I’ve been having fun noodling around chatgpt. But I’ve found clear instances where it is surprisingly fallible.
For example it recommended that I take my power amps out of the wall and power them via a Shunyata Typhon T2. It produced a lengthy, seemingly well reasoned defense of it’s recommendation that went beyond Shunyata’s marketing material.
So I borrowed one and my amps clearly didn’t like it. The bass was deeper but the overall sound was muddy, thick and dull.
Then I asked why that happened and it gave me an equally lengthy explanation for why my amps are incompatible with the Typhon and should go to the wall. doh
And I found other examples of questions worded differently get conflicting answers.
Those mistakes aside, I’m still having fun discussions about things like average and instantaneous current demands vs varying spl.
I like it better than shoveling snow. But I won’t be guided by it’s output.
That’s a powerful endorsement - “I like it better than shoveling snow” 
Can I borrow that phrase please?
It’s yours.
This is an excellent example of how ChatGPT does not “know” anything.
I had a little conversation with my daughter lately. She uses AI in her latest use case for research on a paper she’s working on for school. The ChatGPT responses are not as apt as they used to be (
). That’s why she moved to Copilot instead. As for myself, I’m trying multiple AIs for anything serious and compare results, also refining my queries along the way.
Remember, it’s good to understand that we first need to choose the right AI before asking them on whatever they’re trained for. They’re not intelligent, at most knowledgeable.
My suspicion on recent ChatGPT’s underperforming might have something to do with their financial focus on securing market and datacenter positions. Resulting in a policy of more focus on energy efficiency by providing seemingly elaborate, but sloppy, less accurate answers.
Interesting! I’ve been curious about Perplexity, but from what I’ve read it seems more suited to finding cited research, and not so much on “assessment and recommendation”.
I’ve been leaning on ChatGPT for a couple of topics lately and have come to the conclusion it (AI generically) should not relied on for anything of real consequence. ChatGPT has been quite helpful but I’ve also found it absolutely makes up s**t.
This is a very basic explanation of how the AI engines work for those who may be struggling to understand what happens.
Basically, these AI engines are word prediction engines. They are trained on the content that their owners feed into them, let’s say a curated version of the internet. The AI engine compresses that information down into what is called a model. This model is basically a lossy compressed copy of everything it learned from the internet. Think of the model as something similar to an .mp3 copy of a music file.
When prompted with text (via chat) the engine looks at the sequence of words in the prompt, combined with the instructions that the model creators have built into it and it then starts spitting out the next most probable word that should follow in the sequence. Then the next, and the next, etc.
It does not think.
Keep this in mind as you use it and it will help you to understand why it hallucinates and also help you to get a better understanding of how to prompt it for the answer you are seeking. The more thorough you describe what you are looking for the more likely the results will be accurate and provide the information you are looking for. When you provide it a longer prompt, you significantly narrow down the pool of words that it has to pick from when it starts generating the sequence of words that it will spit out at you.
I hope you find this helpful and good luck in using this extremely powerful too.
Also, define what specific expertise you expect from the AI and perhaps train it (collecting additional data) to your specific query.
This. 
People who are skilled at asking very clear, detailed, and unambiguous questions usually wind up getting less chaff in the answers.
If you have a PayPal or Venmo account you can get a free 1-year subscription to Perplexity Pro. I believe the offer ends on 12/31/2025.
Thanks for the tip! https://www.perplexity.ai/join/p/paypal-subscription
You’re welcome. It is always best to run things against of few different Chatbots to see how they respond. You can get different answers and some give much better and more details answers too.
I will say things are improving very rapidly for all of the Chatbots. It is getting very competitive and they are all offering new features and services too.
I believe Google Gemini out in front for now and offers some amazing features like NotebookLM and Nano Banana. Watch some YouTube videos that explains them in more detail. Very impressive what they can do.
Exactly. In many respects, the LLMs are simply aggregators of information. That said, ChatGPT is an excellent tool if wielded with care.
The ole adage, Garbage In, Garbage Out, still applies though; so one has to be careful with their queries.
Love it!
One of my favorite pastimes is to ask ChatGPT and Gemini the exact same question. I have not only had responses from the two that are diametrically opposed, but I even noticed that they began dissing each other.
As far as I know, they’re still fighting with one another.
I hate it when the kids don’t get along.
In a larger sense, has anyone noticed recently (just the last week, maybe) an unusual lag on this forum? I wonder if it’s due to the ‘bots scraping this forum (along with everywhere else) to gather more data. The same thing was happening on WBF, and the web creator discovered exactly this - a tremendous increase in ‘bot scraping. They instituted two new features: non-members can no longer see site content, and new members are blind to the site for one week before the doors open. Right after they did these two things, site performance miraculously went back to normal.
Hmm… 