Snorkel AI raises $350M as demand for precision AI training data replaces high-volume data factories

Snorkel AI raises $350M on the thesis that hard data is getting harder

Snorkel AI has closed a $350 million round, a milestone Alex Ratner frames not as a bet on data volume but on the opposite: the argument that as frontier models improve, the marginal value of any additional training data shifts almost entirely toward precision.

Ratner, who co-founded Snorkel a decade ago out of Stanford and UW, describes the shift as moving from "data 1.0 to data 2.0." Early in a model's development, almost any data helps — so volume wins, and the rational approach is crowdsourcing and staffing. Once models are genuinely capable, that stops being true. The data that matters is the data that targets specific gaps, misalignments, or failure modes. Finding and building it becomes a research and engineering problem, not a headcount problem.

“When you're early on in the learning curve, the model doesn't know much, so any bit of information is additive and your focus is on volume. Then you naturally get to the asymptote phase where it's not about any data — it's about the right data that targets some area where they're misaligned... You can't get this hard data with just staffing or sourcing human hours.”

Synthetic data and recursive self-improvement

The obvious challenge to Snorkel's model is the recursive self-improvement thesis: if models can generate their own training data, does the human-in-the-loop data business eventually shrink to zero? Ratner is skeptical of the more expansive version of that argument, noting he has worked on synthetic data for nearly a decade and has heard "we're one model advance away from solving everything" more than once.

His actual position is more nuanced. RSI is real, but it runs into bottlenecks — compute, energy, human expertise, and real-world grounding — that pure model self-modification can't resolve. Even within domains like math and coding where automation is furthest along, the economic logic runs in Snorkel's favor: as model accuracy rises, the demand for easy data falls, but the value of each additional decimal point of improvement goes up sharply, and so does the difficulty of producing the data needed to get there.

The result, Ratner argues, is that you can't get to the frontier with staffing alone, and you can't get there with a model generating its own training data in a closed loop — what he calls "synthetic slop." You need a sophisticated blend of human expertise and specialized AI working together. Snorkel's platform is built around that compounding loop, where human feedback improves the specialized AI models used to accelerate data production, which in turn raises the ceiling on what human experts can produce.

The commercial bet is straightforward: if Snorkel can keep pace with the frontier's rising difficulty, its data becomes more valuable than ever. If it can't, it becomes irrelevant. The $350M suggests investors think the former is more likely.