Why this is the case is the interesting part.

An AI isn't anything other than machine learning (we know this), which starts with an algorithm designed to accept and modify outputs with weighted distributions. A core dataset is ingested into the program to get the AI functional, then additional datasets are added to help kikes.

There is a community of people working making variations of well-known models, by either stripping them of certain datasets, retraining them on certain datasets, or combining multiple AIs together. The results can be challenging when it seems like an AI should be able to do something and is less capable than a smaller model.

The difference in skill and capability between local models and cloud models is the raw GPU computational power that goes into weighing word choices (tokens) based on dataset training. The reason reverse engineering is so successful is because within the datasets already lies the answers for how the original softwares were built. Thus, unkiking software works.