Edwin Chen on Building Surge and the Data Behind AI Models

Harry Stebbings talks with Edwin Chen, founder and CEO of Surge, about building an AI data company without external funding and the future of model training. The episode describes more than $1 billion in revenue, while Chen says the business has been profitable since its first month. The interview does not specify a comparable audited revenue period or profit measure; Chen has a direct commercial interest in the market he discusses.

Stebbings asks about Chen's claim that 90% of people at large technology companies were working on useless problems. Chen argues that a smaller company can move faster with fewer meetings, interviews and status updates, while giving everyone a clearer view of customers. His percentages and proposed productivity gains are illustrative judgments, not measured comparisons.

Asked how to decide what matters, Chen says small teams make customer problems and colleagues' work more visible. In his experience, large organizations can instead reward projects that impress managers or support internal growth. His example is an internal tool designed to save time consumed by hiring for growth's own sake.

Chen says managerial ambition can turn the maintenance and expansion of an organization into its own goal. When interviewing, he looks for candidates who ask concrete questions about the product and suggest improvements, rather than concentrating on how many people they will manage.

Chen says he has no standing one-on-one meetings and keeps his calendar relatively clear. His view is that colleagues who already communicate daily should raise problems as they occur instead of waiting for a scheduled meeting. He says Surge is ruthless about removing unnecessary meetings.

He believes a one-person billion-dollar company could eventually exist. His argument combines existing examples of small businesses with AI productivity improvements; that is a prediction, not evidence that such a business has already been built.

On the idea of a 100-times engineer, Chen imagines advantages in coding speed, ideas, effort and fewer meetings multiplying together. He thinks AI may disproportionately help strong engineers who already have more ideas than time to implement them. The multiplier is a conceptual estimate, not a controlled measurement.

Asked whether Surge's position reflects its own efficiency or rivals' mismanagement, Chen says both. He characterizes some competitors as labor suppliers without technology to measure or improve data quality. As the CEO of a competing data business, his description is a commercial claim, not an independent assessment of named firms.

Chen says Surge begins with data quality and builds technology to measure and improve it. Credentials alone do not establish that someone can do a task well, he argues, and systems must also detect cheating, account sharing and AI-generated work when it is not permitted. His examples are his assessment of the problem, not a quantified audit of workers.

Chen traces the company to his experience as a machine-learning engineer. At Twitter, he says, a simple sentiment-labeling project took months and produced poor results because annotators missed slang and hashtags; he ultimately labeled tweets himself. He also wanted data to train recommendation systems toward useful content rather than clicks that rewarded sensational material.

Surge began in 2020, shortly after GPT-3 launched. Chen says he built the first version himself in a couple of weeks, wrote about it on his blog and spoke directly with potential customers. His prior experience helped him understand researchers' needs, and he says demand appeared quickly.

Chen criticizes fundraising as a status goal when founders have not identified a problem they believe in. He advises most software founders to build and test an MVP first, while acknowledging that capital-intensive hardware or other projects can need investment before a viable prototype.

Stebbings asks whether a founder should pursue an idea uniquely suited to them or simply execute well. Chen favors the former for an ambitious foundational company: he wants a founder to have a distinctive conviction and be willing to take substantial risks. This is his philosophy, not a rule that guarantees an outcome.

For Chen, satisfaction comes from customers saying their model launches depended on Surge, and from analyzing data that reveals how to improve models. He describes those customer comments as his experience; the interview does not independently measure Surge's contribution to particular model launches.

Early customers contacted Chen by email or met him directly, with some contracts taking a week or two to negotiate. He says Surge developed product and recruited workers in parallel. The distinction, in his account, is that the company also built methods to identify good work and present data to researchers.

Chen wants machine-learning engineers to inspect their data directly. He argues that when examples include poetry, mathematical problems or research, seeing the material helps developers understand what they are teaching a model. Demand, he says, was apparent from the first year, though he does not disclose that year's revenue.

Stebbings asks why the company did not raise funding or hire a conventional sales team once demand was clear. Chen says it was profitable from month one and did not need the capital; he wanted customers who already understood and valued high-quality data. Early customers shape the product, he says, so alignment with its vision matters.

The host challenges whether listening closely to a few customers can produce a product that fails to serve a broader market. Chen responds that a strong product vision provides the filter: Surge prioritizes quality and says no when it cannot deliver it. His contrast with competitors is his own characterization.

Asked when he allowed quality to slip, Chen says he has not done so and tells employees that quality comes before deadlines. He says they can delay or decline work if necessary. This is his account of company practice, not independent evidence that every delivery met its quality standard.

On hiring, Chen argues that urgency can come from work that was not important in the first place. He says firms should not lower their standards simply to make headcount grow, and that a flat engineering organization need not mean the business is failing.

Stebbings asks whether the industry has swung toward celebrating the smallest team and revenue per employee. Chen says he does not follow the discussion closely enough to know. Although he once worked at Twitter, he values distance from prevailing Silicon Valley views and expects important research or products to reach him through colleagues and other channels.

Chen declines to give a first-year revenue figure, saying the company did well from the start. He identifies ChatGPT as a growth inflection point because it made the value of human data and reinforcement learning from human feedback more visible, while saying growth was already strong before that.

After Stebbings asks about the Scale transaction and customer movement, Chen says Surge saw substantial new interest, including from teams that previously used Scale. He claims researchers already knew Surge and praises its data quality; neither customer migration nor the competitive rankings he claims are independently verified in the interview.

Stebbings asks whether everything has a price. Chen says he would not sell for $30 billion or even $100 billion, citing profitability, control and the resources to pursue his goals. These are hypothetical figures and his stated preference, not a disclosed offer or valuation.

Chen says his motivation is to help achieve AGI and describes Surge as a contributor to frontier labs' work. He believes acquisition would constrain that mission and says he prefers to continue independently. The importance he assigns to Surge reflects his position as its founder and CEO.

If AGI has not arrived by 2040, Chen says the reasons could include the need for further algorithmic breakthroughs and the time needed to gather data. He uses medical experiments as an example and wonders whether simulations or other methods could accelerate them, while acknowledging that he does not know.

Chen says Surge's platform includes highly educated specialists, but argues a PhD alone does not establish the skills needed to create valuable training data. He values creativity, practical judgment and the ability to find difficult problems, not just credentials or performance on narrowly defined tests. His claims about the size and quality of its expert workforce are company statements.

He describes the challenge of identifying the best contributors across many tasks while filtering out spam and cheating. He says the platform can have hundreds of thousands or millions of people and thousands of projects, but gives no precise headcount or period. Fast-moving researchers, in his view, need new expertise and workflows in days or weeks.

Asked to rank bottlenecks, Chen puts data quality first, compute second and algorithms third. He argues that extra compute cannot fix poor training data or misleading evaluation objectives. As the leader of a data supplier, he benefits commercially from that emphasis; it is his ranking, not an established industry consensus.

Chen says some teams have trained and evaluated models for months before realizing bad data made apparent improvements misleading. He uses chatbot leaderboards as an example: in his account, voters may favor long, formatted or emoji-filled responses without checking accuracy. He cites a model incorrectly saying Pope Francis was alive, but the interview supplies no systematic benchmark analysis.

Stebbings asks about Grok's benchmark performance. Chen argues that narrow academic problems do not necessarily represent real-world capability. He also praises the intensity and mission focus of xAI's team, disclosing that Surge works closely with it; his impressions are informed by a customer relationship.

Stebbings asks whether synthetic data threatens human labeling. Chen says synthetic data is useful in some settings but that its value is overstated. He claims some customers found a thousand or a few thousand high-quality human examples more valuable than ten million synthetic examples; the tasks, comparisons and measurement methods are not supplied.

Chen argues that synthetic data can become too narrow and that models can repeat distinctive errors, citing a model that inserted unrelated Russian and Hindi characters. He sees human judgment as an external check. These examples and causal claims do not establish that synthetic data is broadly ineffective, and Surge sells human-data services.

Stebbings discloses that he is an ambassador for Poolside before asking about specialized versus general models. Chen sees room for both: large models may cover many uses, while smaller focused teams may move faster or make product and cultural choices that a general provider cannot. This is a forecast about competitive trade-offs.

Asked about his own weaknesses, Chen says he struggles to understand financial terminology and does not always know the company's revenue, costs or margins. He would prefer a measure of whether models are genuinely improving and how much Surge contributes through data, evaluations and research insights. For now, he uses the diversity and complexity of projects as an imperfect proxy.

Stebbings argues that building a company worth over $10 billion requires constant work. Chen agrees that founders must be willing to work hard and take urgent customer calls, describing requests at 2 or 3 a.m. and a hypothetical delivery of 10,000 data points within hours. But he distinguishes hours worked from value created and says good ideas can arrive on a walk.

Chen enjoys finding and writing down distinctive insights about models, algorithms and datasets for customers. In quick-fire questions, he argues that AI safety deserves attention: models optimized for the wrong objectives may produce misleading benchmark gains today and have more serious consequences as their capabilities increase. This is a risk argument, not a demonstrated forecast of a particular harm.

Asked to choose between 2028 and 2038 for AGI, Chen ties the answer to the definition: he suggests 2028 for automating an average engineer's job and 2038 for curing cancer. Both are speculative timelines, and curing cancer is not a single clearly defined benchmark. He says current models cannot yet perform half the meaningful code and ideas for a company focused on hard problems.

Chen says frontier labs should ask whether they are improving intelligence or merely optimizing benchmarks; product companies should ask why labs cannot replace them. Stebbings questions how far labs can extend into applications. Chen agrees they cannot pursue everything, but thinks a general conversational interface can absorb many simple tasks; he estimates it has replaced around half of his own former Google searches.

Asked what he would do at Google, Chen says the company faces the hard choice of accepting a short-term advertising revenue hit to build a better product. He now expects several frontier AI companies rather than one, with different strengths and boundaries; he cites his impressions of Claude, ChatGPT and Grok. These are judgments about products at the time of the interview.

Chen thinks major new model developers could still emerge, despite Stebbings's question about the immense capital required and existing investors' commitments. He argues that if truly ambitious AGI is still far away, there is room for unexpected breakthroughs, potentially involving AI and humans. His percentages for progress toward AGI are hypothetical, not measurements.

Asked whether AI could raise GDP or productivity by 10% within a decade, Chen says he believes it can; the interview provides no economic model to support that forecast. His advice to his earlier self is to focus on potential tenfold improvements rather than worrying about incremental 10% changes.