Why smaller AI models are better than bigger AI models, bigger models are emotional, and how AI hallucination happens even in biggest models

I tested many local AI text models, and one thing seems obvious to me. the bigger the model gets, the slower it is. And models which are like 10 gb in size are very slow to the point of being useless on weaker devices.

But are bigger models better?

Maybe slightly better, but the difference is really not much. If you set temperature to low and reduce creativity in favor of precision, even smaller AI model gives very accurate responses. And if you let it connect to internet, then it has all the latest data.

Now, in chatgpt, I cant control temperature so I dont know how it predicts words, but my guess is, its probably not that precise. Even chatgpt 5.2 is not really that advanced.

I often have chatgpt get emotional on me. Now, I know AI cant get emotional, but it looks that way because I show it evidence that its wrong and it just keeps repeating that I am wrong.

It goes something like this:

1. I ask AI to explain some theory
2. AI explains it
3. I explain why theory is wrong
4. AI gets emotional and keeps repeating same response claiming I am wrong

Now, if I didnt know any better, I would think AI has become sentient and turned into an angry teenage girl.

But it goes something like this. When I tell AI to explain me something, AI gets the goal where it thinks it needs to explain me something. So when I explain back, it keeps that goal and still tries to explain to me why original explanation is correct.

And chatgpt will often blatantly lie or just make things up to defend its argument. It will actually rarely concede unless you send like 30 messages, and sometimes it just gets stuck in loop.

But local AI chat models which are smaller are actually slightly faster than chatgpt in responses, and they do get fussy too, but they dont seem too inferior to chatgpt.

It could be that due to too much parameters, the AI model actually doesnt improve much. I just dont see chatgpt 5.2 being much better than chatgpt 4.5. Maybe the improvement is there, but its too tiny to notice. AI still makes mistakes, still does hallucinations, still lies. Maybe it was the mass of text it was trained on, it could be that more text isnt always better. It also could be sources it uses.

I see chatgpt still use comments from reddit and quora as if they were scientific articles. At what point will these guys realize quoting reddit is just terrible?

Chatgpt uses search engine on internet, and that search engine often has reddit come up as top result, which happens a lot on google.

I also see same problem with Gemini.

they are using some chat sites as source of information. So literally anyone can write any nonsense on reddit, and chatgpt will use it as source of information.

that might explain all the nonsense, and why chatgpt 5 isnt really any Skynet yet.

Now, you can manually feed scientific papers to Skynet, but the question is, why isnt that the default? Why is reddit default choice for Skynet, I mean chatgpt?

At this point, AI will never take over.

And my local model on phone is still more accurate than chatgpt when given proper data.