Sequoia Capital

Josh Gao on the slow creep to general robots

Josh Gao· Founder and CEO at Mecka AI
·~45 min·English·Sequoia Capital
RoboticsTrainingMultimodalAI Infrastructure
TL;DR

Mecka AI's Josh Gao explains why robots have no internet-scale head start, how his company manufactures the pixels, actions, and force data to close the gap, and why deployment revenue — not one breakthrough — funds the slow climb to general-purpose robots.

01The Data Problem

The Cold Start Robots Never Got

<strong>Robots inherit no internet-scale corpus</strong>, so every hour of training data — the pixels, actions, and logic — has to be manufactured from scratch.

Fundamentally, a lot of this robotics data is three things. It's pixels, its actions, and its logic.

— Josh Gao, Sequoia Capital
Key Insight
The real moat is not the algorithm but whoever can produce clean pixels, actions, and logic at scale. Robotics inverts LLM economics: data is a capital expense you must fund, not a corpus you can scrape.
02The Founding Bet

Betting the Company on Egocentric Data

<strong>A person wearing a first-person camera is a data source you can scale</strong>, so Mecka turned willing humans into a training fleet the way Tesla turned cars into one.

Humans and humanoids look the same. Why couldn't we turn humans into this fleet style learning that Tesla has to be able to all contribute into understanding and ground the physics of the world.

— Josh Gao, Sequoia Capital
Key Insight
Betting on egocentric data is a bet that human demonstrations can be gathered cheaply at scale. If a person's viewpoint transfers, every contributor becomes a data source the labs never had to build a robot to capture.
03Hardware

Sensorizing Humans to Capture Force

<strong>Video never records how hard you squeeze</strong>, so Mecka builds tactile gloves that log force in Newtons — a data type the internet does not contain.

So we've been developing things like gloves for humans to wear and in those gloves being able to collect the information in how much force it takes to do your daily life and different tasks.

— Josh Gao, Sequoia Capital
Key Insight
Force is the variable video silently discards. By manufacturing a data type that exists nowhere online, Mecka is building exactly the kind of moat that commodity web data can never erode.
04Operations

Inside the Data Factory

<strong>The hard part is trust, not collection</strong>: novel data invites creative fraud, so adaptable quality gates become the real product.

I think at one point we have folks uh recording YouTube videos like POV working at McDonald's and trying to submit that as as contribution

— Josh Gao, Sequoia Capital
Key Insight
Because this data is created rather than filtered from an existing internet corpus, catching fraud and adapting the quality gates — not the cameras — is where the engineering effort and the defensibility actually live.
05Scale Reality

Still Orders of Magnitude Short

<strong>Millions of hours sounds like arrival but is only the first rung</strong> — Tesla needed that much for two or three degrees of freedom, and a robot hand has dozens.

I think we finally entered a world where we are in the millions of hours of data and coming from the world where Ego 4D of 4,000 hours 2 years ago where we were then uh is a huge step up.

— Josh Gao, Sequoia Capital
Key Insight
The gap is combinatorial, not linear. Each added degree of freedom multiplies the space to cover, so moving from a car's handful of controls to a hand's dozens pushes the target orders of magnitude further out.
06Thesis

A Slow Creep, Not a Switch

<strong>Robotics will not flip on in one day</strong>; it creeps up task by task, each commercial deployment paying for the data that unlocks the next.

I think it's about unlocking commercial use cases as we go. So it doesn't have to be so much binary in that oh one day we've unlocked robotics or not it is going to be a slow creep up into the world.

— Josh Gao, Sequoia Capital
Key Insight
Framing progress as a creep is also a go-to-market strategy. Each task that earns revenue produces the correction data that unlocks the next, so capability and cash compound instead of waiting on one breakthrough.
07State of the Field

The Data Pot of Soup

<strong>The field stopped fighting over one data religion</strong> and converged on a recipe: a base of human video, a layer of teleoperation, and a thin top of in-distribution demos.

all of this is essentially a data pot of soup. And in that pot of soup, the all the ingredients are going to be important at some point.

— Josh Gao, Sequoia Capital
Key Insight
Convergence on a shared recipe signals a maturing field, but it also commoditizes the recipe — which shifts the advantage toward the ingredients no one else can yet supply, like tactile and simulated data.
08How Robots Learn

AlphaGo and the Missing World Simulator

<strong>Robotics is stuck between imitation and self-play</strong> because, unlike Go, the physical world has no closed ruleset and no simulator detailed enough to practice in.

Then the next step was the self-play version of with the rules of the game, can you ultimately simulate all the different ways until you come up with the magic of move 37 ultimately.

— Josh Gao, Sequoia Capital
Key Insight
The bottleneck is not model scale but a world simulator. Go's rules let self-play run forever; the physical world offers no such closed loop, so robots cannot yet cheaply dream the way AlphaGo could.