The 7 Content Types Most Valuable to AI World Models
Fei-Fei Li, Yann LeCun, and Meta are betting World Models will be bigger than LLMs — and they’ll need real-world data to build them. Here are the 7 content types AI labs will likely want to license.
If you own high-quality visual or non-textual content, your phone might be ringing soon.
There’s a new kind of AI called World Models and at least two of the top 20 AI thinkers recently launched startups have launched startups:
Fei-Fei Li (aka The Godmother of AI) helped kick off the deep learning boom with ImageNet — now she’s betting AI needs to see and simulate the world to truly understand it.
Yann LeCun (one of the Godfathers of AI) — Meta’s Chief AI Scientist quit Meta last month to create a still-to-named World Model business — and Meta immediately agreed to partner with him on it.
World Models could have a major impact on content & data providers.
World Models Watch the World (not Words)
World models aren’t your typical text bots — they learn by “watching” the actions in the world.
Think of it like this: if GPT models were trained on books and articles, World Models want everything before the book is written — reality itself.
LLMs predict the next word.
But World Models predict the next action.
World Models Learn Like Babies
LeCun compares World Models to how babies learn (from physical activities (“Action”, “Physics” and “Objects”):
LeCun provides tons of data in his lecture to underscore how limited LLMs are:
A couple of my favorites:
“A four-year-old child has seen more data than an LLM!
and
“We still don’t have robots that can do what a cat can do or…what a 10-year-old can do.”
LeCun et al. believe that AI’s own path forward to true intelligence is to train them on actions (not just words).
Another nice World Model visual is this one from AI Researcher Isaac Kargar (it summarizes LeCun’s presentation).
Whose data will train World Models?
Since World Models are different than LLMs, they need different data too.
They benefit from visual (non-textual) content.
Here are seven industries I see AI labs needing to license from:
Why This Content is So Valuable to World Models
If you look at every company in my table above, you’ll notice something in common: they capture data that reflects the real world — in motion, in space, over time.
They go beyond text with multimodal data such as:
sensor logs
video
images
audio
panoramic scans
scientific readings
All of it is grounded in real-world physics and interaction.
It’s the kind of input that doesn’t just complete a sentence, but predicts what happens next.
“There’s enough new data that will come not from the web but from the real world — from sensing, from measuring, from gathering images, from collecting enough video outputs — and that’s I think just the beginning of it.
— Cristóbal Valenzuela, Co-Founder & CEO of Runway (in Interview by Yazhou Sun of Bloomberg)
World Model Content Licensing Deals?
I haven’t heard of any deals yet.
But Ben Dickson of TechTalks guesses that World Labs’ Marble is initially trained on “a lot of 3D renderings”.
I’ll do a new post as soon as I hear of anything.
How World Models Learn (3 Steps)
From what I’ve heard so far, I envision three steps to train a world model:
Current state of the environment — This is what the AI sees or senses at a given moment — such as images, video frames, or any structured sensor data. Think of it like a snapshot of the world.
Action taken (or simulated) — This is the event or movement the AI makes — for example, “move forward,” “turn left,” or “grab object.” It’s the input that changes the state.
Resulting state (what happens next)
This is the outcome after the action — the next image, frame, or data point showing how the world changed. It’s used to evaluate how accurate the AI’s internal model is.
Content and Data providers will most likely play the largest role in Step 1.
Here’s Me Creating World Models
I gave the most popular World Model tool (Marble from World Labs) a whirl.
It’s early, and some output is beautiful — mine are crude.
But this screenshare gives you a quick taste:
What will World Models use the data for?
It’s early days, and the applications of World Models are endless.
But here are the types of World Model applications I envision:
Training — Take real-world videos and generate step-by-step onboarding checklists (e.g. Teaching Public Speaking — enter a video of someone at your company who presents well, and you can then ask the World Model to give step by step instructions of what makes them great)
Create Video Games — Generate fully responsive game worlds from a few images or prompts.
Safety Prediction — Spot injury risks in a factory from raw video footage of the factory (you don’t even need to show an actual injury. You can just give it a picture or video of the factory floor or ask the World Model to envision it).
VR Filmmaking — Build interactive films from a few environmental scans.
Robotics — Let bots test tasks in simulation before trying them in real life.
Architectural Tours — E.g. a wannabe architect prompts the World Model to show “a walkthrough of a 3BR/2Bath home in a certain neighborhood” (before the home is built).
Criminal Investigations — E.g. Show it a picture or video from the scene of a crime and it can guess/predict the location and surrounding area and turn it into a video walk-through of what may have happened.
Final Takeaways
For Content Owners:
If you own unstructured real-world data (video, sensor, images), your value is about to spike.
Tag your content, prepare licensing terms, and be ready to pitch — sometimes quality means even more than quantity.
Content Creation — have your team play with the World Models for creating new content. This is obvious in videogames, but there may be new content creation opportunities too.
For AI Executives:
Get strategic: partner early with data-rich industries before prices go up.
Aim to get data with clear ownership so that customers of your World Models have confidence they can use it without getting sued.
Thanks for reading!
Rob Kelly, Creator & Host of Media & the Machine
p.s. How to reach me? —The best way is to subscribe below and reply to any of my emails. There’s a free option — and your reply comes directly to my inbox.