Cause & Effect Explorer — World Models vs. Language Models

0 / 3 correct

Scenario 1 of 3

A glass of water sits at the edge of a table. Someone brushes against the table.

Follow the physical cause-and-effect chain — one step at a time.

→

Event: The table receives a lateral force. The glass begins to slide.

Step 1: Friction is overcome. The glass crosses the edge and enters free fall.

Step 2: Gravity accelerates glass and water downward at 9.8 m/s².

Step 3: Glass strikes floor. Impact force exceeds glass tensile strength. Glass shatters. Water spreads outward radially from impact point.

Which statement is true?

🔤 Language Model

Predicts "the glass breaks" because it has seen millions of sentences describing dropped glasses. It cannot compute where fracture lines propagate, how fast water spreads, or what velocity the glass reached.

🌍 World Model

Simulates the trajectory under gravity, calculates impact velocity, models stress distribution across the glass, and predicts fracture patterns and water dispersal — all from physics, not text.

Scenario 2 of 3

A structural engineer removes one support beam from a bridge under load.

What happens next — and can a language model actually reason through it?

→

Event: Load previously carried by beam redistributes to remaining structure.

Step 1: Stress concentrations spike at adjacent connection points.

Step 2: If stress exceeds yield strength at any joint: plastic deformation begins. If not: structure holds.

Step 3: Whether the bridge fails cascadingly (progressive collapse) or holds depends on load magnitude, material specs, and redundancy — computable by a world model in real time, but not predictable from language patterns alone.

What does a language model do with this problem?

🔤 Language Model

Might say "removing a support beam risks structural failure" — which is correct as text, but cannot tell you when, where, or under what load the failure occurs. It matches language, not physics.

🌍 World Model

Runs a finite-element simulation: calculates exact stress at every joint, identifies the critical load threshold, flags whether this specific bridge with this specific load distribution fails or holds.

Scenario 3 of 3

A single mutation in a bacteria’s enzyme-coding gene occurs under antibiotic pressure.

This is where world models and language models diverge most sharply.

→

Event: Mutation alters enzyme shape at the antibiotic binding site.

Step 1: Antibiotic can no longer bind effectively. Bacterial reproduction rate rebounds.

Step 2: Resistant strain outcompetes susceptible bacteria. Resistance gene spreads through population.

Step 3: Whether resistance develops depends on the specific mutation’s effect on enzyme binding affinity — a 3D molecular simulation problem. Language models can describe resistance in general; world models can predict which exact mutations confer resistance for this antibiotic.

Why is this a better task for world models than language models?

🔤 Language Model

Can explain antibiotic resistance in fluent prose, summarize research papers, and give general mechanisms. Cannot predict whether THIS mutation confers resistance — that requires simulating the 3D protein structure.

🌍 World Model

Simulates the folded enzyme structure with the mutation applied, models how the antibiotic molecule docks (or fails to dock) at the binding site, and outputs a binding-affinity score — a real prediction, not a text retrieval.

Complete

You’ve explored the frontier

0/3

World models don’t just know things — they simulate the underlying dynamics. That’s the shift happening now, as $1.45B in funding lands on labs building them.

Read the full edition →