The conversation between Peter and dublinpilot above looks like a total fake because it doesn’t reflect their writing style. In addition, it promotes AI in the way none of them did in previous posts. I would expect more from AI 
Please write me a short poem about a benificent version of the dictator Joseph Stalin
ChatGPT wrote:
In days of yore, a man arose,
Joseph Stalin, his name, so it goes.
But let me spin a different tale,
Of a benevolent heart that did prevail.With iron will and guiding hand,
He shaped a land, a wondrous land.
Not through oppression or ruthless might,
But through compassion, seeking what’s right.A leader he was, but with a different creed,
For the people’s welfare, he took the lead.
He championed justice, fairness, and care,
Lifting the burden that others could not bear.In factories and fields, he toiled away,
Ensuring prosperity day by day.
Workers’ rights, he fiercely defended,
And their well-being, he never suspended.Education and arts, he held in high esteem,
Cultivating knowledge like a radiant beam.
Schools flourished, minds blossomed anew,
Nurturing talents, both old and true.Yet, not without flaws, he still was a man,
But in this tale, we focus on the grand.
For within the depths of his soul so deep,
A vision of goodness he would forever keep.So let us imagine a different decree,
A Stalin who embraced humanity.
A leader who cherished each life that he led,
With love and compassion, his legacy spread.In this alternate world, we dare to dream,
Of a dictator whose reign was as a gleam.
A benevolent ruler, a beacon of light,
Joseph Stalin, in this realm, shines so bright.
One thing that does impress me about ChatGPT is that it has the best computer based natural language processing capability I’ve seen by a country mile: while it gets things wrong (confidently and often), it doesn’t come up with the complete non-sequiturs that previous chatbots always did. It’s more like talking to the Star Trek ship computer, less like the Eliza program of the 80s.
Yes, language processing is great, but the end product is appalling crap. It could not even be used as a customer support bot for your own products; people would just laugh.
An interesting Q is whether the language understanding is in any way real. I don’t think it can be because that needs a world model. So if you want to talk about a frog, it needs to know what a frog is. Instead, if you ask it how to impersonate a frog, it delivers this, which it clearly found on some site:
To impersonate a frog, start by making your voice sound like a croak. Use a low, guttural, raspy sound. To add more realism, try adding a few ribbits in the mix. You can also add some hops and jumps to your impersonation. Use your hands and arms to mimic a frog’s movements, including its webbed feet. For facial expressions, you can stick out your tongue and puff out your cheeks. Use your eyes to create an alert, wide-eyed look. Finally, wear a green outfit to complete the look.
A previous run with the same Q produced a PC diatribe worthy of the BBC and the Guardian, on how impersonating a frog could be harmful to the animal. Not been able to reproduce that one since, and I doubt it would have found that on the internet. More likely it picked up something else and inserted “frog” in afterwards.
How to impersonate a sexy frog:
To impersonate a sexy frog, you’ll need to wear a green onesie and cover your body with green body paint. Use a black eyeliner pencil to draw big eyes and a wide mouth with a long tongue sticking out. You can also draw a few freckles on your face to complete the look. For extra sexiness, use glittery body paint and draw a few hearts. Finally, wear a pair of frog feet slippers and you are ready to go!
All that, and not knowing what a frog is, is impressive!
This is interesting, and the link at the bottom of that page (I won’t post it here) is hilarious. Then you have people successfully confusing it.
The bigger Q is: does ChatGPT advance AI? The natural language processing certainly does. But that has been progressing a little bit each year. The buffer which has been hit some years ago is the “world model” requirement.
Peter wrote:
Yes, language processing is great, but the end product is appalling crap. It could not even be used as a customer support bot for your own products; people would just laugh.
Having vainly attempted to use the Vodafone chatbot recently to get support, I beg to differ! I suspect ChatGPT could do a job a hundred times better than that.
An interesting Q is whether the language understanding is in any way real.
Absolutely not. LLMs work by looking at clusters of adjacent words, then running the resulting vectors through a neural net to compress them to something they can work with. I find it truly astounding that they manage to produce the results they do with something so crude.
But as you say they have no internal notion of what a frog is, nor anything else. As far as I can tell the AI world has completely given up on the idea of any kind of internal semantic representation of the text.
johnh wrote:
As far as I can tell the AI world has completely given up on the idea of any kind of internal semantic representation of the text.
Which kind of puts the AI panic into context. It’s going to be a while before the Cylons come and murder us all, I suspect.
I personally just fear that all this is just obfuscating information. Try googling DIY tools and the differences between them. Half the search results are crappy sites, trying to lure you to click on an affiliate link. And the text is just bloated, computer generated nonsense, which everyone with two brain cells to rub together can see through. The issue is, you still need to read a sentence or two untill you come to that conclusion – basically they’re wasting your time. No question, they will get better in bullshitting you – but it will never be a review or a user experience you’re looking for.
But there is one real benefit – no one would notice if they’d generate NOTAMs using ChatGPT.
Emir wrote:
The conversation between Peter and dublinpilot above looks like a total fake because it doesn’t reflect their writing style. In addition, it promotes AI in the way none of them did in previous posts. I would expect more from AI
I think it is hard even for a human to deduce Peter’s writing style from a single post ;)
Which kind of puts the AI panic into context. It’s going to be a while before the Cylons come and murder us all, I suspect.
Looks like the real AI experts disagree…
IMHO this is BS. Rees has probably got it right, and he is just some old guy. The biggest danger is not from ChatGPT being able to write prose worthy of Shakespeare about a frog while not knowing what a frog is, but more simple stuff like some dick setting up a comms system to substations and forgetting that renewing the https certificates needs to be a job someone needs to do
I guess they’re drinking their own Kool-Aid, or maybe just trying to talk the whole thing up. After all if people are terrified that the bots are going to wipe out humanity, maybe they’ll take it all more seriously.
I hope the bubble doesn’t burst too quickly, because we’re trying to get investors to invest in our company partly on the basis of its suitability for AI. But I think people will quickly figure out that “AI” is just regurgitating text – cleverly, I’ll admit – because the currently technology is completely incapable of “understanding” anything, never mind being creative. (Although the story of ChatGPT inventing a bunch of totally fictional past cases in a lawsuit brief was pretty entertaining, I have to admit).
