Wordware AI, an IDE for programming using natural language - interview with Filip Kozera

So could you tell us more about what is Wordware trying to solve?

Wordware is like software but with words, so Wordware. And we are essentially solving one kind of the same problem, but in long term it means that AI engineers who will be programming with English as their main programming language and using it precisely in order to achieve kind of AI agents, they don't have the right tooling. And we are providing them with the right kind of schema on how to work on this and giving it to them in a web-hosted IDE where they can interact with loads of different foundational models and orchestrate them in the right way. We are not a chatbot building company. Rather, we create kind of complex AI orchestrations. In the short term, what we are solving is all of the prompting is happening inside of codebase right now. Which means that the domain experts, in order to close a feedback loop, need to communicate with engineers. And that can take anywhere between half a day and a week, depending on how efficient your organization is. We close that loop to be a couple of minutes.

That's pretty impressive.

Yeah, sounds good.

Can you show us a demo?

Sure, let's go. So as mentioned, prompting is the new programming. Let's sign in.

Nice.

And let's just, as in any programming language, we're gonna do hello world first. So, here we'll just have, you know, typical— Shh.

Hello world.

And we're gonna say hi to name. Name becomes an input to this function, and now we do a slash command. This on purpose looks a little bit like Notion.

Yeah, yeah.

And it gives us different, um, different functions. So generation is basically an LLM call. You get to create another input, and those things— loops, conditional statements, and code execution— essentially enable us to be like a complete Turing machine, you know? Okay, okay. So loops are kind they're very important concepts that we take from normal programming and we move them and we make them work with inherently non-deterministic nature of LLMs. So let me just start with a generation. We have a bunch of different models to play around with.

You can select them on the spot, right?

Yeah. So we have Command+R+ online, Mixtral on Groq. We have Claude 3 Summit and some of them are image enabled. So we have this generation, right? And I'm gonna click go right now. I'm gonna say hi to myself because I never remember people's names.

So you're providing the parameter, right? You're inputting that.

Yes.

So it's inferred the parameter and then—

okay, so this is super simple, right? Hello, Filip. Now this application is an API that you can input into your code base so you don't have to play around with prompts inside. You can share it and have a hosted version of this application, which looks like this, and it will always just show the outputs. But it gets a lot more interesting once we are starting to say, for example, generate a JSON in the form of name, @name. We're going to rename and greeting at new generation. You know what? This actually is not looking pretty good, so I'm just going to call this greeting. All of these have the same ID, so once you change one, you change all of them.

You're referencing the previous block. Exactly.

You can see how this is a little bit closer to actual programming. From here— But in English. But in English, correct. Now I can call a separate function. And the way that we actually work in companies is engineers create these functions often with code. Here we actually have ElevenLabs speech synthesis.

Okay, pretty cool.

This is written in code. Yeah.

So it's like foreign function calls, right?

Exactly. It's like assembly in C.

Yeah, you're trying to ingest or I don't know, you're incorporating actual code into this environment.

Exactly.

So normally the people who can judge the outputs of the LLM get to work with pre-written functions. These functions can be database call, vector database call, an API connection, and it can also provide inputs and outputs, right? So now I reference the output, and from here we've got a kind of a super simple application that will generate, say hi to Filip, and it will generate the right JSON, and then it will do the ElevenLabs speech synthesis and output an MP3.

Oh, nice. So you created kind of like a pipeline, right? Right, correct.

We call this, we call this prompt chains or cascades. Um, and this is a fully functioning Word app, right?

Is there a plan, for example, because there's like a lot of things you could like publish in a sort of like a directory, right? A community marketplace laid out.

We are doing a product-led growth with community at its bottom. The community is providing the fun examples of what is even possible. We are swimming a little bit ahead of the wave here, and it's important to have community to show what's even possible with AI agents. Then there's SMBs, which use our product as a backend for the whole LLM. So they put the API inside of the codebase. And the third one, third kind of category of people we work with is enterprises that they essentially take our product, create POCs, often they're consulting. And instead of just outputting a PowerPoint presentation to their clients, they will give them a working product, right? So let me show you just something a little bit cooler. I don't know if you've read this paper called REACT. It is basically prompting an agent to correct itself, write code, and use tools in order to solve an arbitrary task. So it's formatted in this specific way of question, thought, action, input, observation, and it repeats that n number of an amount of times in order to actually answer your question, right? So you can see that here we're using the things that you're familiar from normal programming, like looping, like conditional statements, and actually trying to achieve something here.

I really like the interface. It's like very fluid and like very convenient. Very minimalistic at the same time. And it seems like it's like easy for people who are new to programming to start operating that.

Also, it's inspired by Notion, but it seems to work better than Notion.

So here we— I actually asked it, what is the Bitcoin price today? The thought was, I need to find an API that provides the thing. It reads Google, it realizes CoinGecko API is a free, uh, API, writes the right code, uh, executes it on our servers, and, uh, it actually knows when it's done. That's very important actually. It's a conditional statement that actually works with the LLM nature of how non-deterministically it's working. So, and price of Bitcoin is $60,000. We, uh, yeah, I've got plenty of cool examples. Maybe I'll just show one more so you can cut it. Um, this is because of the way that we execute code. I can give you an input as an idea and a name for any startup. It will then use HTML to write Opus code and it will host it So it immediately gives you a website of your business. So let me just— idea, I'm gonna say help customers move their pets abroad, and name is going to be Doggo. It will now generate the right HTML.

So it builds whole website?

Yeah, so you know Devin?

Yeah.

So we are like a more general version of, uh, the infrastructure behind Devin. So AI engineers are always going to write their own tools because they have a hammer and they see nails. They're going to always code. But a gentle system like Devin have a huge opportunity to help a lot more people. We're giving the people who are writing these kind of tools a place that they can execute it and they can— let's see this website. So it works. Get started. Why choose Doga? Our service is very simple. You'd have to do a lot more prompting to make it work.

Of course, but it shows the possibility.

Yeah, yeah, it's just the beginning, right? It's going to be better over time. So pretty amazing. I really like the aspect of like combining different things because you're opening yourself to, to the community, right? People can provide those like blocks. Yeah, and then we can build on top of that.

It's a big company to build, uh, big company to build because right now people think about AI as retrieval of knowledge either from documents or from the weights of the network itself, from the training dataset. I think that's going to change in the coming years to actually think about it as a reasoning engine. And these reasoning engines can be connected in many different ways. So for example, I, when I do brainstorming, I use 3 different models that argue between themselves and they have a different perspective on each one of these. Is it—

maybe last question— is it possible to use it now?

It's possible to use it now, and I've actually enabled anyone who is— anyone who is— this might go wrong, but anyone who has an IP address from Poland gets unlimited free credits for now.

Ah, wow.

So just use VPN and— Bastard!

Damn it! You guys are going to ruin my company, you know? Yeah, but enjoy it. And we're super happy to hear any feedback. It's a tough tool to like, at the beginning, we do say it's a programming language, hence people expect to need to learn a little bit more about this. And yeah, it looks pretty impressive.

Thank you for showing us around and thank you for inviting everyone to register.

Yeah, thank you.

Come join us.