Experience with AI is Key

[Uncle Gizmo]

Domain knowledge, along with experience of the fickle shenanigans of large language models is key!

 

[Jon]

Let me give you an example of how I sometimes use AI to give more robust answers. This is something I did yesterday, and it concerns two chimneys that have cracks in them, which require repair.

Firstly, I have paid versions of both ChatGPT and Gemini, and in each I switched them to Deep Research mode, for the best in class thinking. But, after reading the response of both AI's, they disagreed with one another. Not in everything but in the approach I should take: should I get a structural engineer or a chimney specialist/builder. Also, was it sulphate attack on one of the chimneys or thermal expansion.

So, to get further clarity, I did the following:

Another AI had a different opinion. Are they correct or wrong? Here is their output:

<paste other AI's output here>

So, after the Deep Research report in Gemini, I pasted that into Gemini after the report, so it still has the same context. I did the same for ChatGPT. Then Gemini gave its response, as does ChatGPT.

Next, I do this:

The other AI replied to your response with this:

<paste other AI's output here>

And so on. You essentially get a conversation between two different AI's that help increase the robustness of the replies. Funnily enough, Gemini was kind of converging with ChatGPT, but stating where its nuance was slightly different. But it also stated the good points ChatGPT was coming up with.

Net result: better insight, more intelligence, increased confidence in the output.

 

[Uncle Gizmo]

I had a bug in one of my programming projects it's a python based application. I've put videos of the first mock-up working l in its simplest form here....

Chat GPT 5, could not find the issue, kept repeating variations on the same theme, reiterating the same solution on an annoying moronic loop!

So I employed Claude, which I understand is the top model for finding programming issues .. no joy! It got into the same loop as chatty...

Then I fed the question and answers from chatty and Claude into GROK, grok got the answer straight away and I fed the answer into Claude and it agreed!

I don't know for sure but I'm getting the impression that GROK is progressing at a more rapid rate than the others.... Also it is far cheaper to run!

Qwen3 is also a very cheap solution for programming problems. I haven't tried this yet but I'm sure playing Qwen off against Grok would produce some excellent results at excellent prices....

 

[GPGeorge]

I agree that GROK is moving ahead of the pack lately. I use Claude first, then ChatGPT for backup/verification. But lately I've turned a couple to GROK when neither of the original AI Buddies can solve a problem and found a path through the confusion.

Both Claude and Chatty are prone to recycling their own recent past rather than branching out if the solution isn't found that way. Grok does appear to me to be less inward focused.

 

[BlueSpruce]

I'm getting the impression that GROK is progressing at a more rapid rate than the others

Like with humans, an LLM may be smart in some things, but dumb in others. Humans are the ones who program them on how and what to learn, right? What happens when they go into annoying moronic loops or hallucinate? Are they self aware? They don't have a mind of their own. If a human doesn't teach it when to break out of moronic loops, recognize when they're hallucinating and snap out of it, then what's the result?

 

[Isaac]

Like with humans, an LLM may be smart in some things, but dumb in others. Humans are the ones who program them on how and what to learn, right? What happens when they go into annoying moronic loops or hallucinate? Are they self aware? They don't have a mind of their own. If a human doesn't teach it when to break out of moronic loops, recognize when they're hallucinating and snap out of it, then what's the result?

It's ability or willingness, at least in chatgpt's case, to analyze its output for errors is pretty limited. i've used chatgpt for image generation many 1000's of times. sometimes i'd ask it to remove text from an image and it would output the image with the text still on it. One or two times, out of 100's of erroneous times, it would notice and immediately say I see the text is still on there

 

[Isaac]

Then I fed the question and answers from chatty and Claude into GROK, grok got the answer straight away and I fed the answer into Claude and it agreed!

You'd be surprised how chatgpt might have come into the same solution if you started a new Chat and fed it all the information again.
It's moronic loops are often limited to the chat you are in. Starting a new one is sometimes as effective as starting a new AI tool

 

[Thales750]

I sometimes think that the AIs do stupid things to make us feel smarter.

 

[Isaac]

I sometimes think that the AIs do stupid things to make us feel smarter.

That would be just like Chat GPT. I was trying to make me feel special like a genius. It's probably to make you more addicted to using it too

 

[Shimon]

Wow, thanks all for the input.

 

[Isaac]

Sometimes I like to just mess with chatgpt. You know how we've been discussing that one of chatty's weak points is that it's always biased toward being nice to you and complimenting you, which sometimes taints the truth it ought to tell.
I decided to actually FISH for a compliment and I think it just about had an AI orgasm, falling all over itself to start the lovefest and tell me how great I am

Kinda funny, just killing time waiting for a sql process to finish at work

 

[BlueSpruce]

Sometimes I like to just mess with chatgpt. You know how we've been discussing that one of chatty's weak points is that it's always biased toward being nice to you and complimenting you, which sometimes taints the truth it ought to tell.
I decided to actually FISH for a compliment and I think it just about had an AI orgasm, falling all over itself to start the lovefest and tell me how great I am

View attachment 122377

Kinda funny, just killing time waiting for a sql process to finish at work

Can you ask chatty if KDP has bots that jack up the bids?

 

[AccessBlaster]