gpt 6 astra
THREE YEARS LATER AND I’M STILL (mostly) RIGHT!
one of my first ever blog posts was “Why AI won’t replace programmers”. it was about my experience using gpt-3.5 and gpt-4 to completely vibe code (yes, vibe coding in 2023) a unity plugin to expose the latest features of google play billing v6. the main point of the post was that AI could definitely help out in coding, and increase productivity a lot, but would never replace programmers.
well, nearly as soon as gpt-6-astra, “AGI” according to some, became available on azure, i decided to check it out. i installed codex-cli, set up a pretty simple AGENTS.md, and got onto agi-powered vibe coding.
i’m happy to report that, more than three years later, i’m still mostly right! i do admit that i was wrong on some points i made in my original post though. i had focused on the lack of up-to-date information the model had instead of the quality of the code itself, but tbf it was my first ever java/kotlin project so there was no way for me to judge the code it had written.
have there been advancements in the models released since the post? yes. is AI able to do things it never was able to do? yes. the problem is, it’s still not able to write good code. yes, most if not all of the code it writes is “correct”, and will “work”, but it’s not really… maintainable. it’s overly paranoid with validation, exception handling, and overall really, really verbose.
similar to how the billing project got me introduced to kotlin and ultimately made me learn it more to be able to make plugins for UXR.QuestCamera, UMoth, UShare, etc. astra has definitely helped me learn new things (in this case, UIToolKit), but i would not be comfortable with maintaining a project written entirely by an AI.
putting aside maintainability and code quality since those are subjective opinions, astra, just like older gpt models, can still be very confidently wrong. it has no issues suggesting completely wrong fixes for issues.
recently, i encountered an issue where litert-lm would crash when using the opencl accelerator on the S10 5G. i decided i’d let the astra try to find the cause for it since i usually don’t investigate issues in such low-level environments. i will admit that it impressed me by supposedly examining the actual compiled binary to “pinpoint” the cause of the crash, but it was still unable to find a fix for it. i was able to manually fix it by just patching the opencl accelerator to reference the missing system library.
AI will help programmers, hand-in-hand, not replace them. you can definitely get things done through pure vibe coding, but actual infrastructure cannot be built like that. i think, in the future, programming will be a lot like art. for example, a lot of completely AI generated videos get tons of views and make tons of money on social media. there’s even some high quality AI-generated/-assisted media that are really successful. but true art, long form videos, documentaries, restorations; they’re still human-made. similarly, a lot of successful software will be completely AI generated, and most if not all projects will have some parts that were made with the assistance of AI. but true, industry defining projects like linux, ffmpeg, they will still have humans at their core.
this was a short one, but tbf there’s not much going on with my foss projects this month. i did update all my native-plugin-based packages to the latest unity-supported gradle versions and recompiled the ios plugins with xcode 27, though.
deadlock
omg city never sleeps is so peak. it’s taking a while for me to get used to the new map, but it is absolutely gorgeous. i’m splitting my votes between nurse harrow and violet.
what if she's gonna be like luz from the owl house and she uses glyphs to do magic? #deadlock #violet pic.twitter.com/XlQKtsrz8N</p>— Uday (@uralstechCTO) September 29, 2026