It's hard to let them ruin wild because unreliability. I've built a simracing tool that's about 50% ai code now, and ai is mostly the boilerplate, accelerating

It's hard to let them ruin wild because unreliability. I've built a simracing tool that's about 50% ai code now, and ai is mostly the boilerplate, accelerating prototyping and most of the data structure packing unpacking needed It never managed to did a pit stop window prediction on its own, but could create a reasonable class to handle tire overheating messages All in all what I can say from this experiment is that it enabled me to get started as I'm unfamiliar with pygame and the ux is entirely maintained by Ai. Working on classes togheter sucks as the ai puts too many null checks and try catches making the code unreadable by humans, I pretty much prefer to make sure data is correctly initialized and updated than the huge nest of condition llm produce, so I ended up with clearly defined ai and human components. It's not prefect yet but I can focus on more valuable thing. And it's a good step up from last year where I just used it to second check and enrich my technical writing and coverting notes into emails. With vision and image generation I think we're closer to create a feedback loop where the Ai can rapidly self correct it's productions, but the ceiling remains to be seen to understand how far this will go.