In the frantic, capital-intensive race to build thinking machines that can operate in the physical world, a philosophical chasm is widening into a canyon. On one side are the pragmatists, who believe in leveraging the colossal power of existing Large Language Models. On the other are the purists, who argue that true physical intelligence can’t be bolted on—it must be built from the ground up. This week, humanoid robotics firm 1X Technologies planted its flag firmly in the second camp, launching the 1X World Model Lab with a declaration that might as well have been fired from a cannon.
“You can’t fine-tune your way to AGI,” declared 1X CEO Bernt Bornich in a pointed announcement. “And you definitely can’t fine-tune your way to robots that can operate in the physical world.” The statement is a direct shot across the bow of competitors who are enthusiastically adopting Vision-Language-Action (VLA) models—AI systems that essentially “wrap” a powerful VLM like GPT-4 with motor control capabilities. 1X is betting the farm on a different, more arduous path: embodied world models.
The Great Divide: Fine-Tuning vs. First Principles
To understand the gravity of 1X’s move, you have to appreciate the two competing doctrines for building a robot’s brain.
The Vision-Language-Action (VLA) approach, championed by companies like Figure AI, is the path of least resistance. The logic is seductive: take a multi-billion-dollar foundation model that already understands language and vision, fine-tune it on a dataset of robot actions, and voilà, you have a robot that can take instructions. It’s an approach that leverages the immense progress (and investment) in LLMs. The problem, critics argue, is that these models lack a genuine understanding of physics. They are sophisticated pattern-matchers, not physics engines. They might know from training data not to drop a glass, but they don’t intrinsically understand that gravity will make it shatter.
Then there’s the World Model approach. This is the hard road. The goal is to build a foundation model that learns an internal, predictive simulation of the world. Before it ever learns a specific task like “pick up the apple,” it must first understand concepts like space, motion, object permanence, causality, and physics. Proponents believe this is the only way to achieve true generalization—the ability for a robot to act intelligently in novel situations it has never encountered in its training data.
Bornich’s stance is unequivocal. “The frontier is not better VLA wrappers,” he stated. “The frontier is embodied world models.”
1X’s All-In Bet and a Key Hire
The new 1X World Model Lab is the company’s answer to this challenge. Its mission is to build the most generalizable foundation model for humanoids from the ground up. To lead this ambitious effort, 1X has poached Sam Sinha, a founding research scientist from the generative video AI darling Luma AI.
The hire is a strategic masterstroke. Luma AI specializes in creating highly realistic video models, a technology that is conceptually adjacent to building a world model that predicts future physical states. Sinha’s entire career has been at the frontier of scaling multimodal generative video models. As he put it, for too long robotics has been treated as a “second-class citizen” in AI, with robot data being a “thin fine-tuning layer bolted onto a model.” The new lab aims to reverse that, treating embodied data as a first-principle ingredient.
1X’s strategy relies on a virtuous data-collection cycle, or what they call a “data flywheel”:
- Start with: Web-scale media, egocentric human videos, and simulation data.
- Add: Dexterous data from remote-operated robots.
- Deploy: A fleet of NEO humanoids to collect on-policy, real-world data.
- Repeat: The robot collects data, the model gets better, the robot gets better.
An Alliance of World-Builders
1X is not entirely alone in its philosophical conviction. The world model camp has some heavy hitters, even if they aren’t all building bipedal robots.
Tesla’s Full Self-Driving (FSD) system is perhaps the most famous real-world application of this concept. FSD relies on a “World Model” to predict the likely future actions of every car, cyclist, and pedestrian in its vicinity, running an internal simulation of plausible futures to inform its driving decisions. It’s not just reacting; it’s anticipating.
AI luminary Yann LeCun, now leading AMI Labs after a storied career at Meta, has been a vocal proponent of world models for years, arguing that LLMs are “fundamentally incomplete” because they lack an internal model of how the world works. His work on Joint Embedding Predictive Architectures (JEPA) aims to build models that learn common sense by observing and predicting video, a core tenet of the world model philosophy.
The Road Ahead is Paved with Petabytes
1X’s move is a high-risk, high-reward gambit. Building a foundational world model from scratch is an astronomically expensive and data-hungry endeavor. While the VLA camp gets a massive head start by building on the shoulders of giants like Google and OpenAI, 1X is choosing to dig its own foundation.
The success of the 1X World Model Lab hinges on its ability to execute its data flywheel strategy at a massive scale. If it succeeds, it could create a powerful data moat and a generation of robots with a far more robust and generalizable intelligence than their VLA-powered counterparts. If it fails, it will be a cautionary tale of eschewing a pragmatic shortcut for an elegant but impossibly difficult ideal.
The battle lines have been drawn. Is the future of robotics a clever extension of the LLM revolution, or does it require a completely new beginning? The industry is now watching to see if 1X’s bold bet on building the world from scratch will pay off, or if they’ll find themselves stuck fine-tuning their balance sheets.