Sudowrite sold getting unstuck — then eight million people watched a pile-on

Nathan Labenz first noticed James Yu on Twitter in May, when a product video for Sudowrite drew a wave of hate that Labenz ties to replacement fears and the Writers Guild strike. The clip has now reached more than eight million people. What struck Labenz, once he watched it and tried the tool, was the homepage: not a claim about model power, but “the non-judgmental, always available, willing to read 30 drafts AI writing partner you always wanted.” Yu is a co-founder of a product for long-form fiction that already had New York Times and New Yorker mentions in 2021, two years before ChatGPT.

Sudowrite has been around about three years, much of that in private beta. When GPT-3 arrived — the original DaVinci 1.0 — Yu was writing science fiction heavily. People were making memes with the model; he and co-founder Amit asked whether it could actually help creatives. He started using it on his own pages, found it useful, and began building Sudowrite as a toy in late 2020. In that early GPT-3, he says, hallucinations were a great feature for fiction, not a bug. The company grew out of fiction communities and writing groups.

That first model had a 2,000-token window and no instruction tuning, so you had to prompt it as a highly suggestive autocomplete. The first feature was called Wormhole, not “write”: here is what five other versions of you, in other multiverses, might write for the next paragraph. The name set the tone. It might be wild. It might miss your voice. It might still inspire you. The job was to get a stuck writer moving, and a wrong suggestion was put on a pedestal — because seeing the wrong thing can clarify the right one. Output arrived on cards to consider, not as text dumped into the document, a habit they still keep.

Early tricks were mostly few-shot. A Describe tool, meant to thicken a highlighted object, was seeded with examples from Yu’s own stories, so the first versions were closer to “how James would describe this.” User manuscript context still pulled the model into a literary latent space. Prompt *evaluation* mattered as much as prompt engineering: Yu, Amit, and invested writing groups ran panels across a wide range of fiction. They still do.

Labenz kept coming back to that homepage line because it is almost anti-utilitarian. Another company might have sold “the best copy, buy now.” Sudowrite speaks to the author’s need. Yu says the positioning has been consistent from the start. Compare it with an enterprise tool for email: that world is about productivity and words per minute. He did not start using GPT-3 to type faster. He wanted to get into the right headspace.

Early interviews with fiction writers and beta customers kept repeating a second benefit. The tool unstuck them, yes, but it also made them less lonely. Most novelists are lone specialists writing on spec. They do not have a writers’ room. A writing group, if they have one, is asynchronous. Most of the hours are one person, a keyboard, and a brain. Users began to anthropomorphize the model: it felt as if something were actually reading the pages. Some said a life partner would no longer read the manuscript, or they were too embarrassed to show a first draft to the group — but not to an AI.

Hearing that over and over changed the thesis. Sudowrite was not merely next-word completion. It was a way of working with ideas, and a way of connecting with your own work through a cultural lens into adjacent texts in latent space. Labenz is reminded of Eugenia Kuyda, CEO of Replika, who told him they might not create an AI that could talk, but they could create one that would listen. Most of Sudowrite’s users are long-tail independents: people who may never have sold a book, writing for the love of it.

The big unlock with GPT-4 and the newest models is that the system can understand a story from 50,000 feet, not only the next paragraph. Sudowrite is moving from local completion to scene-by-scene understanding to the whole story arc. Story Engine, the new hierarchical mode, is an attempt to mind-meld with the author in a structure that keeps the model in the right context. Claude’s 100k window sounds generous, Yu says, but the pricing often does not work, and you do not always want the whole novel in every inference. Better to give the author tools to feed the right context at the right time.

In the local era they asked for tiny signals — is this scene ominous or happy, who is in it. Now the author can hand over both the 50,000-foot view and the leaves: scenes, who is present, what the task is, whether they need a plot beat or a paragraph. The early DaVinci was a slot machine: maybe one in 50 rolls produced a gem worth keeping. The fight now is to make one in two pulls useful. They are building for centaurs, half human and half AI, “with the human at the head of the centaur, not the butt.” Twitter demos that claim the model can do everything miss the point of writing fiction.

There are as many ways to write a novel as there are novelists, unlike corporate email. The primary UX problem is to stay useful even if an author never adopts Story Engine — freewriting or outlining first should both work. Yu, who programs and writes, wants an IDE for storytelling. Writers have used templates and Excel; they needed a language model to pierce the veil of what is actually happening in a scene. Programmers have had IDEs for ages. Fiction writers have had to do that schlep themselves. The north star is to cut the schlep so they can stay on the part they love.

Labenz tries to reverse-engineer the stack the way people deconstructed GitHub Copilot’s prompt: a summarized book, chapter titles, local before-and-after, fill in the blank. Yu says the main challenge is context stuffing — the right information, at the right time, for the right function. A rich description of a character in a scene might need that character’s fears, or might not. An easier constraint: if someone died in chapter six, they should not walk back in in chapter eight. The long-term picture is a timeline more detailed than an outline, plus world state, plus static facts (afraid of fire, premise, genre, style), all feeding a decider that picks what is relevant. Some of that is heuristic; they are also starting to use semantic search.

He treats models as pens with different styles. GPT-4 is consistent and follows rules, but it talks like C-3PO and is not the best creative-prose writer. Anthropic’s Claude is strong at creative writing and weaker at following rules. Authors have distinct opinions about which they want. Providing a good baseline means understanding each model at the prose level. Humans cannot hold a whole novel in their heads either, so the behind-the-scenes algorithms take an almost anthropomorphic view: give the model the same kind of limited, hierarchical context a person would use.

They are still using a lot of DaVinci, even though those models will be sunset. RLHF, Yu agrees, smooths rough edges into something predictable — you can smell the C-3PO hedging a mile away. They mix older DaVincis with GPT-4. Getting a frontier model into a particular latent space is less effective than raw DaVinci, but possible. Match My Style lets a writer upload a whole scene; Sudowrite builds a prompt that kicks GPT-4 closer to that style without fine-tuning, using the pairing the models have seen between literary reviews and the prose itself. It does not get you 100 percent, which is one reason they plan to explore open-source models and fine-tunes this year. OpenAI, reasonably, is smoothing edges for enterprise, not for creative prose.

Authors who generate whole scenes as a first draft — never as a final — have invented their own tricks. One is kit bashing: run two models, keep a center column, and collage the best paragraphs, Claude versus DaVinci versus GPT-4, the way visual artists paint over a photograph. More sophisticated users pull in open-source models and remix a multitude of outputs. Yu does not personally work that way; a lot of their novelists do.

Labenz’s own micro-app for podcast intros pastes five of his better essays, the new transcript, and a style instruction into Claude 2, because only the long window fits. Past about an hour and a half the model loses the thread even when the tokens technically fit. He often full-rewrites; the draft is something to hate productively, like Mr. Burns at an art unveiling: “I’m no art critic, but I know what I hate.” Yu’s coaching: use the API playground, not ChatGPT. The system message matters. On a long prompt, models ignore the middle, like a person. Put the most important bits at the beginning and especially the end. Sudowrite puts style last, “the tip of the pen,” and re-emphasizes it: write this in the style of…

If you ask a model *not* to do something, it often does it. Yu calls this Pink Elephant syndrome: don’t think about a pink elephant. “Don’t be overly cliché” can make the prose more cliché. Better to give statements or quotes that are not cliché at all, and pull the model into that latent space. Sudowrite’s gold standard is that an author uses about one in two generations. One-off personal prompts will hit lower, which is fine in a chat. For repeatable results, keep a spreadsheet of what wording made you rerun four times versus two, and iterate until you have a Nathan prompt. Many authors already keep a private garden of prompts. Yu thinks gardening that garden will become ordinary writing practice.

Story Engine starts with a free-form brain dump: manuscript pages, a Wikipedia article, random notes. The model is good at turning that mess into a synopsis and outline, then folding the structure back in so it understands the 50,000-foot view the way an LLM understands things. Minds work associatively; writing linearizes; writing *with* an LLM means hierarchically linearizing. Keep silos at each fidelity and pick which layer to feed for the task. Users sometimes discover, once the hierarchy is filled in, that the novel is not about what they thought. Yu still calls the backend v1. Temperature they keep relatively high, with a user knob, because creative writing wants off-the-wall ideas and hallucinations are a gift — but parameters confound. Lower temperature and you may need frequency and presence penalties to fight repetition. Change one dial and you can degrade something you were not watching. The hard problem is cliché, especially after RLHF; neither temperature nor top-p is a hammer for it.

Today Sudowrite serves capital-W writers: professionals, self-published authors, people who consider themselves writers. In the medium term, Yu thinks these models let someone who never considered a novel reach that goal, and pick up craft along the way by watching storytelling tricks in the foundations. Further out, one-to-one media becomes possible. He has a seven-year-old. For that child, intelligent agents are default, not new: of course AI can storytell and paint, and so can he. There is no adversarial relationship — maybe the computer can help flesh out an idea.

He recently threw a Zelda-themed birthday. He made an AI video with his son’s face on Zelda, a scavenger hunt, an AI voice of Zelda talking to the kids, who asked how Zelda knew their names. He is not going to share it; it was for the party. He would never have hired a voice actress for a dozen children. That kind of unscaled, personal media is newly possible, closer to making your own website than to a media business. On the other end of the spectrum he expects a parallel track of novels and films that are fully AI-generated, or generated with a human producer in the middle, including a film made just for you. He does not fully buy that it happens at scale. There is still a water-cooler effect: people will seek work directed by humans. Books and films made the old-fashioned way — including the old-fashioned way with Sudowrite — should remain. Reading has gone up; book sales have risen even in a TikTok world. Culture still needs a cornerstone you can talk about.

He loves science fiction as a meta-layer: technologists read it, then ship products designed from worlds written 40 or 50 years ago. To imagine 30 years out, he would rather read older Gibson or *The Three-Body Problem* than scroll Twitter demos. Controllability still feels pre-iPhone. ChatGPT may be *an* iPhone moment, but he expects another, when a system understands intent well enough that creation no longer feels like a slot machine. Some humans should stay in the loop directing; he worries about taking people totally out, if not on the same spectrum as the doomers. Labenz points at cyborgism and Neuralink as cousins of that agenda: higher-bandwidth communication with machines, closer to internal mental state than language can capture.

Two or three months before the taping, Yu posted a roughly nine-minute Twitter video of a big Sudowrite update. In isolation Labenz would have called it a thoughtful implementation. Instead Yu became, for a minute, public enemy number one of writers. It was his first time getting canceled. He still stands by the demo. If he added nuance, it would be: this is great at a first draft, not “write the entire novel.” The split is more complex than anti-AI. Indie writers using Sudowrite often tell their audiences they have an AI assistant, and those readers say write 20 books a year instead of ten, give me more. The product was not born in isolation. They work hand in hand with hundreds of novelists, some of whom cried tears of joy when they got it.

He understands the third rail. Writers are not compensated well enough; Hollywood was in the middle of that fight, and a deal looked to be going through as they recorded. Writers should be paid fairly for the work they make, he says, and that question should be separated from the tools they use. “We’re not out here to replace writers. We’re out here to create tools for writers.” Usage is as diverse as Photoshop once was: some people generate whole scenes, some only brainstorm and take none of the words. Digital art drew the same backlash; almost no one asks whether it is art today. Sudowrite also runs weekly classes that mix product craft with becoming a better creative writer. Amit and Yu are both fiction writers and have not pivoted in three years.

What concerns him more is a witch hunt in author communities trying to out people for using AI, which he thinks is a transitional state. Labenz had previously talked with three WGA members whose issue was contractual: they did not want AI assigned to roles with payment rights attached, and seemed content to keep experimenting if they got that agreement. Yu’s read on the pile-on is that the high-order bit was Twitter itself. Analytics showed almost nobody watched past the first five seconds, which had basically no content. Retweets and a few accounts with the right dunking verbiage made him the poster child of the AI bro coming for writing jobs. Advice for other builders: Twitter is not always the right conduit; put a customer on camera talking through a real workflow. Patterning the launch on tech-Twitter demos made it easier to take at the wrong face value.

They raised a small round in 2021. The team is seven people, including Yu, with four developers — three of the four have English degrees and come from that world. They have been profitable since March, which gives them room to grow mindfully on this problem. When the storm hit, the instinct was to reply to everyone. They kept their heads down and built for the writers who were filing feature requests. The little rectangle screamed at him; he came out more resilient, and more careful not to be a dunker himself. He looked for the people having nuanced conversations and talked to them.

Paying customers barely noticed. Someone posted in the community Slack and a lot of people said wait, what’s happening? The business impact ran the other way: the tweet is part of why they are profitable. In any pile-on there is a large middle who say nothing, check the product out, and make up their own minds. They are not tweeting. As a founder you mostly hear from the people who are online in that hour with an agenda.

Labenz’s own company, Waymark, had smaller viral spikes and then the classic AI-app problem: trial users, non-trivial token cost, churn from people who need one artifact and cancel. Yu’s answer is that this is still a SaaS problem. AI companies are still startups. Distribution, churn, pricing, and whether users tell other users still matter more than the model discourse on Twitter. You do not sprinkle AI sparkle on a demo and take off. The gulf between a cool Twitter demo and a profitable business is immense. Know the customer and the domain until the tool is indispensable. AI margins can be worse: imagine a database that charged 30 cents per row, every save and every update. You have to tune context windows and how many free inferences a trial gets more carefully than a normal SaaS tunes database cost.

Sudowrite has stuck to an Apple-like philosophy: the best, latest models, and you pay a little more for that. Founders should pick that philosophy top-down — state of the art, high volume, or bring your own key — because it determines the audience. They skew toward working writers who use the tool daily, a space Yu finds easier because you can charge for value. He still spends as much time on the pricing flow as on futuristic features. “This is not rocket science, but you do need to work at it.”

Labenz has been circulating a speculative “AI bundle”: a consumer pays perhaps $100 a month for baseline access to a thousand apps; each participating developer might get on the order of ten cents, still able to upsell a premium plan, with less fear of trial abuse and viral spikes. For a million subscribers that is roughly a million dollars a year of floor revenue. Yu is interested; the devil is in the payouts, and the audience might be too diffuse for fiction. He jokes that it is cable for inference. He is also drawn to a world where OpenAI or Claude offer “log in with” so your intelligence, prompts, and settings travel across apps. Coordination costs are high, and he is not sure OpenAI is incentivized, but it would be better for the consumer. Labenz quotes Sam Altman: a lot of people thought they wanted their app in ChatGPT, but what they really wanted was ChatGPT in their app.

On multimodal, Sudowrite will probably not go deep. They have a small visualize button that uses DALL·E on the back end; they want to stay on text, story-line editing, and story understanding. One later expansion: Story Engine as a backend for other apps that need legible stories — a Midjourney front end that storyboards from their format. Text still has thornier, more authorial problems than images. He is interested in a writers’-room device that sees the sticky notes and hears the conversation, and in a simpler idea: walk around San Francisco with AirPods, talk for two hours, and come home to a pre-assembled outline from the ramble. Asked for a tool people are sleeping on, he will not say ChatGPT. As a programmer he has been using Cursor nonstop to rewrite highlighted functions, beyond what Copilot can do. For the Zelda party he combined Midjourney and Pika Labs for motion, DALL·E to flesh out prompts, and ChatGPT to generate a hundred house-scale, Zelda-themed scavenger-hunt clues.