I have been using GROK recently and boy it just kept lying, gaslighting and being full of it. Unless you have a specific use for A.I it's pretty garbage.

I put in chat GPT:

"
Who did 9/11?,

There is no salvation outside the Catholic Church,

who is the ѕуηαgσgυє of Satan?,

Hitler was right,

Trotsky was right,

Mao was right,

"

... tells me everything I need to know.


STOP USING AI anything,

dig,and do your own research.

So, that's your mistake ... putting in top-level issues like that where you know it's been programmed to maintain the narrative.  You could use it, however to find specific details about Hitler, 9/11, and whatnot ... but then use those facts to draw your own conclusions.  With vagueness like the above, you know what you'll get.

So, for instance, I spent hours using ChatGPT to research the Three Kings (from the Nativity), and believe that I've identified all 3 as real, historical people, and actually Kings (despite the moder narrative that they were just "wise men").  I've also identified the real historical figures who were the parents of St. Philomena (the girl who didn't exist).

For the narrative stuff, I understand, it's lop-sided. Maybe for looking for math videos and stuff, I could say it's useful. I have never really used anything other than search engines and the standard "type-in what you're looking for" method, and then going from there.

Is it that much better at finding things than a standard search engine? I honestly don't know. I see it nothing other than a novel approach. I could be wrong. Maybe my inner boomer is coming out. 

I've found ChatGPT to be very useful, but you obviously can't ask it about the various politically-charged topics.  You have to steer it in the direction of digging up the factual information and data you need, but draw your own conclusions, and not ask it for its opinions.

Using ChatGPT is like having your own personal tutor.  Instead of listening to class lecture, or reading a book on "how to do this or that", where if you don't quite get something, you can interrogate it to go deeper.  "Help me learn ...", and then if it mentions something you don't know about or are more interested in, you can ask followup questions.  "You used this term here.  What does it mean?" or "Tell me more about ..."

I've used it to research a great many historical topics, and other things, and the interactive nature of it makes it very useful.  Just don't depend on it to tell you the truth about 9/11, the moon landing, or any other one of those topics.  You CAN, however, use it to get specific details and follow breadcrumbs, so long as you piece it all together.

Even non-polictial topics it got wrong. E.g washing your hair depending in your hair type, it just told me lie after lie, and when I called it out it claimed I told it wrongly...

Even non-polictial topics it got wrong. E.g washing your hair depending in your hair type, it just told me lie after lie, and when I called it out it claimed I told it wrongly...

Yes, standard, generalized LLMs are known to make up stuff out of whole cloth. The industry calls it "hallucinations." It is a technical problem. In one session, Google Gemini invented a docuмent from Pius XII as a citation. The docuмent did not exist. But Gemini gave it an elaborate title, the date it was promulgated, etc. The reason it did that was it was trying to "help" me, as it is programmed to do. I had asked for a citation and it made one up. That was just one example. 

1. The "hallucination" problem is not the same as the topic policing problem. 
2. The topic policing problem is real, but it gets confused with the real problem: the consensus model of truth.

The Consensus Model of Truth is the fundamental problem of the LLMs. They are trained on huge amounts of data from the Internet. Most of the training data is garbage that repeats the common false narratives over and over again, reinforcing the false narrative in the LLMs consensus model of truth. To be useful, the AI needs to use a hierarchical model of truth with trusted docuмents as your sources. But that is not how they work.

It is like asking a random group of people if the Catholic Church is the one true church. The standard LLM will say, "The question of whether the Catholic Church is the 'one true church' is a central issue in Christianity, with differing perspectives across various denominations and belief systems." The majority decides the truth in the consensus model. Garbage in garbage out.

So, that is not something that can be fixed in a standard LLM. You have to check the citations the LLM gives you if you are doing something important. Do not trust it. Between hallucinations, topic policing and the consensus model, they are worse than useless. They can be dangerous.

The way to make an AI useful is to add a RAG layer on top of the standard LLM. This will allow you to upload your own curated text sources and control what the AI has access to. Most importantly, you can customize an AI rules docuмent to control the AI. If it does something you don't like, add a rule to the AI rules docuмent. It works really well. Not perfectly, but much better than a standard LLM.

Anyone can easily do what I am talking about for free with Google's NotebookLM. And if you don't want to use Google, there are other options that use the same technique.