
1M+ Data Points: How Snorkel Builds the Data That Frontier AI Labs Train On
We sat down with Kartik Mathur, Director of Engineering at Snorkel AI, to talk about the data behind frontier AI. Snorkel is a research and data lab that delivers high quality datasets, evaluation environments, and benchmarks to frontier labs. Kartik leads the AI infrastructure team and is building the agent factory behind Snorkel’s data platform: a fleet of hundreds of agents that work alongside human experts to produce hundreds of thousands to millions of data points.
Kartik explained why scraping the web is no longer good enough, why code generation is only a small part of what an engineer solves, and why frontier labs partner with specialists instead of doing everything themselves. He also demoed the agent factory, including shared agents that learn from every reviewer and answer questions like why token costs are rising.
We also covered Google’s new agent that can make phone calls and purchases, and Jordan’s story of his bot calling Sam and leaving a voicemail.
We cover:
- What Snorkel does and how it differs from other data providers
- Which datasets frontier labs want most in 2026: coding, long horizon tasks, multimodal, finance, legal, and physical AI
- How agents are replacing handholding in the data creation process
- Why web scraping is no longer enough to push models forward
- Why solving the whole problem matters more than code generation
- Why frontier labs work with trusted data partners
- A demo of Snorkel’s agent factory, built for control, governance, and auditability without vendor lock-in
- Shared agents that learn from every user, and agents inside Claude Code, Claude, and Slack
- Google’s new agent that makes calls, and why Kartik wants to see how it handles failure modes
- Jordan’s bot calling Sam and agents joining video calls
If you are building with AI agents, working on AI infrastructure, or curious what frontier labs train on, this episode is for you.
Find Kartik and Snorkel AI:
Website: snorkel.ai
LinkedIn: https://www.linkedin.com/in/mathurkartik/
New episodes every Friday at BuiltThisWeek.com
TIMESTAMPS
[00:00] Intro
[01:14] Meet Kartik: AMD, Blue Data, Domino, and Snorkel AI
[02:17] What Snorkel does and how it differs from other data providers
[03:02] Which datasets frontier labs want in 2026
[03:55] From human feedback to agents building the data
[05:19] What all this data is used for
[05:49] How long a dataset takes
[07:09] Where Snorkel specializes
[08:12] Why code generation is only a small problem
[09:27] Why frontier labs do not do this themselves
[10:27] Demo: Snorkel’s agent factory
[11:22] Why they built it: control, governance, and auditability
[12:05] Shared agents that learn, and tracking token costs
[13:29] Rolling agents out in Claude Code, Claude, and Slack
[14:02] Where to find Kartik
[14:25] News: Google’s new agent that makes phone calls
[16:18] Why Google has a superpower with its stack
[16:47] Jordan’s bot called Sam
[17:16] Agents joining video calls
[17:46] Wrap up
#AI #Snorkel #AIAgents #AIInfrastructure #FrontierAI #Google #BuildWithAI #BuiltThisWeek
Otros episodios de "Built This Week"



No te pierdas ningún episodio de “Built This Week”. Síguelo en la aplicación gratuita de GetPodcast.








