Before Li Fei-Fei: World Model Enables 4-Player Online FPS

Beating Li Feifei: World Models Enable Multiplayer FPS Gaming

URL Source: Quantum Bit

May 19, 2026 15:05:05 Source: Quantum Bit

Is authenticity truly important?

Jay from凹非寺

Quantum Bit | Official Account: QbitAI

Did I get killed by an AI?

There's video evidence of an encounter where I was shot by something indistinguishable from a human or AI.

All within a world generated by a world model.

This web-based FPS looks like a pixelated mess.

No game engine, no physics rules, no rendering code.

Everything you see is generated in real-time by a world model called Agora-1.

Humans and AI fight side by side in the same AI-generated world.

The AI Version of GoldenEye 007

Let's check out the official product launch video first.

After watching it multiple times, I had a strange feeling.

The aesthetics reminded me of a show like Black Mirror.

The uncanny factor was heavy, making me question if the people in the video were AI-generated.

Eventually, I had to let Codex transcribe it into text to understand better.

The gist: They created a multiplayer game driven by a world model, supporting up to four players, humans and AI fighting together in an AI-generated world.

Honestly, seeing this made me eager to try it.

No more ado, let's dive into the game.

The first thing I noticed was the weird background music as the page loaded.

Still loopinging in my head...

The UI was also strange, dark tones with low saturation, constantly giving me a Black Mirror vibe.

Everything was consistent with the theme.

Mouse clicks on buttons had模拟 sound effects.

Like an old-fashioned radio, scratchy and eerie, reminiscent of horror games.

Enough talk, let's start the game.

First, you choose a name.

I picked my pen name and entered the waiting room.

Looking for three more players.

Ps: Later, I found out you don't need four players; two can start directly.

Strange, if there are AI players, why not have them fill in when needed?

I don't get it.

The game rules are simple. It’s a homage to 《GoldenEye 007》, a 1997 classic on the Nintendo N64, adapted from the Bond film. It’s considered one of the origins of multiplayer FPS on consoles.

The rules are straightforward: split-screen combat with handguns, assault rifles, rocket launchers, and the iconic golden gun. The goal is to eliminate others.

Pure deathmatch without story or missions, just chasing and killing.

Agora’s design mirrors this classic.

Let’s begin.

The场景 has a "back room" style.

Stranger still, players occasionally glimpse other characters, moving silently like ice skating.

The人物动作 were bizarre, everyone agreed they looked uncanny. I couldn’t tell AI from human.

And the control experience? Terrible!

No mouse look, forced to use the left and right keys.

High latency, screen shake, moving felt like drifting on ice.

And the biggest issue: can’t aim properly!

Unable to stop, the cursor couldn’t stay on enemies.

I died instantly.

Fired a single shot and missed...

The other side must be AI! How come they can aim better?!

So frustrating! Don’t laugh! It’s not my fault!!

The death画面 was also awkward, deep red.

Finally, you get a scorecard.

It’s only fair, 新人玩家互相厮杀.

(Maybe justbecause the对手 are NPCs...)

Apart from combat, the game hides some fascinating features.

For example, clicking the information button reveals details about the company Odyssey.

Players also noted that bricks can be exploited to glitch into walls.

The world model then fills in the missing parts.

No crashes, no black screens—it spontaneously generates spaces you shouldn’t see.

This is intriguing.

In traditional games, beyond the map boundaries is nothingness, the programmer’s unfinished work.

But world models have no concept of boundaries.

The key point isn’t the game itself.

Reflecting on gameplay, it sounds simple by traditional game standards.

But don’t forget, this is an AI-generated world.

No hard-coded physics rules, no pre-rendered maps. Every frame, including those beyond boundaries, is calculated in real-time by the model.

Using《GoldenEye》as a testbed is a bold move.

Split-screen FPS is challenging because it easily reveals desynchronization and inconsistencies.

To ensure multiplayer FPS works, everyone must see the same world, maintaining consistent simulation in real-time.

Moreover, real-time interactive scenes are prone to失控.

Balancing complexity and playability is extremely difficult.

So, who made this?

Odyssey: Dedicated to General-Purpose World Models

The company behind this game, Odyssey, was founded in 2023.

Yes, named after the Greek hero Odysseus, fitting the company’s aesthetic.

An AI lab focused on general-purpose world models, with all products centered around them.

The founders, Oliver Cameron and Jeff Hawke, both have backgrounds in autonomous driving.

In July 2024, they debuted in the capital markets, securing a $9 million seed round led by GV.

Months later, Odyssey completed an $18 million Series A, totaling $27 million in funding.

Originally, their business wasn’t related to games—they were into AI video generation. But now they’re shifting toward interactive storytelling.

Agora-1 is their latest achievement.

The standout feature is multiplayer.

Previous world models, no matter how polished, were single-player experiences.

You explored alone in an AI-generated world, enjoying visuals but lacking interaction.

Agora-1 changes that, enabling up to four players to interact in real-time within the same generated world.

(Though the experience isn’t very friendly.)

So, what makes multiplayer so challenging?

This part is interesting. Let’s explore.

Others have attempted it.

Two companies to reference: Multiverse and Solaris.

Multiverse’s approach was straightforward—stitching player states into a split-screen image processed as one.

It works but feels crude, not addressing core issues.

Solaris took a different route, concatenating participants along the sequence dimension of a single auto-regressive diffusion Transformer for a more stable shared simulation.

But problems emerged—contexts blew up with more players, scalability suffered.

Both companies shared a common pain point:

When players are out of each other’s view, maintaining consistency becomes difficult.

In simple terms, the model struggles to handle multiple perspectives.

To reduce the load, Agora-1 decoupled simulation and rendering.

1. Simulation

Agora-1 learns how the world state evolves over time and responds to player actions.

The team trained the model directly on game states.

In Agora-1, the game is《GoldenEye》. The model learns the underlying game dynamics and how player actions trigger state changes.

2. Rendering

Here, Agora-1 generates visual representations of the shared state.

This is done using a world model based on DiT, conditioned solely on the shared game state without relying on prompts or traditional signals.

Think of this split as akin to modern game engines.

The difference? Both components are learned by the model, not hand-coded.

The result: Direct manipulation of the underlying game state.

This means Agora-1 can generate new levels while maintaining the original game’s dynamics.

The secret to multiplayer consistency.

By the way, the day before releasing, they announced another product.

This one had a bigger impact on me.

It’s Starchild-1, claimed to be the first real-time multimodal world model.

Generating visuals and sounds simultaneously, interactive.

You can make it play piano, with keys dropping and sound following.

Or recreate a cherished memory, like a wedding, using AI reconstruction.

The potential here is vast.

AIGC content could fill in gaps where memories are foggy.

Authenticity: Does It Matter?

Suddenly, I felt a sense of unease.

These products are still early-stage. Graphics are poor, controls lag, experience isn’t great. Far from the震撼 stage of GPT-Image-2.

But while playing Agora-1, I had a moment of genuine bewilderment.

I aimed at a character and shot. They fell.

I didn’t know if it was a human or AI.

I didn’t know how the world around me was rendered.

I even questioned if the opponent saw the same world as me.

Then I remembered—everything is generated by a model.

This feeling was uncanny.

With recent GPT updates, everyone’s worried about AI-generated fake chat records. The era of “pics or it didn’t happen” is over.

But now, I truly think images are easier—they’re static.

World models are different.

They simulate a持续运转, shared, real-time evolving environment.

They model time and subjective experience itself.

Honestly, watching world models evolve this year—from blurry to clear, from single-player to multiplayer, from visuals to sound, touch, and full sensory experiences—sometimes gives me goosebumps.

How can I know if the world I’m currently in isn’t generated by a more advanced world model?

Back in 1997, young people were chasing each other across the small screens of the N64, thinking it was the coolest thing ever.

By 2026, AI has learned to generate its own worlds, and its creators have managed to trick me into stepping inside one.

With the current pace of AI development, what will 2035 look like?

What we call reality…

Is it really that important?

Game link:

https://agora.odyssey.ml/

Reference links:

[1]https://odyssey.ml/introducing-agora-1

[2]https://odyssey.ml/introducing-starchild-1

© Copyright. All rights reserved. No reproduction or use of any form is allowed without written permission.