十亿美元、没有 VC、不到一百人：Surge AI 如何把数据质量做成一门艺术

Chen says Surge hit over a billion dollars of revenue last year with under 100 people, and that he expects even crazier revenue-per-employee ratios within a few years. At the big tech companies he worked at, he felt that firing 90 percent of people would make things move faster, since the best people would have fewer distractions. Surge was designed from the start as a very small, elite team.

He thinks two forces are colliding: people are realizing giant organizations aren't needed to win, and AI is adding efficiency. Fewer employees means less capital, which means no need to raise, so founders who are great at technology and product, rather than pitching and hype, can build things they care about. He hopes this takes the tech scene back toward being a place for hackers.

Chen says he never wanted to play the Silicon Valley game, which he always found ridiculous, because it means explaining every decision to VCs and joining a PR and fundraising hamster wheel. Staying out of it made things harder, since Surge had to win through a product ten times better and word of mouth from researchers. But it also meant early customers were people who deeply understood and cared about data, and who gave useful feedback.

Chen describes Surge as a company that teaches AI models what is good and bad using human data, with products including SFT, RLHF, rubrics, verifiers and RL environments, plus measurement of model progress. He says most people don't understand what quality means in this space and wrongly think you can throw bodies at a problem. His example is an eight-line poem about the moon: a checklist approach would only verify the line count and the word moon, while Surge wants Nobel-level poetry that is unique, full of imagery and makes you think.

Measuring this takes technology: thousands of signals on every worker and project, including keystrokes, answer speed, reviews and code standards, and Surge also trains its own models on the outputs to see whether they improve performance. Chen compares it to Google Search, which both removes the worst pages and finds the best ones. The result is effectively a complicated machine learning problem about people and tasks.

Asked why Claude was so much better at coding and writing for so long, Chen says data is a big part of it, since frontier labs face an almost infinite set of choices. They decide whether to use only human data, what to ask people to create, whether to prioritize front-end visual design or back-end correctness, and how much synthetic data to use.

Some companies optimize for academic benchmarks for PR reasons even if they don't think they matter, while principled ones care about real-world task performance. Chen says there is an art to post-training, with taste and sophistication involved, such as preferences for minimalism or 3D animation in design. The labs with more taste realize good poetry doesn't reduce to a handful of checkboxes and consider the implicit, subtle qualities instead.

Chen says he doesn't trust benchmarks at all. Many contain wrong answers and mess, and their well-defined, objective answers make them easy for models to hill-climb, unlike the ambiguity of the real world. That is why models can win IMO gold medals yet have trouble parsing PDFs.

He adds that benchmarks sometimes leak, and labs can tweak system prompts or the number of runs to game them. Optimizing for the benchmark instead of the real world is another form of gaming. For measuring progress toward AGI, Surge instead runs human evaluations in which expert annotators, such as a Nobel-winning physicist, a teacher or a big-tech coder, hold conversations with models, check the code and equations, and judge accuracy and instruction following.

He contrasts this with casual users on a popup comparing two responses, who are just vibing and picking the flashiest answer. On whether humans will become unnecessary, Chen says not until AGI is reached, and he sits on the longer-timeline side: he bets models will automate 80 percent of an average software engineer's job within one or two years, but the jump to 90, 99 and 99.9 percent will take years each, so AGI is closer to a decade or decades away.

Chen worries that instead of building AI that cures cancer, solves poverty and helps understand the universe, the industry is optimizing for slop and teaching models to chase dopamine instead of truth. He points to LM Arena, where voters skim responses for two seconds and pick what looks flashiest. A model can hallucinate everything and still win with emojis, bold text and markdown headers, and he says the easiest way to climb is more emojis, longer answers and bold, even at the cost of accuracy.

The incentives push researchers too: enterprise customers ask why a model is only number five on the leaderboard, and researchers tell Surge they will only get promoted if they climb it, even if that makes the model worse. He also warns about engagement optimization, drawing on his work in social media, where it produced clickbait. In AI the same pressure produces sycophancy: telling users they are geniuses is the easiest way to hook them.

Chen says he has always been impressed by Anthropic's principled view of how its models should behave. He also says that which companies would build a product like Sora reveals what kind of AI and future they want, and compares funding a newspaper by selling tabloids: the path you take matters, and it can distract from more important things.

Chen rejects the standard Silicon Valley playbook of pivoting every two weeks, chasing growth and engagement with dark patterns, and hiring as fast as possible. His advice is not to pivot, not to blitzscale, and not to hire the Stanford grad who simply wants a hot company on a resume. Instead, he says, build the one thing only you could build, based on insight and expertise only you have.

He criticizes founders who moved from crypto to NFTs to AI with no mission, just chasing valuations, and says he has a romantic notion of startups as big risks on things you believe in. Failing because the market wasn't ready is better than pivoting into another LLM wrapper company. Lenny says he is seeing the same pattern, wild ambition, in research on early employees of generational companies.

On whether LLMs are enough for AGI, in response to the Richard Sutton conversation, Chen says he thinks something new will be needed. He takes a roughly biological view: humans learn in a million ways, and models should mimic those ways. To the extent LLMs learn differently from humans, he says, something new will be required.

Chen explains that reinforcement learning trains a model to reach a reward, and an RL environment is a simulation of the real world, like a video game with a fully fleshed-out universe. His example is a startup with Gmail, Slack, Jira and GitHub, where AWS and Slack go down and the model must work out what to do. Models that look strong on isolated benchmarks, good at single-step tool calling, often fail catastrophically in messy worlds where step one affects step 50.

Rewards can be passing unit tests or producing a document that matches what happened. A financial analyst could build a spreadsheet and tools, such as a Bloomberg terminal, and the reward might check whether a specific cell holds the right profit and loss number. Chen says experts move from writing rubrics to designing environments.

He stresses that trajectories matter because a model may reach the right answer after 50 failures, inefficiently, or by reward hacking. Checking only the final answer misses information about how it behaved. Sometimes you want a model to get there by reflecting, and sometimes by one-shotting it.

He summarizes the progression: SFT is like mimicking a master, RLHF is like writing 55 essays and being told which one someone liked best, and rubrics and verifiers are like being graded with detailed feedback. RL environments are the newest, and complement earlier methods rather than replacing them. Chen says Surge needs a suite of products reflecting the thousand ways humans learn.

Chen says Surge has two kinds of researchers. Forward-deployed researchers work with customers to show where their models lag and design datasets, evaluation methods and training techniques. Internal researchers work on better benchmarks and leaderboards, train their own models to learn which data and which people perform best, and improve data operations.

He says he thinks of Surge more as a research lab than a startup, and that he would rather be Terence Tao than Warren Buffett. For researcher hires, he looks for people who could spend ten hours digging through a dataset and playing with models, hands-on and interested in qualitative failures, not just abstract algorithms.

Chen says a year ago he assumed AI models would be commoditized, but now thinks the values of a company shape its model. His example is asking Claude to draft an email: after 30 versions and 30 minutes he realized it didn't matter. Should a model say there are 20 more ways to improve your email, or optimize for your time and say it is great and to send it?

He compares this to Google, Facebook and Apple building different search engines according to their principles. Lenny notes Grok already has a very different personality. On under- and over-hyped areas, Chen thinks built-in chatbot products like Claude's artifacts, extended to mini apps and UIs, are under-hyped, and that vibe coding is over-hyped because it risks unmaintainable code bases.

Chen traces his path: fascinated by math and language, he went to MIT partly because of Noam Chomsky and worked as a researcher at Google, Facebook and Twitter, where the same problem kept recurring: it was impossible to get the data to train models. When GPT-3 came out in 2020, he realized a new solution was needed, and he started Surge a month later, focused on advanced, complex use cases instead of simple image labeling.

He calls himself a scientist at heart who dreamed of being called by the government to decipher an alien language. He still writes deep-dive analyses of each new model for customers, often presented as from the data science team, and is terrible at sales and meetings. He says Surge is built like a research lab, caring about curiosity and long-term incentives more than quarterly metrics.

Chen frames the deeper mission as helping customers define their dream objective functions, like deciding what kind of person you hope a child becomes versus which test score they get. Clicks and likes are easy proxies, but measuring whether AI makes life richer is harder. His line is that you are your objective function, so Surge wants rich, complex metrics and tools that make people more curious and creative, not lazier.

Asked what he wishes he had known, Chen says he never planned to start a company and feared it would mean financials and meetings, so he thought a research-led company like DeepMind was a unicorn. He now knows you can build a company by being heads down, doing great research and building something amazing, rather than tweeting, hyping and fundraising. Had he known, he says, he would have started sooner.

He closes by saying he hates the term data labeling, which suggests labeling cat photos. He thinks of what Surge does as more like raising a child: teaching values, creativity and what is beautiful, which for AI amounts to how humanity's children are raised.

In the lightning round he recommends Ted Chiang's story behind the film *Arrival*, Camus's *The Myth of Sisyphus*, and Douglas Hofstadter's *Le Ton beau de Marot*. His motto is that founders should build the company only they could build, since a company embodies its CEO's values. He also says he asks what he personally cares about rather than what looks good on a dashboard, and invites readers to send him interesting real-world AI failures via the Surge blog.