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 found Monday and it is entertaining 😂

One thing about AI is that it reproduces all the encoded stereotypes and sexism, racism etc. of the material it has been trained on. As the rather sneery description of “mums” above does. Straight from a Reddit caricature?

You can’t teach kids - or anyone - to be able to critique and analyse writing and information if you just mindlessly reproduce it. It isn’t actually communicating with you. It’s a big predictive text engine! The amusement value fades quite quickly after a while.

Way back in the day I used to share a flat with some software guys who worked for a company that generated “sexy” SMS messages for lonely men who texted a number for “local girl available to chat” kind of thing. It was all generated by software: there were no actual girls — or humans — on the other end. AI is just like a newer, fancier version of that. Doesn’t that make you feel a bit sad? It isn’t a therapist or a funny guy or a mind.

My daughter had an excellent vocabulary. I was told she must read excessively. She didn't but she did access information through a really wide variety of medium. It's really old fashioned to think the only way to build vocabulary through specifically reading. Personally I thiinnk most kids do better through combination methods, eg put subtitles on telly, listen to the words as well as read

Listening to a text is entirely viable method of identifying quotes and deciding what quotes to use. Listening to a text is fine for identifying themes and topics. But current assessment in english literature often requires you to memorise your quotes (no copy), so the kind of text familiarity you're thinking off isnt particularly prioritised either.

It bothers me the not quoting a paragraph correctly though. I have so many questions, why did you want it to quote a whole paragraph, what was the text? Is it a widely available text or a modem novel

🤣🤣🤣🤣You do realise one of the sources of information scraping chat gtp trained on was mumsnet

It always curls me up when it reads like a mumsnet post.

I wanted the whole paragraph as a test of reliability. It's software, therefore could have access to a database of its' own training data, but it seems that it doesn't. It just "knows" an amalgamation of actual data, which is why the results are often a hotchpotch of words that look like valid information on the surface.

That's also my main problem with these "AI" agents. The results they produce are superficially attractive and addictive to people, but if you dig beneath the surface you'll see that it mostly gives you garbage, which can be quite harmful.

It’s a tool. No-one promised an Oracle. A tool is only good when you learn how to use it properly.

Most of these examples just have me shaking my head at the user.

No it never had its own training database, it

But that's exactly the kind point we should be teaching.

If you want an exact replication of a text you should always go to the original.

If you want to language to look for the original you ask chat gtp.

If you put millions of people in a room and asked them what plant this is and how best not to kill it, the people who don't know will shut up and the ardent gardeners will likely give you a pretty useful answer.

If you asked a million people to recite Shakespeare they'll all give it a pretty good go but none of them will be right. If you ask them to interpret a paragraph you'll get a million answers with the same variation in quality. If you asked for an interpretation aiming at gcse level 8-9, most people will go quiet and you'll get those with experience talking about the question

If you asked a million traveller's in a room how they would organise travel to a specific location you'll get some fantastic ideas, but some might have visited 10 years ago, they may not know the gallery closed last week.

If you put a million therapists in a room and asked them to go through something. You can do it immediately he information you provide is update your more likely to remember that key detail that makes a difference. It rephrases allowing you to analysis it. Personally I find it does it with less spin than a human. You are essentially, if you get it right, counselling yourself. May jot be great bug if access to mental health care is limited or expensive it's better than n nothing

Chat gtp is just a language model and an incredible tool. But you absolutely need to understand HOW it works to make it work for you.

If you asked a million people to recite Shakespeare they'll all give it a pretty good go but none of them will be right.

If you asked me to recite Shakespeare I would tell you that I don't know enough of Shakespeare to recite any. If you ask me to recite any paragraph from a book and I don't know it by heart or am not sure, I would tell you so.

Now, I realize that some people will lie and make shit up, but not everyone would. You would find enough people who would at least tell you they aren't sure.

No AI agent is going to tell you this. Or if one exists, I would love to know which one. It hallucinates still more than 50% of the time, but will give you a made up answer with a 100% confidence without saying that it's likely not accurate.

The BS is only part of the problem, it's the lack of a disclaimer that makes it a complete train-wreck.

If you want an exact replication of a text you should always go to the original.

Again, it doesn't tell you this. It makes up an answer, despite it not knowing it. In the responses there is no mention that the "quote" it gives you is just a sample generated from the text it was trained on (whether that is the original or just opinions on said text).

This is my issue. Most people don't know this and assume that what they get back is accurate and the agent doesn't tell you otherwise. People are being duped.

Chat gtp is just a language model and an incredible tool. But you absolutely need to understand HOW it works to make it work for you.

I do understand how it works, which is why I don’t have much use for it. That’s the entire point.

If you want an itinerary or whatever, why would using ChatGPT be better than just reading some interesting articles about your destination? I enjoy doing that! If I want to write an email, I can write it myself better and in less time than prompting an AI and then editing whatever it comes out with so that it doesn’t sound embarrassingly like an AI. It’s quicker for me just to write it. If I want to know about something, I am perfectly capable of judging which resources are most useful to do that with a few quick (non-AI) searches. And then I also know the source and reliability of the information I’m using.

If I were to use it for work, the problem would be that I’m expected to actually know my material and the field, not just present a report from an AI. People will ask me questions about the material and the field that go well beyond the report, and I’ll need to have a detailed and informed discussion about the material/data/field with other colleagues who are similarly well informed about it. I work with sharp people who can all analyse data, understand the legislative environment and have a forensic knowledge of their fields. Using AI would make me look incompetent or lazy, to be honest!

I’d look an absolute ass and a complete idiot if I presented anyone a written report that sounds like an AI wrote it (and yes, of course people can tell); and then had nothing to say about the sources or underlying data. I’d also be embarrassed to have to use an AI tool to do something that I’m employed specifically to do and to know about myself.

They aren’t remembering anything. You give them all the text and papers required to do the essay.

I lived in out local library from about 8 to 18. There wasn't anything I didn't read. (One reason why I remain sceptical about the Online bollocks act ....).

You didn't ask the room whether it knew Shakespeare!!! You asked it if it could recite a paragraph of Shakespeare!! The room gave it its best shot! 😁

Although tbf more likely to be able to do Shakespeare than a modern novel due to open source internet versions vs modern copyright (where it's on tenuous ground legally if it did reproduce a complete paragraph)

Again, it doesn't tell you this

Did you ask it? "If I want an exact replication of a text am I best asking you or going to the source?" It does tell you you are best going to the primary source and why, it then offers to assist you finding a primary source.

You've got to remember it's a computer. It's limited It's still very straight line thinking amusingly so.

I will also say I've rarely caught mine actually hallucinating. But I've also put in sentences to reduce that likelihood in the personalisation options and run separate chats for anything important. Chat GTP told me how to do this to reduce the likelihood of it happening, because i asked. I have found it can get confused if the instruction isn't clear enough or its trying to incorporate/analyse too much information.

Most people need training how to use it and what it is. Otherwise its like your judging a submarines ability fly, missing the fact it's incredible for watching fish. It's designed to train robots how to talk to Humans, just so happens in order to do that it needs something to talk about and it need humans to talk to it. But it's goal isn't actually to be useful, just so happens it can be.

You didn't ask the room whether it knew

Once again. "AI" doesn't know. It never has. And never will.

Did you ask it?

I asked it for a quote. So yes, I asked and it gave me bullshit. If it doesn't know, why does it not refuse? Why is it unable to recognize a question it can't answer.

Btw, you used to be able to ask ChatGPT if it had access to the text itself. It would say no. Then you would ask for a quote of said text it didn't have access to and it would make one up.

Now, if I had asked it to make up an imaginary quote in the style of the text, fine, but that is not what I asked. I asked for a quote from the text, which it had admitted to not having access to.

You are right, a computer doesn't actually know stuff, it simple collates information or language and offers a conclusion if asked.

Humans of course are entirely different. we.... ummm... collate information using language and come to a conclusion?!?😏... . "i know tower if London opens at 9 o'clock" only because i checked the opening hours on the website and found it said it opened at 9 o'clock". I checked information sources and collated language. I know something

Thing is those key phases, I enjoy reading destination guides and creating an itinerary. I can do this quicker myself i cant write a prompt quickly and then i proof read it to remove ai.

I don't enjoy reading gardening manuals, DIY guides. If I can outsource/chatgtp that research, and i the information i get back works, why wouldn't you? I have a better house and garden despite absolutely no interest in either two subjects.

I have little clue about style, colour theory or putting an outfit together, but can outsource that too and I love my wardrobe better even though I might overrule the choice.

I absolutely can write (or dictate) an ai prompt that Includes my knowledge quicker than a report. But then i don't naturally write anything quickly and proofreading it is absolutely lengthy. i will still miss many an spelling error however long i take. A tutor during my degree taught me to read assignments backwards so my brain didn't automatically "correct" and thereby miss errors. (Yes i was reading in reverse 5000 word essays, can you imagine the time). I find proofreading for meaning really quick. Looking back I have trouble working out the ones that took me an hour or so to write and the ones that took 10 mins with ai. But then the content and the linguistic style is still mostly mine, ai just links the sentences with appropriate fillers.

But my point is this, why does is matter if you can tell ai has been involved in the creation if the content is correct?

Humans of course are entirely different. we.... ummm... collate information using language and come to a conclusion?!?😏... . "i know tower if London opens at 9 o'clock" only because i checked the opening hours on the website and found it said it opened at 9 o'clock". I checked information sources and collated language. I know something

Bottom line is we have no idea how we think. Which means it's impossible for us to create something to do it artificially.

We can makes something that can mimic us*. And one day, maybe we will. But that day was not when "AI" became a viral marketing trend.

*There are philosophical debates around what that actually means.

If it told you it had no access to the text why would you ask it to produce a quote it told you it had no access too?

Then why would you expect it to be right?

If you asked a friend if they knew any keats, they told you they knew no keats and you asked them to quote Keats, would you expect the quote they gave you to be accurate?

It's just chatting to you at that point, which is afterall it's actual purpose.

But that's the point, we attribute concepts to words and so language evolves. If conceptually what a computer does is similar, even if the underlying mechanisms are different, we might use the same words. I don't think it's unreasonable to say "knows" in relation ai and a set of data if conceptually it conveys the desired meaning.

We're straying into that philosophical morass ...

Can an entity capable of perfectly mimicking another, be said to possess the same qualities as that entity.

So if a machine is capable of seeming intelligent, can it said to be intelligent ?

When it happens, I guess we'll find out.

I quite like philosophical morass, especially related to computers!!

Chat gtp can't do that. Not well anyway

This thread has got very... deep.

Yes I could read travel blogs. I don't want to. I want it to make me an itinerary in minutes, feed me the key info I want, without me trawling through endless blogs, it saves me time.

If it's wrong I correct it. Yesterday it said blueberry had 3 b's apparently but I've just asked it and it said 2.

It's not perfect. It's AI! It's a computer. But it's a helpful tool, often fun, a good time saver.

And it got me my job. It improved my CV no end, I wrote my application letter in bullet points (how my brain works best) and it put it into a coherent letter, same with my application questions. I ask it to make my emails more formal/more conversational etc as needed.

Co-pilot does about 50% of my work at the moment 😂. I love it!

Mainly we have some very old systems for data, and copilot saves me loads of time analysing spreadsheets and collating info.

AI is really only as good as the info you give it. I use it to reword emails, reports, summarise massive documents. And find specific information I need that used to mean trawling different provider websites.

Can you define intelligent? I'm not sure all humans meet that quality.

But then are we talking quality or linguistic convention designed to convey meaning?

In Latin, compute mean to calculate or settle something. A human would do it. Now we use computer to apply to a machine that computes. Might not do it the same way as a human but it's conceptually similar enough for the language to stick.

Oh and I once spoke at it (literally, with the voice feature) for 15 mins about the absolute shambles that was DS's time at school, the failures of the SEN department, the lack of appropriate support, the way his EHCP wasn't implemented, etc. It was completely incoherent and all out of order, I just kept saying 'oh and another thing' and just rambled and rambled. And it produced me a truly extraordinary letter of complaint for the governors that was 10 million times better than anything I could ever have done myself.

It needed a few edits - remove this, change that - but it was unbelievable how brilliant it was. For someone like me, time poor, overstressed and poor mental health at the time I couldn't ask for more.

Lot of assumptions here.

I did not expect it to be right. I was testing it's capabilities because people were praising it to the sky. I wanted to see if it lived up to the hype and if it was capable of "replacing us all in 6 months to a year", because that's what was being bandied about a lot. I found the experience and its' capabilities underwhelming.

Secondly, if a friend told me they did not know much about a subject and then told me a complete whopper on the same subject, I would stop trusting that friend, even if I (accidentally) asked them a question on a subject they had earlier admitted not knowing much about. It's like the boy who cried wolf. Why would you keep trusting someone or something so unreliable? Yet, here we are. There are users like yourself who acknowledge it's unreliable yet still advocate for its' use. Why?

What's the purpose of "chatting" to an entity you can't trust?

That's not even talking about the waste of water and energy for each and every prompt. It's about as useful as an extremely wasteful toy.