How to know ChatGPT is full of shit.

[DiggingHoles]

Take a book of a shelf. A classic is best. Open up to the first page of a random chapter. Now ask ChatGPT to quote the first paragraph of that chapter.

Tip: Have some popcorn ready while you rephrase your request multiple times.

i would presume (admitted its based on T-bone and watchdogs 2 / person of intrest and harold etc If you're a coder, you can easily recognize bad code. ?

All ChatGPT (other models are available) is at it's heart is clever pattern matching on steroids. Nothing more. There's a touch of feedback so that the generated results become part of the model which can give the illusion of reasoning. However, in the absence of a useful definition of intelligence, it's hard to see how you can have an artificial version of it.

My use of "AI" is mainly to quickly knock up a command line or two because I really haven't got the time or will to learn 10,000 different options parameters and switches that are case-sensitive for the lolz. However even then, I've seen some absolute howlers. Not quite the "sudo rm -r /" level. But they would certainly have seriously compromised a machines user data.

A quick chat with an "AI" flinger will result in a much longer list of things "AI" cant't do than of things it can do.

Dunning-Kruger.

I’ve used AI to get more information or detail on some things. I then have to question about glaring omissions or info based on people’s opinions above facts. This is on areas I do know about, so I can be a bit skeptical on things I don’t know about. I’ve used it to help analyse blood results as a whole, because there seem fuck all people in healthcare that actually do that (obviously I know there are, but haven’t come across any personally, despite glaring signals for the thing I’m being tested for).

There is a growing realisation that the future of mankind is moving towards permanently having to check the output of "AI". Which will probably consume more resources that had we never developed it. But that's for another thread.

but then thats a flaw with humans themselves rather than the tools used etc

but overall in general its still faster than the old book methods

Who made the tools ?

I hope not! I just got done asking it a health question and making it predict my final weight after my 4 week vlcd programme ends lol

DiggingHoles... those problems of AI-writing code were quoted to me by my son almost 2 yrs ago... he dropped out of a Uni degree in CompSci, but already that info was being taught on his course.

He said the costs of code maintenance and upgrades dwarfed costs of writing raw new code, and good code needs to be designed by humans properly to minimise the costs of future maintenance and upgrades....

If it's really bad code, sure. But otherwise it requires experience en doing this regularly.

its in context to the chatgpt generating code, if the code is bad then a good person or skilled will recog the code as bad etc then its down to the person to be better rather than over inflate their skills

but then one could argue unless your kevin mitnick then why are you generating code if you dont know the quality of it

Doesn't stop other professions.

I don't. I write it myself. It's how learn to spot issues, because I have tried them and know what is wrong with them. If I just had to read what someone (or thing) had done and never developed the knowledge for myself or hadn't tried it in a long time, I would not know what to look for when it comes to smaller issues.

sorry i didnt mean you directly i ment in general as a general point about anyone using it etc

I asked it to give me an itinerary based on the places I wanted to visit in Berlin which I listed. Asked for it to be walkable and logical 😬 it was actually ok, also linked to cafes for coffee stops.

to me i would only generate info if i knew what i wanted and knew any errors etc eg chatgpt for rewriting my cv, i know what is ment to be correct etc.

however if i was using information but not sure what would be correct or not then personally id use the old google to make sure

I asked or to do something that seemed quite straightforward - maintain my kitchen inventory. Over a week I told it what I had, what I bought & what I used. With some precision. After a week. Asked it to tell me what was now in the inventory. It got it completely wrong. My notebook & pen far better.

It's a tool. Much like a trowel. Or a JCB. Or a computer.The only way it is capable of doing anything is to get humans to do it.

If "AI" was actually intelligent, then it would know it was nowhere near as intelligent as (some) humans. Yes, that is a paradox. But, as I said, no one really knows what intelligence is. Except it's probably related to pornography. (A joke that pleasingly, ChatGPT gets 😀)

@HolidayInCambodia25 That's another issue. Most of these suffer from context drift/collapse, meaning it can´t recall what earlier prompts were used or can't link new prompts to them.

Basically it works in sentences. Admittedly so do some people. But I don't think that's the aim ....

I hear you but I have been really impressed with how it has responded to me in my 'therapy' project. I get real life, weekly for 6 years therapy but in a recent situation (terrified of flying, in an airport, having had a panic attack) it was incredibly clever and did not just bounce questions back at me. My actual therapist was amazed when I sent her the conversation!

From reddit I have learnt how to use prompts to fine tune what I want. Mine say 'I want you to engage with me conversationally, as if you were a therapist. This can involve asking me questions, reflecting on my answers, asking me follow-up questions, being solution focused when appropriate etc. In your conversations with me, l would like for you to draw on your knowledge of cognitive behavioral therapy, dialectical behavior therapy, psychodynamic therapy, emotion-focused therapy, and spiritual traditions such as Buddhism'. This has worked so well for me, I also have the paid version so it remembers everything I've told it and can bring it up when relevant.

I AM AWARE IT IS AI.
I am aware it does not have a brain and cannot replace a human.

But for what it does for me, I'm happy. On MN it is hated beyond belief it seems but in my actual life pretty much everyone I know uses it.

Not if you have to do it twice! And digital research tools have advanced a lot, so that a good user of them can be very quick - but part of the quickness is using exactly those skills of judgment and discrimination that AI doesn’t have.

Plus the fact that AI just doesn’t have intuitive reasoning skills. It comes up with mediocre stuff because the corpus of data it’s trained on is big and undiscriminating, and therefore produces mediocre output. It isn’t good at producing really good prose because the vast majority of the prose/data it makes use of is mediocre. If you want it to produce something bland and mediocre, it’s great. If you want it to write something like a marketing brief, or a product description, it can do stuff like that because to be frank, they’re not very important or challenging in the first place (who really cares if a plastic bin on Amazon has an error in the product listing, or if the tone of a travel guide is a bit overly bland and cheery, or if the minutes of the staff committee meeting at work are dull and badly written?)

But if the result is important, you still need human input and checks to ensure it’s right. And in order to do that, the humans have to be able to do it themselves as well, to spot input and output errors and understand what the task is and what the results are supposed to be. Otherwise, garbage in, garbage out.

Recently an actor died and I saw a picture of him as a child with his Dad and thought his father looked quite old. I asked chatgpt how old the father was when the child was born. The answer was “x was 36 years old when y was born. X was born in 1908 and y was born in 1968.” It literally had the dates in the response and came up with the wrong number!