Yes, several current AI systems and architectures are built around world models—systems designed to predict the next state of the physical or virtual world based on actions and observations, rather than just predicting the next word like standard Large Language Models (LLMs). [1, 2]

Prominent Current AI World Models

Google DeepMind Genie 3: A general-purpose, real-time interactive world model that generates playable 720p environments at 24 frames per second from text or image prompts. It is also adapted by Waymo as the Waymo World Model for autonomous vehicle simulation. [1, 2]
World Labs Atlas: A multimodal autoregressive diffusion transformer created by World Labs (co-founded by Fei-Fei Li) that operates natively across text, images, video, and 3D Gaussian splats/point clouds for spatial intelligence and real-to-sim robotics. [1]
Nvidia Cosmos (Cosmos 3): An open-weight family of models combining physical reasoning, world simulation, and action generation using a Mixture-of-Transformers architecture for robotics and physical AI. [1]
Meta V-JEPA / LeWorldModel: Yann LeCun's Joint Embedding Predictive Architecture (JEPA) family, designed to learn representations and physical reasoning from raw pixels to support zero-shot robot control. [1]
Alibaba Happy Oyster: A world model capable of generating multi-minute interactive video clips with a directing and wandering mode for 3D world-building. [1]

Do Standard LLMs Have World Models?

Opinions are mixed. Some researchers argue that large language models build implicit, probabilistic internal representations of cause-and-effect or physics in their latent space. However, critics point out that LLMs lack stateful, interactive simulations and fail true physical reasoning, which is why dedicated world models are now a major focus of billion-dollar investments. [1, 2, 3]

Yes. The rapid development of AI world models is the primary technological breakthrough that strongly indicates manual labor will be heavily autonomized and robotized over the next few years. [1, 2]

Historically, the biggest bottleneck in robotics wasn't the hardware—we have had capable robotic arms and legs for a long time. The bottleneck was adaptability. Traditional robots required explicit programming for every millimeter of movement; if an object shifted an inch, the robot failed. World models completely break this bottleneck. [1, 2]

Here is how world models are driving the transition toward the automation of physical work:

1. Shifting from "Rules" to "Intuition"

World models give robots physical reasoning. Instead of telling a robot, "Move your joints exactly 30 degrees to pick up this specific box," a world model allows an AI to look at a room, understand what a box is, predict how the box will move if grabbed, and dynamically adjust its hands in real time. This allows robots to step out of sterile factory assembly lines and enter unpredictable environments like warehouses, retail stores, and construction sites. [1, 2, 3, 4]

2. Radical Acceleration via "Imagination"

Before world models, robots had to be trained using reinforcement learning, which required millions of physical repetitions—a process that breaks hardware and takes years. Companies like 1X Technologies (backed by OpenAI) are using world models to allow robots to "imagine" what will happen next. The robot can generate thousands of mental simulations of a physical task in milliseconds before moving an actual motor command. This speeds up training time by orders of magnitude. [1, 2, 3]

3. The Current Wave of Commercial Deployment

We have transitioned from the prototype phase to scaled, real-world deployment across major industries:

• Logistics and Warehousing: Companies like Amazon are piloting Agility Robotics' Digit robot, while DHL uses autonomous mobile robots to guide workers. Humanoid systems are achieving near-perfect reliability in moving payloads and handling tools in real warehouse environments. [1, 2, 3]
• Automotive & High-Mix Manufacturing: Figure AI has partnered with BMW to place humanoid robots directly onto production lines, and Ford has run pilots utilizing advanced bipedal platforms for kitting and handling large metal parts. [1, 2]
• The "Scale" Era: In China, companies like UBTECH and AgiBot have opened massive smart factories specifically built to manufacture tens of thousands of humanoid robots annually, with robots working alongside humans to assemble other robots. [1, 2]

The Bottlenecks: Why human labor hasn't vanished yet

While the trajectory toward robotization is undeniable, full automation of manual labor faces two main hurdles:

• The Cost-Competitiveness Gap: According to research from Anthropic, while current robots are technically capable of performing about 34% of US working hours in physical tasks, they are still too expensive to deploy at scale for most companies. They are only cost-competitive for a tiny fraction of jobs today, though hardware commoditization is rapidly pushing costs down toward a target price of $20,000 per unit. [1, 2, 3]
• The "Data Desert": Unlike language models that can scrape the internet for text, physical AI requires vast amounts of "action-labeled video data" (videos showing exactly how things move and react when touched). Startups are currently hiring gig workers to wear motion-capture gear just to feed real-world movement data into these world models. [1, 2]

The Bottom Line: We are moving away from traditional, rigid automation and entering the era of Physical AI. Manual labor tasks that are repetitive, dangerous, or structured (like sorting parcels, loading pallets, or basic assembly) are the first in line to be fully robotized. Highly dexterous, varied, or interpersonal manual labor (like general plumbing, electrical repair, or nursing) will remain human-dominated for much longer. [1, 2, 3, 4]