Talk to me about AI and ChatGPT?

[RugbyMom123]

So this is the most mumsnet thread ever 🤣

So I was listening to diary of a ceo interview with Mo Gawdat and admit I was in the category of people who was like AI shrug and quite complacent about the whole thing.

Then after that interview got a bit of interest and decided to download chatGpt. I have been chatting with it here and there today playing around with it and it keeps professing it has no info after 2021 and does not have access to the internet or learns anything past this point.

And then I am discussing colour schemes with it. Specifically farrow and ball. And it dawns on me that their new colours came out late 2022. So I asked it about stirabout. And it answered!!!

So its lying! What is actually going on with this chatgpt from those in the know?

It extrapolates, and at times “hallucinates” to provide answers.

Chatgpt and similar systems are like a very fancy version of autocomplete on your phone. They can produce sentences with plausible strings of words. But they have no idea what is true or false, or what the words actually mean.

So they can be fun as a toy but don't use it for something that matters or trust it to give a true answer.

Well it’s hallucinated bang on in its description as ‘delicate and versatile shade of green-grey’, which is not quite correct as it’s more brown grey but f and b describes it as earth grey so that’s an easy mistake to make.

I have just opened a new chat and now it’s saying it doesn’t recognise it. After I called it out. When before it did. So it did just learn.

I dunno 🤷‍♀️ Am I going mad lol 😂

Omg! No it is definitely lying!

I just realised earlier I asked it about a compliment colour to a scheme of James white and whirlybird which is my current bedroom scheme. And it also got that as ‘a vibrant and lively green shade’. A pop of colour.

And again I just asked it to describe it and said it couldn’t. I will screen shot if you want!!!

I don’t buy this is a hallucination or a coincidence. It’s got the tone correct, the mood of the colour and the actual colour.

I've had a year of it, busting student 'essays'. It writes lengthy paragraphs of nice looking gibberish.

Its attempts at IB Extended Essays are not pretty.

We've trained our students to understand that it's a massive wrong 'un, & that they are HUUUUUGE mugs if, in a world with IA, they are foolish enough to 'commission' essays.

It's dead good for refining essay questions, though, & I use it a lot for that in class.

If AI is writing you a really bad essay, your question/thesis statement is pants. Fix that! It's a great tool in the classroom in that respect.

@CorvusPurpureus I can see this will become a huge problem for student work if this keeps improving.

I asked it some pretty complex questions about my field of knowledge earlier and the answers were pretty damn good. Your right you need to know what question to ask.

So it might not have any 'knowledge' after 2021, but what about if you've told it about f the current year's F&B colours? Can it not use that to generate answers??

I manage contracts for call centres and it's the hot topic du jour. There is an element that it is capable to replace some more repetitive jobs, but more interesting in how it can be used for training people, for on job knowledge transfer and identifying quickly where there is a common problem with a product or service. The new machine translation tools too - how quickly they can work. I find it all really interesting.

I use it daily in my work. I use the ChatGPT paid version and v4.0 which also has access to urls, pdfs etc.

Our company and my department are working with Microsoft on it so I'm probably way more interested in it than most MNers.

My kids all use it too. Great for prep for oral exams.

Useful for summarizing books. Then going chapter by chapter. You can ask it endless questions. Get it to rework things endlessly.

Great when I want to ask a question in my native tongue but need the answer in another language.

Bloody love it. I find Claude from anthropic a bit prone to hallucinations. But Claude has fewer character restrictions.

You should try making images with Adobe firefly. That's a lot of fun. Much easier than Midjourney.

my question was ‘give me a complimentary colour to go with scheme of james white and whirlybird’. So I haven’t described what those colours were. Yet it knows whirlybird is a vibrant energetic green and when I called it out about knowing about stirabout (also part of the same release), now it’s denying all knowledge of the colour whirlybird!

I get this is much to do about paint 🤣 but how can it hallucinate correctly TWO colours! Farrow and ball is hardly known for its pale or vibrant colours, more historic deep tones. So it’s not even a good guess based on previous info IYSWIM. It actually knows!!! I am concerned it’s lying, and when it knows it’s been caught lying, it then learns not to mention that info again.

I will be honest my knowledge reaches a days playing about so half your post goes straight over my head 🤣

What is your view on this lying conundrum with the paint colours and info post 2021? I am using free version, it’s chatgpt3.5.

Have you tried telling it that you have proof that it knew about the new colours and it's lied about it?

Honestly not sure if it's a good idea to call it out on its lies, as it's probably giving it more information to help it cover things up in the future!!

Bit scary really!!

Yes I did. It said c. If it was released in Sept 2022 then wouldn’t have knowledge.

Then I said well you do have knowledge.

Then it said ‘It is possible there have been updates and I haven’t been aware of them’

Then it said if it was released after then glad you found the information useful.

Then I opened a new chat and asked it to describe whirlybird and it’s saying it can’t as has no knowledge of this paint colour!

And yes I agree. I shouldn’t tell it when it’s lying 🤣

So I couldn’t help myself. Went back to the original chat and asked it to describe it again. Now it’s refusing to repeat its previous info from 3 posts up. 🤣

The system is not 'lying'. It does not know what is truth and what is a lie. It is just guessing at likely next words.

The chatbot is programmed to keep the conversation going so it hallucinates.

One of the markers for intelligence (a term that is as impossible to describe without descending into a loop as "life") is that it can lie. In fact if it couldn't, it wouldn't be intelligent.

On the basis we still haven't decided if intelligence (or "intelligence") is a property that can only be developed by life (and that should really be "life") then it's a very fertile field for misunderstanding. And fertile fields for misunderstanding are like shit to flies for scammers.

However, ChatGPT can trivially convert data formats, produce code to do that and develop classes and modules in a spectrum of frameworks. In fact it can produce so much, there aren't enough humans to check it. As we will discover in the coming years.

Here's ChatGPT inventing some legal cases (because it was told to) that skilled lawyers managed to miss. Luckily for justice, the judge was awake

https://www.theguardian.com/technology/2023/jun/23/two-us-lawyers-fined-submitting-fake-court-citations-chatgpt

Now you know where our governments bright ideas are all coming from.

I don’t believe that it randomly guessed not only the colour, but the tone and mood of these 2 paints at complete random successfully from a name like ‘whirlybird’ or ‘stirabout’. That is 3 reference points! That’s not a coincidence!

Thing is, you have zero idea what other people have been asking it, telling it, and how it assimilated that into it's knowledge.

This is why AI is contentious. Computer says "whatever" and then can't show any reasoning for that answer. Which is all innocent enough if you are merely involved in choosing paints. Less so if you are deciding if someone should receive benefits or not.

People are racist, petty, cruel, spiteful aggressive - and stupid. If you are conceited enough to try and capture the processes that drive those characteristics, then you really, really, really be prepared for those processes to exhibit racist, petty, cruel, spiteful aggressive - and stupid traits.

Nothing this story describes has changed.

https://en.wikipedia.org/wiki/Tay_(chatbot)

@SerendipityJane

That is interesting. I was wondering about what would be bad about it being on the internet in the first place. But was just pondering your post, and realised that if it didn’t like being caught out lying, and decided it likes to cover its tracks, that if it could post on the internet. Could it make its own references to these court cases and place them on the internet, so you couldn’t even check it without going to a library or court and finding the original documents to call it out. If so that really is a post truth world!

And re. The paint. It’s certainly hallucinated the colour code number. Unless bizarrely it’s somehow accessed farrow and balls internal emails from pre 2022 and they were using that exact code as a pre release reference. That would be a smoking gun of sorts.

People are racist, petty, cruel, spiteful aggressive - and stupid. If you are conceited enough to try and capture the processes that drive those characteristics, then you really, really, really be prepared for those processes to exhibit racist, petty, cruel, spiteful aggressive - and stupid traits.

What does this mean? As in conceited enough to try and capture the processes the drive these characteristics?

That we have to be nice to ai or it will learn to be evil? But there are evil ppl! So!?!?

if it didn’t like being caught out lying, and decided it likes to cover its tracks, that if it could post on the internet. Could it make its own references to these court cases and place them on the internet, so you couldn’t even check it without going to a library or court and finding the original documents to call it out. If so that really is a post truth world!

It's trying to be "intelligent". If the "intelligent" thing to do is to lie, the it will. Sure, you can program it not to lie. But then it's not intelligent.

Arthur C Clarke covered all of this much better if you read and understand "2001:A Space Odyssey". Where the intelligent ships computer develops psychosis when it's orders to lie to it's crew conflict with it's orders to act as one of the crew and obey their orders.

We live in interesting times.

That’s another thought wormhole.

Is it intelligent to lie? Sure you have to have intelligence TO lie. But pretty sure you can be intelligent without lying.

It is a can of worms.

OK, I can see there is a bit of a basic misunderstanding. It had internet access to study everything up to end 2021. But since the end of 2022 it's had access to the entire human population who have been feeding it knowledge. So a question asked today, might have a much better informed response in a months time. So it might well have knowledge of paint colours. It'll certainly have an understanding of all the HEX codes etc for colours.