From Software Engineers to AI Word Artisans: Filip Kozera of Wordware

Contents

Filip Kozera: We have 30 million software engineers in the world, and we have 750 million active users of Excel. And now you’ll ask, like, “How does Wordware compare to Excel?” And what Excel did in the ‘80s to data analytics and numbers is what we are trying to do to AI. So in the ‘80s, you either had to have a team of data analytics people or engineers, or you were using a calculator. And I would say the calculator equivalent here is the ChatGPT that every time you need to redo the conversation, and every time you need to instill your own needs into it. What Wordware is trying to do is saying, “Hey, a lot of the things that you do are repeatable, similar to Excel, and you can encode your taste into it with Wordware.

Sonya Huang:  Today we’re joined by Filip Kozera, co-founder of Wordware, who’s building tools to help bridge the gap between human creativity and AI.  Philip shares his vision for why English is becoming the new assembly language for LLMs, why he believes that the future belongs not just to coders but to people he calls “word artisans” who can communicate effectively their creative vision to an AI system, and what that means for the future of programming computers.

Filip, thank you so much for joining us today. I’m excited to learn about you and Wordware and get your take on how, you know, programming computers is fundamentally going to change with large language models.

Filip Kozera: Thank you for the invite.

The hottest new programming language is English

Sonya Huang: Let’s dig right in. I want to start with Andrej Karpathy. He had a tweet in 2023 that went viral.

Filip Kozera: Yeah.

Sonya Huang: “The hottest new programming language is English.” What do you make of that, and what does it mean relative to what you are trying to build at Wordware?

Filip Kozera: I think syntax will not be as important. You know, everyone will be somewhat of a coder. People used to have to know Python; now English is enough. However, you still need to know what you’re trying to say. And in that way, I would say, you know, not everyone will be able to use it because some people don’t have that much to say. [laughs] And I would maybe rephrase it a little bit in a little bit more of an exciting way for us, is that the assembly language to LLMs is English, but you still have to structure it in the right way, and you still have to use some of the concepts even from typical programming, in order to actually make sure that it does what it’s supposed to do.

Sonya Huang: You heard it here first, okay? English is not the hottest new programming language—it is the new assembly language. And at Wordware, you are trying to build that new programming language to work with that new assembly language.

Filip Kozera: I think bringing structure to human-AI collaboration is something that I’ve chosen to spend the next 10 years of my life on. And it’s an exciting problem, because right now we are trying to mix the structure of programming, which is very rigid and deterministic, with something that’s intrinsically fuzzy. And that marrying of these concepts and choosing the right affordances and the right abstraction layers for that communication to be incredibly easy, yet enable people to do complex things is hard. And this is what we are trying to do with Wordware as the engine to enable the right mix of the two.

Sonya Huang: Let’s talk a little bit more about no-code. I remember back when I first got to Sequoia in 2018, I remember no-code was the hottest thing ever, and it was like, we’re not going to program anymore. And obviously some no-code companies have done extremely well, like Retool, for example. But by and large, we still have software engineers today and so far that no-code promise hasn’t really come to fruition. What’s different now? Like, what has changed?

Filip Kozera: I think coming back to what I just said as well, the assembly language is English. And I kind of have this small insecurity when people call Wordware a no-code tool.

Sonya Huang: Okay.

Filip Kozera: Because we have not yet reached a ceiling with any of our clients. So you can achieve absolutely everything. We’ve created an ability to put in code execution blocks, which if you want an escape hatch and Wordware is not fully capable of everything, you still can do it. And in that way, you know, the document format of how to structure agents and how to write them down one by one is still very similar to how code works. We still have loops, we still have conditional statements, and we have flows calling other flows, which really is function calling, you know? So in a way we don’t see ourselves as a no-code tool, and we kind of believe that the word crafters, the word artisans of the future, are still coding. It’s just, you know, they are structuring the English in a very precise way to make sure that the prompt is populated in the right manner.

Sonya Huang: Hmm. Word crafters and word artisans. I like it.

Filip Kozera: [laughs] Wordware engineers. You heard it here first.

Not no-code

Sonya Huang: How are you using the English language then to make—you know, since you bristle at the no-code term, how are you making no-code more code-like? And maybe this is a good time to just give a 30-second overview on the Wordware product and how people use it to build things.

Filip Kozera: Sure. So by trying to bring that structure to the intrinsically fuzzy English language, we’ve created an editor where you can use the similar concepts to software engineering—looping, conditional statements, functions calling functions—and marry it in that editor in that natural language, IDE, in a way that people can construct agents, something that we are not, we are not codegen. We go straight to agents. And in my definition, agents are almost—like, they are still software. They are still a little bit like software, taking in inputs and outputting outputs, but some of the stages of what they do is fuzzy.

So marrying that structure in the editor, whatever you build there, whatever you iterate on there, you then have three different ways of deploying it. You can deploy it as an API that will power your product, or that AI button, or that AI chatbot that is a little bit more complex than just doing a vanilla API call to Claude or OpenAI. That’s number one.

Number two is you can deploy it as a workflow, where the main part of the workflow is not like Zapier, it’s more AI native. And really, the brain of the AI is a little bit more complex than a couple prompts strung together. And the third thing that we’re building is the GitHub for AI, let’s say, for people to share what they’ve done, and other people are able to then fork it and to use these things as components. Again, marrying some concepts from software engineering, you need other people’s components and libraries, and you want to be building on top of the shoulders of the giants of these AI thinkers. So again, Wordware engine, editor, way to actually create these agents, and then three different ways to deploy.

Sonya Huang: I think one of the beauties of code is just its expressibility and its precision. You know exactly what you are telling the machine to do, and you are expressing it in some languages in the most precise way possible. English is not like that. To your point, it’s fuzzy. And so as you are turning the English language more code-like, should I think about that as helping with the controllability and the steerability of that? Or what is the, I guess, abstract thing you were doing with English to make it more programmable?

Filip Kozera: Yeah, I think you’re exactly correct. We are trying to bring a little bit more structure. It’s not all the way, because if you go all the way to a programming language, you lose the fuzziness and you lose the power of it. But it’s hard. And right now, most of the people—you know, we had this wave of evaluation software companies at some stage. And what we’ve realized with a bunch of companies that we work with is that they don’t know, like, they don’t have these data sets to use for evals. And what we came up with is that editor very quickly gets you to understand and develop an intuition of what works and what doesn’t work.

And for now, this is the most important thing. It’s clicking around a hundred times quickly, and making sure that what you’ve written here has enough structure in order to output things on the right side that you want. And in that way, you know, as long as you know what you’re trying to achieve—and this is very hard. You know, we have a lot of companies coming in and just saying, “AI, predict weather.” I literally had a big customer say, “Can you guys, like, predict weather for me?” And that’s not the case. You need a document where you outline what you’re trying to do. And even that document, that has enough structure, if you have done an intro, those are the inputs. You’ll be playing around with images and PDFs, and then you will manipulate it in this way, and then you will get some outputs.

And developing that intuition is just enough structure for today. Soon maybe we’ll be able to do better evals. But for now, a lot of people really don’t know at the moment when they start playing around with Wordware, they don’t know what they are trying to achieve. And one of our customers coined this term of “speed of creativity with Wordware is higher.” So they learn what they are actually trying to do as they encounter problems with the underlying models, and they realize Gemini 2.0 Pro might be better. And maybe Gemini can take huge PDFs, and Claude can kind of take PDFs but smaller, and then GPT4 cannot take PDFs. And, you know, they develop that understanding, and that helps them to structurize their thoughts.

Sonya Huang: How do you instruct the machine to go from intent to outcome? Like, let’s say I’m a brilliant filmmaker and I want to use Wordware to create the next hit. What do I do in Wordware in order to make that happen?

Filip Kozera: Yeah. So for now, we focus on knowledge workers.

Sonya Huang: Knowledge workers. Okay.

George Lucas and GPT-7

Filip Kozera: Where that is a little bit easier. I think, using your example for figuring out how the future will be, you know, if we have, let’s say, George Lucas playing around, hypothetically playing around with GPT-7 and trying to create Star Wars, he might type in just the prompt being like the two sentences, and he would just say, “Hey, create a movie about wars between star systems.” And that’s just enough to give a model an ability to run on its own.

And this is actually not what you want. You want to convey your creative vision, which for now in Wordware is just knowledge work. But soon it will be all work where it needs that human sprinkle, that human taste. And these are the things that I really value is like, people say, “Oh, nobody will have a job whatsoever.” I don’t agree with it at all. I think the human taste and how you do things will matter even more. And I use this George Lucas example because it’s a little bit easier to understand how taste is influencing that. But everywhere, writing a good email is dependent on good taste. Figuring out—I was just hiring for an executive assistant and, like, everyone in our company needs to show that taste, and needs to show a little bit more conviction in the way that they do things. So for an executive assistant, like, she needed to choose a right restaurant for our off site, you know? And that also has taste. I don’t want to trust an AI with this. So yeah, that’s kind of that sprinkle of human touch is very important.

Sonya Huang: So taste as the last bastion of humanity.

Filip Kozera: I think so. I think creativity and taste.

Sonya Huang: Do you think machines can learn human taste?

Filip Kozera: I think they can, but that’s not the point. There is an interesting analogy here. They put humans into an MRI machine, and they’ve shown them two different pieces of art. Both of them were done by AI, and they told the people, “Hey, one is created by a human artist and another one is created by AI.” Our brains work completely differently, and our different parts of the brain fire when we assume human intent behind something. So, you know, I can create a song and just, like, it will be a good song with Suno or whatever and send it to my friend and he’ll be like, “Yeah, it’s a cool song.” But if I ingrain my intent and I’ll create a song about our, you know, skiing trip to Chamonix, and I’ll make it funny and that intent will be there, I’m pretty sure his brain will be firing in a completely different way, giving him a completely different experience in that way. Does that make sense?

Sonya Huang: It makes a lot of sense.

Filip Kozera: Yeah.

The next billion developers

Sonya Huang: I love it. I want to transition to talking about the next billion developers. And you’ve alluded to this a few times in the conversation, and I really want to just pull on that thread. So you started this conversation by saying there’s a certain set of people in the world that know how to code, but there’s a difference of the people in this world that have creativity and have ideas. Do you think that set of people is larger? Is that how we get to the next billion developers?

Filip Kozera: There’s an interesting analogy here. We have 30 million software engineers in the world, and we have 750 million active users of Excel. And now you’ll ask, like, “How does Wordware compare to Excel?” And what Excel did in the ‘80s to data analytics and numbers is what we are trying to do to AI. So in the ‘80s, you either had to have a team of data analytics people or engineers, or you were using a calculator. And I would say the calculator equivalent here is the ChatGPT that every time you need to redo the conversation, and every time you need to instill your own needs into it. What Wordware is trying to do is saying, “Hey, a lot of the things that you do are repeatable, similar to Excel, and you can encode your taste into it with Wordware.”

And I believe that taste will be important, as mentioned before, and I think the next 500 million or a billion users of AI might be calling them—I don’t know what will be the term. Hopefully it’s ‘Wordware engineer,’ but, you know, word artisan or whoever. And the really important part here is that they need to know what the AI is supposed to do. Many, many, many times, you know, we are a horizontal tool, and people come to us and they say, “Hey, what can I do with AI?”

And I tend to explain it as if for now it’s an intern, but an intern after university. And you need to write out on a piece of paper a couple of things that you would want to do with the intern. So you need to say, “Hey, this is your job. This is the title of what you’re trying to do. Here are some of the documents or input that you will be working with. Here are the data sources, and here’s the output that I’m expecting of you.” And the important caveats here that people don’t often understand is that the data sources has to be something that you trust. You can’t just say, “Hey, go and search the internet,” because often you end up with things that you don’t agree with. And if the intern works on top of that, that’s a problem.

And another one is—and this is very important, is that you’re going to trust the intern with this. So, you know, if you want to send a thousand emails to every person that needs a response in your inbox, with some people you just won’t trust an intern to do this. And this is how AI works right now. So as long as you have a job right now that you say, “Hey, if I have one intern, I could easily explain it to them.” And a lot of our work is like this right now. We often read an email, go search in Dropbox, go search Notion. And then we create a response that is essentially based on this database that we curate. Then you can be using AI. And I think more and more, this knowledge work is going to be automated, and I think in that way, the next billion people are going to need two things: Intent, what action needs to happen, and taste, how do you want to do this? And all of the rest will feel like CEOs of the biggest enterprise, because we will have a thousand knowledge workers working beneath us and trying to actually execute on these two things.

Sonya Huang: What I heard just now was a lot of automation about knowledge work. I mean, the thing that I’m most intrigued about within AI and within generative AI is its generative capacity, including the ability to create. You know, you mentioned the George Lucas example, but also to create new applications, new marketplaces, you know, new products. And so do you imagine—do you see Wordware primarily serving, you know, making the knowledge worker more productive, or do you see it also assisting in kind of the creation of new products, services, you know, pieces of art?

Filip Kozera: For now, it’s mostly about the productive work, I would say. It’s—you know, the AI engine is—the AI heart of your product is Wordware. Currently we have not dipped into the generative UI part of things. We’re not Lovable. We’ve actually used Lovable to wrap our AI heart for some of our customers, and that has worked great. I’m just so impressed with their product.

Sonya Huang: Talk more about this. Lovable, I think, also sees themselves as enabling the next billion developers. You have a similar vision. How do you think your view of how the world will go fits with their view? And, like, why didn’t you choose to make that style of no-code tool?

The UI for AI

Filip Kozera: Yeah, because the way that I see the word ‘developer’ is a little bit different. They see the word ‘developer’ as what developers do today, which is a lot of SaaS is a wrap around a database with some dashboard and ability to manipulate that data. They are creating much more personalized dashboards, you know, and a lot of people are going to create incredible vertical SaaS based on Lovable. And I think that’s incredible. And the one thing that was missing through all of this is that they not only grab the UI part, they also grab the database part, which many people do not know how to manipulate. And hence, they unlocked a lot more use cases. But what we are trying to say is that this part of, like, creating a generative—creating a UI on top of a database is not the future. The future is to actually utilize this reasoning engine that an LLM is in a productive manner, and we focus on that substance of AI at the beginning.

In the future we might want to expose that engine in a UI, maybe it’s a Chatbot, maybe it’s digesting some images, et cetera. But the real important part is the AI engine.

Sonya Huang: So if you think about an app as there’s the UI, there’s the application logic, and there’s a database, what you’re saying is you really want to just knock it out of the park on the application logic, so to speak.

Filip Kozera: Yes. And I think the databases that we used to work on were discrete, and right now we are able to work on a lot more data which is not structurized. And this is the big, big difference right now. A lot of SaaS’s right now will still work in a similar manner, just the database is fuzzy. And the database might be what you see every day, and how the hell would you put that in a typical normal database? And I think working on top of that context is the really exciting part for me.

Sonya Huang: Yeah. Let’s go back to this concept of word artisans or Wordware engineers. If everything goes right, what’s your vision of what a Wordware engineer looks like in 10 years?

Filip Kozera: Oh, that’s a tough one. I have been thinking a lot about what does work look like in 10 years for human beings. And I was struggling with this at the beginning because it’s really hard to understand people’s jobs, even today. And often I boil it down to the software that they use. They can talk a big game about strategy and, you know, I set the mission, and in the end of the day, I ask them, “Do you do meetings, do you do email, do you do PowerPoint presentations? Do you work in Excel? What exact—or do you work in code?” And I just want to understand what does work look like in 10 years and, like, what are you really working with, you know? Is it like the movie Her, like interface, when you just talk to the AI and it does a lot of work for you?

I think, to be honest, voice is kind of not the best modality to express that. Hence, I kind of think that in its simplest form, Wordware is a document where you jot down your thoughts, and you do it in a more structured way, Wordware copilot AI is helping you throughout, structuring it. And in the end of the day, you behave like a CEO, which sets the strategy, intent, and all of that on that piece of paper, essentially on this blank canvas, and you draw that vision. Maybe it’s even more than words. It’s just, you know, you generate this vision of how your own enterprise works.

And, you know, I look at different things around us, and I see furniture or shoes or whatever. And I think there is taste ingrained into what kind of shoe would you like to make. So in 10 years, if somebody wants to become a creator of the best brand of shoes, it becomes about—that shoe becomes a luxury object which has ingrained taste and intent in it. And then a bunch of things in the end will be—will happen on its own. The really tough parts, even manufacturing it and so on will happen on its own. But what’s your job in the future is talking to other CEOs, I think will not—humans don’t want to lose that control. So you will talk with other CEOs about maybe doing a partnership with your shoe brand and somebody else. You have to be still critical about the intent of the other person, and you have to instill taste and your own creative vision into that shoe. [laughs]

ICP as analytical creative

Sonya Huang: Yeah. Do you think that—you know, do you think a billion people globally will be capable of programming a machine in English language the way you describe, or in Wordware documents the way you describe? Because it does require—you know, it’s almost like pseudocoding. And, you know, there is logic, there’s loops and things like that. I guess maybe talk about today, like, do you need to be technical in order to use Wordware? Like, who is the ICP today, and what is needed to move that ICP so that you can reach a billion developers?

Filip Kozera: Yeah, I think right now what we’ve enabled is people who are somewhat technical—CEOs, technical PMs—high up in the org chart to ingrain their own, like, kind of—think that they know what needs to happen and get there quicker. You know, so Max from Instacart, for example, he is a founder, and he spent four days just, you know, refining his idea in Wordware instead of hiring a whole team. But he is somewhat, you know, analytically minded. And for now that’s the case. We did not want to make too much magic because the models were not there. Right now what we’re doing is we are moving more into that blank canvas when you just describe the idea and we take care of guessing the right structure. And you still will be able to, you know, in a very fine-grained way, edit it, but you will start playing a lot more. We’ll probably use o3 to get you to the first draft of how that flow works.

And when we kind of look back to the future of how it will all look like and really whether we’ll have one billion developers, you know, working in Wordware, it becomes a much bigger question here. It becomes a question of, like, will a billion people want to do productive work? Like, you know, we just talked about the shoe. How many people will have the drive to put out something to the world, and they will want to express that creative vision? Maybe in, you know, post-resource scarcity world, most of us won’t want to work, but I think we’ll still have the equivalent of billionaires. And it will be about influence, it will be about taste, and it will be about how you utilize your own resources and how do you multiply it to have the equivalent of future money? And I went a little bit deep here.

Sonya Huang: No, but for what it’s worth, I think the innate drive to create is like a deeply human drive. And I think that exists in a post-capitalistic world.

Filip Kozera: I also have that opinion, and I really believe in humans. Like, I want them to succeed. Like, somebody asked me, one of our prospective employees asked me, like, “Filip, in 10 years, what do you want there to be—like, what have you done?” And I want to save, like, the human creative vision. I don’t want everything to be AI. I really have the pleasure when I go to an artisan shop on my holiday, and I know that somebody put in the intent and put in the work, and I want to interact with it and I want to interact with the story of it.

Sonya Huang: Totally. Okay, so today your ICP is the analytical creative, which is a little bit of a unicorn. And over time, as you kind of lower—as the models get better, as you iterate on your interface, you’ll lower the bar. So it’ll really be just more of the creative as your ideal user. You’re going to lower the bar of how analytical you need to be in order to use Wordware.

Filip Kozera: Yes, but at the same time, my use of the word ‘creative’ is not to what most people associate it with right now. I think a good creative is also, you know, using growth channels in the right manner. They are creative about everything that they do in this new, uncertain world of AI, where everything is changing. And, you know, I’m not thinking only about an artist that’s painting on the canvas, I think creativity can basically show itself in so many different aspects of work.

Transformers are the new transistor

Sonya Huang: Yeah. Let’s talk about user interfaces and, you know, the future GUIs—the GUIs of the future. Right before we filmed this podcast, you made the analogy that, you know, transformers are the new transistor. Maybe say a little bit more about that, and what you think the new GUI is going to be.

Filip Kozera: So I think the analogy here is that if the LLM is the—well, if transformer is the new transistor and it’s being packaged as the model, the model is kind of the mainframe, let’s call it, and then we took our sweet time to utilize the power of that mainframe in a GUI that’s accessible by billions of people. You know, there has been really two big spikes there. The first was the desktop, you know, and Apple coming up with their GUI. And the second one was mobile. And, you know, right now we are almost exposing the numbers and the logic in a chat-style thing, and nobody has had a better idea. We think that the document style is better for doing kind of more complex work, because often when you’re trying to achieve something, you just give it two sentences and the model just runs on its own. And it’s just enough, the two sentences. Our lead investor recently said that the two sentences is just about enough for a model to hang itself on. And, you know, you will get something completely different than what you actually wanted. And this is a problem of, like, the Lovables and Devins of this world as well. But I basically think that there are better GUIs coming and, you know, whether they will be based on AR or, you know, there will be an assistant that’s listening to everything that we do. That was actually my first company, augmenting human memory of always-on listening devices using GPT-2 and BERT. I’ve been in this …

Sonya Huang: Ah, you’ve been in this since the GPT-2 days.

Filip Kozera: I mean, my research was into LSTMs which are the precursor to the transformer architecture. And I’ve been in this for a while. I think nobody has yet delivered on this. I want everything that I hear to be somewhere in a searchable database that also has the perfect context about me, you know, the way that I want to do things. And I think those affordances and those—like, we called it GUI but it’s really the underlying, like, way of interacting with intelligence, it’s not going to be mainly chat. I just don’t believe it.

Programmable documents

Sonya Huang: Totally. Programmable documents. Do you think that is fundamentally what Wordware looks like, UI wise, in, call it, five years?

Filip Kozera: I think there is more and more magic in it, and I would believe that I want people still to be able to do that fine-grain work. You know, we’ve linked it with George Lucas doing the movie, you know? In a way you almost want to firstly start with the high-level thing, the two-sentence description and then zoom in and zoom in and zoom in, and create modules which make the best scene that is five seconds, and then combine them together in that way.

So what I would like Wordware to be is to transcend abstraction layers and be able to zoom it all out. Start with a sentence and have it run, maybe see whether it’s working in the right manner. And then as you’re seeing that some things are not doing the thing that you want them to do is to be able to zoom in and see maybe four sentences of exactly what it’s doing. What are the inputs to this, what it’s trying to do in the middle, and what are the outputs? You know, that’s kind of the most simple one level in. And then you want to zoom in more and more and more as you redefine and reiterate on your idea of how this should be done.

Sonya Huang: Yeah. How did you arrive at the current user interface? I think it does feel really novel compared to how others are enabling AI builders today. How did you arrive at the current user interface? Was it more experimentation, listening to users? Was it you philosophizing about what it should be?

Filip Kozera: I think currently, the—currently and before, the approach to creating these agents was a block based on a 2D canvas. And once—I’ve been building agents for a long time, I think March, 2023, I put out the first article about how to build agents, and me and Robert, my co-founder, we’ve been in this for a long time. And the more—the better the models got, the prompting became more difficult because you can do more complex things with it. So at some stage there was this movement of, like, the prompt is going away, and so on.

We actually really disagreed with it, and that that idea is gone a little bit. It’s like, you know, we came back, did a loop again and be like, actually communicating your vision is really important. And when we tried to communicate our vision, which was a little bit ahead of what the models could have done at the time, we started to notice that the 2D canvas is just not enough. Like, if you do a reflection loop inside of a reflection loop, you run out of dimensions. And we basically really like the way that code is structured. Code has an ability to express very, very complex concepts in a way that is still—like, you can still manipulate it and understand it.

Think about trying to structure the whole Uber app with all of, like, everything in it on a 2D canvas. It would become so cluttered and so messy. You know, you can do the big picture thing, but not really the—you know, you don’t want engineers to be interacting in that way. You want the engineers in the future—Wordware engineers to be interacting with something that’s easy to grasp the structure of very complex systems.

Sonya Huang: Whereas the Uber app actually could probably be described in pseudocode. And it seems like you’re getting people closer to that vision versus the 2D canvas.

Filip Kozera: Yes. And I think the most important part here is that Uber has an agent equivalent, and this is what we are trying to build, you know? If you want an agent to decide where is that person going and where are they starting their journey and where they will accept that charge, or, you know, you want to maybe make sure that the charge is right for that particular person, there is an agent equivalent there.

And, you know, people are going—like, people can build that agent on Wordware. It’s not like you’re going to create that whole UI for Uber. And I think, you know, probably Uber is the right abstraction layer. You don’t want to be ordering an Uber through a chatbot or through, like, a voice-based thing or, you know—but you might want an Uber to be ordered for you if you have a calendar invite. So, you know, in a way that, like, for your personal use, Uber is nice because you can click around and the agent will not always know. But I was coming here, and I wanted a Waymo. Actually, Waymo can’t get that far yet, but I wanted the Waymo to be ordered and to be ordered perfectly when I need this. And it’s almost like a personal assistant would do this for me. And now that capability is open to everyone. So we’ll soon have these kind of affordances and these kind of abstraction layers there.

Lightning round

Sonya Huang: Totally. I think that’s a great note to end on. Should we end on a lightning round?

Filip Kozera: Let’s go. Yes.

Sonya Huang: Okay. One- or two-sentence answers only. Okay, first question. What is your most hot take or contrarian take in AI, not related to Wordware or everything we just discussed?

Filip Kozera: Pre-training is still going to matter. And DeepSeek is a little blimp that people liked to—people jumped on because people love a good drama, and it was connected to China, and actually it doesn’t matter that much.

Sonya Huang: Okay, I know I said lightning round, but you have to say more. What do you mean it doesn’t matter that much?

Filip Kozera: I mean, they utilized some cool techniques and the rest of the community is going to learn from that. However, you know, the fact that they, like, trained it for a little bit cheaper, for a lot cheaper, does not involve all the experimentation that they did before that. And, you know, I don’t know if I’m supposed to say that, but I’m pretty sure they had access to the best Nvidias as well for that experimentation. And it’s not that novel. Like, people jumped on it because they were like, “Oh my God, China is taking over the race!” And so on. And Nvidia stock price, like, plummeted. And I just think it’s another place where some models were trained that were open sourced, and it’s not going to—you know, we’re not going to remember it in a year or, like, even six months. Or maybe they will take over, but the model doesn’t really matter that much. How you kind of work with that best model out there, that’s what matters.

Sonya Huang: That is a hot take indeed. Okay, next question. Who’s going to have the best frontier model next year?

Filip Kozera: Oof. I think OpenAI is always super bullish and they always promise a lot. And then I was just on a talk with Sam Altman on the YCAI retreat, and the o3 the way that he pictured it sounded great. But I think we both know that they overpromise a little bit—a lot. And I love Antropic. I think their kind of vision and their kind of the way that they’ve created this is great. But recently, Gemini 2.0 Pro with their abilities to ingest 6,000 pages of PDF is really blowing my mind. So end of the story is I have no clue. This is a place where it’s super fragmented and people have zero loyalty.

Sonya Huang: Pre-training is hitting a wall. I think famous people including Ilya have been quoted saying something to that extent recently. Agree or disagree?

Filip Kozera: Disagree. Right now I think, you know, it’s the intelligence of a model is linked logarithmically to the resources that are needed to train it. But, you know, doing a 2X of intelligence is, on its own, exponential. Like, if I’m smarter 2X than somebody else, it doesn’t mean I’ll do 2X of the work. It means that I’ll find ways that probably mean I’m a 10X or even more.

Sonya Huang: Favorite new AI app, not Wordware.

Filip Kozera: I would say I started to edit content because we need to explain and educate people a little bit more about both Wordware and AI. So Descript is something that I’ve been loving. And I use Granola every day. And the newest model that I’m really impressed is the Gemini 2.0 Pro. I really like it.

Sonya Huang: That’s a hot take as well. I haven’t heard much of that from people.

Filip Kozera: I think it came out, like, four days ago, so people have not been playing around with it. Their PDF capabilities are awesome.

Sonya Huang: What application or application category do you think will really go mainstream and hit this year?

Filip Kozera: I would love to see—I’m personally very, very involved with that whole, AI having the context of your life, and being able to basically make better decisions based on the context. And, you know, Rewind, which, you know, I think they are called limitless right now.

Sonya Huang: Yeah.

Filip Kozera: I’ve ordered their pendant, by the way. It’s been, like, a year and a half and I still don’t have it. I don’t know, send it to me. And I had to change a color because they didn’t have the color. But I would love for there to be a provider which has a lot more context and can do the personal stuff for me.

Sonya Huang: Don’t you think that’s Apple, over time?

Filip Kozera: I was just about to say, I think ideally that N421 model, or whatever it’s called, of the AR glasses that they are trying to push out there, which I think Facebook has taken over a little bit, maybe we’ll see early stages of that. And I think they’re the only ones where the privacy really—like, they have a good brand around privacy. And two, even if your new AR glasses run out of battery, it’s still cool to be wearing a $5,000, you know, piece of hardware. And maybe that’s the UX, but I don’t know what’s that UX. And, like, a microphone so far failed. Yeah.

Sonya Huang: Single piece of content that an AI aficionado should read or watch.

Filip Kozera: I would say all of the DeepLearning.AI resources. Everyone—like, we have a bunch of candidates apply for jobs. By the way, we are hiring. Whatever. I should be looking very, very aggressively. So come join Wordware. But the DeepLearning.AI resources are awesome, and they explain everything from the bottom layer all the way to the practical layer of how to actually get it done. I also think if you don’t understand the underlying technology, go see 3Blue1Brown, an incredible channel on YouTube. And they explain everything super well.

Sonya Huang: Wonderful. Your lightning round was full of hot takes. I didn’t even have to ask you for a specific hot take. Well, Filip, thank you so much for coming on. I really enjoyed chatting about how you see the world evolving from developers, word artisans or Wordware engineers, if everything goes right. And appreciate you sitting down to share your vision and your hot takes.

Filip Kozera: Thank you for having me.

Sonya Huang: Thank you.

Mentioned in this episode

Mentioned in this episode:

• Lovable: Generative AI app that builds UIs and web apps
• Her: 2013 Spike Jonze film that Filip uses as an example of how voice will not be the best modality to express knowledge work.
• Descript: AI video editing app that Filip uses a lot. 
• Granola: AI notetaking app Filip uses every day.. 
• Gemini 2.0 Pro: Google’s newest long context model that can handle 6000 page pdfs.
• Limitless pendant: Wearable device for collecting personal conversational context to drive AI experiences that Filip can’t wait for to ship.
• DeepLearning.AI: Andrew Ng’s amazing resource for learning about AI
• 3Blue1Brown: Grant Sanderson’s incredible channel on YouTube that explains math and AI visually.