The CEO Who Laid Himself Off Is Building a Business Operating System

Nathan Labenz opens with Siqi Chen, a four-time CEO, three-time founder, math major, and prolific commentator on X, now running Runway — a finance platform, not the video-generation company of the same name. Chen's pitch is dry on purpose: make the business understandable not just to people in finance, but to everyone else. He calls it the most interesting problem of his career. The insecurity behind it is older than the product. He never thought of himself as a business person, and that identity gap got worse with every company he ran.

It peaked when he was appointed CEO of Sandbox VR shortly before Covid. A financial model, he reminds Labenz, is not just a Wall Street artifact. CFOs and founders use it as a simulation of the business: raise more, change prices, spend on marketing, hire, and watch what happens to cash and margin two years out. When the pandemic arrived they had to scenario-plan for three months, six months, or twelve, all in spreadsheets. They laid Chen off. Headcount went from more than 400 to around ten. The founder hibernated the company.

Half an hour later he was on the phone with the CFO and the Andreessen Horowitz team that had invested in Sandbox. Figma existed. Notion existed. Surely, in 2020, there was something better for this. There was not. Spreadsheets can do everything, he says, which is why they are not really good at any particular thing. Employees get used to treating finance as a mystery they are not supposed to understand. If they could see it more clearly, they would make better decisions and collaborate better. That is how Runway started, and AI is supposed to make the whole thing accessible.

Labenz had taken a demo the week before and sent the link to the CEO who replaced him at Waymark. The moment that shifted how he thought about finance was a text prompt: ramp marketing enough to 3x revenue over the next twelve months. Runway spat out a full scenario. Chen walks through what a CFO actually does with that question. You talk to whoever owns the model. You guess which drivers most affect revenue — sales headcount, marketing spend, pricing, conversion — you change the numbers, you look at the output, and you iterate until the constraint is met.

Language models, he says, are good at the human part of that loop: they can understand a constraint, apply common-sense reasoning about which variables probably relate to revenue, and judge how far the result is from the goal. They are bad at the part humans are also bad at, which is mental arithmetic. Almost nobody can do complicated math in their head, which is why spreadsheets exist. Runway's move is to hand the LLM a calculator. The model proposes which levers to pull and when; the model engine does the actual computation; the system compares the output to the constraint and tries again.

That loop is much harder on a spreadsheet because the context is missing. Product roadmaps, hiring plans, fundraising plans, and marketing plans usually live in Notion and Google Docs, not in rows. Runway had to invent new primitives — they call them plans — that can hold business intent and sit next to the driver logic. On the backend those drivers start as UUIDs. The team hydrates them into semantic descriptions, then uses AI itself to write explanations of the submodels: how runway is calculated, how burn is affected by expenses and revenue. The more of that context the LLM gets, the better its first guesses. Labenz recognizes the pattern from Waymark, where the hard early work was defining a lightweight syntax the model could emit and the backend could rehydrate deterministically with code.

Chen then zooms out to how AI has shown up in products since GPT-3. First came cool demos that called the API. ChatGPT was the interface people actually understood. Then every product added a copilot: a slide-out pane and "chat with whatever." Now the industry is building agents, except the models still cannot self-reflect well enough to do real work. The next step-change, he thinks, is better reflection and planning. Until then, the chat interface has two structural holes. It usually has almost no context about who you are or what your company is doing. And despite how many people use ChatGPT, surprisingly few of them understand what they can even ask it to do.

Runway's answer is what they call ambient intelligence: the system does the work in the background without being asked, and the result shows up as ordinary interface. Hover a driver and there is already a description. You did not prompt for it. It does not look like AI. He thinks Apple is aiming at the same place — surface the capability as text or suggestions inside the normal workflow, not as a conversation with a message window. Labenz asks when it became obvious that models would actually help a finance product. Runway started a few months before GPT-3 shipped. Naval Ravikant wrote the first check and named the company.

Chen had asked Naval what he thought of the name cfo.ai. They still own the domain and do not use it. Naval's 2020 advice was not to put AI in the name, because it would date quickly: everything was going to be AI. In retrospect Chen finds that almost painfully accurate. Nobody describes a product as powered by JavaScript or AWS. AI should become the same kind of invisible capability. The mission never changed: make business accessible and understandable to everyone. That is also the thesis of Universal Primer, the GPT Chen currently has ranked number one in the ChatGPT store's education category. Large language models, he says, are already good at taking something complicated, explaining it in plain English, and filling in the prerequisite concepts you are missing. He confirms he is getting paid for it. The amount and the method are under NDA.

Labenz wants the interface between the language model and the deterministic product. Is it tool use? Is it an iterative hunt for a goal? Chen says yes to the second. You can give the model the entire dependency tree, ask it which drivers will perturb the constraint most, then ask it *when* to apply them — three months out, six months out, not five years out if the goal is 3x next year. Each guess runs through the model engine. The system reads the revenue line, compares before and after, and tells the model how far off it was. The LLM is not doing the math. It is pulling levers on a black-box calculator, the same way a CFO would. It does not always converge. As models get better at intuition, planning, and self-reflection, both the guesses and the orchestration around them get better.

Labenz wonders if this is just OpenAI's Assistants API. It is not. Every customer's model is different, so Runway built custom orchestration. Asked what will drive the most user-facing improvement — scaffolding or raw scaling — Chen says both, then adds a more basic layer. Orchestration sits on a substrate. It would be very hard to run this workflow on a spreadsheet, because the abstractions in a spreadsheet are cells and rows, and those are not well suited to hold a product roadmap, a fundraising plan, or a marketing plan. Job one was building a substrate that can connect that context to the business model and pass it to an LLM. Larger context windows help after that. The substrate is not something an LLM will invent for you out of the box.

That is why he does not think applied-AI companies can skip straight to the model, at least for the next year or two. The unfair advantage is a tool for thought with the right abstractions, already in use, that can hold more and more of the context people already have. Once you have that, feeding an LLM becomes a much better starting point than dumping a grid of cells into a prompt.

Labenz has the integration count in his notes: about 680. He assumes they started for non-AI reasons and turned into a Swiss Army knife for context. Chen's framing is larger. Finance is a simulation of the entire company. Every person and every department eventually shows up on the bottom line. Today the sales team has a model, the product team has a model, growth has a model, marketing has a model, and they live in different spreadsheets that do not talk to each other. Runway integrates widely because you want one view of how the marketing plan affects the sales plan affects the financial health of the company. A good number of customers even enter plans directly.

The strangest integration is Jira. A finance platform ingesting tickets is unheard of, not least because finance people barely use existing finance software. The reason it works is an abstraction called plans that can encapsulate business intent. Attribute a ticket to conversion, conversion to growth, growth to margin. Labenz imagines asking: if ticket A lifts signup by some percent and ticket B lifts retention by another, which one moves the two-year P&L more? Chen's nuance is that you can already do the single-ticket version today without AI — click the thing, see runway, revenue, margin, burn. AI gets interesting when there are a hundred candidate tickets and you want the highest-ROI one. They do not ship that yet. What they do ship is budget-versus-actual: revenue is off the year-ago plan by 20 percent, drill in, was it pricing, leads, or conversion, using the warehouse and analytics already wired in. Traditionally you trace that dependency tree by hand.

How big does the vision get? On LinkedIn Chen describes it as disrupting the $70 trillion business industry. Stripe's mission, as he hears it, is to increase the GDP of the internet; Runway's is to increase the GDP of the world, by letting everyone in the company understand the plans and the impact of their own decisions. He wants a world where a designer, a PM, and an engineer can say: if we work on X instead of Y, here is how margin and growth move two years out. People work in businesses and still accept that almost nobody is supposed to understand them. They market Runway as a finance platform because that is the understandable frame. Internally they think of it as a business operating system. Chen's video-game version of that idea is Civilization or RollerCoaster Tycoon: understandable, visual, interactive, immediate feedback, a mental model you can actually hold. Real companies do not feel like that, and he argues the tools are the reason.

Labenz had just been at the AI Engineer World's Fair, where the running theme was that demos are easy and production is hard. Has Runway tamed scenario generation? Chen is blunt: they have not tamed it very well at all. The models hallucinate, misunderstand, and get things wrong. He would not trust that use case out of the box. It is their version of the shiny AI demo, and it will get good as capabilities and orchestration improve. The work that produces value today leans on what LLMs are already good at: making complicated things understandable. Explain mode is the example. Hold Control, mouse over any part of the product, and it tells you what you are looking at. The design premise is that most of the time you do not even know what question to ask, so the product should not wait for a prompt. Pull requests between models get the same treatment: the diff is just explained.

When labs ask companies what they need GPT-5 for, Chen's answer is reflection. Look at the output. Is it good? Is it accurate? How would you change it? Then write a better prompt. He treats that as the highest-leverage gap in orchestration and, personally, as the barrier to AGI. Below a threshold the loop goes nowhere: it hallucinates and loses track. Once a system can continuously inspect its own orchestration and prompts and get better, you are approaching self-bootstrapping. Today's models, he thinks, already answer many questions better than half the population. The real bar is the extreme end: the humans capable of improving the LLMs themselves. Labenz's coordinate is similar. He dates the frontier model passing the average person on the average task to PaLM and BIG-bench in the summer of 2022. Now the models are closing in on expert performance on routine tasks, and the word *routine* is doing a lot of work.

The hover-to-explain UI looks like it might have required a giant labeled screenshot set. It did not. Chen calls it mostly a design problem, not a multimodal science-fiction problem. They walk the DOM of whatever the mouse is over, classify the object — a driver, a report, a pull request — and show a pre-generated explanation. The industry, he thinks, systematically undervalues design. The iPhone did not win on new foundational-model capabilities. It won on a well-considered design. Applied AI has a few labs with the most capable models, and a much larger unsolved problem: how to express those capabilities to customers in useful ways. Right now the bulk of that expression is still "chat with a bot."

Runway has an internal goal of never having more than 100 employees. AI is supposed to keep raising leverage, and a small company is where people still have freedom and ownership. More people, Chen says, more problems. Two workflows show what that looks like in practice. When the website went viral, leads poured in faster than the team could qualify them. New signups get enriched through Clearbit, then an LLM scores them against criteria Chen would give a human SDR in ordinary language: growing companies, a key decision-maker or someone in finance, largely based in the US, ideally more than 50 people. Once a lead clears, another prompt — heavily tweaked — drafts a follow-up from the website, the value proposition, and whatever else they know, with instructions to be clever, not cringe, not mid. The draft lands in the salesperson's Gmail drafts folder. Morning inbox: pre-qualified, pre-written, press send.

The other workflow is culture interviews. Chen used to type notes while talking, which made it hard to connect with the candidate. Grain records the session. He pastes the transcript into an internal GPT and gets notes in the format he wants, pulling out motivation, values, and the other traits he screens for. The model is not good at the hiring decision. It is very good at extraction. Tooling is pragmatic: ChatGPT and Zapier when no-code is enough, custom code when it is not. He built a Notion bot that watches for a mention of Steve Jobs or Jeff Bezos and comes back in that voice to critique the document. Most of the code was written with ChatGPT. The quality, he says, lives in the prompt: do not say anything generic, make the insight novel even to an expert, and imitate how the person actually speaks in private, not the media persona. The more stringent the quality bar, the slightly better it gets.

What the models still cannot do is write his tweets. Good writing, in his view, is not about the prose. It is about the ideas, and the most interesting ideas sit at the tail of the training distribution. He has a human ghostwriter for LinkedIn, which is a different register, and the bottleneck there is the same one an LLM would hit: imparting lived experience and stories. Labenz offers a counter-example. For podcast intros he feeds Claude 3 thirty handmade essays plus the current transcript; three episodes out of four he could ship the draft, unless he has a specific take he wants in the open. Chen agrees that tracks. Enough examples plus a bounded task, and the model can sound like you. A timely, surprising tweet is a different object. You can tell when someone posted something ChatGPT-shaped. It is, as he puts it, very mid.

Labenz asks how a near-term AGI outlook changes investing, parenting, and the company itself. On investing Chen talks his book: the two pillars are proactivity and context, mostly at the application layer. He backed Automatica, an AI SDR platform, and Maestro, an AI chief of staff. He was a small investor in ElevenLabs and a few other companies, and he expects relatively few new foundation-model labs. How does AGI change that? The same way it changes Runway: on some level all bets are off. A short hop from AGI to ASI would dwarf his company. The impact on him, in absolute terms, is probably much smaller than the impact on Microsoft or Amazon, because the questions become what work even is. He does not spend much time there. The one company-shaped thought he does entertain is a Runway agent: a CEO agent that holds a holistic view of everything that happens in the business. That complete context is scarce data. Very few companies have it.

For his kids the advice is the same as it would have been before AI: be makers. Learn to code, make things with your hands, make games. Over Christmas two nine-year-olds — his and a friend's — decided to build a Transformers-meets-Gundam game. They asked ChatGPT for art, for rules, for how to get Unity, then downloaded Unity and started. If AI can eventually make the whole thing instantly, why bother? Chen's answer is that things provably made by humans will still be valuable. He keeps coming back to Charlie Munger's observation that people think greed drives progress, but it is actually envy: social and status competition that has been with humans and other animals all along. Even after a singularity, people will invent new status games. The valuable software might be the version you can prove was coded by hand, without AI. Nobody watches AlphaGo play AlphaGo. They watch Magnus Carlsen, because a human is interesting. There will always be something for humans to do, he adds, unless AI kills us first.

On regulation he is generally skeptical of government intervention, not because he thinks nothing will be tried, but because he is not sure anyone knows the right move. Compute is already being regulated, and he does not know that is the right target, or that today's architectures could not still reach ASI. He is more bullish on mechanistic interpretability: understand how the systems work through better technology. One concrete worry is that if AI is as powerful as people think, regulation puts the rule-followers at a disadvantage against less scrupulous labs and nation-states. If OpenAI decides to be super safe, someone else just goes faster. If America regulates compute, another country will not. Labenz would rather go down trying to cooperate with China than racing it to godlike AI. Chen does not argue the other way. He just calls it a difficult problem. His closing plug is simpler: if you run a midsize company, pay for Runway.