Snorkel AI raises $350 million at $3.5B valuation

Snorkel AI started in 2019 as a Stanford AI Lab spinout whose first product, Snorkel Flow, automated the costly work of labeling datasets for supervised learning. Labeling is slow because it requires prompts paired with correct human-written answers, and the company says its founders' statistical methods fixed the accuracy problems that undermined earlier automation.

Last year the company changed its business model, moving from selling tools that help developers build training data to selling ready-to-use datasets. It also moved into reinforcement learning, a harder approach where models face unanswered questions and learn from feedback given by human reviewers or automated systems. Snorkel relies on tens of thousands of human experts to create those tasks, and it supplies customers with evaluation rubrics and training sandboxes alongside its datasets.

With the new funding, Snorkel says it will hire more engineers and put money into AI safety and open-source model evaluation benchmarks. How far that goes may depend on whether customers keep buying packaged data and rubrics, rather than building them in-house, as reinforcement learning becomes a bigger part of AI training.