ChatGPT

[Morkin]

Hmmm.....

[patters]

NSFW - contains lots of swearing. I tend to agree with Adam on this.

[Einar Saukas]

I tend to agree with Adam on this.

This video is spot on. Except the problem goes beyond the inability to distinguish false information. Attempting to merge accurate texts still produces lots of garbage when you don't understand their meaning.

I tested ChatGPT at work, at the request of an executive manager. The produced texts were typically fairly convincing but full of errors. The problem is, you can only identify the errors if you already knew the correct answer. However if you already know the answer, why do you need ChatGPT? Fixing errors and rewriting the text (to address exactly the points you want) takes more time than simply writing the entire text yourself. And if you don't already know the answer, the last thing you need is a tool that gives you the wrong answer in a convincing way.

I didn't try using it to generate code, but I'm sure it would be even worse.

[Ralf]

I didn't try using it to generate code, but I'm sure it would be even worse.

It works better with popular languages like C#, Java or SQL and fails totally with things like Z80 assembler or Zx Spectrum Basic.

The problem with Chat GPT is that it never admits it doesn't know something. If it doesn't know something, it will make it up.

It will invent unexisting commands in the code, will describe unexisting events and places in details and even will give you unexisting links for that.
If you correct it, it will agree with you, say sorry but will keep talking rubbish. It will be all expressed in nice, grammatically correct sentences but it will be still rubbish, like a crap packed in a pretty, shining paper.

[AndyC]

It helps to remember that all it is really doing is picking the next most likely word, like constructing sentences by just clicking whatever autocorrect suggests. It has a massive dataset of probabilities and it's some pretty clever algorithms behind it all, but there is no knowledge, just random word generation.

That it works so well might say a lot more about people than about AI.

[patters]

The current buzz is all about 'generative AI' but it can only ever be derivative as Adam Conover explained in that last video. As Einar said, you have to be competent in the subject field to spot its errors, so AI will never truly replace competence. It will for example produce code by consensus (via its training data), which will replicate common errors and misconceptions.

There's another video in which a coder rates GitHub CoPilot, and in his own chosen examples it makes error after error, and he spots them because he's a career software dev:



But put this stuff in the hands of inexperienced young noobs and we'll have planes crashing and cars running people over within weeks. And guess what, inexperienced young noobs with AI copilots are a lot cheaper than seasoned professional developers.

[JMcGibbitts]

I can't really look past the folk who have pushed and funded AI for private sector, corporate purposes. Elon Musk tasked Open AI as being an answer to Demis Hassabis' Deepmind (because - Google), but Demis' approach seems more measured, starting with games, then science applications. It's a completely different thing to give AI tasks that would take humans decades to find results, like curing diseases.

Open AI's investment-led agenda means all bets are off. Some plausible AGI might still seem a way off, regardless of this "exponential growth", but it's certainly possible to train bespoke AIs to do very specific tasks well. With the right plugins, AI will have an effect as it can learn when it tries to do stuff that doesn't work. Give a dedicated Z80 AI coder a code editor, emulator, assembler, list of opcodes, the plugins to be able to assemble, execute, test the output vs projections and reiterate... It's going to do a better job than ChatGPT.

I think Governments should mandate companies run all AI server compute off-grid, generating electricity, minding emissions - everything they need without cutting into public demand for general resources. That should see growth that's more sustainable and should focus minds on what's most important.

[equinox]

I tested ChatGPT at work, at the request of an executive manager. The produced texts were typically fairly convincing but full of errors. The problem is, you can only identify the errors if you already knew the correct answer. However if you already know the answer, why do you need ChatGPT? Fixing errors and rewriting the text (to address exactly the points you want) takes more time than simply writing the entire text yourself. And if you don't already know the answer, the last thing you need is a tool that gives you the wrong answer in a convincing way.

You're talking about work stuff rather than the Spectrum BASIC attempts, but I think it's bad at Speccy BASIC because that's rather under-documented in internet terms. It should, and does, perform much better at (say) JavaScript, even though JS is much more complex than Spectrum BASIC, because it has seen way more of it.

It's still just a clever trick, though. It's possible to get something frequently right by chance, but you can do that by just repeatedly flipping a coin and saying "heads", you know? Creativity minus understanding is fun (fractals!) but should not be relied upon to drive anything useful.

[Nienn Heskil]

But put this stuff in the hands of inexperienced young noobs and we'll have planes crashing and cars running people over within weeks. And guess what, inexperienced young noobs with AI copilots are a lot cheaper than seasoned professional developers.

Let's face it, modern IT is pretentious sh*t run by idiots. They leave bugs the size of Red Square in mainstream processors that remain since the 90s. They think returning an 'int' from a 'byte on byte' operation is reasonable thinking. Hell, they are unable to fix a single messenger app so that it doesn't crash/hang at least 5 times a day.

As long as the critical areas are left to the few that actually know what they're doing and not just a bunch of buzzwords, I see no difference whatsoever for most applications.

[equinox]

Let's face it, modern IT is pretentious sh*t run by idiots. They leave bugs the size of Red Square in mainstream processors that remain since the 90s. They think returning an 'int' from a 'byte on byte' operation is reasonable thinking. Hell, they are unable to fix a single messenger app so that it doesn't crash/hang at least 5 times a day.

As long as the critical areas are left to the few that actually know what they're doing and not just a bunch of buzzwords, I see no difference whatsoever for most applications.

get off my lawn

[Nienn Heskil]

get off my lawn

pls no cri

[Joefish]

One thing to remember is that ChatGPT is based on the contents of the internet from BEFORE ChatGPT existed. The minute it starts churning through its own output plastered all over the internet and using that as input is the moment it starts spiralling into nonsense. There are plenty of example of this where one Chabot talks to another and they go completely mad. But then you can't do that with ChatGPT at the moment because it never actually asks questions; it only does answers.

Now there are already 'protective' / 'malicious' algorithms designed to modify artworks in ways imperceptible to people, but that throw any AI into chaos if it tries to 'learn' from or mimic elements of them. Just like those T-Shirt prints from a few years back that were designed to jam up facial recognition cameras. And remember there are whole profitable industries based around trying to 'game' Google's website search and ranking algorithms. So it won't be long before there are algorithms that target these 'A.I.s' by seeding your prose with specifically crafted phrases and word patterns that trigger them to give priority to all sorts of dubious ideas and ideologies, or just prompt them to spout utter nonsense.

As many commentators point out, it's a cleverly constructed adaptive algorithm, but there's nothing approaching an aware 'intelligence' in it. Calling it an 'A.I.' is like calling something that still has wheels on a 'Hoverboard'.

[stupidget]

One scary thing I discovered at the weekend is the general acceptance by people in their 20’s that everything they get from a ChatBot is correct! An example is that my nieces boyfriend, who works in IT as a support engineer said, and I quote:

‘If there’s something I don’t understand or need to find something out, I’ll just ask ChatGPT. It’s better than google as it gives a far more detailed description.’

He now takes EVERYTHING he gathers from a ChatBot as 100% correct.

Having worked in cybersecurity for the past 6 years and about to start working for an external cybersecurity company that focusses on Deep/dark web analytics I find this pretty scary. How soon before these bots star spewing out nonsense based on their own ‘learning’?

[SteveSmith]

How long before he makes a mistake and has to make the excuse to his superiors "ChatGPT told me to do it".

[equinox]

One scary thing I discovered at the weekend is the general acceptance by people in their 20’s that everything they get from a ChatBot is correct!
‘If there’s something I don’t understand or need to find something out, I’ll just ask ChatGPT. It’s better than google as it gives a far more detailed description.’

The artist formerly known as Wikipedia.
You purely aren't allowed to cite it in papers, because your teacher just hates it for reasons unknown. (Tell 'em: Wikipedia is like an academic paper, but a really bad one. It still has sources/references at the bottom, and you can always use those if you can find them.)

See also: "video removed due to copyright", they think copyright is some kind of weird devil like the Terminator.

Whether people feel a certain way about a certain source (hell, Wikipedia won't ever cite the Daily Mail, they feel it doesn't count as a newspaper, unlike thousands of plebs) -- doesn't really matter. What you're supposed to get from school is some kind of "critical media literacy", i.e. the ability to read and evaluate a text without just going with the herd of sheep. We seem to have lost that.

And NEVER trust anyone else to read and evaluate a text for you. NOT EVEN ONCE.

[stupidget]

How long before he makes a mistake and has to make the excuse to his superiors "ChatGPT told me to do it".

He said that that is actually what he’d say if he did something wrong!!!! He was serious

Looks like the days of trawling through O’Reilly books to ‘learn’ something are well and truly over!!

[equinox]

How long before he makes a mistake and has to make the excuse to his superiors "ChatGPT told me to do it".

AI war crime trials

About 10-15 years ago I remember posting a comment on The Register to the effect that the entire story and headline (some kind of financial hacker nonsense) read like a bad cyberpunk novel, and wouldn't have made sense in the real world even 15 years ago. Hello AI, hello.

[equinox]

about to start working for an external cybersecurity company that focusses on Deep/dark web analytics

do not hire this man, he talks about his job

[stupidget]

do not hire this man, he talks about his job

Like it’s so hard to find Pii these days

[AndyC]

The artist formerly known as Wikipedia.
You purely aren't allowed to cite it in papers, because your teacher just hates it for reasons unknown. (Tell 'em: Wikipedia is like an academic paper, but a really bad one. It still has sources/references at the bottom, and you can always use those if you can find them.)

See also: "video removed due to copyright", they think copyright is some kind of weird devil like the Terminator.

Whether people feel a certain way about a certain source (hell, Wikipedia won't ever cite the Daily Mail, they feel it doesn't count as a newspaper, unlike thousands of plebs) -- doesn't really matter. What you're supposed to get from school is some kind of "critical media literacy", i.e. the ability to read and evaluate a text without just going with the herd of sheep. We seem to have lost that.

And NEVER trust anyone else to read and evaluate a text for you. NOT EVEN ONCE.

Ah, citations on Wikipedia:

[equinox]

Ah, citations on Wikipedia: [xkcd comic]

I mostly hate xkcd (it tends to have the smug "I'm better than you because I memorised some f?ck?n vim launch parameters"), but sometimes he is absolutely on point. For example in this "cyclical Wikipedia talking sh*t about itself case" I've seen it actually happen. It's funny in the comic, less funny outside. When you spot the mistake, see if you can remove it without five teenage kids putting it back with Twinkle because SOMETHING WAS REMOVED. Yeah it was removed because it was wrong, genius. I had another word but I'm not allowed to say it here.

EDIT: reminds me, a local "news" site had a list (yes, listicles) of places with a rooftop terrace, and they included a little pub in my town that definitely doesn't have one, never did and never will, but was for some reason ranked in search engines under that term. Clearly this author had done some bad googling and "phoned in" the entire thing. Jesus. Imagine if your job is going out to booze & eat food and you still stay at home. (I wrote to the site and told them, but they wanted to keep their TOP 10 or whatever, so they just edited to say "it might not have a rooftop terrace BUT..." bullsh*t.)

kisses

[Joefish]

The worst thing I find about Wikipedia is any explanation of any mathematical principle gets edited over and over by people who think they know better until eventually it gets to the point where the article itself is so arcanely and esoterically phrased that no-one anywhere, however knowledgeable, can sufficiently understand the piece to make any more edits. By which time it's long passed any point where it may have been readable or remotely useful to the layperson.

But I do like XKCD. Although I can see it always carries the risk that you might be called upon to learn something new to get the joke.

[equinox]

The worst thing I find about Wikipedia is any explanation of any mathematical principle gets edited over and over by people who think they know better until eventually it gets to the point where the article itself is so arcanely and esoterically phrased that no-one anywhere, however knowledgeable, can sufficiently understand the piece to make any more edits. By which time it's long passed any point where it may have been readable or remotely useful to the layperson.

But I do like XKCD. Although I can see it always carries the risk that you might be called upon to learn something new to get the joke.

I assume we can blame this on the white man, and colonialism, etc. but I love this Indian dude who has gone all around the Internet trying to promote his alternative to trigonometry, and calling himself a genius. Sorry, I know I'm helping him to promote, but it's just too good. Errm, yeah, if you actually try to USE the formula, it doesn't work in most cases, but that doesn't stop a lot of Indian nationalists saying "look, we invented a cool thing".

https://artofproblemsolving.com/wiki/in ... 0functions.

All we have to do is strongly prioritise formal logic and falsifiability in education. But people would scream OOHHH NO IT'S THE "NEW MATH" OR EBONICS. Idiots. If we could teach people how to follow and prove and disprove an argument, instead of long division, we would have terraformed Mars in 1982. With a Spectrum. ON TOPIC.

[Joefish]

Partly agree, but I also think we'd need to do some sums to land on Mars, and division might constitute some of them. Nothing wrong with starting with first principles, even if you do have a free calculator watch to hand.

[EdToo]

https://arstechnica.com/tech-policy/202 ... an-ex-con/

Looks like ChatGPT doesn't just make up dodgy code examples.