Snorkel AI Inc., a provider of artificial intelligence training data, today disclosed that it has raised $350 million in funding.
Insight and S32 led the Series E deal. They were joined by more than a half-dozen other backers including Alphabet Inc.’s GV startup fund.
Snorkel AI was founded in 2019 by researchers from the Stanford AI Lab. Its first product was Snorkel Flow, a software platform that reduced the amount of work involved in supervised learning projects. Supervised learning is an AI development method that trains neural networks using labeled datasets. Such files comprise prompts and correct, human-generated answers to those prompts.
Creating labeled datasets is a highly time-consuming process. Snorkel Flow automated the process using statistical methods developed by the company’s founders at Stanford. According to Snorkel AI, those methods addressed the accuracy issues that made earlier automation approaches ineffective.
Last year, the company changed its business model. It pivoted from selling software that helps developers create training data to providing ready-to-use training datasets. Additionally, Snorkel AI expanded its focus beyond supervised learning to a second, more complex AI training approach known as reinforcement learning.
Whereas supervised learning uses datasets that contain prompts and correct answers to those prompts, reinforcement learning datasets contain unanswered questions. The AI model being trained must figure how to answer them without any human assistance. Once it produces a response, human reviewers or an automated system verify its accuracy. They then provide the AI with feedback that helps improve its reasoning.
Snorkel AI relies on tens of thousands of human experts to generate reinforcement learning training tasks. The company also provides customers with other technical assets that are needed for AI training runs.
When an AI model completes a reinforcement learning task, human reviewers check its work based on predefined evaluation criteria. Those criteria can span several pages. In the case of a programming task, for example, the evaluation guidance must cover all the cybersecurity and performance requirements that AI-generated code must meet.
Snorkel AI develops AI evaluation rubrics for customers. Furthermore, it improves those rubrics over time based on feedback from the human AI reviewers who use them. Some of that feedback is generated when two human reviewers give different scores to an AI prompt response. Such differences usually point to an inconsistency in the underlying evaluation criteria.
AI models are often trained in specialized virtual environments. A code generation model, for example, might require a simulated version of a developer workstation. Snorkel AI provides such as training sandboxes alongside its datasets and evaluation rubrics.
“Since launching our new data-as-a-service offering nearly a year ago, we’ve grown over 18 times, and this week crossed an annualized revenue run rate of $375 million,” co-founder and Chief Executive Officer Alex Ratner detailed in a blog post.
The company will reportedly use its newly raised funding to hire more engineers. Snorkel AI also plans to invest in AI safety initiatives and support the development of open-source model evaluation benchmarks.
