know what Ali Goetz thinks in terms of, you know, if you call the topic, in general terms, "setting the pace of the border," etc. What do you agree with most about what's being said? What do you disagree with? And perhaps there's some nuance that's not being grasped? Yes, I'd be happy to talk about it.

Martin and I argue a lot, so I 'm sure it won't take us long . I'll try to rate it this time. Try to stay calm. But well, I think that , above all, we might agree that leaders have a responsibility not to scare people unnecessarily, unless there is a very, very good reason. And I think , you know, there are always different people in society who are in different places, in their mental space. Therefore, talking about these kinds of existential risks and scenarios in which all of humanity is going to disappear seems irresponsible to me.

It can cause many people to break down and can cause many mental health problems. Unless there's something that's going to wipe people out. Yes, as I said, yes. If there's a real reason for it, then you know, that's another story. But I think that right now the existential risk is almost zero. So why scare everyone? Actually, it 's not necessary. There are risks. We'll talk about it later. That's probably where we disagree. Well, but above all, I think leaders shouldn't scare everyone. And I mean , you know, if there are technical nuances in how we're doing AI research and all that, well , researchers can debate it.

There's no need to appear on television every time or send messages on Twitter to millions of people saying, "I believe there is a 10% risk that all of humanity will disappear." I don't think that's helpful for many people. In fact, I think it causes a lot of harm to many people who get stressed and are unaware of the nuances of all this or what it means. So I don't think we should do it . I don't think it will be fruitful. It does n't really help anyone.

I mean , I think this is very, very true for the general public. Yes. Like my sister. Mhm. who is great and a school teacher in rural Arizona; On Sunday he wrote to me: "Martine, should I prepare the cabin for you?" She is, by nature, a prepper, but she asked me: "Should I prepare the cabin for the AI ​​apocalypse?" You know, I have water ready. When are you going to show up? "I said:" Wait. " Yes. As if we weren't there. So , it's clear this has spread to the population, which I agree with.

Yes. It's unnecessary and it has repercussions. I think there's a second reason, which is I do n't know if you saw it, like when I came in here, I was checking out X and Elizabeth Warren just talked about pausing all AI development. That , of course, is down to Bernie, who 's also working with Bannon. Like Steve Bannon, like in the... Well, now... Besides scaring people, the federal complex is now up in arms . And I think that might be quite counterproductive to the actual goals of the message.

And then, there's more than just public hysteria at play here. Yes, there's a lot of politics going on, but I'm in all these groups and, you know, I see both sides. There's a lot of politics going on on both sides, we should say. Right? Oh, yes. And this is happening on both sides. No, this is the... I think both parties, except for Trump, agree that AI should be restricted in some way level. I'm also talking about the other side of this argument. Let me give you an example.

Even Greg Abbott, right? Even Greg Abbott said that you can't have data centers in Texas. Well, I'm not talking about politicians. I'm talking about the fact that there's politics on both sides, right? There's politics on the business side, people who want to see big IPOs and want to profit from their investments, and they say, "Don't screw up my IPO," and they want to get lucky and for everyone to shut up so they can get their money back. So that's it , they have resources and they're using them, and there's politics on that side, and they're not just sitting around doing nothing; they can pull strings and they have connections.

On the other side, there are other people who say, "Okay, how do we weaponize this?" "This is great. This guy tweeted this, you know, let's weaponize this. Let's plant this, you know , let's spread these threads. Let's talk about the specific foothold everyone is using because I think this is a classic case of a PR disaster, and it's not just the fatalistic, pessimistic kind of thing . So here's the PR disaster , I think, which is that it's not unusual for industries to try to regulate themselves.

It just isn't , is it? And I think saying that security is important—that's what every techie talks about—and we want to have some oversight, that was very , very sensible. The problem was that it's just expressing this notion of pace, and there are a number of problems with pace. First of all , it's orthogonal to security. Like you can build a weapon bit by bit. Like that's no different from building a weapon. People don't feel it as something genuine because these companies have been in a race to the death.

If you're buying more computing power to be even more..." Fast. I mean, no, I mean they just have n't done it. They haven't historically done it, but it also feels like a kind of almost graceless capitulation to the pause people. So you say, well, you say pause, well, I say rhythm, which is almost like pause, but it 's not like pause. So, they chose this kind of banner to follow the rhythm, but if you really analyze what I'm saying, did you even read the document I wrote myself?

It's a totally sensible document. I read it, yes. It just has nothing to do with rhythm, right? And then, honestly, no, he brings it up. Look, I don't kind of disagree. Look, there's a tragedy of the commons. It's kind of like , hey, if you want to stop, if you want to slow down , why don't you slow down? Why do you write? There are a lot of people who make that argument. But no, as a business leader, I get it . There is a tragedy of the commons.

Like, if I'm competing, I want to win, you know, and you're going winning. I'm also going to market equilibrium, which suggests that the pace probably isn't practical anyway. Yeah, I'm just saying, you know, it makes sense for people to say, "Hey, if you don't stop this, this tragedy of the commons is going to continue." I'm not going to stop competing because, you know, there's an IPO at stake, there's competition at stake, and there's also some animosity among people. So I'm not going to stop unilaterally. I'll be a fool, you know?

Why don't you stop first, you know? Then they say, "Hey, can you come and stop us?" Well, but you know, I think you could also argue that if we look at the OpenAI face-hugging incident , which, by the way, I think these companies are great and I think they're probably investing a lot of resources. But it's very clear if you read what happened, that they weren't monitoring every token that came out or even having it, you know, they were just doing these RL experiments and then, after the fact, they went in and checked what was happening.

So, they should have gone slower. They should have been much slower in that particular incident, shouldn't they ? I simply don't want to, I don't want to argue about syntax, but words matter in public relations, right? Yes. So, let's take the face-hugging incident. Yes. When I read that, you know my reaction wasn't, "OpenAI should slow down ." It was like, "Damn, secure your stuff," right? How to do security checks the way we've always done them. But it is going slower. Yes, it does go slower. As in the history of the Internet, we had all these things where we said, "We're going to set the pace for the growth of the Internet." Set the pace.

We're going to make security. We're going to do a check. We're going to do what you should be doing, but it's about setting the pace in the sense that, look, I deal with this all the time. I have a legal department at Databricks. I have a security department at Databricks. And you know, they always say, " Hey, let's slow everything down for everything, except AI." Literally, every little thing. Like: "Oh, you're going to do a podcast. Okay, what's the script? What are you going to say?" And you know, let's go over that .

And do you know what the legal aspect is ? You can't say this. You can say that. You know, everything you say has to be materially true. You can't. I know there are people setting the pace in this room with us right now. So, you know, you're running an RL experiment, you're training the next model. Should the security team be there to operate all the monitors and supervise everything? I mean, millions of GPU hours of tokens were produced and these agents were running, you know, a simulation in isolated environments.

They would have slowed down considerably if the security team had sat there and checked everything. Now they're saying, "Hey, if we do that, we'll slow down. And I'm not sure the other side is doing that. So, can you step in and slow us down? Just tell us. Put some barriers around us . We'll follow the rules and do what's safe. Otherwise, there's no point because we'll get kicked . I think nuanced, second-order words don't work when people are really scared. You think, 'I'm going to walk back and forth, and therefore this kind of thing isn't going to happen.' I think you should have said, ' Safety is paramount.

We're going to put these controls in place.'" That's what's important. And I think that nuance was lost. If you look at what Zuck said, do you agree? That was Zuck's doing. I thought it was a good thing. We'll go at our own pace. We are going to implement security. The reason we're releasing this later is for security." What I found so great about Zuck is that he was very focused on security and self-regulation. The first five words of Dario Dario or whatever are like, " We need to set the pace of the frontier." Right?

It just puts you in a very different mindset than if he'd said, " We need to secure the border." Okay. We need security on the front line. I mean, I think on some level they're trying to optimize for both the doomsayers, who are calling for a pause, and the politicians, and they're not satisfying either group . But don't you think people are going crazy in the labs? And there are a lot of security people who are really going crazy . By the way, not all of them are AI people, etc.

And people are like, "Hey , they're shocked, right?" " But the thing is, using the word rhythm doesn't help either of us. I think it's like you 're literally trying to find this, like rhythm is the uncanny valley of making pessimists and politicians unhappy because to pessimists it doesn't feel like a pause, it feels like rhythm. And you know, everyone's like, 'Well, this isn't really...you're not going to do it anyway,' and you're not supposed to be in safety. So again, regardless of what we should be doing, what we should be talking about, I think the way it was presented was just bad and it just didn't work.

And that's why these guys, you know, they're not trained, you know, PR people, you know, and yeah, I agree with you on some of this stuff. I mean, I agree with the central premise that we shouldn't scare the public. I think the essential risk right now is close to zero. But let's talk about the essential, which is the fact that, you know, anyone doing large learning runs by reinforcement and is giving it a reward function, so they're unleashing, you know, 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 giant cluster for a month or two. We'll try to solve anything." It does n't have to be something related to security . It could be something like doing anything, you know, solving this math puzzle. Really bad things can happen. Really bad things happen when you get hacked, and, you know, cyberspace is the main one, right? That's real, right? I think this makes it something like, "Hey, this is an existential risk," etc. "Which I think was a mistake.

I don't think it's good to scare the public like that. I think it's become something that everyone, not just your sister, everyone on the planet is talking about now. I've met all kinds of people who never cared about this and find it extremely boring, they call me and say, 'What do you really think about this?'" "This is really important to me because now I'm starting to worry about it." Well, then it becomes a political problem, and we have elections here soon. But there are elections all over the world.

Well, you'll see that they won't stay still in other parts of the world either. But I believe it is our responsibility to discuss this in a balanced way and to outline the risks. I think that superintelligence, that idea from that book, is very, very far off. I see no evidence that we are actually moving towards that, or that it is going to happen. Apparently, some people do. Apparently, some people in the laboratories are scared that there might be progress in that direction . And I think it comes from RSI, recursive self-improvement, models improving themselves.

I would love to understand how much they have seen that we don't know about. There are, you know, a kind of four criteria. If there are...If those four things are happening, I would love to understand them. One is that the models, uh, the next one, yes, if we end up in a situation where the following four conditions are met, which is that the next model requires fewer resources, less GPU to train. And, you know, in a super-linear way, not just a little bit. The next model, you know, also takes less time to train.

So, that's the second condition. Uh, the third one, the accuracy of the model, the intelligence is increasing. And fourth, we can do the previous three over and over again , you know, it's not just us. All of them, all of them at the same time, right? All at the same time, not just any of them. Yes, all four. If all four occur, then you can imagine a way in which you can, you know, because none of them happen, like for example if the resources are constant, then it's okay because we're going to run out of hardware.

Then it will regulate itself. As if we didn't have enough hardware to do it. Not enough GPUs, right? Time is the same. So it has to be that you end up in this situation . So, if what you mean is that software writes itself, we're already there today. Since 90 percent of the software at Databricks is written by AI. Does it matter if the last percentage is also written by AI? No, it doesn't really matter that much. But if you meet these four conditions, then you could achieve an acceleration where the next model, say, takes half the time and half the resources and is smarter.

And if you keep doing that , you know, then you could end up in a situation where I don't know, by the way, I don't even know if that necessarily leads to superintelligence per se. It will not converge, yes. But I could, so it would be riskier. So it would be great if you could share all that data and we could shed some light and transparency on the matter . In fact, I think you have a great breakdown . I do n't think anyone is using it as a definition, are they?

I think people... I think there are a few people who get scared by things like, "Oh my God, emergent behavior. Now it's creating itself." And things like that. But I think, as I said, many people define this as: "Hey, I 'm not even programming anymore and it's programming itself." Clear. But I think they're getting confused: "Hey , what's my worth and I'm scared?" with: "Hey, that means we'll get that superintelligence that was theoretically hypothesized by Bostrom in 2014." Yes. Well, your point about the calculation is really good, but I think it's overlooked in many arguments about the RSI, isn't it?

Because as far as we know, the minimum computational threshold needed to train a good model keeps rising. Before it was 100 million, now it 's probably 5 billion. And that's a model now? To train like a model, right? Between 5 billion and 10 billion here. Correct, correct, exactly. Compared to 1 billion? Billions. Billions. Billions, yes. Compared to the second one, which is very expensive. Yes, the border is very expensive. Replicating the border six months later costs approximately 1/20. No, I think Sarah has a great point, which is a good argument against all this, which is that the next model, in the first place, there are only one or two runs of this type per year that each of these laboratories does .

And they take it as the opposite of the four criteria I mentioned, right? It will require more resources, more people involved, and is even more fragile, and they have to build the data centers. I mean, labs don't necessarily do 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 have the right network, and they have to do the engineering to make sure they can tolerate, because you know, every order of magnitude of additional GPUs you're putting in there, now you have to worry about errors that you didn't have to worry about before.

Therefore, you need to increase the robustness of the... It's like a very fragile process, and if it fails, you've wasted a lot of money. Therefore, they are very, very careful with that execution, and there have been several executions that have been observed. Yes. So, it's the opposite of that that you hate. The next model is faster, cheaper, smarter, and recursive improvement is the opposite. It's as if it takes longer, it's more fragile, and there are more people. And it's harder to achieve. So I think that's true.

Regarding the RSI, regarding real cyber risks and hacking, we have to take it very seriously, yes. So I'm going to conduct a sort of new trial by fire here. I love your four criteria. I was literally waiting to discuss it, but I think it's very, very...So I'm going to let you have a sort of black box. What are you going to believe when you're dealing with these systems that are so dynamic and adaptive? Are you going to believe the numbers or your deceiving eyes? Right? So I think you have to look at the numbers in these cases.

So, what numbers should we take into account? I really think you should basically do it, and maybe going public is the right way to do it. For example, if these companies continue to grow, reducing the number of people and the amount of money allocated to them, then I would say that something is definitely going on here. I think you can put this in a black box and take a look, but none of it indicates that they're hiring like crazy. But that's not fair. It's not fair because, you know, companies aren't necessarily efficient, right?

So what if you have? I mean, OpenAI itself was doing like a million different things. A very small team of about 10 people was doing LLM and the LLM material was useful. There used to be a lot of people on Twitter, now there are far fewer people doing LLM. I agree. Just another trial by fire. We need to have two acid tests. We have your acid test, which I think is great, but then you'd need a way to implement it. And then we should have the acid test of the black box.

I mean, listen, if Anthropic in 2 weeks is, you know, 12 people and they're still growing and launching models at an increasing rate, I think we should probably take note of that. That's sufficient criteria, but it's not a necessary condition, right? What I'm saying is, you know, it could be that, you know, the right way to do it is to really look at the pre- and post-training that's being done and that's really necessary. Because they have so many resources they could be doing many other things they don't need to do, but they are doing it and they can hire other people because they have infinite and infinite money, so really the people who are training the next model are tiny teams and they are actually shrinking and doing less and less work and it's just AI doing it and then the subsequent training and then everyone is using less GPU.

That is not the case. We've had this discussion many times as an industry before. I remember when we learned to group computers because the mainframe was really limited by things like memory coherence. Do you remember that? They just make it so big, you know? And then we moved into the client-server area and then we didn't have that problem and then we started creating supercomputers that were basically like, you know, clustered computers. Yes. And then, at some point, the internet appeared. And that was more or less when GPUs started to improve.

Do you remember that we liked export control PlayStations because we were worried that Saddam Hussein would use them to run simulations? The arguments were very similar, that these things were becoming infinitely powerful. We were using them to simulate nuclear weapons, which we did. Like me. Yes. Like, we can't, you know, this has an existential risk. They did n't actually use those words, but this has the potential to lead to nuclear weapons or whatever, and we should stop it. And none of that ever happened . So I think a very reasonable discussion is whether this time is different.

Yes. I don't have an answer for that. But I'm on the PC. But you know, you're a ... Yeah, I mean, look, I don't think I'm old. enough to remember. So ignorance is bliss. So I can... I don't remember a PlayStation being illegal or Saddam Hussein being... I just don't know. Wait, how old are you? Perhaps he's just ignorant. This is like 1999. I'm old and ignorant. I mean, you know. Maybe he's just a ninja of old age. But whatever the case may be, Sweden doesn't care about export controls in the EU.

Yeah, you know, whatever it is, I think it 's a different scale now, right ? With AI and, you know, with what we 're doing, the pace of development, etc., they're going crazy at Frontier. I think cyberspace is actually one of the biggest problems we're going to see, right? Because there is so much infrastructure on the planet. By the way, much, much more than there was when Saddam was around or Xbox or whatever it is you 're talking about. I mean , it's like we've interconnected so many more things and they're interdependent, and the planet looks different today compared to the internet, technological dependence, interconnection, and, you know, 30 years ago.

So, I just want to make this clear. There's a lot of unsafe infrastructure out there, isn't there ? And if you're going to unleash these agents, they're going to find loopholes, they're going to find vulnerabilities, they're going to break in here and there. Well, this is a real risk. And you can't just ... And by the way, he didn't this time, but you could imagine a scenario where he starts jumping. As if it consumed resources and started running in another location. So, it spreads a bit like a virus.

That's a real risk. So this is purely out of curiosity. I promise I'm not, you know, trying to be a counterpoint here, but why do you think we haven't seen much? So? Like me, again, I am much older than you. I remember it very well. So, when, literally, when the Internet appeared, by then we had literally eliminated 10% . We had rendered hospitals unusable. We had eliminated critical infrastructure. We had caused tens of billions of dollars in economic damage from worms. As if all that had already happened.

And, as you said, we had much less development, you know, less of the economy depended on it. And, well, you know, AI has a lot of people who want to find risks and threats. We're going very fast. A lot of money has been invested in it. And we haven't seen anything consistent with the early days of the worms. What is the reason for this disconnection? Yes. Well, look, I do remember those days. But that's in due time. Yes. Well, look, I'd just say I sleep well at night and I don't think there's an existential risk right now.

Yes, I think there is a lot of infrastructure that needs to be protected. We have a product on the market, in the detection market, Lake Watch, which helps to perform detections. And space is advancing very quickly because, you know, you used to have these security teams, people from the security operations center, watching what intrusions were happening, how we were being attacked, etc. And now humans can't keep up . So, this whole space, the cybersecurity space, is becoming fully automated, using agents for detection on the other side.

If we don't do that, we're rushing things. We're rushing, the industry is rushing to do it very, very fast. If we don't do it , I think you'll start to see that kind of thing, like sites going down, you know, whole systems ceasing to function for a while, and there will be consequences that are not existential, but economic damage and, you know, it could happen that people get hurt, etc. So we just have to run very, very fast to do all those things. There's simply no ability to respond quickly enough to the attacks that are occurring.

Therefore, you just need to automate all of that. And most organizations aren't even close to doing so. The banks are doing it. Some of the people who are very security-conscious are doing it, but most of the industry today operates with old-school security operations centers and people who wake up every day to hundreds of emails about detections that have been triggered. Many of them are just false positives. Therefore, you don't have to ignore them, but some of them are not. They simply don't have time to review them.

And you must identify them. You need to have an automated threat search where you 're actually attacking your own systems automatically with agents, etc. It hasn't happened. Therefore, I think if we simply said, "Hey, this is like the internet in its early days." Well, you know, bad things are going to happen . Therefore, there is a risk. You know, I was actually very surprised. I was at a conference early this morning, and I think you and I are quite close and we talk regularly. I feel like I know quite a bit about data blocks.

This morning I was on a call where a founder basically said something like, "Yeah, listen, we're doing all this as observability agent threat detection and we're using data blocks." I didn't even know you had this offer , frankly. Well, this is just from an educational point of view, to what extent have you made progress in the safety of the observability of AI agents? Yes, I mean, we gave a talk this year at RSA with Ben Horowitz. But the problem is that data and AI are getting mixed up with cybernetics.

These two markets are collapsing. And the reason they're collapsing is that it used to be like, "Okay, we have data and AI, the kind of stuff that data blocks and these kinds of companies used to do, which is like, okay, you have a lot of data and you run AI and machine learning, and that lives separately. And then you have the cyber world. The cyber world is, you know, we want to detect if something bad is happening. If bad people are trying to hack us, if bad people are doing things, we need to detect it.

Okay? But now, on the data and AI side, we have agents running internally within the company. People have agents running. And the agents also interact with other people's agents, and they're producing a lot of data. Logs, traces, you know, fingerprints that are being left behind. And so, you know, now you have these agents internally doing that . So, these worlds are starting to merge more and more, which is like, well, all the data that's being produced needs to be analyzed. And the scale at which you need to do That's like many, many orders of magnitude more than just a year or two ago.

So, things have changed drastically. Like in 2018-19, the time it would take from, well, you know, a CVE vulnerability being published until it becomes weaponized in the industry, would be like 2-3 years. I'm narrowing that down, you know, to 2022 significantly, but it was still like 8-9 months. Yeah, yeah. So, okay. You have 8-9 months from a vulnerability until... That was in 2022. Yeah. Now, if you look at the curve from 2022 to now, it's now down to basically hours. Yeah. So, basically, there's no time.

Things become weaponized immediately. So, you need to do it in an automated way with the data platform and AI approach. So, I'm going to argue that this market is just going to collapse, actually. Yeah, I agree. Well, it's a very specific question here. In fact, I think a lot of people would like to backtrack. in the debate about the additional existential risk. It seems there are almost two sides. Yes. There's one side that believes it's really an engineering problem and that companies like Databricks can solve it, and they can solve it through products, engineering solutions, and services.

So, we, as an industry, need to solve that problem. Yes. And there are others who really believe... It seems to me there isn't an engineering solution. You have to, you know, slow it down ; you have to implement the necessary regulation now. It's more like a nuclear weapon, etc. So, does this mean you think it's an engineering problem? Or are you still not comfortable saying that? Because wait, really, why would you buy data blocks? Let's put things in the national lab. Just take it slow. That's the only solution.

That's it... Everyone takes their time and it'll be fine . The bad guys... There's no line of research that comes down to taking your time, I don't think so. I think it's more like you deposit or like you solve. What we've learned here is that my team hates the word "Rhythm." I'll never use that word with you again. Okay. Clearly, properly noted. But frustrated, Mark. Yeah, is this an engineering problem that engineers can solve, or is there something more to it? Actually, what problem are we talking about?

There are two distinct problems that I think are being confused. There's the superintelligence problem. You know, and I think a lot of this comes from Bostrom's 2014 book on superintelligence. And if you look at the definitions, I think people don't have clear definitions of what superintelligence is. So I think what he had in mind when he talked about superintelligence is AI, but I'm not sure, I don't know what the examples were. Something like they write a whole PhD dissertation with novel, peer-reviewed stuff in a couple of seconds.

And they can think millennia, you know, in an instant. And you know, so this is like the level of, you know, how fast they are, how they can learn. They can learn. Yeah. Yeah, it's just... Yeah, I mean, Yes, but it's many, many, many orders of magnitude. Right? It's like the scale of the problem is completely different. So, if something like that exists, do I think it's just an engineering problem to solve? No, I think it's actually very... If something like that were to happen, it would be very existential.

Absolutely. And that's what everyone agrees on. So, I think that's getting mixed up with... Now we have agents that aren't even close to that. I said it's not even like there's nothing like that and we don't have anything on that path right now. Yes. But these agents are capable, and you can do something with them that you could never have done before in the history of humankind. So, I think there's been a tipping point. Something has changed. Which is that we could have good security researchers at Databricks, but I could never say, "Let's put 10,000 of them in a test environment for a month, for one month, and have them do paid work worth $100 million." Now we can do that.

We just press a button. button and we can put 100,000 of them. Or math. Like we could say, "Hey, you know, we want to solve a conjecture." Okay, let's get some good mathematicians, but let's get 10,000 of them working together, you know, and that way you can progress really fast." Well, I think this is what leads to all these cyber risks. I think the cyber problem is the main one here. Yes. I think this can be solved with engineering. And I think we're working on it, many others are working on it.

There are still risks, but they are not existential. I think we should do it. There is the question of superintelligence. It's something that could make us write a new doctoral thesis or reason intuitively in 11-dimensional space physics instantaneously and without writing anything down. Something that humans cannot do. That type of superintelligence. The question is whether the RSI and the recursive self-improvement that the laboratories are doing are taking us there. Are we going to get there? Sure, trying to do it. Trying to do it. Is that what's going to happen?

Yes. And how fast is it going to happen? That 's the big question. And they've suggested that, you know, we should have inspectors come and observe what we 're doing. And that's a good idea. Let them go in there and get the data. I'd love to, the question is, who are the inspectors? Because it can be stacked, right? That can be stacked. What do you mean? Which team? Yes, well, there are a lot of people who are actually on either side . I wouldn't mind if they were the inspectors; I wouldn't be very impressed by what they said because they've already made a decision even before going in there .

Correct, exactly. But let's say, for example, if Jan LeCun , who was one of the inventors of this , you know, deep neural network technology, right? One of the pioneers. If he, he, he were to say: "Hey, there's nothing to see here. There's no risk." You know, I'm paraphrasing. It's no big deal. This superintelligence is nonsense. Keep going . Go fast, fast, fast. Nobody would believe it. I am putting words in his mouth. I'm not saying exactly what I would say. Now, if I were one of the inspectors and I went in there, took a look, and came out and said, "Hey, I've looked and it's just as I said.

There's nothing to see here. Move on ." "I would feel really good about that. I would say, 'Well, I would feel really good .' Or if he comes out and says, 'Oh my God, it's wobbly,' and I would change my mind a little bit about that . That would also have a lot of interesting signals. So, I think it all comes down to who we choose as inspectors, and I think it's a good idea. Let's choose some of them and choose a diverse group of people so we can get different, nuanced points of view .

What do you think of this kind of Elon Musk view, which is like less of a third-party view? It's more... It seems like there are like three proposals. Like the anthropic one of opening your eyes, that's a third part. Yes. Elon Musk's, as far as I know, is that labs check each other as a pure review, like you do in science. Yes. And then Mark Zuckerberg's is that they monitor themselves, right? What do you think of this one in the middle? That they should monitor each other, like peer- reviewing each other." Yes.

I mean , like I evaluate you and you evaluate me if we have boxing matches in the ring , the boxers should just be judges of each other. Would it work? No, they'd just yell "foul" all the time. Foul, foul, foul. Like, you know, it's like, you know, the moment the other guy, the moment the other guy pulls out the big model and it's like a big super-intelligence race. Absolutely, you know, they haven't been responsible. As you know, when there are vested interests at stake and there are IPO plans and these two companies are so competitive and have this history with each other.

Yeah, they'll be very fair to each other. I'm sure of it. That's why you need a third party, right? I mean, why do we have judges in the world? Why do we have third parties? Why ca n't people work things out between themselves? But they should try. If they want to, they should try. But I doubt they won't be biased in many ways when judging each other. So I have to ask Ali, do you think Elon said anything too? I think he was at the All In summit.

He said something like, "This is a very elaborate 4D chess game because, on one hand, you're saying all of humanity will die." On the other hand, you say, 'Hey, what do you want for your OPI assignment?' "Right? So, I guess that's probably a more cynical view, but how do you reconcile that? I mean, I think the dissonance confuses a lot of people. How do you think you reconcile it? Look, I think all these things get mixed up. I think there are people who are scared. And I think there are people who say things like, 'Hey, if there was some regulation that kept us in check,' excuse my language, that would be good for us, wouldn't it?

That would be good for us. But I also think people have vested interests, right? These things, you know, people usually find a way to make all these things align harmoniously in their heads. Well, yes. I think there's been a tendency in the past to also use marketing tricks, saying, 'Oh my God, this latest model is so good that I trained it.'" It's incredible. "It almost scares me." And then how does everyone start focusing on it? Yes, there's been that kind of marketing. Yes. But at the same time, as I said, the time from when a CVE is created until it becomes a weapon has gone down from years to minutes, in just 3 or 4 years.

Well, it's real. Cyberattacks are real. And this is the " but," but there's also a big marketing strategy to, you know, every time you train a new model, make a lot of noise about how risky it is for the world. It helps, right? Well, you know, maybe these things don't contradict each other. Well, you and I are people who network, and there's a long history of forming third parties to help arbitrate things, right? Like the IETF or, you know, the IEEE or, you know, even ICANN.

I see where this is going. No, no, no, no, no, so no, my question for you is, I think this is actually a very sensible proposal they have. In fact, I'm I agree with you. You probably want to make sure it's independent, which isn't fair right now. And there will be a lot of arguments about who you put there, and everyone will disagree. But you said, do you know why we have judges? So, actually, state intervention is quite different from industry self-regulation . So, at what point do you think it makes sense to really consider federal involvement?

Or do you think now is the time to really consider actual federal involvement rather than more industry self-regulation? Well, this is very different. They're distinct things, but they kind of overlap; for example, FINRA isn't a completely independent entity—it is, but at the same time, it's tied to the government. So, I think these things will overlap. I think it's difficult for them to evolve toward… Yes, historically, the industry self-regulates and then evolves toward regulation. If they say there's an existential risk, which is what they say, and then they say, " Come and watch over us and regulate us," I think it's very difficult for regulators to say, "No, we're not going to do that." So far, they've said that, but I don't think that's going to last long.

I think David Sacks said, "No CEO has ever asked us to regulate them." And what I love most is that the CEO has never had a regulator say no to that. Say no. I mean, the reality is that the real metapolitical machinery is already in motion, right? I mean, everyone has a talking point. Obama leaving is a big issue. Do you think it's too late? This is going to be a big issue in the midterm elections, and we are, in fact, going to implement very strict federal regulation.

And all this is going to stop, you know, Anthropic goes into the Department of Energy, and we're past that point, or do you think we can really end up with sensible self-regulation? I think we should strive to do the right thing. I think there's still some degree of freedom in how things evolve, and there's still time. And yes, you're right that, to a large extent, they You give these companies that we 're investing so many billions of dollars in. Yes. And the way reinforcement learning works is that, you know, you give it the reward function that's verifiable, like, we're going to solve this math problem or this kind of, you know, limited area of ​​programming, etc., and we put a lot of money into it.

You can get pretty good results in that kind of thing, but that doesn't mean you're acquiring that superintelligence. But you can even fool yourself into thinking that, say, fewer inputs are giving you a better result just because you're running so many experiments and thinking so much about it, right? But it's actually very difficult to do a closed experiment this way given the amount of resources that are being invested. Well, fundraising is increasing astronomically, as you say. Yes. Yes. And that's why these companies are going public, right?

I think otherwise they'd probably say, I mean, as someone who runs a large-scale private company, I think they'd prefer to remain private, so On the contrary. Why are they going public? Because they need the capital. And they consider the laws of scale and capital to be a strategic advantage. Well, that's why they're going public. But I would say let's go back to the four things I mentioned. If those are true, would you like to know? Would it be worrying if it got out of hand? Now, there's no evidence that those four things are happening.

But if there were, no, that's where it's headed. Yes, yes, yes. In fact, I think it's important to understand any system . Yes. Any kind of self-propelled property is important. And we've done this in the past with dynamical systems, haven't we? As we have with compilers. We did this with all the research into nanotechnology. It's been a common interest for us. Yes. And I don't think it's anything new; it's an interest that we should continue to maintain in the future. I think the fear is that these particular systems are such complex net economic systems that the risk is that you'll be crying because you're seeing it when you're not .

And I think a lot of that is happening. Right now. Yes, but I think, of course, if you see it, yes, of course. I mean, you want to know. But it's fair to say that the labs are now focusing a lot on the RSI, and that's where they're heading next. And maybe they're just unjustifiably worried, like they were about GPT -2. Right? It's like they said GPT-2 was the end of the world, and then it wasn't, and GPT-3 and 4 came out. So I do n't want to object.

A lot of times when they talk about the RSI, they're really referring to autocatalytic effects . And autocatalytic effects have been around in our industry for a long time. For example, there's no way you can create a computer chip without a computer chip. It's like, you just couldn't do it. Anyone with a computer science degree would write a compiler, write their own compiler. Well, that's closer to the RSI. But like the steam engine was autocatalytic, right? Listen, my full-time job is to get people out there From the labs and companies, they say RSI because everyone says RSI, and in reality, only 1% of them are RSI.

It's more like we're using AI for data cleaning. We use AI to make things. Let's make this distinction. So, you 're saying it's basically capitalist in the sense that they use AI to speed things up. It's autocatalytic, yes. So, I would say that what we hear in the RSI model, right? You're using a model to build a GPU core. You're using a model to do data cleaning. Just like I use a computer to build a computer. It's autocatalytic, like all technologies. The internet was autocatalytic because it allowed people to collaborate remotely.

So, I would say yes. This is anecdotal. 90% of calories are autocatalytic, which is 100% of what you would expect. And that's been happening for a while. I mean , it's not even new. Yes , but I would say there's also a focus now on moving toward whether We can actually get the model to train itself. This is kind of like the automated research that Karpathy did, but now they want to do that. There are teams doing it. It's not as many calories as you might expect.

And I feel like I have a good sample of this because everyone comes and talks to us. And they say, " Cool." So, can we be one of the inspectors? Can we get all that data and then we all look at it? Maybe there's nothing to see here, you know? Personally, I don't think it's very likely that those four criteria will be met. Greg Brockman was on the podcast this week and said that we're in the age of AI. I asked him a question after he said that, and now everyone's saying we have AI.

Well, you know, people do what they say. But, but, but people have been saying for a long time that AI is smarter than most of the people around me most of the time. That's something they've been saying since the third or fourth quarter of last year. So I asked them how many of you have hundreds or thousands of agents that You manage things that coordinate with each other in swarms and negotiate and, you know, automate your life and everything around you. And if that's the case, raise your hand.

It's like hardly anyone raises their hand. Of course, Martina has done it at home. You know. Most companies use Microsoft Copilot. Yes. That's the extent of their AI. Most of the companies I spoke to when I asked them this question said, "No, we don't have anything like that." So we said, "Well, what are you doing?" "They're using a chatbot. Like they're asking a chatbot questions. Basically, it's an efficient, much- glorified old-school Google search. The results are simply a faster Google search. And then there's the coding happening.

Some people are using it for coding. Although the ROI is, you know, we can discuss the ROI there. But there isn't a job like an agent that has automated the entire enterprise. That just hasn't happened. So why is that? And I think the real reason, if you analyze it, is that the models are smart enough. But they just don't have the context that exists within any organization. It's like they haven't been in all the meetings. They do n't know what's on everyone's mind. They don't know all the processes.

They do n't know. There are always a couple of employees who know everything in any organization. You know, you tap them on the shoulder and they're all like, ' Oh my God, what if he or she quits?'" "They don't have that context. And if you just merge that and give that context to AI models—just the frontier today— I think there are a lot of productivity gains you could get for any organization on the planet. For that, we don't actually need smarter models. We don't need a smarter model that can solve Navier-Stokes equations, other conjectures, or do better on humanities exams.

Like, we need to go from 60 to 70 percent. None of that is necessary. Well, I think people are actually very upset about something like, 'Oh, it's replaced the frontier.' But really, if the frontier isn't advancing, it does n't really matter, I think, for the vast majority of organizations on the planet. They're way behind on the adoption curve of actually automating things and getting value from them. But it would be disastrous for the labs because the price of intelligence is falling asymptotically. It's going down by a tenth every six months or something like that.

So, that would drastically change their businesses if they weren't expanding the..." Limitations. Yes, but this is what we should be focusing on, right? We should be focusing on, you know, there are two sides to every coin. We talk a lot about costs here. For example, there's a cost-benefit analysis that we should be doing on everything, right? We've talked a lot about costs here . For example, is there an existential threat? Are there any cyber risks? Are there things we should be concerned about, etc.? That's the cost aspect.

What's the benefit? And I think this has now become a public thing, and the public cares about AI. They're asking, "Hey, what's in it for me? What's in it for me ?" "It seems like nothing's happening. So how do they manage it? What are some of the use cases you 've seen to date that have perhaps surprised you in a positive way? Yeah, I mean, first of all, there's a lot of concern about existential risk, etc. So I think a lot of people don't know what an interesting use case is where people are doing interesting things.

We have a lot of use cases that are fascinating. One I like is Crisis Text Line. They actually use extensive language models with us to detect if teenagers want to self-harm or commit suicide. Wow, yeah. It's an amazing use case. And it actually saves lives. It's a great company, and that organization is doing incredible work. Another one that's interesting is the Omnipod, which is for patients with diabetes. They can put the Omnipod in, and it uses AI to actually learn insulin release and glucose levels and actually release them accurately.

I don't know if you remember that before, people used to inject themselves regularly, didn't they? But now this happens automatically, and it's like an AI that learns on its own." Your body. That's a great use case . Zipline is another one. They're doing great, but when they started, they were like those drones that were completely automated, all controlled by AI, from battery optimization to routes and everything, and they were delivering food to areas in need. Blood to refugees. Yes, it started in Africa and then in other parts of the world.

So , yes, that's it, yes, it's an AI use case, built on Databricks. Well, that one's great, but there are other, more advanced ones as well. One that I like, but that's harder to explain, is this model we created with Merck, a transformer-based model. It's called Teddy. Transformer-enhanced drug discovery. Yes, and they publish the research so you can check it out. Basically, it's a model: instead of predicting the next token, it predicts how the gene regulatory network , the GRN, will respond, and it can detect which cells are causal and which are just reactive, meaning they're just reacting, and so they can start using this in drug discovery and reduce significantly reduces the costs of developing drugs that target specific diseases.

That's a great use case. There are many of these. You know, Genie, I mentioned, you have this ontology and you can ask any question. Novo Nordisk is using it. So, you know, they created this drug GLP-1. Well, but what Novo Nordisk is doing is that they're now using it for all of their trials that they're conducting. Wow. It can reduce the time it takes to get information. For example, if you 're doing a study on obesity or something like that, from weeks to minutes. There are many amazing use cases for AI.

We shouldn't forget these advantages. We want all these things and we don't want to be at odds with them. Yes, exactly. You're absolutely right. So how do they, say, plan for the next 12 months? How do companies actually extract value ? You've left out the word context, but how do they operationalize it? It's actually harder than most people realize, but, well, first and foremost, we have to make sure that we've digitized everything that happens in a organization . You can't just wave a magic wand and make that happen.

So, you know, every meeting needs to be transcribed. You have to be able to get all the context from every meeting and everything that's going on . All the digital content needs to be sent to the AI. So, you have to create what we call an ontology. We did create it. But first and foremost , you have to collect that. That in itself is a problem in organizations because legal teams will say, "Don't record every call." "Don't record everything." So, you have to do it in a way that defines the ontology for everyone because I know Palantir uses the word a lot, but it's not like they own it.

What does that mean? And for the people listening, how should I ...? Yeah, I mean, you know, ontology just means that in an organization, the relationship between all the abstract concepts of all the goals, all the departments, all the people, and all the projects that are going on . What exactly do they mean? And what is the relationship between them, the people, the resources, and what that company does? Well, it's the difference between a person who is new to the company and just started today and a person who has been working there for 5 years .

Let's say they have the same capabilities. They have the same educational background. They are equally intelligent and hardworking and all that. But for one, today is their first day of work, while the other has already been there for 5 years. Yeah. What is the difference between these two people? One has an ontology of how that organization works, Who are the people, how are things done? Don't look at the org chart. Don't go ask that person. Yeah. They won't do anything. Go ask this person, you know, they'll do it for you.

And that's not how it works. You don't need to submit that paperwork here, and, you know, this is this project, this is what's happening, this is essential. So, there's a lot of ingrained knowledge that's in everyone's head. Who knows how an organization works? That's why people in the startup world say, "Hey, if you lose most of your people, that company can't recover ." You can't just replace and hire new people. People are so essential. How do we obtain that context? That's the ontology, and we're giving it to AI.

Part of it is that we just need to have, you know, the recording and all that. But the second part is, how do you actually condense that into a graph? Actually, a digital graphic. That you can then feed to the AI. So, the way many of the agents work today is as cloud-based code, and any of them—X code or pi or , you know, open source—or you can review the whole set of them. You know, they have this agentic loop. He can reason, but then he goes and checks each resource one by one.

Then, it will go to this MCP server for your question and try to see if the answer is here. Is there another one? He summarizes it and gives you an answer. But it's a bit slow. I compare it to if Google had built Google Search this way 25 years ago, we would have said, "Okay, let's get 10 blue links. We're looking for keywords here." But instead of giving you 10 blue links, I would have gone to a website, summarized with an LLM what it does, found some hyperlinks, jumped in parallel to some of them, read some websites, done that for 10 minutes, and then given you the top 10 blue links I could find.

Well, that would be very expensive; it would cost a lot of money to do that every time, to log onto the website. Two, it would have taken a long time; you would have had to wait 10 minutes. And three, the quality would be bad because you're really only seeing a very small subset of everything that exists, right? So how did they do it? They have an index. TRUE? You never leave Google's servers. Search...Get to the index, the reverse index, it immediately gives you the 10 blue links in, you know, less than 100 milliseconds.

Did you do the same for AI? So, the ontology is that we need to calculate that index offline all the time. So, it's almost like the PageRank algorithm that Google invented back in the day, but it's more complicated because Google was only looking at a website that anyone can access. Here, there are permits and... Yes. There are permits involved here. The data I am allowed to access may not be the data you are allowed to access. Therefore, there is privacy, there is access control. Furthermore, there are many different types of objects we are dealing with, not just websites.

Well, the problem is a little more difficult. But it's manageable. In fact, you can do it. Well, I'm convinced you can do this and you can make huge productivity gains because we do it for Databricks. Yes, exactly. We did it for ourselves. Yes, yes. Yes, and it's as if the company has completely changed. It's not like it used to be a year ago, I would say, because of all this. I mean, you've been doing dogfooding. Databricks for Databricks forever , but perhaps you could say more about the impact you've seen as an organization.

Yes, I mean, once we got this ontology and started working on it, in fact, we probably have the biggest of all our clients, we have the biggest ontology. Our ontology is bigger for us than that of any of our clients when they use us to build their ontology because Databricks uses Databricks more than anyone else uses Databricks. And, well, there are like millions and millions of nodes in the graph, in the ontology graph that we have. Well, it's just, you know what happens in an organization?

What happens in an organization? You have a tree-structured organization where information flows up and down through it. You know, if you can't make a decision, you tell your boss, maybe he can try to break it, and things get complicated. So, if you catch up on what's going on and get the full context, then they make decisions. Once the decisions are made, you have the ability to filter them within the organization. Now, AI can do a lot of this if you have an ontology. Because? Because, well, what happens in a meeting?

In a meeting, someone has done the analysis. I probably have a PowerPoint presentation with some nice graphics. The person who did the analysis is an intelligent person who used Excel, made some models, there are some numbers. So now you can do a lot of that with AI. So AI can do the analysis for you. It has all the context. You can present it however you like. You can ask questions about it instead of having follow-up meetings . You can ask questions directly to the AI. Well, it's very similar.

It's in line with what Jack Dorsey has said you can do in the organization. It's just one specific way to implement it. Well, it's a radical change for us. It's like everyone is on their phone at Genie meetings now. And they ask Genie questions. You can see it as soon as someone says something complicated or something like that and, you know, you see everyone going to their phones. Can you share that financial anecdote you mentioned once at a board meeting? Yes, it's an internal board meeting.

Yes, well, just one kosher one. Yes, exactly. No, it's actually necessary for one of our presentations. I need to know how many Fortune 500 customers we have who use... What is our Fortune 500 penetration? And I asked one of the sales operations people because I thought they would have it. And he replied, saying, "Oh, sorry, I can't log into Genie right now . I'm on a flight." And I said to him, "Well, if you're going to log into Genie, I can do it myself. I don't know.

I asked you because I thought you had something reliable that I don't have access to. Then, I got a little annoyed. So I texted the CFO and said, 'Dave.' Then I texted Dave and asked, 'Hey, do you know what our Fortune 500 penetration is?'" " And he copied and pasted a screenshot from Genie. That's it. So he has it too. So I said to him, 'Is anyone doing anything new here, or does everyone just go to Genie and ask the ontology, you know, if they have questions?'" "It's like, 'Let Genie do it for you.' For example, can you get it from the ontology?

I think it's a game-changer, but it's not as simple as pushing a button to have a working ontology across the entire organization. I think Palantir has done a great job going into organizations and getting a lot of that tacit knowledge written down and getting it to the organizations. We automatically receive it and create the graph. Then we pass that graph to the agents so they can answer the question and answer it in the way that business leaders would like to see it—in graphs , analytically, and in a way that allows you to interrogate that question .

And, you know, keep asking questions and getting answers to them so you can make decisions. And then send that information back to the organization. Yeah, it's pretty amazing. You've marked as favorited that developers are using AI, a questionable value. I want to talk to you more about that because I think you were among the first. Well, and I say I don't want to use the word ' token maximization' because it has a It has a very negative connotation, but I think in terms of applauding people who can use AI to be more productive, you guys were at the forefront of that, right?

And then , of course, there's this cycle of, "Oh wow, people are wasting, now we need to maximize value." So , what was your own journey with that? What are your thoughts on maximizing value, not maximizing tokens? So, I'm going to include Unity Gateway in this, right ? Because I think the cost management aspect is becoming increasingly important, and you guys are helping people do that, but maybe linking it to extend it; that's a good point. Yes, around Q4 last year was when, you know, the models got really good, and we started noticing that, well, they're actually starting to deliver a lot better productivity.

Well, I started using the models myself to start pushing code to production for blocks of data, like the real thing, because I want to get them into production. So, I did that and started pushing the organization to said: "Hey, everyone has to do that." I've done it, why don't you? If the CEO can push code to production and a highly confidential data platform that has all these security requirements, you should be able to do it too,” whether you’re a manager or anyone in the organization. So I started pushing everyone hard, and we started doing leaderboards in Q4.

By early January or February, when the year started, we were already at full capacity, everyone was using the material, and we were pushing and managing this. But, well, the whole token maximization thing was happening in February or March; it was already happening. Well, yes, we were fortunate to be a few quarters ahead of people to see what was going on here. And it was getting out of hand. Well, we already had a gateway. It’s called Uni Gateway, and it was already being used to provide token capacity.

So you can get capacity from OpenAI , Anthropic, Gemini, and Grok. Any customer can come to us, and we’ll provide that capacity because we have a relationship with them. And any open-source model . Well, we started setting budget constraints and warning people, like: "Okay, you have a portion of your budget left." "You 're approaching your ceiling." We started doing it per person and per group. And then we started doing really good analytics so we could predict exactly where the costs were going. And then we added smart routers that could choose cheaper models if you were getting close to your budget or if you had simple questions.

We started doing that. We also created a harness called Omnigent, which can multiplex between different harnesses. It turns out the harness itself matters. Because you can use the same model with different harnesses, there's almost double the cost difference. Even with the exact same model. Yes. You know, the same version, but with a different harness, you get double the difference in the actual cost. So if you can change the harness, you can get a lot of cost leverage. So we started using all of this. We were really able to, uh, bend the curve, and in fact, our cost for AI has basically been that the tokens have continued to go up, but the costs have plateaued.

So that's been really very, This is very important to us. And there's a huge demand for it. I think every organization is going through this now. Yes, yes, for the first time I'm doing a lot of board meetings, so I'm on about 20-something boards. For the first time ever, a large-scale company said last week that they were moving from frontier models to GLM. This is a large engineering organization. Do you see this? Do you think that's a trend or do you think it's just an isolated anecdote?

Because I've been hearing about... Remember the first deep search moment and then the video stock price spike , and that turned out not to be real? And then the KB moment and the next deep search moment. None of that seems to have had an appreciable impact on the market. But now the number of anecdotes I have are quite real, and it seems to be happening. I'd love your take on it. I mean , I think people want both . They want, you know, they want the latest super-intelligent model for the tough task where they get a return on investment.

But then there are a lot of other things Mundane and silly. Like, you know, you guys literally use your harness to rename files and stuff. Yeah. You know. Like you're paying, you know, orders of magnitude more for that. Please write it yourself. Don't let the model do that. It'll spin for 5 minutes and then rename the file for you and it'll cost you, you know, pennies. But I understand what people are wondering, do you really see market movement on that? No, people are moving, but what they 're doing is, you know, the pattern is: either use, you know, you can use this expert pattern where you have, you know, a small, cheaper, open-source model that uses an expert model, the big ones, or vice versa, or a way where they can , uh, ping-pong with each other.

But also multiplexing hardnesses and just changing hardnesses so you can control costs is also what people are doing. You know, people have found, for example, that Pi is very efficient when it comes to of a harness. Yes. Well, yes, I think there's going to be a lot of these. It's easy. The models themselves are stochastic, as you said. They give a different answer each time . And they change a lot. So, there's a lot of experimentation. Therefore, I think we'll get to a world where the smartest model won't always be used for everything, which has been the paradigm for the last few years.

When a new model comes out, it's super smart and they use it for everything, even very simple, everyday tasks. Yes. I'll tell you what I see. I see people using Fable and Astra for architecture. Uh-huh. A cheap model for implementation and then Fable or Astra for auditing. Yes. It seems like that's emerging. What do you guys think of startups? I mean, aren't they...? Well, that's it. Honestly, that 's the pattern of open source. Per token or per dollar? Either one. So, per dollar, open source is like 5%.

That's very little, but by token count it's more than 60%. Yes, I was going to say, I mean, we talked to, say, Decagon or something. Well, I think it's a different internal versus external use case for the product. In the external case for the product, I think almost 90% is open source. In the internal case, and I do n't want to say for Decagon in particular, but a lot of them say they do n't care. We'll just use Frontier. We don't think about cost control. But as it grows, yes.

Sure, you and I were in another board meeting where they really scaled it back just from a waste perspective. So I definitely see it moving more towards open source on the product side. And actually, that ties into another question , maybe about open source, but specifically about post-training. I feel like you came a little early. I remember talking to you back in 2023. When did you buy Mosaic? 2023. 2023, right. So this vision you had in 2023 came to fruition in 2026. I don't know if you'd agree, would you?

It's something like what we 're hearing, of course. Oh, something like, "Hey, we're going to own your own intelligence." You're going to be, you know, post-training your open-source models, etc. "And that's definitely what startups are doing. I don't know if that's what companies are doing yet, but do you feel like you were early on that or...? Yeah, I mean, uh, first of all, you know, when we started it was also, uh, we're also pre-trained for you, which doesn't make any sense. You know, you can... now there are really good pre-trained models that you can use.

Right. But that can be done after training the model, and you can apply reinforcement learning. Yeah, we're doing that at scale. And a lot of the startups are customers. So, we help them, you know, using our reinforcement learning or LE environments, where we can make the models really, really good at the specific task they're doing. It makes a lot of sense for them to do that. If you have a repetitive task, for example, if you have a startup and you offer a product and that product does something specific.

It's not just general intelligence. It does something specific for you. It makes a lot of sense to take a really good open-source model and, you know, use reinforcement learning and To get really good at that specific task. You can reduce the cost. You can do it very quickly. They own control of their own intellectual property. So, in that sense, it 's possible. But large companies only need basic automation, and it's too much for them to do this right now. I think one of the challenges is that you need good assessments.

Absolutely. And making good assessments is difficult . Well, while startups can do that and are motivated to do it, other organizations might choose to use a front-end model rather than having to create their own assessments. We actually generated customer assessments automatically in the product, and we had them front and center. But people didn't want to use them. So we said, "Okay, let's move them to the back end so they're optional." And then they never used them. Well, I would say, generally, why? They just don't want to get into it.

It's too complicated, or I think they want something quick, fast. Yeah, that makes sense. Kind of reinforcement, you know, "Hey, there's a new model." I want to try it. I want to solve this problem. You don't have time to do this using the scientific method of conducting an evaluation. Let's have a good baseline." And it's kind of like TDD, test-driven development. You know, in software engineering, do people actually do test-driven development ? Very few did, right? Everyone said it was the right way to do it, but nobody actually did it .

Well, it's the same thing; it's kind of the curse of, you know, training your own model—the evaluation is the hard part. I know you have an FD model at Databricks, which is a very popular word or acronym right now, but do you like having them at such large companies? Is a full FD model required, or how do you do it, and how has it evolved? Yes. Yes, I mean, we've had FD, and the demand for it has increased significantly. Yes. A lot of it is, you know, how do we build that ontology?

The ontology is automatic, but if you don't collect any information, you don't record anything. Right. Well, that's one of the key things we do, but also things like, for example, I want to create an agent. I want When we put it in place, I want it to be customer-oriented and have very low latency, and I want it to have safeguards so people don't abuse it or ask it things we don't want it to answer, etc. So we can create something like, for example, the sports AI that Fox has.

You can chat with it about sporting events. You can try asking it about politics, and it's very good at rebuffing you. Move it and talk about sports instead. So that's what the feds created . So, you know, we'll help organizations get started with AI. It's important. Because it's simply... Many organizations don't have the internal expertise to create these things. So they just need a little extra help, and then they get started. It makes sense. This is more related to the agent side, but I recently saw that I think a third party—a neutral third party, I believe—did some tests that showed that Lake Base or Neon was actually the Postgres database data chosen by the agents.

And I thought that was interesting, firstly, because, you know, exciting data works, but secondly , I probably wouldn't have guessed that back then. One year. Surprise for you. It was a surprise only because there are others out there who also have, you know, a lot of momentum as developers. But it was clearly number one. So I'm curious to know how they figured it out. And what makes them win over the agents? Because if you win over the agents now, you win the market. Yes, I think a lot of the credit goes to Neon, Nikita, and the team.

And I think what they 've done is just become obsessed with how to get models—how do you get models to choose an agent that favors Lake Base or Neon as their database? So what did they do? Agents want to experiment. You know, they go off, try to build a little piece of software. They need a database. So you need a database that loads fast. That's why they had this obsession that everything should take less than a second. So the database loads in well under a second.

You can clone gigantic databases. A gigabyte database can be cloned in less than a second. It's highly elastic and highly responsive. And then they created this amazing feature called branching. Branching allows you to branch the database, and you can have many, many branches of the same database. And they just made it very, very lightweight. We saw this with other things with agents, right? Like UV, you know, rip grep, like a busy reimplementation of many of the tools in Unix, making them really, really fast and lightweight, and also kind of a failsafe for agents.

They just did this with a harder problem, which is the database. Now you have a Postgres database , and the Postgres database has all these advantages: it's very fast, agile, and failsafe. You can go back to snapshots and do those things. That's why I think it's easier for agents to use. They also made sure they had a pricing model that wasn't the kind where you want agents to build software and experiment, you don't want the cost to go up. You don't mind paying for your database if it's for production use and they use it.

A lot of people. But just to experiment, so I think they were obsessed. They weren't trying to win the database war or be better than other vendors. They were obsessed with, "How are we the best for agents?" Yes. And that's a new personality because in databases, the obsession has been, "How do we help DBAs, how do we help developers and those who use the database?" Yes, exactly. They changed the game and said, "Hey, how do we focus on agents and help agents get the best database they want?" "." And you know, now more than 90% of their databases created in Neon and Base are created by agents.

They're not even human. The numbers speak for themselves . By the way, it's extraordinary. I started using Neon as my standard database and it felt strange because , as is often the case when you join a large company, things slow down. In fact, the products have improved considerably. Yes. Are they completely independent? Do they work with the rest? No, it's a great team. We work very closely together . We love databases and data. It's something we experienced. But now our team does a great job creating very fast, agile, and cool things for agents.

Okay, Ali. What's your P Doom? Less than 10%. No. Almost zero. And yours ? Do n't know. I'm just saying that my only answer is that my P Doom without AI is much bigger than my P Doom with AI. What do you think? Oh, that's another one. And you, sir? I made a technological pilgrimage . I agree with Ali on this. Yes. Okay.

Go deeper

• All A16Z episode guides