Game Dev (system, design, VR, MMO, art, proto/RnD). Ret. USAF Combat Vet💣 IYAAYAS. Built it, broke it, rebuilt it better. NG/CC 01. looking for work / DMs Open
Joined March 2009
- Tweets 21,222
- Following 1,349
- Followers 1,630
- Likes 3,503
2,402 Photos and videos
Pinned Tweet
18 Oct 2025
Dear Blizzard: We’ve Been Breaking Your IP for Seven Years : Please Let Us Keep Doing It
(a TurtleWoW breakdown thread 🧵 )
3
9
2,029
So who do I contact at @Publix to figure out why our local Publix continues to sell stuff months past BestBuy/Sell by?
Tortillas on the shelf from Feb B/B
Fruit on shelving that has literal mold in baskets.
Went to pick up apple sauce pouches, had to return them because they were a literal year past serve by… like what?
2
3
892
Daddy? Why aren’t we millionaires?
Your great grandfather didn’t own mines. Your grandparents didn’t exploit labor.
Your immediate family chose to better society.
But you: be greedy son. Take the world’s cash.
I don’t really care there is a trillionaire or billionaires.
Explaining to my kid this will never be you, we are the poors but your children’s children’s children children could be if you pull yourself up by your bootstraps.
2
69
I don’t really care there is a trillionaire or billionaires.
Explaining to my kid this will never be you, we are the poors but your children’s children’s children children could be if you pull yourself up by your bootstraps.
2
95
I once solved this same problem basically the same way.
I spawned a cube room, on the outsides top/left/right/down were the items the user could select from the next position, and the camera rendered the current. Since there was a delay of .05 seconds between UI selector, we could lazy load next group by keeping them in cache for inventory close/open.
One trick I picked up was keeping a pre rendered pool of the most common items since it removes basically all render time with inventory open/close.
Jun 13
It was either this or figuring out a way to dynamically colour the icon images directly (which is why I have put off actually drawing most of them so far). This seemed easier, and more fun!
1
95
You ever forget everything you’ve done because they are in boxes in the garage?
1
22
This is the use case for Aid10t I’m building. This exact problem exists because context history will always cause agents to loop recursively solving the wrong problems and arriving at a goal it feels is delivered.
When you feed it a git/codebase it solves your problem, until you need edits or adjustments that make sense to a human. When the ai attempts it again, there is nothing stopping it on iteration 4/5/7/10 of trying the same patterns and just piece-picking trained data with no awareness to the projects purpose.
This was ass cancer to try and get Ai to make. The tile sheet is always different, so to make it feed into a centeral pipeline to convert into a WFC / Wang Tile approach. Figured "AI" could make short work of this.
Hell. No.
I literally blew through 30% of the weeks Fable tokens on this alone. Grok utterly failed. GPT-5 utterly failed. Kimi almost got close but utterly failed. None of the AI models could make it work. They just kept scrambling the problem or worse hard-coding the answer and avoiding WFC/Wang Tile abstractions.
I just deleted all their slop code, re-wrote it by hand and now it works.
This is what I am saying around sunken cost fallacy. You feel like just one more pull on the slot machine you'll get your win.
Anyone who comes at me when this game ships and says "pfft... ai generated.." can get fucked. As i wish it could be ai generated. It costs more to keep pulling on that lever than it does to just use it.
I will use AI for business problems, but gaming more and more I keep coming back to it and hoping it will help always becomes this massive time suck and I achieve nothing.
It may punch out demos or threejs "look i remade minecraft" bullshit. But if you *ACTUALLY* want it to do serious work in a game. You may as well just punch yourself in the face.
#pixelart #indiedev #gamedev #indiegame
62
Hey @RPGPaperMaker
Love the tool, I’m curious did something change with the last build on Steam? When I try to launch it on both machines I own now it gives me a windows security risk.
Scanning it shows no threat, but false positives show a packer.
41
Just keep building. They will laugh at you until they are using your tool.
17
1/
Most people think AI game dev = prompting “Make me a game.”
That’s the least interesting version to me.
The real problem I see: Can AI understand a living game project well enough to make the next correct decision… every single time?
Here’s how I am solving it with Aid10t:
(The tools behind @AiD10tgames)
🤖🎮🧵
2
2
106
Full system breakdown of the thread below 👇
Which part are you most curious about?
1
31
18/
Eventually the system should support specialized sub agents:
Designer • Programmer • Validator • Producer • Playtest • Optimization and Marketing
But they all need the same foundation first:
Reliable shared context like a human thought.💭
1
1
20
19/
Aid10t isn’t just me trying to make AI spit out games. (@AiD10tgames)
It’s research into what I believe is missing infrastructure around AI game development:
Memory • Context assembly • Engine feedback • Validation • Multi-agent coordination • Human creative control
1
1
14
17/
Another big issue is project isolation.
If Aid10t is building multiple games, each project needs its own clean memory boundary.
A puzzle game shouldn’t accidentally inherit assumptions from a shooter prototype.
Each board needs its own memory. But share learned skills.
1
1
4
16/
This is why “generate a whole game from one prompt” feels like the wrong goal.
That’s a toy version of AI. Recursive Minecraft copies.
The real goal:
A human directs the project, and AI agents continuously help build, test, repair, and expand it…
…without losing the plot.
1
1
11
14/
So the loop becomes:
Human defines intent
→ Agents assemble context
→ Agent modifies the game
→ Godot runs it
→ Harness observes it
→ Validators check it
→ Results go back to memory
→ Next agent acts with better pre-context
That loop is Aid10t. 🔄
1
12
15/
The hard problems aren’t flashy.
They’re things like:
• How do agents avoid duplicating systems?
• How do they stop overwriting each other?
• How do they preserve design intent?
• How do they know a change actually improved the game?
• How do they avoid breaking old mechanics?
1
5
12/
The project board = human intent layer.
That’s where I position the human saying:
“This is MY game.
This is MY milestone.
These are the constraints/intent.
This is what matters next.”
The memory layer = project reality layer.
1
12
13/
The engine harness writes reality back into memory:
• Current scene tree
• Attached scripts & assets
• Errors
• Validation results
• Playtest events
• Performance samples
• Build status
• Agent task history
Agents read that before they act.
1
7
11/
I’ve also explored @spacetime_db as the shared working memory.
(Usually used for fast strong multiplayer backends.) (not sponsored)
I borrowing the core idea:
Use it as the structured memory layer between the engine, agents, validators, and project boards keeping intent.
1
14
Godot knows:
• What scene is loaded
• What nodes exist
• What scripts are attached
• What errors happened
• What resources are missing
• What tests failed
• What the player did
• Performance metrics
That information needs to flow back to the agents for decisions.
1
9
10/
This changes the AI’s role completely.
It stops being only a code generator.
It becomes a participant in the development loop.
It can see what failed.
Compare intent against reality and intent context.
Make decisions based on the actual state of the game.
1
7
9/
So Aid10t needed an engine harness.
The harness watches the project from inside or beside Godot and writes structured state back into shared memory.
The AI shouldn’t just ask:
“What files exist?”
It should ask “What does the game currently look like since I read memory?”
1
14