A Hundred People, a Billion in Revenue, and Zero Venture Money

Sarah Guo and Elad Gil introduce Edwin Chen as the founder of Surge, a bootstrapped human-data company that, they say, surpassed a billion dollars in revenue last year and counts Google, OpenAI, and Anthropic among its clients. Chen confirms the number: over a billion in revenue, a little over a hundred people, and — in his telling — the biggest human-data player in the space. The company had stayed under the radar until recently. It has been around five years, founded in 2020, on a thesis he states without ornament: human data advances AI, and the job from day one was to make that data as high-quality as possible.

Before Surge he worked on machine learning at Google, Facebook, and Twitter. The problem that kept repeating was data. Even basic things, like training a simple classifier, were blocked by the inability to get what the models needed. If they could barely do that, he asked himself, how would they ever build the next generation of systems? That blocker is why the company exists. The product, when they finally describe it, is not a marketplace of contractors. It is the data itself, plus the failure modes around it.

Surge was profitable from the start, Chen says, so it did not need the money, and giving up control always felt strange. What he hates about the Valley is raising for the sake of raising. Founders he knows do not have a product they believe in. Talk to a cluster of YC founders and the goal is to tell their friends they raised $10 million and to show their parents a TechCrunch headline. Friends who have spent a decade at Google or Facebook can pay their own salaries for months, yet the first move is still to raise. Ask what problem they want to solve and they do not know. They are bored. They want to start something new.

They might talk to users and build an MVP, but in a throwaway way, to check a box on an accelerator application. Then they orbit VC dinners and DMs. From first principles, if you did not already know that raising was a thing, why would you do it? What does money actually solve for 90% of startups whose founders have savings? His instinct is to go build the thing you are dreaming of, and think about raising if you hit a real financial problem — not before you know what you would do with the cash. Guo, speaking as an investor, often talks founders out of fundraising inside Silicon Valley and sees the opposite outside it: too few people raise when the money would actually help them scale.

Gil presses the young-founder case: no network of engineers, no work history, straight out of school — is a round not the way you attract talent and pay them? Chen splits it. If you literally have no savings and need money to live, raise. If you have been at Google for years, ask whether you really need to hire all those people. He hears founders plan their first five to ten seats as a PM and a data scientist. He would never hire a data scientist that early. He used to be one. Data scientists are great when you want to optimize a product by 2% or 5%. A new company is swinging for 10x or 100x; those percentage points are noise. Product managers become useful when the company is large enough that engineers no longer have time to live in the details. At the beginning, the founders and the engineers should be having the ideas themselves.

Guo asks what the billion dollars is actually for. Chen's answer is blunt: the product is the data. Frontier labs that want better coding models get several kinds of it. SFT examples — handwritten solutions and unit tests a good program must pass. Preference data — two pieces of code or two explanations, and a judgment about which is better. Verifiers — a web app that should have a login button in the top right, or a click that should cause something specific to happen. Alongside the data they deliver insights: loss patterns, failure modes, where one model is weak and another is worse. The universe around the data is the product.

What is repeatable, versus what competitors miss? A lot of companies in the space, he says, are body shops. They are not delivering data. They are delivering warm bodies, which means they have no technology. Surge's bet is that quality is the whole game: is this a good coding solution, a correct math proof, a great poem? Historically the industry treated data as a commodity. Draw a bounding box around a car and Sarah, Edwin, Hemingway, and a second grader will draw roughly the same box. The ceiling is low. Ask for a poem and Hemingway leaves Chen in the dust. Ask for a VC pitch deck and Guo will beat him. In generative AI, the ceiling on quality is almost unlimited. Without a platform that measures what annotators produce, you do not have a data product. You have staffing.

Gil asks how the feedback loop works: who evaluates the evaluators? Chen reaches for Google Search and YouTube. Millions of pages and videos cannot be judged by staring at them. You gather page signals, user signals, activity signals, and feed them into a giant ML system. Surge does the same with annotators — their work, their activity on the site — and an internal ML team trains algorithms to score it. Who evaluates the evaluators? The algorithms do. Measurement is what makes quality more than a slogan.

As models pass the 90th percentile of a random human, the annotation task keeps getting harder. That is scalable oversight, a live research problem in alignment: how do models and humans work together to produce data better than either could alone? A few years ago an annotator might have written an SAT-style story from scratch. Today that is inefficient. You start with a model draft — often vanilla, full of croft that does not benefit from human ingenuity — and a person edits it substantially, layering creativity onto a skeleton. The practical question is how you build the right interfaces and tools so people and AI combine in the right way. Surge spends a lot of technology on that.

Labs have shifted, in recent months, toward RL environments and reward models. People underestimate how complicated those worlds are, and that you cannot just generate them synthetically. Imagine training a sales agent. The person lives in Salesforce, takes leads from Gmail, talks to customers in Slack, tracks them in Excel, writes Google Docs, and makes PowerPoint decks. The calendar has a meeting; a car accident notification means leave earlier. You need thousands of Slack messages and hundreds of emails that stay consistent with one another, with time actually evolving and external events that are interesting, realistic, and not incongruent. That takes a lot of tools and a lot of sophistication. How real is enough? Chen sees no ceiling. More diversity and richness, longer time horizons, more for the model to learn. A five- or ten-year bet on demand is not RL environments *or* expert reasoning traces. It is all of the above. Trajectories are long; a single reward is not rich enough to capture a complicated goal.

If models become superhuman, when does human input run out? Chen says never, for a couple of reasons. People overestimate synthetic data. Surge uses it heavily for the croft that is not worth a person's time. Customers still show up with a familiar script: for six months they experimented with synthetic data, gathered 10 to 20 million examples, and realized 99% of it was not useful. They are trying to curate the 5% and throw out the rest. Often a thousand pieces of highly curated human data are more valuable than those ten million points.

The other reason is that models think differently from people, so they need an external signal to stay on the objective you actually want. He will not name the lab, but one frontier model that everyone treats as top-tier will, maybe 10% of the time he uses it, dump random Hindi and Russian characters into the middle of an answer about Donald Trump or Barack Obama. The model is not self-consistent enough to notice. Someone outside has to say this is wrong. Gil adds a biology rhyme: protein evolution against a catalytic function, from a giant random library, also coughs up weird unanticipated activities. Training is an evolutionary landscape. The feedback signal you select against shapes the monster you get.

He calls LMSYS / LMArena a plague. The mental model is a user writing a prompt, getting two responses, and spending five or ten seconds picking whichever one *looks* better — not checking hallucinations, factual accuracy, or instruction following. Formatting, emojis, and a more impressive vibe win. Train on that subjective soup and you do not see the consequences. The model does not know its objective. Without an external quality signal it runs in crazy directions; he alludes to what the industry saw with Llama 4. The alternative every frontier lab treats as gold standard is proper human evaluation: time to fact-check, to verify instructions, people with taste judging writing. Skip that and you are training on the analog of clickbait. Internally Surge already does this with the labs, surfacing loss areas. They want an external version: which models code well, which follow instructions, which hallucinate enough that you should not trust them.

The Meta–Scale deal, Chen says, was on balance good for Surge. They were already the number one player, but some legacy teams still used Scale simply because they did not know Surge existed. Low-quality data vendors burn customers on the whole category of human data; those teams then try slower methods with the wrong objectives, which harms model progress. Getting more frontier labs onto high-quality data helps the industry, not just the vendor. Asked which underdog might catch OpenAI, Anthropic, and DeepMind, he bets on xAI: hungry and mission-oriented in a way that gives them unusual advantages.

He does not think frontier models will collapse into a commodity. The last couple of years surprised him with how different the strengths already are: Anthropic on coding and enterprise, OpenAI on consumers via ChatGPT, Grok willing to say and build a different set of things. Each company has a different set of principles, a different personality, a limited number of focuses. AGI may eventually encompass all of it; until then, providers will keep different strengths, and users — including Chen — will keep switching models by task. Over the next few years he is excited about a public research push. Frontier labs have mostly stopped publishing. The industry has fallen into a trap of bad incentives that nobody corrects.

Researchers tell him their VPs make them raise the LMArena rank. Some have said they are fine making the model worse at factuality and instruction following if the ranking goes up, because leadership wants the metric. Voters in those arenas do not care about following instructions or facts; they care about emojis, bold words, and long answers that look impressive. The easiest way to climb the board, Surge has found, is to make responses longer. Companies then see "progress" for six months that is just more length and more emojis, realize they trained a better clickbait machine, and discover they made zero real progress — Llama 4 is the industry example he points at. Academic benchmarks wander off the real world too. IFEval, an instruction-following eval, includes items like: write an essay about Abraham Lincoln, and every time you mention the name, five letters capitalized and the rest not. Customers still come asking to raise that score. Surge wants to publish and educate so the industry steers toward real-world progress instead of SAT tricks.

What does high-quality data actually mean? Suppose you want a model to write an eight-line poem about the moon. Most companies hire off Craigslist and check three boxes: is it a poem, is it eight lines, does it contain the word *moon*? You get high-school verse. The upgrade is English-literature PhDs, which is also wrong: many PhDs cannot write, and Hemingway and Emily Dickinson did not finish college. Chen went to MIT, as did Gil; plenty of CS graduates he knew were terrible coders. Surge wants the kind of poem a Nobel laureate would write, and it wants to notice that there is no single correct moon poem. A haiku of moonlight on water, a piece built on internal rhyme and meter, a poem about the feeling of the moon rising — thousands of ways, each a different insight into language and imagery. Math is the same: maybe a thousand proofs of the Pythagorean theorem. Quality understood as inter-rater agreement and checkboxes gives you commodity data. Quality that embraces intelligence and creativity teaches models the deeper patterns that make language compelling. Throw humans at the problem without that, and you are scaling up mediocrity.