750 Million Excel Users, 30 Million Engineers: Wordware's Bet on Word Artisans

The cold open of Sequoia's Training Data puts the wager in two numbers: 30 million software engineers in the world, and 750 million active users of Excel. Filip Kozera, cofounder of Wordware, says what Excel did in the 1980s to data analytics is what his company is trying to do to AI. Back then you either hired a team of analysts and engineers or you used a calculator. The calculator equivalent today, he argues, is ChatGPT: every time you need the work done, you redo the conversation and re-instill your own needs.

Host Sonya Huang introduces Kozera as someone building tools to bridge human creativity and AI, and as a believer that the future belongs not only to coders but to people he calls word artisans. He comes out of research on LSTMs, the architecture that preceded Transformers, and his first company tried always-on listening devices with GPT-2 and BERT. With cofounder Robert he has been writing publicly about how to build agents since March 2023, which puts him among the earlier practitioners in the current wave.

Huang opens with Andrej Karpathy's 2023 tweet that the hottest new programming language is English, and asks what Kozera makes of it. He agrees that syntax will matter less, and that people who once needed Python can now get somewhere in English. He immediately qualifies the slogan. You still have to know what you are trying to say, and not everyone will be able to use the new stack, because some people do not have that much to say.

His preferred rephrase is more precise than the tweet. English, he says, is the assembly language of LLMs: you can command the machine with it, but you still have to structure the request and still have to borrow loops, conditionals, and other old programming ideas if you want the model to do what it is supposed to do. Huang jokes that listeners heard it here first: English is not the hottest new programming language, it is the new assembly language, and Wordware is trying to build the language that sits on top of it.

That grafting problem is what Kozera says he has chosen to spend the next ten years on. Programming is rigid and deterministic; model output is intrinsically fuzzy. Mixing those two, and choosing the right affordances and abstraction layers so the collaboration is easy enough for people and still powerful enough for complex work, is the hard part. Wordware, in his telling, is the engine meant to get that mix right.

Huang remembers arriving at Sequoia in 2018 when no-code was the hottest promise in software, and notes that we still have software engineers. Kozera admits to a small insecurity whenever people file Wordware under no-code. No client has hit a ceiling yet, he says, and the product keeps an escape hatch: code execution blocks, for the moments when natural language is not enough. The document format for writing agents still looks like code. There are loops, conditional statements, and calls from one flow to another, which he describes as function calling.

The people he wants to serve are still programming. They are word crafters or word artisans, structuring English precisely enough that the prompt is populated the right way. Huang offers "Wordware engineers" as a third title. The product is an editor that marries those software-engineering concepts with natural language, an IDE that goes straight to agents rather than generating application code. An agent, in his definition, is still software: it takes inputs and produces outputs, except that some of the stages in the middle are fuzzy.

Whatever you iterate in that editor can leave in three ways. You can deploy it as an API that powers a product, an AI button, or a chatbot more complex than a vanilla call to Claude or OpenAI. You can deploy it as an AI-native workflow whose brain is more than a couple of prompts chained together, unlike a typical Zapier-style glue job. The third path is what he calls a GitHub for AI: a place to share what you built so other people can fork it and treat those flows as components, standing on the shoulders of other AI thinkers.

Huang asks whether making English more code-like is really about controllability and stability. Kozera says yes, but only a little more structure: go all the way to a programming language and you lose the fuzziness that makes the models useful. A wave of evaluation startups taught him a more practical lesson. Most companies he works with do not have datasets they can use for evals. What the editor is for, today, is clicking run a hundred times until you grow an intuition for what works. He has had a large customer walk in and ask Wordware to predict the weather. Without a document that names the inputs, the manipulations, and the outputs, there is nothing to evaluate. One customer coined the phrase that the speed of creativity is higher in Wordware: people learn what they actually want as they hit model limits, noticing that Gemini 2.0 Pro will take huge PDFs, Claude will take smaller ones, and GPT-4 will not take PDFs at all.

Huang asks how you instruct a machine from intent to outcome, and offers the thought experiment of a brilliant filmmaker. Kozera says Wordware is still focused on knowledge work, where that translation is easier, but he uses George Lucas and a hypothetical GPT-7 to show the failure mode. If Lucas typed two sentences — create a movie about wars between star systems — the model would have enough to run on its own. That is not what you want. You want to convey a creative vision, the human sprinkle of taste that, for now, lives in knowledge work and will eventually touch all work.

He rejects the claim that nobody will have a job. Human taste, he thinks, will matter even more. Writing a good email depends on it. So does picking the right restaurant for an offsite, a test he used while hiring an executive assistant, because everyone in the company has to show conviction in how they do things. He does not want to trust an AI with that choice. Huang boils it down to taste as the last bastion of humanity. Kozera agrees: creativity and taste.

Can machines learn human taste? He thinks they can, but says that is not the point. He cites a study that put people in an MRI machine, showed them two AI-generated artworks, and told them one was made by a human artist. Different parts of the brain fire when we assume human intent is behind the work. The same gap shows up in music. A decent Suno track sent to a friend is just a cool song. A song about a skiing trip to Chamonix, with the jokes and the shared history written in, gives the friend's brain a completely different experience.

Huang pulls on the next-billion-developers thread: is the set of people with ideas simply larger than the set who can code? Kozera returns to the Excel numbers. What Wordware is trying to say is that a lot of what people do is repeatable, and that they can encode their taste into it. He hopes the next 500 million or billion users of AI might be called Wordware engineers or word artisans. The important part is that they need to know what the AI is supposed to do. A horizontal tool keeps attracting people who ask, instead, what they can do with AI.

His working metaphor is an intern who has just finished university. You write on a piece of paper the job title, the documents or inputs, the data sources, and the output you expect. Two caveats matter more than people admit. The data sources have to be something you trust; "go search the internet" often returns things you do not agree with, and then the intern builds on a rotten foundation. And you have to be willing to trust the intern with the task. Plenty of people would not let an intern answer a thousand emails in their inbox. As long as you have work you could explain to one intern, he says, you can use AI: read an email, search Dropbox, search Notion, write a response on top of a database you already curate.

The next billion people, in that picture, need only two things: what actions should happen, and taste — how they want those actions done. Everything underneath will feel like being CEO of a large enterprise, with a thousand knowledge workers executing. Huang hears a lot of automation of knowledge work and asks whether Wordware is also in the business of creating new products and art. For now, Kozera says, it is mostly productive work. Wordware is the AI heart of a product. The company has not dipped into generative UI.

Some Wordware customers have already used Lovable to wrap a UI around that AI heart, and Kozera is impressed. Huang notes that Lovable also talks about the next billion developers, and asks why Wordware did not become that style of no-code tool. He sees the word "developer" differently. Lovable, in his telling, sees developers as people who do what developers do today: a lot of SaaS is a wrap around a database with a dashboard. Grabbing both the UI and the database unlocked a wave of vertical SaaS that people could not have built if they had to manipulate the database themselves. He thinks that is incredible, and not the future he is building toward.

An app, Huang restates, has a UI, application logic, and a database. Kozera wants to knock it out of the park on the application logic — the reasoning engine an LLM actually is. The databases people used to work on were discrete and structured. Now you can work on unstructured, fuzzy data: what you see every day, the documents inside a company, context that would never fit a normal table. A lot of SaaS will keep working in a similar manner, he says. The database is just fuzzy, and working on top of that context is the exciting part.

Asked what a Wordware engineer looks like in ten years, he starts from a more basic confusion: it is already hard to understand people's jobs. He boils work down to the software someone actually uses — meetings, email, PowerPoint, Excel, or code — rather than the strategy language they talk. He does not think a *Her*-style voice interface is the right modality for complex knowledge work. In its simplest form, Wordware is a document where you jot thoughts in a more structured way, a copilot helps you structure them, and you behave like a CEO setting strategy and intent on a blank canvas.

The shoe-brand thought experiment is the ten-year job description. The shoe becomes a luxury object with taste and intent ingrained in it; manufacturing happens on its own. What remains is talking to other CEOs, staying critical about their intent, and instilling your own creative vision. Today's ideal customer is narrower: somewhat technical CEOs and technical PMs high in the org chart. Max, a founder at Instacart, spent four days refining an idea in Wordware instead of hiring a team. The product is moving toward a blank canvas that guesses structure — probably with o3 drafting the first flow — while keeping fine-grained edits. Whether a billion people will do that work is, he admits, a question about drive, not syntax. He still wants to "save the human creative vision," the feeling he gets in an artisan shop on holiday when someone has put intent and work into an object.

Just before taping, Kozera had told Huang that Transformers are the new transistor. The model, packed around that transistor, is the new mainframe. It took a long time to put mainframe power into a GUI that billions of people could use, and there were two big spikes: Apple's desktop GUI, then mobile. Right now the industry is exposing the numbers and the logic in a chat box because nobody has had a better idea. He thinks a document is better for complex work. Give a model two sentences and it runs on its own. Wordware's lead investor said those two sentences are just about enough for a model to hang itself, and you get something completely different from what you wanted — a problem he also pins on Lovable and Devin.

He does not believe chat will remain the main way of interacting with intelligence, and he does not think voice is efficient enough for this kind of work. His first company was already chasing a different affordance: always-on listening devices, using GPT-2 and BERT, to augment human memory. He still wants everything he hears in a searchable database that also holds perfect context about how he likes to do things. Nobody has delivered it. The interface he wants for Wordware is a document you can zoom. Start with a two-sentence description, let it run, and when a step is wrong, zoom into four sentences that name the inputs, the middle, and the outputs, then keep zooming until you are editing a five-second scene. George Lucas again: high-level vision first, then modules, then the cut.

How they arrived at that UI is less philosophy than two years of building agents. In March 2023 Kozera published an early article on how to build them; he and Robert have been in it since. The dominant approach was blocks on a 2D canvas. As models got better, prompting got harder because you could do more complex things, and a fashion grew that "the prompt is going away." They disagreed, and the industry looped back to the idea that communicating your vision matters. When they tried to communicate a vision a little ahead of what models could do, the canvas ran out of dimensions. Nest a reflection loop inside a reflection loop and you have nowhere to put it. Code's gift is that it can express very complex systems in a form you can still manipulate and understand. Drop the whole Uber app onto a 2D canvas and the wires become a mess.

The point is not to rebuild Uber's UI in Wordware. It is to build the agent equivalent: decide where the rider is going, where they start, whether they will accept the charge. You probably do not want to order an Uber through a chatbot. You might want one ordered for you when a calendar invite appears. Kozera was heading to the recording and wanted a Waymo at the right time, the way a personal assistant would handle it. That capability, he says, is now open to everyone. The UI you click is still the right abstraction for some jobs; the agent is the one that notices the calendar and places the order.

Huang switches to a lightning round of one-or-two-sentence answers, and Kozera does not stay short. His hottest take, unrelated to Wordware, is that pre-training is still going to matter and that DeepSeek is a little blip people jumped on because they love drama and because it was connected to China. Cheaper training, he says, does not count the experimentation that came before, and he is pretty sure they had access to the best Nvidia chips for that work. The rest of the community will learn from the techniques. Six months later, he bets, it is another place where some open-source models were trained. The model does not matter that much. How you work with the best model out there does.

Who ships the best frontier model next year? He has no clue. OpenAI is always bullish; he had just heard Sam Altman talk about o3 at a YC AI retreat, and the picture sounded great, but he thinks they overpromise. He loves Anthropic's vision. What had been blowing his mind for about four days was Gemini 2.0 Pro ingesting 6,000 pages of PDF. The market, he adds, is super fragmented, and people have zero loyalty. On the claim, associated with Ilya Sutskever among others, that pre-training is hitting a wall, he disagrees. Intelligence scales logarithmically with training resources, but a 2x of intelligence is not a 2x of work. If he is twice as smart as someone else, he does not do twice the work. He finds ways to do ten times, or more.

Favorite new AI apps, not counting Wordware: Descript, because the company now has to edit content to explain Wordware and AI, and Granola, which he uses every day. The newest model that impresses him is again Gemini 2.0 Pro. The category he wants to go mainstream is an AI that has the context of your life. He ordered a Limitless pendant — Rewind, renamed — a year and a half earlier and still did not have it. Huang asks whether that product is Apple over time. He was about to say the same: AR glasses with a privacy brand, and the extra property that even if they run out of battery you are still wearing a $5,000 piece of hardware. For AI practitioners, his reading list is Andrew Ng's DeepLearning.AI, from the bottom layer to the practical one, and 3Blue1Brown on YouTube if you do not yet understand the underlying math.