Pushing models to their limits: The "Bad Apple" benchmark
retained text
I first wanted to do something with GPT-6 Astra, but every time it completely missed what I wanted and constantly tried to cheat, it just couldn't get it right. So, I started from scratch with Opus 5.5, it succeeded surprisingly well, it immediately got what I was looking for, however I had to give few feedbacks because it often misinterpreted what the characters were doing. I didn't corrected every detail, but more the striking things only. So, you might say the experiment could be biased with human intervention instead of one-shotting it, but really, I didn't have to do that much. It built very good methodology and tooling, it defined its scoring and measurement/comparison systems unprompted, however what it lacked really was vision, and that showed more and more around Remilia, it was hallucinated stuff. That's when I brought GPT-6 Astra, something surprising is that it immediately one-shotted good portions with zero need for feedback (most striking example is 00:41-00:55), despite that it couldn't do something proper when starting from zero, but continuing on the existing work by Opus 5.5 turned it into a beast and it was also very good at figuring out transition, I'm impressed how well the Flandre-Youmu transition is done. I tried again GPT-6 Astra alone and it still did terrible a job, so, this could only be achieved by both models working together... Maybe one day we'll see 0.999 accuracy ? GitHub : https://github.com/Ni-Cobra/Bad-Apple-ThreeJS
document fetched 2026-09-30 00:33 · canonical https://reddit.com/r/singularity/comments/1wsyqcd/pushing_models_to_their_limits_the_bad_apple