What iPhone Packaging Reveals About the Robotics Opportunity
Josh Gruenstein argues robotics is the ultimate systems problem: winning requires owning the full stack, deploying robots into underserved greenfield customers, and using that real-world data flywheel to build general robot intelligence.
Watch on YouTube ↗Charlie's summary
- ▶ 3:01
Robotics is the ultimate systems company: skipping any layer of the stack eliminates 90% of your trade-off space
Josh thinks robotics is not just model, hardware and software but also supply chain, customers and product. Even with infinite robot intelligence, robots are still a cottage industry with no software and no customers. If you refuse to touch a layer, you “eliminated 90% of your trade-off space.” He says this is why robot startups start horizontal and end up going vertical.
- ▶ 9:06
Tutor builds an internet of robot training data by pairing deployed robots with human teleoperators
Josh notes the internet created LLM training data as a byproduct, but physical work leaves no record. Tutor deploys robots with human teleoperators, called tutors, who teach them new tasks, creating data while delivering economic value. Whether tutors can be removed is a unit-economics question, like the cost to serve a model on an H200. Josh says Tutor isn't a purist and uses simulation and internet-scale data too.
- ▶ 13:11
Data from paying customers is guaranteed to be economically valuable, unlike data collected speculatively
Josh contrasts robots working at real customer sites with strapping cameras on people's heads and sending them around Indonesia. With the latter, “there's no invisible hand of the market” ensuring the data is useful. If customers pay for a robot to do a task, you know the data is valuable. He thinks the long-lasting value is the flywheel, not any one architectural trick, and Tutor has survived every architecture shift since 2021.
- ▶ 15:12
Frontier labs training on 3D spatial data has catapulted robotics years ahead
Josh says general knowledge was baked into big models but lacked the spatial reasoning to connect to robot action heads. Labs now post-train on CAD and 3D modeling data because there's a real market for it, and that is the spatial reasoning Tutor's models need. Because robotics is a systems problem, he sees this as accelerating Tutor rather than threatening it.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.