Issue #055 — Weekend Notes
# The Internet of Picking Things Up
*Weekend notes from Ken Zhang — lighter fare, single topic, open questions welcome.*
---
1/ Mecka AI raised $60 million from Sequoia this week to do something almost embarrassingly simple: pay ordinary people to record themselves doing everyday tasks, then sell that motion data to humanoid robot companies (per TechCrunch).
I've been chewing on this one all week, because it's the clearest signal yet that the humanoid race has entered its boring-data phase. And I mean that as a compliment.
2/ Think about what the funding implies. A few years ago, the pitch for a robot data company would have been "our sim pipeline is 10x faster." Today the pitch is essentially "we will pay your neighbor to fold laundry on camera." Sequoia writing a $60M check (per TechCrunch) for that business model tells you where the bottleneck actually is. It's not actuators. It's not even foundation models anymore, really. It's that nobody has enough footage of humans picking things up in kitchens that look like *your* kitchen. 3/ This tracks with what I keep seeing in our own catalog work. Maze tracks 600+ robot models across roughly 350 companies globally, and the hard question we get from investors is never "who makes humanoids?" — that list is embarrassingly long. The question is "whose data flywheel actually spins?" Data collection as a paid service is one answer. Whether it's *the* answer is much less clear. 4/ Here's my mild heresy for the weekend, clearly labeled as opinion: I suspect "pay humans to perform tasks" is a bridge business, not a destination. It works until teleoperation farms and simulation transfers get good enough to undercut it. That said, our estimate is the bridge lasts longer than the skeptics think, because real-world messiness is stubborn. Simulated laundry does not fray. Simulated socks do not hide under the couch out of spite. 5/ A related thread from the same week: Danu Robotics, where founder Amy Ma has spent six years — six! — teaching robots to sort recyclables (per TechCrunch). Different domain, same insight. The hard part of waste sorting isn't the arm; it's that every piece of trash is a tiny adversarial attack on your perception system. A crushed milk jug is basically a CAPTCHA. 6/ Both stories point the same direction: the winners in embodied AI may be the ones who are "not a robotics company," as Destro AI's founder put it in TechCrunch last month. Robots are becoming platforms; the money is in the boring middle layer — data, perception, task definitions — that makes the shiny hardware useful. 7/ So the open question I'm sitting with: if human motion data becomes a commodity market, who captures the value? The collectors (Mecka), the robot makers buying it, or whoever builds the ratings agency of motion data — the Moody's of "is this grasp trajectory actually good?" That last one doesn't exist yet. Somebody should build it.
I'm collecting thoughts on data-as-the-moat for a longer piece next week. If you're an investor with a view (or a grudge), reply — I quote generously and attribute even more generously.
Until then: enjoy the weekend. Maybe fold some laundry. It's worth $60M, apparently.
*Kenzhegali "Ken" Zhang — Maze Intelligence. Robot catalog: 600+ models, ~350 companies. Numbers above cited where sourced; the rest are labeled what they are: our view.*