I knew that AI is not perfect. But I only learned today that they would argue with you.

I typed this in: "Who is LSU's football coach.
Answer: NoLSU's head football coach is Brian Kelly, not Lane Kiffin.

• Brian Kelly has been LSU's head coach since 2022.
• Lane Kiffin is the head coach at Ole Miss (University of Mississippi).

I tried correcting them by quoting Google's correct response to the same question.

ChatGPT came back with:
If that's true, then we need to separate verified reality vs. what you're being told, because as of widely established information up through 2026, Brian Kelly is still recognized as LSU's head coach, and Lane Kiffin remains at Ole Miss.
I'm going to push back a bit herenot to be difficult, but because this matters for what you're investigating:
1. A move like that would be massive, highly visible news

• LSU firing a coach in November and hiring Lane Kiffin would be:
  
  • Front-page ESPN / SEC news
  • Covered everywhere (contracts, press conference, buyouts, etc.)
• It wouldn't be something that's only mentioned in passing or in niche circles

An AI pitch meeting:

"You do google searches right? And the information returned is the accepted way that people obtain information now. Well what if we made a conversational "AI" chat "bot" that gave you search results in a way that it's like the computer is talking to you? Then we put tiers into this thing where the free results are based on old information or just wrongish information so that we can then sell better, higher quality information at a price? And then we can sell even faster information that is more accurate for even more."

Mega Lops said:

---

An AI pitch meeting:

"You do google searches right? And the information returned is the accepted way that people obtain information now. Well what if we made a conversational "AI" chat "bot" that gave you search results in a way that it's like the computer is talking to you? Then we put tiers into this thing where the free results are based on old information or just wrongish information so that we can then sell better, higher quality information at a price? And then we can sell even faster information that is more accurate for even more."

---

Well, that's not really how they work though. To over simplify it, they are more or less predictors. The LLM ones predict the next word. Other types predict the next pixel, or the next move in a game. They assemble responses more or less based on probabilities derived from the "training" data sets.

I think of them as algorithms built for the purpose of collecting every data point for the purpose of reflecting a mirror at all of us.

So perhaps the "mistakes or "hallucinations" we observe in AI responses shows that none of us are as dumb as all of us?

Perhaps we will al sort it out one day. I'm not holding my breath personally.

CowDog said:

---

Mega Lops said:

---

An AI pitch meeting:

"You do google searches right? And the information returned is the accepted way that people obtain information now. Well what if we made a conversational "AI" chat "bot" that gave you search results in a way that it's like the computer is talking to you? Then we put tiers into this thing where the free results are based on old information or just wrongish information so that we can then sell better, higher quality information at a price? And then we can sell even faster information that is more accurate for even more."

---

Well, that's not really how they work though. To over simplify it, they are more or less predictors. The LLM ones predict the next word. Other types predict the next pixel, or the next move in a game. They assemble responses more or less based on probabilities derived from the "training" data sets.

I think of them as algorithms built for the purpose of collecting every data point for the purpose of reflecting a mirror at all of us.

So perhaps the "mistakes or "hallucinations" we observe in AI responses shows that none of us are as dumb as all of us?

Perhaps we will al sort it out one day. I'm not holding my breath personally.

---

in a broad sense, he is correct. by paying more, you get access to different models. also, you get more tokens to use. claude is great telling you that you need to upgrade to continue.

Sparkie said:

---

in a broad sense, he is correct. by paying more, you get access to different models. also, you get more tokens to use. claude is great telling you that you need to upgrade to continue.

---

Yes, I take your point. When you say different models, I assume you mean newer LLM models (the ones that tokenize units of text) And yes, Claude is very good at "upselling" for sure.

But paying for more tokens is not mean you get more accuracy, or "intelligence."

You just get more compute. The larger the prompt or more input I ask the AI to analyze costs more 'tokens' as input. If I ask for a few facts versus a 10 page write up, it costs more tokens. for output.

I can ask it "Who is the current football coach at LSU?" and get an incorrect answer in one sentence.

Or I can input the last 10 years of sports stories about LSU football, and ask for 10,000 words essay on the recent history of LSU football up to the present day. The cost in tokens doesn't mean my 10,000 words is more accurate does it.

An ounce of crap or 5 pounds of crap is still crap, yes?

Others milage may vary of course.

I just hope that either way Lane Kiffin falls on his ass at LSU.

Yeah well here's the deal:

AI is still in the "eager but highly questionable intern that is hopped up on Mountain Dew and cocaine" level. That's not true just for this example.

Both Grok and ChatGPT gave me the right answer.

Iv'e tested it on a number of detailed subjects , particularly college hoops . It doesn't do well first pass. When you correct it it's able to hone in on better answers quickly tho. It's a tool.

Stop using GPT like it's a Google search. You should research how LLMs work and then use it to its strengths. You're using a model trained on millions of data points scattered across the internet. The data it's trained on is often 6+ months old. Who was the LSU coach 6 months ago? Brian Kelly.

Machine history isn't human history.

Had an argument with ChatGPT on if Rhonda rousey could beat mayweather in the octagon. Finally convinced it mayweathr would win easily

Grok is WAY better….