You can’t fine-tune your way to AGI, 1X CEO Bernt Børnich told me yesterday while announcing the launch of the 1X World Model Lab. The goal is to accelerate the path to fully autonomous humanoid robots, and 1X has hired Sam Sinha, a founding researcher at video-generation startup Luma AI, as its Head of World Models to run the new lab.

This is a progression on the existing 1X World Model, which the company launched in January of this year. That AI foundation model was built on video data that let Neo, 1X’s humanoid robot, turn a prompt into an action, even on objects it hadn’t seen. But now the company is scaling actual production of robots and getting closer to shipping at scale. That means much more data is now becoming available, and the goal of the lab is to turn that increasing information into smarter and smarter robots.

Those robots will eventually be fully autonomous, capable of working together in teams, and maybe even truly smart in ways not dissimilar to us.

"You can't fine-tune your way to AGI"

That’s Børnich’s line, and it’s kind of the whole pitch in seven words.

Børnich says that every AI leap of the last five years has come from feeding models richer kinds of data, in the right order. Text first, because there’s an inexhaustible lake of it online. Then text and images, then text and video. The mistake, he says, is treating robot data as an afterthought: a thin fine-tuning layer bolted onto a model that was pretrained on everything else.

"You can’t fine-tune your way to AGI," says Børnich. "You need to actually train the model properly."

I also chatted with Sam Sinha, who spent the last four years scaling multimodal models at Luma. For too long, Sinha says, robotics has been treated as “a second-class citizen.” Most humanoid robotics companies train on web-scale data, then fine-tune on a hundred hours of robot demonstrations. "That principle is so fundamentally broken," he said. "You need to see your most important tokens from step zero."

His summary of the job: "Good tokens in, good tokens out."

The key is that when you have actual robots in the wild doing things every day, your data stream for training your AI foundation models can get extremely diverse and extremely rich. It’s not just video data anymore, or text data, or images. Now there’s the live visual stream. There’s proprioceptive information: where Neo’s joints are and what forces are acting on them. There’s the data coming off 1X’s very impressive hands themselves , including pressure and force data. And there’s everything else available outside on-policy robot rollouts, including web-scale human video for diversity and volume, simulation and other human-centric collection.

As Sinha put it, the difference between gripping a bottle just hard enough to lift it and not hard enough is “tremendous,” and a camera alone can never fully capture it.

Data … and why Neo is so close to human

The bet here is that robotics is the same as any other AI problem: it gets solved by scale. If that’s true, you want to train on all the data. And guess what: in order to use all the human video on the internet, your robot has to be human enough that the data transfers … that’s why it’s relevant to what your robot can do.

"You build your embodiment so close, as small an embodiment gap as possible," Børnich said. "So now you can just pretrain all of the human video out there, and that actually transfers to your robot."

Of course, there’s always a gap, but this is why Neo looks and moves the way it does. Why it’s tendon-driven rather than geared, and why it has a hand with a massive 22 actuated degrees of freedom.

Dar Sleeper, 1X’s head of product and design, showed me video of those hands moving in a separate conversation , and the speed was unlike anything I’ve seen from another robot. Børnich calls the hand "the final boss of robotics." The point isn’t just dexterity. It’s that a more human body – and hand – makes human data more relevant for robot training.

Data as a competitive moat

I’m currently tracking over 400 humanoid robotics companies. New ones are popping up all the time: Galaxea Dynamics yesterday, VinRobotics from Vietnam’s Vingroup just the day before.

How does any of them compete? Or establish a strong competitive position?

For Sinha, the foundation model that will run 1X’s robots is almost downstream of something else: the data flywheel.

1X builds NEO from, as he put it, "literally raw copper wires" at its Hayward, California facility. That vertical integration is what lets 1X manufacture at scale, and manufacturing at scale is what puts robots in the world collecting data. That data feeds better models, which helps robots do more work, which makes them more financially justifiable, which then leads to more robots in the market, which then feeds more data collection.

It’s essentially a virtuous circle.

"I believe 1X has a chance to build a data moat," Sinha said. "And that to me is the most important thing."

His analogy is Cursor. The coding-tool company shipped its Composer model before it was state of the art, because shipping was how it collected the interaction data to climb. "That did not stop them from releasing it," Sinha said. The crappy-but-deployed product was the data-collection engine.

It’s a notable framing for a $20,000 home robot. The early units aren’t just product. They’re sensor units collecting more and more data.

Why the lab lives inside the factory

The obvious question is why a frontier AI lab needs to sit inside a hardware company.

Børnich's answer is that it can't sit anywhere else.

"Mind and body is not separable," he told me. The volume of hardware changes 1X makes specifically so the models can work is, in his telling, immense. And, he says, you can't move fast if the AI team and the robot team are negotiating across a corporate boundary.

"The entire company exists so that we can generate the data and embodiment that can get intelligence."

Børnich frames 1X’s extreme vertical integration on the manufacturing side as the only viable answer to China’s scale advantage. Makers like Unitree and UBTech build their own motors, gears and electronics end to end; 1X does the same, swapping gears for tendons. Its edge, he claims, is iteration speed: 1X takes just “four weeks from major changes in CAD until the robot walks off the new production line.”

1X keeps running in fast batches rather than continuous high volume, so the design can keep changing as feedback comes in. Essentially, it’s the hardware version of agile development in software.

Børnich says it’s the fastest-on-the-planet iteration, acknowledging that this is “a bold claim.”

Of course, there’s plenty of competition

Physical Intelligence, Google DeepMind and NVIDIA are all pretraining robot foundation models. Figure has Helix. Apptronik is working with DeepMind. Unitree and Agibot and dozens of other Chinese humanoid robotics companies are building and enhancing their own models, and for some with existing shipping scale, the data flywheel is already turning.

What’s distinctive in 1X’s version is the insistence on force and action-consequence data in the mix from the start, paired with an as-human-as-possible body. And the fact that 1X will start shipping 20,000 pre-ordered humanoid robots this year – Børnich re-confirmed this on our call – will supply 1X with an ever-growing data flywheel.

The lab will ship this year too

The lab should have early results before the end of 2026, which is good timing because the hardware is also shipping on a similar timeline.

What exactly will ship is the question, of course.

Børnich says Neo will ship “something by end of year that is useful with full autonomy.” He’s careful to manage expectations, though: 2026 and early 2027 are for early adopters who’ll need patience. 2027, he says, is when Neo goes "from useful to this is something I would really want."

(The good news here for early adopters is that Neo’s hardware will support many over-the-air software updates as the AI improves. And that Børnich says if additional bits of hardware need to be improved; they’ll accommodate that too.)

The bet stacking up here is layered: hardware good enough to ship now, designed so that better models — trained on data only this hardware can collect — make the same robot dramatically more capable over the air.

Børnich thinks the climb from "surprisingly useful" to mastering nearly any human task is "a lot shorter than most people think."

It’ll be exciting to see if that’s true over the next 18 months or so. And, given his comment on AGI, or artificial general intelligence, exactly how smart Neo is going to become.