RX-36

Web developer. Building applications and APIs with Django. Lately passionate about creating browser games with Phaser.

[RX-36 @RX36_web]

AI coding and app development have already become normal, but the “90% problem” that people were talking about back in the GPT-4 days still remains.

Especially when it’s combined not only with frontend, but also backend, authentication, containerization, and so on, the last 10% becomes a road of hardship.

The closer it gets to completion, the harder it becomes to adjust the details.

How is everyone overcoming this problem?
Should we expect GPT-5.6 to solve it? 😅

[RX-36 @RX36_web]

After trying over and over again, it still seems difficult to create pixel art with gpt-image-2.

Even when I ask Codex to “make it strictly on a 32×32 grid,” and then shrink the generated image down to 32px × 32px in Affinity to check it, the pixels are slightly off. They don’t snap precisely to the 1px grid.

I also tried using the Canva plugin, but in that case, the quality of the artwork itself wasn’t very high.

For now, if I want to use gpt-image-2 to create pixel animation, the most realistic workflow may be to generate an image first, bring it into Affinity or Photoshop, manually redraw it, and then import that asset into Aseprite.

I thought pixel art would be one of the first things AI would automate, but apparently not. It looks like there are still things humans can do.😀

[RX-36 @RX36_web]

I found something interesting.

It looks like it could be quite effective for solving issues with AI-generated pixel art.

https://t.co/n5tEjtgl2h

[RX-36 @RX36_web]

I’m on GPT Pro, the $100/month plan, but I hit the limit after asking only 7–8 questions in a day...
And now I can’t use it for almost a week. Why?

[RX-36 @RX36_web]

With ChatGPT Images 2.0, I asked for a pixel art tileset themed around office items, divided into 32px tiles, and this is what it produced.
This is amazing.
We’re getting very close to fully automating pixel art creation.

[RX-36 @RX36_web]

Isn’t this just way too expensive?
Hmm...
Still, I’m rooting for Copilot.

[RX-36 @RX36_web]

Image gen-2 is astonishing, but it still struggles with pixel art. In particular, if you ask it to take an existing piece of pixel art—say, 64×64 pixels—and resize it to 32×32, it completely fails. There’s still plenty of room for pixel art designers to shine.

[RX-36 @RX36_web]

GPT-5.4 xHigh and High are back in GitHub Copilot. That’s a relief.

[RX-36 @RX36_web]

GitHub Copilot has met a sudden death.
I hope it comes back.
GitHub Copilot really was the strongest.

[RX-36 @RX36_web]

It looks like GPT-5.4 xHigh and High have disappeared from GitHub Copilot. Why is that?

[RX-36 @RX36_web]

No one is talking about the fact that GPT-5.4 + xhigh is now available in GitHub Copilot.

And it only costs 1x premium requests.

It’s cheaper than Claude Code, and the performance is top-tier.

Why is nobody talking about this?🥲

[RX-36 @RX36_web]

In game dev, Pixel Art animation generation still doesn’t feel like it’s had the same insane leap that coding has.
I keep trying all kinds of tools, and somehow the frame-to-frame consistency always falls apart and turns into a mess.
Can somebody please make a ridiculously powerful tool that actually keeps it consistent?

[RX-36 @RX36_web]

I was working on updates to my own site using GitHub Copilot and GPT-5.4, but no matter how many times I prompted Copilot, the image would not expand to fill the full width of its parent element.

After struggling with it for about 30 minutes and getting completely worn out, I finally thought, “Maybe I should actually check the code for once,” and looked at the CSS. It turned out that one class had its width set to auto. I changed it to 100%, and it was fixed immediately.
Something a human could have corrected in 5 seconds ended up costing me 30 minutes.

What this made me think is that even if you can build something with a coding agent, if the human side lacks knowledge, it is still very easy to get stuck at the very last stage.

In fact, requests like this are already increasing:
“I built an AI agent, and it’s about 90% done, so please finish the remaining 10%.”

Even when an app looks 90% complete, it can still get stuck right at the very end.
Cases like this may increase even more from here, and at least for now, I think you still need not only the ability to grasp the overall structure, but also detailed development knowledge.

What will things look like a few months from now?
Will the current 90% problem disappear?
I do not think so.

[RX-36 @RX36_web]

Comparing GPT-5.2-Thinking and GPT-5.4-Thinking, the gap feels pretty large on problems outside web development. GPT-5.2-Thinking feels like it mechanically breaks an issue into pros and cons, and once it locks onto a downside, it just keeps hammering on it. By contrast, GPT-5.4-Thinking feels like it makes wiser judgments. It’s the first time I’ve had that feeling while using AI.

That said, when I compare GPT-5.4 and GPT-5.3-Codex in GitHub Copilot, I don’t feel that big of a difference. If anything, GPT-5.4 sometimes seems more likely to miss parts of my requirements in implementation. Maybe it’s just my imagination?😜

[RX-36 @RX36_web]

But pair-programming with OPUS 4.5, I finished coding, testing, and even the CI/CD deployment in just 1 hour... with only one single error. Opus 4.5 is absolutely insane. It's honestly scary.

[RX-36 @RX36_web]

While developing a game anti-cheat API, I realized my initial 2-week DRF build couldn't handle a load of 1000 RPS. 😇 So, I decided to decouple the architecture: keeping the Admin in DRF but migrating the API to FastAPI.

[RX-36 @RX36_web]

Normally, this refactor is a 2-week to 1-month slog — and at my skill level, it’s realistically a solid 2 weeks.

[RX-36 @RX36_web]

I built a browser game using Phaser 4 and Box2D.
Music is generated by Suno v5.
It's an idle game where you just hit the Start button and watch the physics.
Good for killing some time. Give it a try.
https://t.co/hahUhHMo4X

#indiegame #Phaser #gamedev