That's not how LLMs work either.

An LLM had no knowledge, but has the statically probability of a token to follow another token, and given an overall context it create the statically most likely text.To calculate such probability as accurently as possible you need as much examples as possible, to determine how often word A follow word B. Thus the immense datasets required.Luckily for us programmers, computer programs are inherently statically similar, which makes LLMs quite good at it.Now, the programs it create aren't perfect, but it allows to write long, boring code fast, and even explain it if you require it to. This way I've learned a lot of new things that I wouldn't have unless I had the time and energy to screw around with my programs (which I wished I had, but don't), or looked around Open Source programs source code, which would take years to an average human.

Now there is the problem of the ethic use of AI, which is a whole other aspect. I use only local models, which I run on my own hardware (usually using Ollama, but I'm looking into NPU enabled alternatives).