Jev 如何把 AI 变成真正能办事的软件

Where the hell is all that automation? AI is incredibly smart, but so useless at everything else. It doesn't matter how many AI-based coding agents you use; the software doesn't actually get better. Maybe you're just running it faster. It seems like you're getting worse. OpenAI has been trying to automate customer service since 2020.

Instead, I want smart software. I want to expand the capabilities of the software itself so that what should be automated finally becomes automated.

What I like most about your slogan is, “We build a product, not a god.” This is so cool, because if we had any other leader of a big lab, even if they were happy, they would hide it. Your view is so different. You say, “No, we will create a much better world.” Due to certain nuances, I don't think we're on the path to an AI apocalypse or anything like that.

1. Meet Diogo and Type Safe

Today we have with us the founder and leader of TypeSafe, Diogo Almeida, who is a true hero to Martin and me. Not only is he creating a really interesting product, but he's also starting what we think is a very important movement. We are very happy about today's meeting. Welcome to the show.

Thank you.

Maybe you could briefly explain what Jev is, what TypeSafe is, and why it is important?

Is this a Chris-friendly format or not?

Yes, yes, yes. Okay, cool.

I was once asked to do an “elevator pitch,” and I tend to get carried away and I'm not good at it, but I realized that my favorite pitch for Jev is, “Where the hell is all that automation?”

It's incredibly tragic. There's so much intelligence. AI is incredibly smart, but it's not that I hate chatbots or coding agents. I adore them myself, but they are so useless at everything else, and it's just a tragedy.

It's tragic because we have such a huge diamond in the rough that's not yet ready for work. That's why TypeSafe builds AI for software. We want to make AI powerful not only for people, but also for creating real software, and Jev is, for us, the first model in this area that will significantly improve automation.

Yes, it's interesting because it's already become a real hit in the programming world. One of the things that made us ask, “What the hell is going on here?” is that every developer we know calls us and says, “Oh, this is incredibly cool. It's fast. This is great. Everything is getting better.”

How is that possible? Everyone thinks, “Well, we have Cloud Code, we have Codex. Don't we already have that?” What's the difference, and how does it lead to true automation?

I wish I had some visuals, because I have a favorite picture for this. I like Cloud Code and Codex. I like Gary Tanenbaum's description of it as just-in-time software. It's an incredible way to describe what they do. It's like creating software on the fly, and you can program it in natural language, but it has the same expressive power as regular software.

I want smart software instead. Instead of automating software development, I want to expand the capabilities of the software itself so that what can and should be automated becomes automated. In more refined language, I want to express things like intention. I want to expand the vocabulary of what we can do.

I can talk about all sorts of weird sci-fi stuff I want, but programming is about hyper-detailing valuable things and then endlessly reproducing them. It's so cool, and I just want more of it.

2. Smart software, not just faster code

One way to think about it is this: instead of a tool that, to some extent, replaces the software engineer with a faster, maybe not even as good, engineer, you say, “No, no, no, we're going to give existing engineers superpowers to write much, much better and more interesting things.”

Yes, I actually mean that. By the way, I think a lot of people miss this point, and it's so subtle and so important to get it out. If you're using something like Claude Code or Codex, which is great, or Cursor, which is also great, they write code, but that code is the same as what a human would write. Maybe it's better, maybe it's worse, but it's basically still code, the same as code from 10 years ago.

The thing about Jev is that, whether you're a Cloud Coder or a human, you get this new primitive—a new thing that you put into the code that actually extends the capabilities of the software. So instead of writing code, it's something you add to your code.

Which one? Continue.

No, you continue.

What's interesting, by the way, is that it's a very powerful primitive, and it would be great if you could explain it, but it's also a little different from how programmers think. For example, there's the concept of probabilities, and therefore, perhaps—

That is, the intelligence level within the software. Like a library where you can describe what you want in natural language, add a state machine, and it chooses what to do with a certain level of confidence, which has never been done before on this scale.

There's a lot of nuance there. I'll first jump into one thing I like about what you said—the question of where the hell all the automation is. I love software so much. I would like to write it all day. I don't recommend people to be CEOs, but I don't care.

It's strange that AI is so cool and the software hasn't changed in 10 years. For me, it doesn't make sense, and the most we can do is add a chatbot on the side that can sometimes perform actions, but not all of them, because some actions are not reliable enough. I just wanted to make that little digression. I like this moment, so I'll go back to the idea that it's a slightly different way of thinking.

Yes, I think “machine-native” doesn't exactly match the bits perfectly. This is, in fact, the art we are working on. On the first day of onboarding, I draw a Venn diagram of what AI is good at and what is valuable in code, and we are in the middle.

We don't output, for example, extrapolated floating-point numbers, because AI simply doesn't do a good job of that. But things like probabilities are nothing new, and this is like the argument over whether Jev is just a classifier.

Jev is definitely a classifier. Classifiers are cool. They were created to be useful. They are designed to be useful. In general, this is the same interface as some natural-language-processing concepts, because they came from practitioners who were trying to make systems work.

What I see now is that Jev, in my opinion, is probably better than a machine-learning team from 2019 that would do it for you, because you can program on the fly. Who knows what can be built? In 2019, there weren't many good machine-learning teams to fill narrow niches, collect datasets, measure them, and do all that. This is just the beginning. I feel like there's still a lot ahead.

Do you think there's a certain slider here, where at one end there's “language at the input, language at the output,” as we have today, and on the other side there's a regular imperative program? You can move between them, or do you think this is a point in the design space—language input, machine-state output—that will become a universal tool for programmers?

That's a difficult question. I'll say what's on my heart. Deep down, I think it's a slider. When I was designing our primitives, I might have made mistakes due to my own preferences, but right now my main guideline is intelligence per dollar. To be clear, this could be a mistake. Intelligence per second may be more valuable in the short term.

Even our interface—we call the input a “state”—was designed that way intentionally. It's the same as with a code patch. This should be an internal part of programs. Deep down, we are truly optimizing for that. Most of my work is about even more complex internal state structures in programs. Can intelligence be added there? I think it will be a constant struggle.

We take design very seriously. Pragmatically, I think certain things happen by themselves. For example, it's easier to create AI that works at millisecond scales, so for a while it will be more like a database than a standard library. But I would like to see this become a standard library as well.

Can I step back for a second? What alchemy creates someone like Diogo? When a man and a woman love each other—

But listen, you're talking like an AI researcher. You're still talking like a systems engineer, and you're talking like a programmer. Usually these things don't overlap much, and you take the AI that we've developed as an entity and turn it into a programmer's tool. Tell me a little about your personal path, which is somewhat unconventional.

My path in AI is somewhat unconventional. I was a mathematician. I was a mathlete. I describe it this way: I was pretty good at math—it's a little embarrassing. I was good enough at math to attract girls, so that's pretty good.

No, that was the point.

You must be pretty cool. What kind of girls do you attract when you're so good at math?

This is interesting. This is something our audience needs to know.

We have to inspire young people.

Don't do this. Young people, don't do this. It's not worth it. Just be cool, calm, and interesting, and don't try to seem better than you are. I can't believe I said that.

I was a mathlete, but I've never actually liked math. I never made any effort. I was just a big fish in a small pond. For me, mathematics was the path I was directed toward, but I hated it because it was all about winning competitions.

Computer science is actually very similar to mathematics. It's essentially math, but cool and useful. It's fun and interesting, and I still love conducting algorithmic interviews.

Is this the best I can do? I don't know, but do I like it? Yes. And does that allow me to judge people very well? Yes, it does. So I love computer science. I consider myself a computer scientist much more than an AI researcher, despite my background.

What really got me into this was that I also won a competition on Kaggle. Not through complex math, but by automating everything with more and more nested loops. It was a systems problem, and that event ultimately led me here. For example, I was forced to speak at NeurIPS. It's usually an honor, but I hated it because I just wanted to be in the thick of things.

Was this from that Kaggle competition?

Yes.

Wow, wow.

Actually, the presenter on Kaggle was Isabelle Guillon, co-author of SVM. In fact, I think she was the first author of the SVM paper, although I'm not 100% sure who the first author was. She saw that I was someone who didn't fit into the research community at all, took me under her wing, introduced me to all the AI experts, and my career went in that direction.

And then OpenAI?

No, it was a startup with Jeremy Howard.

Are you kidding? I adore Jeremy.

Yes. Fantastic. And then Google Brain for a while. Eventually, I got tired of doing nothing. I thought, “You know what? AI is a hell of a lot of fun.” I joined OpenAI for that very reason, and it worked out really well—surprisingly well.

So you said something that's quite unusual in today's world: AI is actually a lot of fun. Your company also has a completely different behavior and view of AI than everyone else. My favorite thing you say is, “We create a product, not a god.” It's so apt. If it were any other leader of a large laboratory, they would try to hide their joy, even if they felt it.

Your approach is so different. You say, “No, we're going to create a much better world, and it's going to be great. There won't be fewer jobs; there will be more jobs, and they'll be much better jobs. Everyone will have a great time.” Just being around you shows that you truly believe it.

So tell us about it, because for us, Safe Superintelligence is more than a company. It's a whole movement toward a positive future that most people in the AI world don't seem to like, or at least don't share.

I think they don't understand it. You know, it's like complaining about a classifier at the level of machine-learning problems while everyone else is having a party with Jeff. It's like, hell, we can do whatever we want, and if you're not a developer, it's hard to understand what's really going on.

I agree with that 100%.

I really think they're painting a pretty bleak picture, which I obviously disagree with. I think it comes from this blind belief in the monolithic model that everyone believes in.

Exactly. One big brain to rule them all.

This is what we did. It sounds much more sinister.

Yes, that's what people hear. Of course, that's what people hear.

But is this one brain really on its way to ruling us all? We still haven't automated very simple things that, in my opinion, people shouldn't have to do. There are a lot of really basic things.

It hurts me when the world doesn't match reality. Part of the pain is, where the hell is all this automation? How can AI be so incredibly intelligent, and why are there so many financial incentives to automate everything? You can find an excuse for diffusion, but I don't believe it at all. It's not worth mentioning names, but this is obviously not true.

Part of the problem is that the mismatch with reality, combined with the fact that AI has enormous potential, makes it tragic to me that we haven't released it. So now it's like a little holiday for me, but I was afraid. And all the Jeff users, it's like—there's a happy AI, the people on Jeff, and there's a gloomy AI, the people who aren't there.

Yes, yes, yes. It's actually quite a fascinating dichotomy.

I'll tell you, on your point about automation, I had an interesting conversation this morning with David George, who runs our growth fund, because we were talking about new tools. I asked, “Have you tried the Muse thing?” He's like, “Oh, she's great.” I asked, “What are you doing with her?” He said, “I finally canceled my subscription to The New York Times.” And I said, “That's hard to do.”

3. Where's all the automation?

This is just the tip of the iceberg of the terrible things we need to automate.

I think if we want to be intellectually honest and truly strive for the “north star” of automation, we can't fall into the same anti-patterns that AI has fallen into—namely, focusing on individual use cases.

Many people ask me, “What are your favorite uses?” And I say, “I'm not sure if they work.” I want them to run in the background so that someone can trust them to run without having to check, and so that people can build something new on top of them.

And, you know, Layout, layout, right? Layout, but also security, right? This is a different type of security where if you want to run something with associated resources and access to things, you need guarantees to do so, or at least statistical guarantees that nothing will go wrong. Break it into pieces, make it fancy, something like that.

I don't think our models will do that anytime soon unless someone writes software around them. That would be a very cool achievement, and we should give credit for it. But not in a way that makes us lose responsibility.

I'm just curious: how long did this intuition take to mature? I remember talking to you, maybe in 2017. It was a long time ago. We really talked about it, and you already had a lot of these ideas. You said that data is important, and you said that you wanted to focus on the task.

Did you know that this would all end up being a classifier, or was it just an intuition that there was a different perspective on this whole AI movement?

There's a funny story about that conversation. During a talk in 2017, my topic was very similar. I think it was called something like “AI: Modular in Theory and Flexible in Practice,” which is very typical of software systems, so I'm sticking with that idea a bit.

I think this all actually started right before ChatGPT. When we released these things, I had no premonition. To be honest, I was very, very pleasantly surprised by the generalization capabilities of reinforcement learning from human feedback.

When was that?

Probably at the end of 2021—maybe the fourth quarter of 2021. It was really very general.

If you read the paper, unlike others who are trying to prove themselves right, we actually tried to disprove it using the scientific method: Is this a scam? My favorite question was, “Why is it important to eat socks before meditation?” We made sure that this was not already on the internet, and the models were able to provide plausible, human-like responses.

For us on the team, it was a moment of truth: this is not a scam. In machine learning, you should always be wary of fraud.

What really blew me away was how we released it. I'm obviously a big fan of the possibilities of AI, and I did a lot to release this model. I really thought the model had a good chance of becoming SHGI. When that didn't happen, my whole world just collapsed.

So you were a little on the other side for a while, like that “Crazy Train” or...

Well, no. I'm just like, no, no, no, no, no. RL generalizes. Maybe we have SHGI. Reinforcement learning from human feedback generalizes quite well. Reinforcement learning from verifiable rewards is something that doesn't generalize that well, from what I saw.

You were behind the scenes at ChatGPT. You were behind those early GPTs. It was a completely different goal: creating a chatbot that would communicate with a human, not a programmer's tool, and so on. So I'm just curious—

Actually, at the very beginning, sometime in 2020 at OpenAI, when we were talking about SHI, people were describing it as Elijah and every statement he made. It's not exactly like that, but we were talking about the definition of AGI at OpenAI.

Part of it is that it's intentionally blurry. It's a wide tent so everyone can be inside. But due to certain nuances, I don't think we're on the path to self-improving AI. I didn't think so before, and I don't think so now.

I really think that what OpenAI has defined as general-purpose AI is quite achievable. Automating most of the world's economically valuable labor sounds like—oh, I don't know. There is so much different work there, but most of it is very routine and simple. In terms of volume, to be able to outsource the work, you need simple instructions that anyone can follow.

As far as I can tell, the level of intelligence required for this has been in these models for quite some time. My main complaint is, “Why isn't this available yet?”

Since the advent of reinforcement learning from human feedback, the industry has been divided: big promises but weak results. I think GPT-3 was pretty balanced at the time. But when humans judge the quality of the models, they seem to be very accurate, because humans are the judges and we optimized for that judge, not for automation. That was the missing link.

I would say that's when it dawned on me: Why isn't this thing more useful?

So you think the measure should be how much we can automate real, productive tasks? When you say “many promises and few results,” is that the dimension you're referring to?

The ability to automate tasks. In my heart, it feels like cool science fiction, and I believe this is the first warning sign for such “cool” fiction.

Are you seriously claiming that the math is solved, or even that 2 years ago GPQA—the questions that even Google doesn't know the answers to—was solved, but we still can't handle ordering a self-driving car? This is very difficult to grasp at the same time.

4. Is it just a data problem?

And I think a lot of people don't have a good answer to that. Can I test one thing? It may not make sense, but isn't there an argument that the real world is different from the digital world? It has heavy tails in its distributions; there are many outliers, and we don't have all the data.

Couldn't it be that the reason we don't do productive things in the real world is simply that we don't have data for this distribution? We don't train on it, and that's why it's all reduced to these lower-dimensional manifolds, like math or code?

I don't quite agree with the data argument, as far as I'm concerned. I believe there is a long tail. Certainly, it would be kind of silly to deny it. But I don't think, in my situation—with the canary in the coal mine—we need to automate this long tail.

I think we need to be extremely pragmatic about everything. Creating reliable software is always an investment, right? Do you know what the three great virtues of a programmer were? Laziness, so as not to do it again, arrogance, and there was a third.

No—yes, I remember. This has been around since the days of Perl.

Yes, yes. There's a third one, dude. I would like to remember it. But laziness means spending 10 hours automating a 5-minute task once and never coming back to it again.

This should be a solution based on return on investment for people who automate things. I just wish it could be automated.

Yes. And I think people will just create new kinds of work, hence the Jevons paradox, when that becomes possible. But as a benchmark, I feel it's useful to see if we can actually automate what AI seems like it should be able to automate.

OpenAI has been trying to automate customer service since 2020. It's not that simple. It's pretty amazing. It's just crazy.

Well, within companies, very little is automated right now. The projects didn't work.

Yes, apart from programming, which works great. Can you perhaps categorize the types of problems that you think are easier to automate? This is quite interesting.

We were actually looking at support even before the current wave of generative AI. It was interesting to meet the company, and the company said, “We respond to 95% of all support requests.”

And I'm like, “That's so much.” But then you really look at the data and understand that it's all about resetting passwords. If you were judging by uniqueness, it was 50% or something like that.

So it seems that when you deal with people and natural systems, there is a very long tail of exceptions.

To what extent did you even foresee such a wide range of use cases? I probably get a message every hour that someone is using Claude for a new task. I'm like, “I had no idea.” Did you expect this to happen? Are you surprised by it?

Extremely surprised. I didn't expect this to happen. This launch wasn't something that anyone would have expected. They'd probably have to be crazy.

I don't think anyone would have expected ChatGPT for developers, because ChatGPT was for regular users, and that's weird. I don't even really know what percentage of people who joined Jeff's party are developers. I can't imagine anyone other than developers using it. I don't know how they would use it.

But even my friends who aren't developers are just part of this Twitter party, joking around and all that. So, number 1, it's phenomenal.

5. Reliability and "do what I mean"

Number 2, this is going to be hard to convey in a short message because it was blood, sweat, and tears over many years. The extent to which I care about reliability is a lot. Reliability is the essence of this thing. If you don't understand this, it will be very difficult to create an analogue that meets the criteria.

It seems to me that every percent of reliability will be valuable to everyone, even if it's not the most valuable thing in terms of market capitalization, because it will simply open up new opportunities. We're fighting for all sorts of strange levels of reliability that we don't even fully understand, because we're just trying to introduce this “electric motor” of AI—of intelligence—into people's workplaces, and they'll decide what to do with it.

What does reliability mean in this context? Is it the availability of the model, or is it that I call the model and it returns the same result? How should I even think about reliability for something that is inherently stochastic?

It's not so much the first; the second is closer to the truth. I would describe the first as uptime or an SLA. I would call the second one closer to determinism, and the third one I would consider more like reliability.

I would describe reliability as the same intelligence every time.

Oh, interesting.

Yes. It's not exactly determinism, because I believe determinism is useful for unit tests, but not for real systems. Think about it: if you add a UUID to the query, the result should be the same because, functionally, it's the same.

Yes, but it is not entirely deterministic.

Right. I think there is another level, the name of which I don't know yet. Maybe it's what I would call a certain form of intelligence, where the result doesn't have to be identical, but it has to be intelligent every time.

If you were in that situation, would that thought be understandable to a human being? The developer can take this into account in the code.

Really, for me, the ultimate measure of reliability would be to reach a level where people can program with Jev without creating example queries. When you just trust it, you're in a constant state of flow, creating great, incredible software.

And many programs are at this stage right now, aren't they?

I don't think it's entirely unrealistic, and I'm perfectly fine with it.

By the way, this is a bit of a strange question, so feel free—if it's too strange, just don't worry about it. But it occurred to me that the value of things like coding agents actually diminishes if you have a primitive like that.

You could say, “Whatever. Codex builds all the software for me, but it doesn't use Jev, so the software it builds is somewhat limited.” Or you could say, “As a human, I'll write a program without using a coding agent, but I have this very universal primitive that makes writing easier.”

Do you see a future where coding agents use Jev and you manage them, and do you have any redundancy? Or do you think people are implementing it themselves? This is more of a question about coding agents than it is about Claude.

Oh, yes, definitely. I feel like I'm not as immersed in the intricacies of programming as I would like. You two may know better than me about this, and that's unfortunate, but my experience is that coding agents are very good at syntax and really bad at semantics.

I would say they're incredibly bad at architecture. I think architecture is the most human, creative part of software development. That's why I like using AI agents for coding.

I think this is almost certainly not part of the data distribution. It would be creepy if they were learning from our users' data. So most likely, that's not the case. But I don't see a problem with instructing them on syntax when it's part of the distribution.

The thing is, the models may not just be terrible at architecture; they may be at the 50th-percentile level. And if you yourself don't know anything about architecture, then this will be enough.

These are all gray areas and compromises that you will have to navigate. Sometimes speed is exactly the leverage your company or project needs. You're willing to choose a 50th-percentile architecture over a 60th-percentile one because you want to move faster, so Codex can run all night or something.

Yes, by the way, there's an interesting phenomenon in the market. When AI agents appeared, a “SaaS apocalypse” occurred and their value fell sharply. And when Claude came along, every SaaS company said, “This is the best thing in the world.”

So explain this.

6. SaaS apocalypse, reversed

I don't even know what else to say. It seems quite natural to me. In this story of the SaaS apocalypse, the idea that, in my opinion, has not been fully realized is that software is very cheap and easy to copy.

I can believe the first part, but not the second, because a lot of things happen under the hood. Maybe I'm too fanatical about software.

We all are.

Yes, okay, okay. I didn't know where this code could be. We have a lot of heritage in this regard.

So I don't think it really worked. The markets seem to disagree, but in my opinion, SaaS provides the same value as before. Maybe the markets are just scared.

But I think SaaS is going to be one of the biggest winners in this whole AI game. And I want to work really, really well with all those big, boring, user-aware SaaS companies because, in my opinion, they know best what workflows are worth automating.

What do people need? This is their main bread and butter. They spend big money because software is always a capital investment, but you invest it up front to make this experience even better, and then it spreads to the entire huge user base.

True.

So I think it's going to be—I’m not going to make any predictions about the financial markets—but in terms of opportunity, it's going to be like a reverse SaaS apocalypse, and I'm just excited about it. I need to come up with a name.

Yes, you should have a name.

SaaS-appaloosa.

That sounds a little too much fun.

All SaaS applications will suddenly become much more useful. And by the way, a significant portion of a SaaS company's capital investment goes into actually reaching all its customers.

So if you've already covered all your customers, and then you've not just made a chatbot in your SaaS product but have really, significantly improved the software itself...

Hmm, this is incredible. I don’t know if this is a realistic dream, but I think there’s a world where forms with options to choose from simply disappear. It seems to me that they just map natural language, which the software already understands, into Java-like output.

It’s literally been around since the 1980s. We used to call it 4GL. Remember the MR2? Fourth generation language.

Yes. I also think it’s from the ’80s. It might even be an insult. I wasn’t born yet.

I think the concept of “do what I mean” will reach a whole new level. If I can mention one developer app, I don’t know how reliable it is, so I can’t promise anything, but it was so cool. Someone was using their voice to control a computer, and they were constantly making decisions: Is this a command, or is this text input? Where is this text input?

This sounds incredibly cool. I think the interfaces might just change completely, and we might have to do it cheaper and faster.

Yes, then you’re already in Star Trek.

There’s a deep intuition here that, if you’re using AI to build software today, you’re still creating the same software as before. But if you look at the average pull request for a large company, it’s about 10 lines, right? Seriously, it really is. I was at Google, and we did some research. It’s about 10 lines.

You automate 10 lines. And, by the way, those 10 lines are part of the learning from the client or something, so you’ve optimized something that’s actually pretty minimal. What it doesn’t do is give the software new capabilities. It’s like automating something that ultimately turns out to be relatively insignificant.

Now truly new opportunities have emerged, and it’s entirely possible that the software will simply get better. It never even occurred to me before that no matter how many AI agents you use to code, the software doesn’t actually get better. Maybe you write it faster, but are you really getting worse simply because there’s less supervision?

So I think it’s often less safe.

Yes, that’s right, without a doubt. But you can now argue that applications will have new functionality because of this, because you’re providing this new primitive. In a sense, it speaks natural language—it can reason—but it combines that with a state machine.

7. New capabilities, not more code

If people took that as the main takeaway, it would be the biggest compliment to what we do. I even think it’s almost too grand a vision: going beyond the 3 logic gates that we have into something where our types are like the same logic gate, but with a little brain inside.

That would be the best compliment to the legacy of type safety, because it’s a huge, nontrivial thing for the world. I’m not going to promise more than I can deliver, but I will fight for it.

I think there are still open questions about how deep this can go in serious areas like state consistency, reliability, or a real system level where you need to provide strong guarantees. This will 100% change things like log analysis, email analysis, user-interface design, and human interaction.

That’s true, but you could argue that, over time, it will become something like a smart database. And also air-traffic-control systems—anything, right? What we really need is a little scary.

I think my philosophy is to automate the easy work before the hard work.

I also think that a whole era of probabilistic programming will open up. By the way, there’s a huge history of probabilistic programming, which actually died in the ’70s. I’m familiar with this.

I actually think it’s similar to what you could call neurosymbolic AI. Your co-founder Eric came from this very field. He told me.

Oh, cool. Yes, he studied biology a lot.

What I mean is, I’m not a fan. My brand is pragmatism—extreme pragmatism. I’ve never been a fan of biologically inspired things at all. Have you ever noticed this? I mean, I don’t think they ever worked.

It’s useful for motivating crazy people to work on something for decades until it works, and then they refine it into an engineered version. That’s the history of AI today: neural networks, for sure. But a lot of the stories about how it worked were inaccurate, right?

The hierarchical features of the applications didn’t really work, because otherwise ResNet wouldn’t work. Long story. I really think it’s promising, but from a systems perspective, I’m not excited about that part. I’m excited for the world—not that I’m going to program it—because it sounds really complicated.

But I think that when we have a lot of intelligence with different trade-offs in cost and speed, system specialists will make choices where developers will be 1,000 times smarter than them. They just need a rough reference to make a rough guess and optimistically push the data here and there.

It’s going to be incredible, these capabilities that will be available in extreme systems.

The good news is that we’re going to be able to rebuild systems again, and that’s great, right? We have a new primitive. No, seriously, we have a new primitive—a new way of thinking about building software.

We’ve done this before. We did it with the internet. We went from mainframes to client-server architectures. We do it periodically. And, by the way, because of cybersecurity issues, we’ll probably have to rebuild almost all systems to make them secure.

I think so. I think it’s pretty obvious that there’s no—or at least no critical infrastructure here, for sure.

8. Apps vs the guts of systems

Do you think of it more as applications, SaaS, and analytics, or more as systems fundamentals—or all of that together?

Do you mean what I’m thinking about, or a general application for this thing?

When you think, for example, about working on Jev and imagining how people would adapt it, do you even have any idea?

I have a little bit of an idea. I think of it a little bit as diving into the very depths of TCP. You know how TCP is a transition from unreliable to reliable? That’s my analogy.

When I think about AI—and that’s why I’m interested in intelligence per dollar, just to be clear—I came to this conclusion by going from the opposite direction, from an AI-based economic revolution. AI is everywhere in science fiction and all that. It’s as if all software has AI everywhere.

I ask myself: What percentage of calls to AI—imagine that’s a function—are intended for human consumption, where style is needed and all that? That will be a lot of nines. And, from the same question, how many will be at the first level, and how many will be somewhere deep in the jungle?

I think there will be a lot of nines in the slums, but it will all start from level 1. If you’re not aiming for the jungle—wait, that’s kind of weird. If you don’t aim for the very essence, it will take you a long time to reach your goal.

People don’t realize how much AI has been built into software in secret. Even if you tried to build AI into the software, it behaved unpredictably, because software doesn’t understand natural language very well. You do all these strange things: You put in the prompt, “Here is the JSON output I need. Here is the schema,” and it never listens to you.

As a result, you just had to take that output and give it to the person. You say, “Well, to hell with it, right?” Or you give it to another language model. That’s the “while” loop—the agent’s “while” loop, right?

Going from the basics, it should be either a human in the loop, which is a chat, or an agent, which is a “while” loop, because natural language needs to be routed back into the next loop.

I’ll put it this way: I watched it, and it was like the 5 stages of grief. People take AI and say, “I’m going to use this in my software.” Then there’s denial, trying to make it work, anger, and then they move to acceptance: “Okay, never mind. I’ll just give it to another language model or a person.”

So it was all very chaotic. I think this is the first time I’ve seen someone say, “You can take a language model, take AI, and actually turn it into a state machine,” and do it productively.

I hope so. I also don’t want to overpromise and underdeliver on this. I don’t know if it’s ready for all the uses it’s been so loudly promised for. I really, really want this, and my team will, of course, fight for it. We really care about reliability.

We could have released this much sooner. I don’t think people understand this, and honestly, I don’t think they will. Judging by the Twitter discussions, people may never understand it intellectually, but they’ll just have a feeling of, “Oh, I can trust this.”

This is the anti-frustration machine.

I hope so. I hope “do as I mean”—for me, it’s about fluidity in the world, so that everything moves more smoothly and connects like gears. I have my own utopia with artificial intelligence in various directions, and “do as I say” is a huge part of it.

Imagine if all technology just did what you meant. This isn’t science fiction. Look how smart artificial intelligence is, right?

Yes, it’s amazing, and perhaps that’s where we should end: “Do as I say.”

Yes, I like it.

Thank you, Diogo. It was a great conversation. I really liked it.

Mm-hmm.