If I flip this coin, you know that the

odds are going to be 50/50. And you know

that as a fact without using fancy

notations because you grew up observing

the laws of physics over time. But

unlike us, Lars language models don't

have this luxury because LLMs don't have

a simulated environment to test out its

theory. The closest thing LLMs have to

that is reasoning models that use chain

of thought to reason its thinking. So

there in lies the question, are large

language models inherently flawed in

their ability to grasp the physical

world? And how exactly does world models

actually overcome this? Welcome to

Kilbright's code where every second

counts. Quick shout out to BCloud. More

on him later. On a recent podcast with

Dario Amade, he talked about the current

approach of pre-training in large

English models. Unlike humans, LLMs are

trained using trillions and trillions of

tokens over months in time. And in

comparison, humans not only develop a

lot slower, but they also experience the

world in modalities other than trillions

of text tokens. For example, we know

that the laws of physics govern the

physical world. And our senses observe

and act in the physical world and we map

our understanding about the physical

world as we embody them in our own

brain. But pure large language models

are trained only using text which is the

highest abstraction that describe the

physical world. So if LLM's never really

experienced a coin flip, how does it

really know whether a coin flip by law

of large numbers will eventually

converge to 50/50 without really

observing them in the physical world?

Around 2018, there was a resurgence of

what's called world model. And world

models approached things very

differently than LLM's in how it modeled

its understanding. What if instead of

feeding an AI model streams and streams

of text tokens, we train them to

essentially simulate the physical world

in its own brain that best represent the

physical world. As you can see, being

able to pull this kind of thing requires

a thorough understanding of the laws of

physics and cause and effect that best

mimic the physical world in its own

model. So, how exactly does a world

model do this? The original paper by

David Ha uses three main components. You

start with an environment and you

introduce a vision model that

essentially observe this environment.

The vision model's underlying

architecture uses variational

autoenccoder that is trained to

essentially compress what it observes

visually into a lower dimension in

latent space. What this compression does

is that it trains the vision model to

extract only the important features and

throw out the rest. So now that we've

set up our vision model that observes

the physical world, we now need a model

that can process this information. They

use MDN RNN as the underlying

architecture. And since recurrent neural

networks are really expressive in being

able to keep track of all of its

previous hidden states, what this

architecture allows is to store all the

previous hidden states that it saw in

the past, which allows the model to then

predict based on what it saw before and

based on what it's currently seeing. And

the MDN RNN architecture works very

similar to this app sketch RNN where

once I start drawing a picture of my dog

Logan, it'll take over and actually

finish my drawing based on its

prediction of what the rest of the

drawing would look like. The final piece

is a controller model where it samples

from NDN RNN's output as well as the

original outputs from the vision model.

And the sole purpose of the controller

model is just to make an action like

passing butter or moving left and right

or picking up objects. As you can see,

these three components in a world model

will learn to interact with the physical

world and map out what it thinks the

physical world looks like after seeing

the cause and effect of the real world.

And eventually its own representation of

the world will be sufficient that you

can just sever the actual environment.

What this allows is for the agent to be

trained solely by the simulation of the

world models inner mapping of the real

world to train the agent to do whatever

you want. And the assertion being made

here is that this kind of modeling fits

much closer to how humans think and gets

us much closer to AGI than LLMs do. And

the initial result was pretty promising

where the world model was able to learn

how to drive on this randomly generated

track by staying on the road while only

using less than 5 million parameters in

total size when we sum up all its parts.

So the real question here is can world

models actually scale to human level

capabilities or even more? And here's a

quick shout out to BCloud. If you want

to learn more about the theory behind AI

and AR research, check out ByCloud's

intuitive AI that is full of learning

materials. You can start from beginner

level to understand all the way from how

tokens work to embeddings, encodings,

and attention mechanism that power most

large language models today. He really

mixes in good illustrations while giving

an easyto- read narrative on how the

technology actually works intuitively.

You don't need to have a deep math

background. It'll just read like a novel

where you can sequentially learn from

the beginning or just use it as a

supplemental tool on areas that you're

curious about. He goes through different

pre-training and post-training mechanism

here as well as more advanced concepts

like Laura for you guys. By cloud has

given away 40% discount on the yearly

plan using the coupon code link in the

description below. One of the biggest

reasons why large language models that

we use today are so popular is because

LLMs scaled quite beautifully. While

world models are still domain specific

to certain jobs, most LLMs are what's

called foundation models, which

basically means that we can use a

generic LLM like GPD 5.2 or Opus 4.6 to

do many downstream tasks beyond simple

chat, but run deep research, software

development, manage our computers, and

more. But ever since the resurgence of

world models in 2018, we also had many

iterations and different flavors of

world models too. One of the biggest

proponents of role models is Yan Lakun

who worked on world models in meta as he

contributed to JEPA based models. He

later left Meta to start his own company

called AMI or advanced machine

intelligence that is seeking up to $5

billion valuation. And Yen has been

patronizing LLMs for quite a while now

saying how LM simply don't understand

the physical world beyond its auto

reggressive nature that predicts token

after token. But we know that language

actually contains more than what Yen

probably gives credit for. Languages not

only represent the physical world in

words, they also contain grammarss that

dictate meanings, figures of speech that

provide abstract understanding of the

physical world and the individual units

of words like nouns, adjectives,

adverbs. They all reveal facts about the

physical world. But meanwhile, the gap

between pure LLMs and world models have

also been blurred as of around 2023 as

models like GPT4 from OpenAI, Gemini 1

from Google introduce what's called

multimodality where vision language

models that can also perceive images

using cross attention help LLMs also

perceive so to speak. And conversely, we

also have VLA or vision language action,

which is a type of a world model that

uses vision transformers with LLMs to

create action tokens. And this kind of

setup is what powers Neo, the humanoid

that was released back in October 2025

that became viral back then. But many

people still criticize multimodal LLMs

because, well, at the core, it still is

an LLM that lacks spatial awareness

about the physical world. And Fei Lee

wanted to demonstrate spatial

intelligence through her startup World

Labs back in September 2024, which

raised more than $230 million. World

Labs released a product called Marble

that essentially creates gajian splats

that produce millions and millions of

these particles to interact with that

are quite beautiful to work with. So,

just like how we saw in the original

paper, what you're seeing here is a

world model's representation of the

actual environment, but mapped in its

own model that we're able to see here.

But what's different with Marble in

comparison to traditional world models

is the absence of controllers that

actually grapple with the physics of the

generated world. This kind of sentiment

is expressed by Py who founded General

Intuition in October 2025 that's trying

to create a closer representation of a

world model that can actually interact

with games and simulations. Google has

also been a huge contributor in this

space when we look at SEMA in March 2024

and Sema 2 in November 2025 and of

course their most recent Genie 3 that

creates a hyperrealistic world that we

can move around in. As you can see, the

world models depiction of the physical

world is what allows us to generate AI

videos on Sora, train cars in a

simulation, and align robots in

factories. Nvidia also has a hand in

this by providing an open-source

platform called Cosmos, which is a world

foundation model. Similar to how

foundation models in LLM allow a generic

model to be used across many downstream

tasks, NVIDIA's Cosmos provide tools for

developers more upstream where we can

use three pre-trained models to cover

various use cases mostly around data

augmentation and data generation for

downstream training like custom

post-training like autonomous vehicles,

robots and video agents. Now that we

covered the landscape of world models

architecture and different approaches

that are taken by other companies, we

are still left with this rather

philosophical question. Do LMS really

demonstrate thinking and understanding

like humans do? Or does it even matter

that they actually think like humans? Do

you think that LM and world models are

mutually exclusive or do they just solve

different problems? What is the best way

to augment intelligence artificially?

What do you think?