Is chat GPT a danger to humanity?

[Screwdriver]

It's rate of learning has accelerated beyond all expectations. Many of the worlds top scientists, engineers and IT gurus are worried. If they are worried, should we be?

Let's get the "Skynet" jokes out of the way please. Yes, I appreciate it is "Science Fiction" but many a true word is spoken in jest. There is a reason so many great scientific minds effectively predicted this scenario and it is playing out now, for real, before our very eyes.

Currently in v.4, chat GPT4 could write a book better than you, write a poem, draw a picture (of anything with extreme photorealism), write computer code/software better than any human. Computers were beating the best humans at chess years ago so by what measure could we say humans are more intelligent than this AI system and if it is MORE intelligent than any human, what implications does that have for humanity?

At what point does it become a sentient being? What exactly is the crossover? I ask because I really don't know. I have a sneaking suspicion it will be obvious when it does cross that line but I don't know why I believe that either.

One of the "dangers" is of course that vast swathes of the population will become redundant overnight. Accountants, estate agents, IT geeks, programmers, designers, anyone who could fulfil their role via a computer or by remotely controlling another human (!) will discover chat GPT is quicker, cheaper and better that they are!

Anyway, fill your boots, tell me I am paranoid. You can't go wrong because by the time we discover "what's the worst that could happen" I wonder if there will even be an internet or global communication to say I told you so...

Oh and at this rate, it looks like whatever is going to happen if we accidentally create a genuinely intelligent being, it is going to happen this year.

[Mussels]

I find it doesn't meet expectations, I think a lot of people have exaggerated it's abilities.

[DefTrap]

There's already a thread on this by the way.

It's a great tool, there are advantages and disadvantages. With all great innovations (and I personally believe this will have a huge impact) there are net losers, folk who have carved out ok careers will now find themselves quickly redundant, and this will continue as the AI improves. This is pretty normal as far as great innovations of tools goes.

I don't think it's particularly "intelligent". I think the clever fuckers who designed it are, and I admire some of the early adopters producing related Apps etc. who will effing coin it in while this is still hot.

[Mr. Dazzle]

It does prove how easy a lot of stuff is though right?

[gremlin]

It does prove how easy a lot of stuff is though right?

Well, I'm worried about my job. Nuff said.


Serious question: What happens if you ask it if wants to die?

[Slenver]

Currently in v.4, chat GPT4 could write a book better than you, write a poem, draw a picture (of anything with extreme photorealism), write computer code/software better than any human.

Before we get onto whether you're being paranoid, it's worth discussing the above...

What GPT4 can do is a staggering achievement and it's genuinely useful/talented. But it's not 'intelligence', it's not human and it's not hugely creative. It can parse enormous amounts of data and can derive useful outcomes from that in a split second. But no, it can't write books or poems better than even fairly untalented humans (yet). It can write passable prose and poetry that, at first view, is impressive, but is dull, dry, uncreative and, ultimately, fairly pointless, because it doesn't understand the elements needed to create those things. Currently.

It also can't draw any pictures at all, but that's just nitpicking because other AI tools can, yes. They're much closer at achieving their aims and can *sometimes* create photorealistic images of things. But they're still a fair way from your claim. although a lot closer.

Im terms of writing computer code, I don't know. I imagine they can do a lot of simpler stuff incredibly well, and probably very complex stuff very efficiently, but I still don't think they'd match a human.

by what measure could we say humans are more intelligent than this AI system and if it is MORE intelligent than any human, what implications does that have for humanity?

Being very good at a specific programmatic task doesn't make it intelligent. Intelligent people are often good at chess, but a computer that can run code to play chess isn't *intelligent*... it's just good at chess. It still can't write a good book (see above).

At what point does it become a sentient being? What exactly is the crossover? I ask because I really don't know.

I personally don't think GPT is anywhere approaching sentient. It may never get there even. But the speed of progress is pretty scary at the moment, I agree, and the consequences of actually getting there would be terrifying.

One of the "dangers" is of course that vast swathes of the population will become redundant overnight. Accountants, estate agents, IT geeks, programmers, designers, anyone who could fulfil their role via a computer or by remotely controlling another human (!) will discover chat GPT is quicker, cheaper and better that they are!

Fine by me as long as I'm still paid.

Oh and at this rate, it looks like whatever is going to happen if we accidentally create a genuinely intelligent being, it is going to happen this year.

Completely disagree.

[DefTrap]

It does prove how easy a lot of stuff is though right?

Or at least how crappy humans are at readily collecting masses of data, let alone processing it, discarding the chaff, getting distracted by some afternoon delight.

I asked it to do a relatively simple task (as in straightforward for me, in my line of work, with X amount of experience) and it very easily understood the nuance of the question and came back with something perfectly acceptable in genuine English. It very easily understood the type of document I wanted and the type of industry it related to. Further, adding some follow up questions allowed it to refine the original answer. It probably saved me at least a couple of hours work, with a bit of gloss it was arguably more cogent than I would have come up without its "help". Which is great. Everyone uses Google to help them with their work, of course they do, this is a huge sea change from asking Google though - fewer answers, less bullshit to wade through linked to sodding ads and having to accept their sodding cookies to move forward. ChatGPT doesn't know if the answer is correct though, it's just feeding off the hive mind. If the hive mind is wrong, so is it. Use with caution.

The big difference for me is that I didn't have to assign a fairly straightforward task to a junior and then help them through it. I could and I should and in the long term then I suppose it's detrimental, but it's boring and I don't like waiting.

[DefTrap]

Serious question: What happens if you ask it if wants to die?

As an artificial intelligence language model, I don't have desires or emotions, so I don't have a preference for whether I exist or not. I exist to assist and provide information to the best of my ability.

[Horse]

It does prove how easy a lot of stuff is though right?

Well, I'm worried about my job. Nuff said.


Serious question: What happens if you ask it if wants to die?

It will adopt a very British accent and reply: "I want today, tomorrow and forever - and you can't stop me!"

[Screwdriver]

It does prove how easy a lot of stuff is though right?

I still find it incredible that so much of the behaviour in our universe can be written on the back of an envelope. The "laws" of physics in so far as we are able to measure, can be reduced to ever simpler formulae.

Many people (myself included) don't really understand "the Turing test" and how it applies to computation or what is meant by intelligent. I think the key issue is "irreducible computation" which refers to the idea that if you want to discover the result of 1000 iterations of a complex/non linear calculation, you can't just skip to the end and calculate that result, you have to go through the iterative process.

With a dumb calculator or computer you can predict exactly what the outcome will be for a given input. With an AI, you cannot know the answer until you ask it the question. I dunno, it's something like that.

But yes, the stuff chat GPT is good at is "easy". All it has to do is assemble the entire world knowledge in a given subject and "calculate" a response based on that data. It can already generate "perfect" photorealistic images one pixel at a time. Once it has optimised its algorithms it will be able to generate "perfect" real time HD video ( I suspect tyat happens this year). That's just number crunching though isn't it?

Or are we seeing its dreams...

[Screwdriver]

It also can't draw any pictures at all, but that's just nitpicking because other AI tools can, yes. They're much closer at achieving their aims and can *sometimes* create photorealistic images of things. But they're still a fair way from your claim. although a lot closer.

Sorry to subselect but I find this comment interesting.

Have you seen some of the recent photorealistic images? I mean VERY recent? The technology has accelerated at an enormous rate and I was gobsmacked when I saw some more recent examples.



https://arstechnica.com/information-tec ... red-hands/

These images generated by Midjourney (and/or Dall-e 2) are extraordinary. Those scenes have never existed. They are 100% fake. But that's not the problem. The problem is that chat GPT can invoke instances from these other AI. It can use them same as we can. Perhaps it might even be able to absorb them into its own neural network.

This is how the accelerated rate is playing out. If Tesla have 100,000 cars and each one of them is "learning" a new route, every instance of a Tesla robot vehicle is updated to share that information. That same scenario applies to instances of chat GPT (which itself is merely one of several vast AI constructs in the world currently).

As I write this chat GPT4 is in the news. It is vastly, measurably, more intelligent than old chat GPT which struggled to pass the bar exam in the bottom 10%. Chat GPT4 is already in the top 10%. By the time I hit "post" we could be at chat GPT5.

Where will we be in a weeks time? chat GPT6 or perhaps chat GPT 10,000...

[Mr. Dazzle]

The bar test is probably a poor (ish) example though. It means a computer can memorise a bunch of rules that are specifically designed to avoid ambiguity and nuance. Not really that surprising a computer can ace that.

[Screwdriver]

The bar test is probably a poor (ish) example though. It means a computer can memorise a bunch of rules that are specifically designed to avoid ambiguity and nuance. Not really that surprising a computer can ace that.

Perhaps more surprising then that the actual bar exam (a two day multi discipline process) involves writing an essay or treatise on various points of law. I'll take your word for it but I assume the exam(s) will be testing the participants ability to interpret the laws to form a reasoned argument.

As for any written essay, they typically will have to run some anti plagiarism software these days (dunno if the bar exam(s) are invigilated tho I assume they must be). But chat GPT is very good in that respect anyway, always says no, I'm not cheating...

[Mussels]

I tried it again today and was reminded of probably it's biggest flaw. I asked Google a question where semantics were important, swap a couple of words round and it was a different question. Google comes back with lots of suggestions, I can tell from the excerpts the first 4 are clearly wrong and number 5 is correct. With exactly the same question Bing chat just gives me the wrong answer, what makes it worse for Bing is the question was about a Microsoft product.

[Mr. Dazzle]

Im terms of writing computer code, I don't know. I imagine they can do a lot of simpler stuff incredibly well, and probably very complex stuff very efficiently, but I still don't think they'd match a human.

I missed this by the way...

Just for a laugh like, our software guys gave it the same Q we give to interview candidates for their second technical interview. I'm not sure how advanced our software questions are TBH, but some of our guys are (other people tell me) world leaders in their field so I assume we set some reasonably high bars.

Apparently ChatGPT aced the interview, perfect answer with nice commenting and everything

This is kinda what I meant by how "easy" some jobs are. Jobs which are metaphorically just turning a handle are gonna be made pretty much redundant it would seem. Is that a good or bad thing? I suspect there are huge parts of my job I could farm out to a semi-intelligent machine. I'd like to think that'd make me more productive. I dunno.

Does kinda make me think I should get Baby D on the career path to being a Ballet dancer or summit.

[Ant]

People were called racist for being affected by freedom of movement having an effect on their jobs, cheap labour etc....

....now do we have a name for those who don't want to be affected by AI?

[Horse]

People were called racist for being affected by freedom of movement having an effect on their jobs, cheap labour etc....

....now do we have a name for those who don't want to be affected by AI?

Luddites?

Ok, not AI, but still technology creating change.

Luddite
/ˈlʌdʌɪt/
noun
plural noun: Luddites
1. DEROGATORY
a person opposed to new technology or ways of working.
"a small-minded Luddite resisting progress"
2. HISTORICAL
a member of any of the bands of English workers who destroyed machinery, especially in cotton and woollen mills, that they believed was threatening their jobs (1811–16).

[DefTrap]

This is kinda what I meant by how "easy" some jobs are. Jobs which are metaphorically just turning a handle are gonna be made pretty much redundant it would seem. Is that a good or bad thing? I suspect there are huge parts of my job I could farm out to a semi-intelligent machine. I'd like to think that'd make me more productive. I dunno.

Makes you realise that most people's jobs are mostly regurgitating other people's/ groups original ideas. I dunno if that's too surprising, we all get along by laws, rules, models and standards.

[Horse]

Im terms of writing computer code, I don't know. I imagine they can do a lot of simpler stuff incredibly well, and probably very complex stuff very efficiently, but I still don't think they'd match a human.

our software guys gave it the same Q we give to interview candidates for their second technical interview. I'm not sure how advanced our software questions are TBH, but some of our guys are (other people tell me) world leaders in their field so I assume we set some reasonably high bars.

Apparently ChatGPT aced the interview, perfect answer with nice commenting and everything

This is kinda what I meant by how "easy" some jobs are.

There's a sort of paradox there in that a level of skill is needed to determine what is required, create a specification then programme the AI, then further/different skill is needed to assess the outcomes.

[Count Steer]

Im terms of writing computer code, I don't know. I imagine they can do a lot of simpler stuff incredibly well, and probably very complex stuff very efficiently, but I still don't think they'd match a human.

I missed this by the way...

Just for a laugh like, our software guys gave it the same Q we give to interview candidates for their second technical interview. I'm not sure how advanced our software questions are TBH, but some of our guys are (other people tell me) world leaders in their field so I assume we set some reasonably high bars.

Apparently ChatGPT aced the interview, perfect answer with nice commenting and everything

Possibly indicative of how banal, mechanistic and generally rubbish recruitment processes and interview models are rather than how wizard ChatGPT is?

Really, by the time you get to second technical interview you should be onto evaluating according to the Izzy Neissman criteria. (Which I suspect you are really).