Josh Gao pitches a data factory that cannot solve the force gap
By Nikhil Raghavan · Reporting from San Francisco ·
Josh Gao pitched Mecka AI's pipeline for sensorizing humans, but until foundation models learn physical force instead of relying on video, general-purpose domestic humanoids will remain sellable rather than shippable.
Josh Gao spoke with host Sonia on the "Sequoia Capital" podcast. He said Mecka AI provisions upload lines across operating sites to move physical human movement data. The reason is blunt and expensive. Large language models grew fat on free tokens scraped off the open web. Robotics has no public commons to plunder. Training a humanoid machine to fold laundry or twist a bottle cap demands custom hardware. Engineers must build custom baseball caps packed with cameras. They must manufacture tactile gloves to measure grip force in Newtons. Mecka must pay gig workers and bakeries to perform daily chores. Finally, the company must verify footage against fraud before feeding it into foundation models.
The pipe dream of cheap physical tokens
On the podcast, Gao outlined the training recipe the robotics industry is converging on. It pairs roughly one million hours of egocentric human video with robot interaction and teleoperation demonstrations. Two years ago, academic benchmarks like Ego 4D hovered at just 4,000 hours. Scaling past that baseline sounds like an inflection point. Yet that claim wilts once you inspect the engineering constraints.
Anyone who manages ingestion queues knows where clean slide decks hit dirty real-world friction. Moving high-resolution sensor streams over residential and commercial uplinks creates severe bottlenecks. When a pipeline ingests video, contact data, and action trajectories from contributors across homes and factories, quality collapses. Gao admitted Mecka caught contributors passing off YouTube videos of McDonald's kitchen shifts as live demonstrations. Sanitizing unstructured physical telemetry with automated heuristics is not a solved software problem. It is a high-latency operational grind. In the end, verification costs threaten to swallow collection margins.
Why driving cars did not solve folding shirts
The strongest case for Mecka's factory model rests on the bitter lesson of deep learning. Data volume and raw compute eventually steamroll hand-tuned heuristics. Autonomous vehicles proved that feeding driving video into neural nets could produce workable perception systems. If scaling laws hold for physical manipulation, bankrolling human demonstration data looks rational.
Yet that analogy breaks on mechanical reality. An automobile operates with two or three degrees of freedom: steering, throttle, and braking. A human hand possesses dozens. In autonomous driving, a camera estimates velocity and lane boundaries. It rarely calculates the tactile resistance of an eggshell or the friction of a damp towel. Video alone cannot capture Newtons of applied force. Gao acknowledged that off-the-shelf gloves cannot reliably record force. Mecka must iterate on bespoke hardware hubs instead. But hardware loops run on physical lead times, not software sprints. Operating inside unconstrained living rooms demands physical observation that capital markets cannot subsidize forever.
Who gets paged when the robot slips
This engineering divide will soon confront regulators and local councils sold on general-purpose domestic humanoids. Gao floated a sci-fi future where robots outnumber humans. Yet he admitted immediate deployments belong to industrial scrubbers, data centers, and hotel delivery carts. These machines rely on rigid operating procedures. The policy stakes do not lie in whether silicon machines will one day mine asteroids. They lie in product liability and workplace safety right now. Who gets paged at three in the morning when an uncalibrated force sensor crushes a worker's hand? If a foundation model relies on outsourced gig labor without standardized benchmarks, liability vanishes into dataset quality gates.
Building a data factory for physical AI is sellable to venture funds hunting the next frontier. It is not shippable as a domestic reality. The industry must first prove it can reliably measure and model physical force at scale. It must avoid choking on data fraud and bespoke hardware bottlenecks. Until then, general robotics will remain tethered to tightly bounded industrial cages.