Brendan Greene, better known by his only pseudonym PlayerUnknown, is launching his next game later this week. He spun up PlayerUnknown Productions after moving away from PUBG, the battle royale sensation he created, in 2019. Six years and several studio leadership shake-ups later, the studio will release survival game Prologue: Go Wayback! Into early access on November 20.

It's supposed to be the first part of a three-game grand plan that will culminate with a world-scale MMO. To accomplish that lofty goal for what the studio currently calls Project Artemis, Greene believes utilizing machine learning is necessary. As such, Prologue: Go Wayback! is being used as a training ground for a machine learning model to generate a height map for a survival game with a nearly 25-square-mile world.

The use of both generative AI and machine learning in video games is currently a hot-button issue. To grok exactly how machine learning is being used in his next game, I hopped on a call with PlayerUnknown himself to learn more about machine learning's role in Prologue: Go Wayback!, Greene's hopes for the tech, and the ethics of its broader use in the game industry.

How Prologue: Go Wayback! Uses machine learning

Billions of height maps, generated on your PC

In Prologue: Go Wayback!, the player's goal is to navigate from a starting cabin to a weather tower on the other side of a generated map. In typical survival game fashion, you have to account for your own character's health to make sure you get there alive. There are no enemies in the game, only the weather and the terrain. It's solely single-player for now, although Greene told me he was open to adding multiplayer in a DLC.

Greene was attracted to this survival genre for this project as it aligned with his technological goal for the project: generating world terrains locally in real time with machine learning. Most survival games use some form of procedural generation to craft the world, using tools like Unreal Engine 5's Procedural Content Generation Framework (PCG). PlayerUnknown Productions also uses those tools for Prologue: Go Wayback!, but builds on top of a terrain layer generated locally on your PC using machine learning.

That comes out in the form of a black-and-white image that determines where elements like rivers and mountains should go and what shape they should take. The machine learning model was created with open-source data from agencies like the ESA and NASA and does not have an always-online requirement.

Above, you can see the images generated via machine learning that Prologue: Go Wayback! builds its world on. Greene broke down how this Guided Generation process works in more granular detail to me:

"What we're using machine learning for in Prologue: Go Wayback! is that it generates the height map of the world. It's called Guided Generation. We have a river schematic. We started off with just splines in Blender of where rivers might look like it could go. From that black and white image that tells you where the water is with white, it generates a height map for the world or for the terrain from that…It's really just generating that base layer, and then we put it into Unreal and use a bunch of typical procedural generation systems like PCG and stuff to populate the world with assets like trees, the foliage, and the cabins."

To create those plants, rocks, and buildings, Greene also still sees a place for environment artists in his company. "We have environment artists because they still have to design the scenes that Unreal's PCG system uses to populate the map," he explains. "All we're doing is just generating the base ground layer, and that's populated using traditional procedural systems, which most studios use."

Why Prologue: Go Wayback! needs to use machine learning

The end goal: Project Artemis

While Prologue: Go Wayback! is primarily developed in Unreal Engine 5, the hope is that these learnings can be applied to PlayerUnknown Productions' proprietary engine Melba. More specifically, Melba will be used to power Project Artemis, a world-scale MMO that is Greene's current white whale as a game developer. One day, he'd love for Artemis to possibly become an open-source engine where you can use machine learningto create not just height maps, but population maps as well.

"Our models need to be deterministic because if the goal is to build an earth-scale world, where do you store the data? You can't, it's just so much data if you're building an Earth-scale world with millions of players, so that world has to be essentially a generative world," Greene explains. He's thinking about making something even more ambitious than Light No Fire from Hello Games, a world-sized MMO that can support millions of players as they explore, create their own economies, and hopefully have fun somewhere in that process.

I've been around the game industry long enough that I'm always a bit skeptical of ten-year plans that seem to prioritize technology over game design. That said, PlayerUnknown Productions is at least not staying in the realm of lofty promises and is actually delivering games ahead of Project Artemis, making its creation in the future more possible. In Greene's mind, being able to generate terrain locally without servers is a necessity for Project Artemis to ever see the light of day.

"Most image generation models require very large server farms to generate these things. We do it all locally on the GPU in about 60 seconds or so. It's a very quick process because it's just a 2048 by 2048 pixel height map that's a black and white image. That can be done locally. With all the stuff we're building with machine learning, it's leveraging it client-side to make stuff more efficient, so you can generate more locally and not have to worry about having servers to do this for you. If you're building a world with millions of players, it has to be done locally with a peer-to-peer connection."

When Greene took a step back and explained his ambitions for Project Artemis, it became more understandable why he embraced deterministic machine learning for Prologue: Go Wayback! He also seems concerned with creating playable builds and finding the fun in his games, which isn't always the case with machine learning and AI-first projects. Still, there are ethical dilemmas hanging over the project's head.

The ethics of using machine learning for game development

Greene doesn't think this tech will cost game developers their jobs

Over the past month, we've seen backlash to the use of generative AI in games like Arc Raiders and Call of Duty: Black Ops 7. Prologue: Go Wayback!'s use of machine learning technically involves a generated image, but it's for terrain mapping based on open-sourced data. It's done locally, too, so it doesn't rely on problematic server farms to generate that terrain map.

In that way, Prologue: Go Wayback! is more ethical than many other AI projects. Still, if Melba's open-source future Greene hopes for becomes a reality, one can't help but question if this machine learning technology could eventually put developers out of a job. Greene stressed "it's not that kind of tech" when pressed about that possibility of Prologue: Go Wayback!'s machine learning techniques potentially putting game developers out of a job.

That said, Greene did go on to lean into some questionable terminology I've heard in AI spaces, saying that this machine learning technology could allow "smaller teams of artists to iterate on the world a lot quicker." It's Greene's hope that doing so would "allow teams to focus on the gameplay loops and systems that are in the world."

For his part, PlayerUnknown Productions has embraced that in recent years during development while actively listening to its community. If a corporation like EA, Ubisoft, or Microsoft began to use the technology, I don't think I'd be able to say the same. Nowadays, any and all advancements in AI and machine learning in game development should be met with a healthy amount of skepticism until the nuances of the situation are hashed out and the safety of game developers is ensured.

Green compares his studio's use of machine learning to being the conductor of an orchestra, rather than a musician within the ensemble. "Think of it like an orchestra. Instead of having someone playing a violin, they now just conduct the orchestra. They know where all the trees and stuff should go, and the machine learning agent just put it together. But we still have to make all the assets, write the music. It just plays it."

While PlayerUnknown Productions is currently operating with a full orchestra, Greene's comments suggest that the conductor is much more important than the violinist, which might become hard to get on board with if the violinist's solo ever comes. Even though PlayerUnknown Productions isn't costing its own employees jobs by developing this tech, part of me can't help but raise an eyebrow when I hear Greene say this tech can make game development "easier" with "smaller teams."

A critical moment in the history of machine learning's use in game development

Who is more important, the conductor or the musician?

Prologue: Go Wayback! is shaping up to be a fascinating first outing for Greene's studio. It's betting big on machine learning technology that many people, myself included, are skeptical of because of its ethical implications. I can see how machine learning may have more of a place in game development than generative AI, although I still see plenty of worrying problems that will always prevent me from blindly embracing it.

PlayerUnknown Productions may be testing machine learning technology in a fairly ethical way for Prologue: Go Wayback!, but will a gaming conglomerate do the same if it were to use the studio's technology in a future where it goes open source? Will it be used to empower design freedom, as Greene hopes, or will it be used as a justification to hire fewer developers who can make real, meaningful art that resonates with people?

I don't expect this game to have all the answers, but that is a question that will continue to loom over games like Prologue: Go Wayback! because of the bad blood left behind by the more egregious likes of Arc Raiders and Black Ops 7. For now, machine learning is just being used for a height map in Prologue: Go Wayback!, which we've yet to see the critical and audience reception to. If Greene's big bet on machine learning pays off, we could be playing a much different tune (or none at all) on our violins ten years from now.