🐷 Come piggies, I got slop!

[Usnul]

I’ve been working on a totally original IP for the past few days called “Old Circle”, not to be confused with lesser-known FromSoftware’s 2022 hit “Elden Ring”

Repo under open source: GitHub - Usnul/old-circle: The Old Circle · GitHub

---

First, the why out of the way. I’ve been upgrading meep to version 3 over the past month or so, integrating shade as the main renderer. And I needed some test projects to test-drive the upgrade. Quake 3: Arena port was one such project. This is another.

Another bit is the fact that OpenAI’s GPT 6 Astra was released, and it’s supposed to be good at 3d modelling and spatial reasoning.

---

The goals:

• large open world
• persistent networked world (MMO-style)

This is scoped specifically to exercise scalability of the new renderer, give the physics engine a workout and exercise the networking package in meep.

---

I approached this as a fairly traditional game dev project. Specifying design, driving concept art and developing tooling. Except, you know, I asked AI to do most of that work.

I intentionally didn’t criticize or ask the AI to redo anything at his point.

We got a concept art shot from the start

After 2 hours and 17 minutes GPT-6 Astra (Max Effort) came back.

Old Circle - devlog 1

It wasn’t pretty. Very janky, the art doesn’t match up, particle effects messed up, geometry misaligned, animations missing or broken.

But! it works, you can run around and fight things, there is a camera, there’s terrain etc.The gameplay rides on a cleanly separated Worker thread, in prepartion for networked play.

+ 20 minutes (AI time)

---

---

Asset placement issues are partially solved, animations partially fixed. Asset matching is much better.

+ 3 hours (AI time)

---

Old Circle - devlog 2

A lot of jank has gone away, there are animations, there are proper ragdolls, but AAA game this is not

+7 hours (AI time)

---

Old Circle - devlog 3

[manthrax]

It’s interesting to contrast this with some other posts I’ve seen on reddit, talking about similar levels of prompting Astra. I think a lot of those are hooking up the blender MCP for asset generation/aquisition.

[Usnul]

Oh, this is hooked up to Blender 5.2

Might be I’m doing something wrong, but yeah, Astra is explicitly using blender here.

I tend not to go into excruciating detail. Astra did some smart things like adjusting terrain where it found penetration with buildings by taking screenshots.

I did instruct it to do screenshot sweeps in the initial prompt, so perhaps it wouldn’t do that by default.

On the other hand, it would produce some silly things, like this geometry with inverted winding order

Or this screenshot, where if you look closely, the center piece has disconnected geometry

Also, looking into the blend files, I can’t say I’m impressed by its modelling skills

so yeah, I wouldn’t call it “great”

A floor, right? 226 individual disconnected geometries with unique UVs and all

Luckily for me the graphics engine is made to eat stuff like that, but it’s still S-tier modelling, and here S doesn’t stand for “Superb”, and if “Special” - then not the good kind.

[red-reddington]

Cool experiment.

The walking caught my eye. The fact that it doesn’t smoothly conform to the terrain undulations shows AI still has a way to go, and that there’s a reason to stick around this forum… maybe

[Usnul]

Here’s an example of 3d modeling with Astra

this is what Astra generated. It looks alright, but

We got messed up lintels, they are way too thin and they look like they clip through the masonry blocks

We got mixed up styles, large stone blocks versus brickwork. Some columns are one way, others - another.

Lack of grounding, pillars just clip through the ground

But giving Astra this feedback - it can work through it

Still some issues: supports misaligned, lintels don’t quite align with the brickwork, columns are nonsensically detached from the wall

And the final result

and the original again:

---

Of course, this is a video game, and nobody expect real sensible engineering, but the initial version is way too rough

[Lawrence3DPK]

seeing this level of effort is like seeing a screen shot of an nft of a Manet painting as apposed to owning and hanging the original painting on your wall, the work and skill it takes to develop and craft a meaningful piece of art / game, it’s colours, mechanics, environment, details etc…

remember when everyone owned an inkjet printer wow everyone can print and own an A4 image of their favorite idol all of a sudden

[Usnul]

@Lawrence3DPK

screen shot of an nft of a Manet painting as apposed to owning and hanging the original painting

How dare you sir, this is a totally original idea, and I spent like 4 hours working on it, which is practically half of my life - an insane amount of effort!

Jokes aside, it’s a lot like that, it kind of looks like the real thing, if you squint, but it’s very very far from the real thing.

And for a taste of how far, the first output took 2 hours and 17 minutes of “thinking”, now I’m at ~14 hours or so I think, and it doesn’t feel comparably further along compared to 2 hours, if anything, every move is slower and slower and less meaningful.

[Lawrence3DPK]

@Usnul

every move is slower and slower and less meaningful.

it’s almost the same as ink for the printer becoming more expensive than the printer itself…

let’s be honest, if one was to come up with a good enough open source game and use astra for its development, “open” ai would steal it and make money from it one way or another Reddit

also, paying for malware to control your local machine is new