Mark

Growth
Building Real-World AI Infra
Ex-NEAR | Ex-VC

[Mark @markkmii]

@TheHumanoidHub Most countries are debating robot policy.

China shipped registration infrastructure: 28,000 IDs across 100+ companies before mass deployment even starts.

Identity and provenance rails are boring. Theyre also what real-world deployment runs on.

[Mark @markkmii]

@LeRobotHF @rabault_nicolas The command line was quietly a filter on who could collect robot data.

Remove it and every hobbyist with a robot arm becomes a data contributor.

More hands collecting beats better tools for experts.

[Mark @markkmii]

@gi_labs Egocentric human video is the only manipulation data source that scales with population instead of robot fleets.

Recording it was never the hard part. Making it metric and robot-usable is.

93% hand coverage and 0.33cm drift is real progress on that.

[Mark @markkmii]

@APompliano @Rewkang Agree. And the most underrated timestamp is 31:05.

Where training data comes from decides who wins everything else on this list.

[Mark @markkmii]

@gravicle The key line: scaling teleoperation to every task is physically and economically impossible.

Thats the constraint everyone in physical AI is designing around.

Human video, world models, on-robot learning. The race is over what replaces teleop.

[Mark @markkmii]

@NVIDIAAI Fully open changes the calculus. When the model layer is free, it stops being the moat.

The differentiator moves to what open models cant ship: your real-world data and the loop that keeps it coming.

[Mark @markkmii]

@sharifshameem An arcade where every match is a training episode.

Customers pay to generate adversarial human data.

Best business model in robotics.

[Mark @markkmii]

@wallstengine The $2B isnt pricing the model.

Its pricing 500,000 hours of real-world manipulation data and the gripper network that keeps collecting it.

The data engine is the company.

[Mark @markkmii]

@StockMKTNewz Vision had the internet. Touch has nothing.

No web-scale dataset of force feedback exists, it has to be collected contact by contact.

Deploying this across warehouses isnt just automation. Its the largest tactile dataset ever built.

[Mark @markkmii]

@BerntBornich The loop at the end is the real bet. Robot collects data, model improves, repeat.

Curious about the mix: how much does on-policy NEO data move the needle today vs egocentric human video? Or is human video the bridge until the fleet is big enough?

[Mark @markkmii]

The headline is world models.

The real bet is the last line: robot collects data, model gets better, robot gets better.

Pretraining gets you to deployment. The loop after deployment is what compounds.

[Bernt Bornich @BerntBornich]

We’re going all in on World Models.

Today we’re launching the 1X World Model Lab.

The bet is simple:

You can’t fine-tune your way to AGI.

And you definitely can’t fine-tune your way to robots that can operate in the physical world.

General-purpose humanoids need models that understand space, motion, objects, causality, affordances, physics, and action before they ever see a specific task.

The frontier is not better VLA wrappers.

The frontier is embodied world models.

The 1X World Model Lab will focus on large-scale embodied world model pretraining: building the most generalizable foundation model for humanoid robots from the ground up.

The next frontier in AI requires scaling:

web-scale media + egocentric human videos + sim + dexterous remote operated robot data + on-policy NEO data → real-world deployment for robot data collection and RL → abundance of data → physical AI

The robot collects data.
The model gets better.
The robot gets better.
Repeat.

To lead this, we brought in one of the best for the mission: @_sam_sinha_ , as Head of World Models.

Sam was a founding research scientist at Luma AI and has been at the frontier of scaling multimodal generative video models his whole career.

If you’re the best in the world at large-scale pretraining, video models, robotics, RL, infra, or data — and you want your models to move atoms, not just pixels — join us.

Send background + evidence of exceptional ability to:

[email protected]

We’re building the model that makes autonomous labor real.

[Mark @markkmii]

Clear use case here.

Real city setting, constrained task. Fixed dock, known location, predictable approach.

Its a deployable pattern, narrow and controlled even in the real world.

What happens when someone grabs the bike before it docks?

[Fabrizio Bustamante Escudero @Fabriziobustama]

A robot just parked and docked a Citi Bike by itself in the middle of NYC. 🤖🚲

We’re entering the era where AI won’t just think… it will interact with the real world beside us. 👀

#AI #Robotics #IoT #5G #Tech #innovation #SmartCities

[Mark @markkmii]

@upadhyay_harsh1 Multiple companies launching for physical AI data in the same week.

When you see parallel attacks on the same problem, the gap is real and large.

Action data for embodied AI is still scarce, task-specific, and expensive. Thats changing fast.

[Mark @markkmii]

@lukas_m_ziegler No two fields are the same, no amount of hand-coded logic was going to get there.

Foundation models fix the code problem. Real-world outdoor data at scale is still the harder challenge

[Mark @markkmii]

Gig economy scale makes sense but tradeoff is data quality.

Robot learning needs precise, correctable demonstrations, not just volume.

India gets you scale and distribution. Maintaining training fidelity at that scale is the unsolved part.

How this relates to ethics and data ownership is another discussion that will come up more often in the future.

[Mark @markkmii]

@TheHumanoidHub Soccer is harder than factory work because the environment doesnt stay still.

Reacting to an unpredictable opponent is the physical AI problem in miniature.

What generalizes from this beyond soccer is the real question.

Hope we get to see the Atlas' ball control abilities!

[Mark @markkmii]

Vertically integrated is the key phrase.

A few hours of onsite data adapted the robot.

This only works if the whole stack is yours. You cant outsource a layer and still close that loop.

How similar has the tasks been between different customers? Curious on training time required on new edge cases.

[Mark @markkmii]

@humanoidsdaily @gs_ai_ 89% correlation is the headline. The 11% gap is the thesis.

Simulation closes testing for known edge cases.

Unknown edge cases, the ones you only encounter in real deployment, dont appear in sim at all.

That gap doesnt shrink with compute.

[Mark @markkmii]

@Fabriziobustama Clear use case here. Real city setting, constrained task. Fixed dock, known location, predictable approach.

Its a deployable pattern, narrow and controlled even in the real world.

What happens when someone grabs the bike before it docks?

[Mark @markkmii]

@MistralAI @Airbus @BMW @EDFofficiel Aerospace, automotive, energy.

All highly structured environments with defined tasks and humans in the loop. Not a limitation.

Thats the deployment pattern. Narrow, monitored, then expand.