Edwin Chen founded Surge AI after struggling to obtain trustworthy human-labeled data for machine-learning work at large technology companies. He says Surge grew without venture funding and surpassed $1 billion in annual revenue; that figure is his reported company result, not an independently audited fact. — No Priors — Edwin Chen Interview.

Infographic for "Lessons from Edwin Chen of Surge AI".

On Data Quality: The Bedrock of AI

Chen says the difficulty of obtaining reliable training and evaluation data at Google, Facebook and Twitter led him to build Surge around high-quality human data. — No Priors — Edwin Chen Interview.

  1. Get the data the model actually needs: Chen says that at several large technology companies, acquiring suitable labeled data was a recurring bottleneck for training even relatively simple models. — No Priors — Edwin Chen Interview.
  2. Make human-data quality a founding principle: Chen says Surge was built on the belief that high-quality human data could advance AI, with quality emphasized from the start. — No Priors — Edwin Chen Interview.
  3. Poor labels teach poor behavior: Chen argues that models trained on inaccurate labels can mimic those errors and give inaccurate predictions. — 5 Q’s for Edwin Chen.
  4. Data quality can be the bottleneck: Chen describes trustworthy human-labeled data as a constraint on model iteration: slow or low-quality labels delayed the work his teams wanted to do. — Surge AI Founding Essay.
  5. Curated human data can outperform volume in particular cases: Chen reports customers for whom a small, carefully curated human-labeled set was more useful than millions of synthetic examples; he also says synthetic data remains useful for some tasks. — No Priors — Edwin Chen Interview.
  6. Spreadsheets are a weak labeling interface: Chen says spreadsheet-based labeling made performance difficult to monitor and motivate, prompting him to build richer human-evaluation tools. — Surge AI Founding Essay.
  7. Bad labels can corrupt evaluation too: In Surge’s review of Google’s GoEmotions data, Chen says mislabels threatened both model training and the meaning of its evaluation metrics. — 5 Q’s for Edwin Chen.
  8. Creative work needs more than majority vote: Chen argues that judging creative outputs requires taste and task-specific quality criteria; merely aggregating simple checklist labels can miss what makes an answer good. — No Priors — Edwin Chen Interview.
  9. Use human creativity in training data: Chen argues that rich human-created examples can teach models more than literal instruction-following patterns, including qualities of writing and problem solving. — No Priors — Edwin Chen Interview.
  10. Data quality matters in deployed systems: Chen points to content moderation, customer support and search as examples where low-quality labels can lead to inaccurate AI predictions. — 5 Q’s for Edwin Chen.
  11. Relabeling sometimes improves an existing model: In a 2021 founding essay, Chen reported a customer case where relabeling an existing dataset improved model performance by 50%. This is a company-reported example, not a general effect size. — Surge AI Founding Essay.
  12. Preserve context for emotion labels: Chen criticizes presenting Reddit comments for emotion labeling without subreddit or author context; words and slang can change meaning with their setting. — Context-Sensitivity in Data-Centric AI.
  13. Use skilled labelers for nuanced tasks: Chen says complex problems such as hate speech and misinformation require motivated human workers with enough skill to judge and measure subtle outcomes. — Surge AI Founding Essay.
  14. Treat quality control as adversarial: Chen says labelers can make mistakes or deliberately spam, so a serious labeling system needs technology to detect and correct poor work. — 5 Q’s for Edwin Chen.

On AI Development and the Future

Chen expects humans to keep supplying feedback as models improve, while model providers develop different strengths and richer training environments. — No Priors — Edwin Chen Interview.

  1. Keep human feedback in the loop: Chen expects human feedback to remain useful even as models become more capable, including for aligning behavior with objectives humans actually value. — No Priors — Edwin Chen Interview.
  2. Use synthetic data selectively: Chen says synthetic data can supplement human work, but heavy reliance on it can produce narrow, benchmark-oriented training sets that miss real-world diversity. — No Priors — Edwin Chen Interview.
  3. Check models against an external human signal: Chen warns that models can optimize toward the wrong objective; human evaluation can test whether outputs meet the intended task rather than a proxy score. — No Priors — Edwin Chen Interview.
  4. Expect different model-provider strengths: Chen says model developers make different product and training choices, giving their models distinct strengths rather than a single uniform capability profile. — No Priors — Edwin Chen Interview.
  5. Choose models by task: Chen says he already switches among models according to the job and expects such task-specific selection to continue. — No Priors — Edwin Chen Interview.
  6. Build coherent RL environments: Chen says useful agent-training environments require consistent tools, events and long-horizon trajectories; simply generating isolated synthetic tasks is not enough. — No Priors — Edwin Chen Interview.
  7. Treat wrong objectives as a present risk: Chen links contemporary benchmark gaming to the broader AI-safety problem of systems optimizing proxies that do not capture what people actually want. — 20VC — Edwin Chen Interview.
  8. Contribute to AGI through data: Chen describes helping build more capable AI as a central motivation for Surge, and says high-quality data is a critical part of that effort. — 20VC — Edwin Chen Interview.
  9. Do not confuse leaderboard rank with model quality: Chen argues that a response can climb a preference leaderboard through length or presentation while becoming less factual or useful on real tasks. — 20VC — Edwin Chen Interview.
  10. Leave room for specialized products: Chen sees a role for smaller or specialized model products when a broad provider’s culture, priorities or constraints keep it from pursuing a particular use case. — 20VC — Edwin Chen Interview.

On Entrepreneurship and Building Surge AI

Chen says he self-built Surge’s first product, stayed profitable without outside funding, and prioritized customers who valued the data quality it delivered. — 20VC — Edwin Chen Interview.

  1. Bootstrap when the business does not need funding: Chen says Surge was profitable from the start, so raising outside capital would have meant surrendering some control without solving an immediate financing problem. — No Priors — Edwin Chen Interview.
  2. Ask what outside capital would solve: Chen criticizes raising as a status ritual when a founder can already build and test a product; he allows that capital is useful when a project genuinely needs it. — No Priors — Edwin Chen Interview.
  3. Build the first useful version: Chen says he built Surge’s initial product himself in a few weeks because he knew the labeling problem and wanted to test it with customers before hiring a large team. — 20VC — Edwin Chen Interview.
  4. Test demand before raising when feasible: Chen advises founders who can build without large upfront capital to launch an MVP and see whether customers care; he explicitly excepts capital-intensive projects. — 20VC — Edwin Chen Interview.
  5. Keep teams close to the customer: Chen argues that a smaller, high-talent team can move faster because it spends less time on hiring, meetings and internal reporting and stays closer to customer problems. — 20VC — Edwin Chen Interview.
  6. Choose early customers who share the product vision: Chen says early customers shape a product through feedback, so Surge preferred teams that understood its high-quality-data approach. — 20VC — Edwin Chen Interview.
  7. Let product value lead sales: Chen says he did not want a broad outbound sales team; he wanted customers to buy because they recognized the quality of the data. — 20VC — Edwin Chen Interview.
  8. Question standing meetings: Chen says he avoids recurring one-on-ones and expects frequent day-to-day communication to surface problems without waiting for a scheduled meeting. — 20VC — Edwin Chen Interview.
  9. Prioritize customers over internal promotion: Chen criticizes large-company projects undertaken to impress a manager rather than improve the end customer’s product or experience. — 20VC — Edwin Chen Interview.
  10. Value autonomy over an exit price: Chen said he was not interested in selling Surge for a large offer because profitability and control let him pursue the company’s mission. — 20VC — Edwin Chen Interview.
  11. Take a distinctive risk: Chen argues that an ambitious startup should pursue an idea its founder deeply believes in, rather than pivoting among ideas for fundraising attention. — 20VC — Edwin Chen Interview.
  12. Invest in richer agent-training tasks: Chen describes Surge’s work on complex RL environments with coherent tools, timelines and varied situations so agents can learn beyond isolated synthetic examples. — No Priors — Edwin Chen Interview.
  13. Build tools that measure data quality: Chen distinguishes Surge’s technology for assessing and improving label quality from simply passing workers to labs without measuring the data they produce. — 20VC — Edwin Chen Interview.

On the Future of Work and the "100x Engineer"

Chen argues that AI can remove routine work for highly productive builders and help them implement ideas faster; “100x engineer” is his illustrative view, not a measured productivity law. — 20VC — Edwin Chen Interview.

  1. The 100x engineer is an illustration, not a metric: Chen argues that differences in coding speed, ideas, effort and available AI tools can compound; his “100x” framing expresses that view rather than a measured industry average. — 20VC — Edwin Chen Interview.
  2. Use AI to remove routine coding work: Chen says current AI tools can take over some day-to-day drudgery, freeing builders to spend more time on ideas and implementation. — 20VC — Edwin Chen Interview.
  3. Help capable people implement ideas: Chen argues that many strong builders have more ideas than they can execute; AI can make it easier to turn some of those ideas into working output. — 20VC — Edwin Chen Interview.
  4. A solo billion-dollar company is a possibility, not a fact: Chen said AI-enabled efficiency could eventually make a billion-dollar-revenue one-person company possible; he did not claim one already existed. — 20VC — Edwin Chen Interview.
  5. Match expert talent to the actual task: Chen argues that high-quality data work requires finding people with the right domain skill and creative judgment for each project, not only a plausible credential. — 20VC — Edwin Chen Interview.
  6. A credential does not guarantee useful work: Chen says a doctorate alone does not establish that someone can create the best examples or evaluations for a particular AI-training task. — 20VC — Edwin Chen Interview.
  7. Combine human judgment with software: Chen describes Surge as a human/AI operation in which people provide judgments and algorithms help detect errors and improve labeling workflows. — 5 Q’s for Edwin Chen.

On a Personal Philosophy

Chen says the work he finds rewarding is analyzing data, helping models improve and hearing when customers consider Surge’s contribution important. — 20VC — Edwin Chen Interview.

  1. Follow the data questions that interest you: Chen says Surge reflects his long-standing interest in analyzing data and using it to improve models and products. — 20VC — Edwin Chen Interview.
  2. Keep distance from status-driven discourse: Chen says he is glad not to track every Silicon Valley conversation online, preferring important ideas to reach him through colleagues and direct work. — 20VC — Edwin Chen Interview.
  3. Take pleasure in useful insights: Chen says he enjoys seeing Surge’s data reveal patterns that help customers improve models in ways they otherwise might miss. — 20VC — Edwin Chen Interview.

Learn more:

  1. No Priors Ep. 124 | With SurgeAI Founder and CEO Edwin Chen - YouTube
  2. 5 Q's for Edwin Chen, CEO of Surge AI - Center for Data Innovation
  3. Surge AI: A Modern Data Labeling Platform for NLP | by Edwin Chen | Medium
  4. 5 Examples of the Importance of Context-Sensitivity in Data-Centric AI
  5. 30% of Google's Emotions Dataset is Mislabeled - Surge AI
  6. Surge CEO & Co-Founder, Edwin Chen: Scaling to $1BN+ in Revenue with NO Funding