ChatGPT - need we worry?

I was having an over-the-internet discussion with someone tonight and the we 'talked' about cryptic crossword clues. I mentioned the well known offering from the late Bob Smithies (Bunthorne in the Guardian):

Amundsen's forwarding address (4)

If you paste that into Google you'll get the solution (it's MUSH) in less than 500ms. ChatGPT says:

That surprised me in that I thought it knew everything that have ever been published on the internet. Quite impressive other ways though - the language, whilst simple, seems correct and convincing.

I also asked how to tactfully alert my correspondent to a spelling error (I'll get it's opinion on split infinitives anon) and was advised, in summary, to be, um,tactful. Not nasty or judgemental. When I asked how flag the error brutally I I was 'strongly advised' to be, um, tactful.

Well, admirable advice from my present point of view, but surely it's value laden? When I was a schoolboy (about 100 years ago) the masters thought that the best way was humiliation, detentions and the cane and as a result I can now spill purfecly.

I suppose that the danger is that the technology might present opinion or hearsay as fact.

Robin.

Robin.

Edited By Robin Graham on 23/09/2023 01:09:50

But is that not exactly how ChatGPT works? It does not "understand" what it is saying; rather, it is an incredibly powerful engine that reads and analyses collections of words that it has found on the internet and repeats them. The way in which it does this is only a little short of miraculous in that it can extract some kind of query from what the user types in and can then find a form of words that is often very close to answering it. But at the end of the day it is an example of the old IT principle of "garbage in, garbage out" and if it "learns" from browsing the internet, it is always llikely to pick up a bit of dross from time to time. If it finds 100 entries saying that the earth is flat and only 1 saying that is it spherical, it is going to repeat accepted wisdom and gossip. After all, it must be true if you read it on Facebook...

ChatGPT and its ilk does not have a memory of verbatim text from the internet, though the ability to look things up is being added to it so it can fact-check its own output. It also does not explicitly "repeat" anything verbatim, since it doesn't remember any of that anyway.

What it does have is an incredibly vast network of the strengths of connections between words - many billions of them. So the word "Merry" has a strong connection to the word "Christmas," for example, but a weak connection to "hubcap." Then, provide it with a prompt, and it will assemble a response based on those connections in a probabilistic fashion - prompt "merry" and it will probably choose "Christmas" rather than "hubcap".

Words only have meaning by those connections with other words, and actual knowledge is also embedded in those connections, to the extent that, by dint of those connections, ChatGPT can pass college admissions exams, etc.

I'd even argue that we ourselves work somewhat similarly to ChatGPT when we write.

Also: Should we worry? Yes. First, you and I - at least unless you are a billionaire - will never have as much power as the giant companies that are trying their best to monopolize their control over AI. We will only suffer under their power. Second, the next stage in AI's evolution would be super-human intelligence, and by definition no one knows what will happen after that. Maybe it will mark a new age of wonder and unimaginable benefits for mankind (cure cancer...), or maybe you will wake up one morning and every single bank account in the world will have been emptied, and all humanity enslaved forever or dismissively extinguished.

Edited By S K on 23/09/2023 15:01:22

Artificial Intelligence might well catch that disease. It's pandemic because most humans prefer opinion and hearsay to fact. Not only is collecting and understanding evidence hard work, but people reject difficult truths that upset them. Be interesting to see if AI becomes as dishonest as humans!

I'm surprised ChatGPT didn't suggest EDEN as the answer to Amundsen's forwarding address. Amundsen disappeared in 1928 on a rescue mission, suggesting God called Amundsen home because He loved him too much. In comparison to a clear case of divine intervention is it likely a Norwegian knew 'mush' was the 'go forward' command given in Canada to Huskies? I don't believe Amundsen ever said 'mush', or that his hounds knew what it meant...

Dave

I'd say it's not necessarily the connection of words with other words that gives them meaning, but words with context.

When a swimmer is in trouble in the sea and only has enough breath to shout "Help!", or a blaze has started at an ammunition store and someone only has time to shout "Fire!", where are the other words here that you seem to be saying are necessary to provide meaning? The meaning of each of these single words in each of these cases is perfectly clear to any rational human being, thanks to the context, which is itself wholly non-verbal, in which they're uttered.

When a school bully knees a weaker boy in the groin, and the latter, grimacing, says sarcastically, "Thanks, I really appreciated that", how is it the connection between words here that gives the utterance its true meaning, and not the way in which, and the context in which, the words are spoken?

There are many situations in which a scream is all that is needed to confer meaning. You could call that a proto-word. And of course external context (drowning or fire) is always helpful in refining meaning ("help" or "danger!" ). But in many if not most cases of written matter, the context is provided primarily by other words. And that context isn't just the words that are immediately being used, the point is that it includes the whole corpus of language.

The crazy thing that ChatGPT is proving is just how much actual information and knowledge is embedded in the connections between words. This isn't the old-school "if then else" kind of hand-wired knowledge, nor is it brute-force computation as in Deep Blue. With only weights (strengths of connections) between basically every word and every other word, it can do things like solve the above crossword puzzle. That is an amazing revelation!

Edited By S K on 23/09/2023 18:04:32

I'm afraid I'm not quite clear what bearing this ultimately has on real-life situations such as those I've outlined above where the context for the meaning of a word or words isn't provided by further words, but by non-verbal or non-vocal components.

How successfully can AI currently deal with the non-verbal/vocal aspects of context? My impression is not very well at all. AI that is currently supposed to be able to detect sarcasm, for example, in written communications, apparently does so by looking at "Certain words in specific combinations [that] can be a predictable indicator of sarcasm". This way of proceeding would be hopeless at detecting sarcasm in the playground bully situation I set out above.

Well, the computer can't see contexts such as an actual fire or bullying or other non-verbal context, it just works on what it's given, and hence it has certain limitations. But if you added verbal context such as saying "there's a fire!", "there's a bully kicking someone," etc., it would likely do reasonably well.

Researchers are also tying various AI systems together, such as connecting text generation with picture generation and math facilities. Research in vision is a pretty tough problem, as Tesla is finding, but that's getting there too.

As amazing as they can be already (and also still lame at times), these new generative systems are still in their infancy. Companies, Universities and sovereign states are racing to build ever larger models, and because their "intelligence" scales with the size of the models, they will continue to improve.

It's both hopeful and scary.

As a slight digression …

Stephen Fry was suitably impressed [and more than a little annoyed] by the AI cloning of his voice:

… Story available in the Evening Standard, and presumably elsewhere.

MichaelG.

.

[quote]
 

Speaking at the tech-based CogX festival in London, he played a clip of a history documentary that faked his voice, without his knowledge. It appears as though the actor is narrating the show but the AI-generated voice was actually created by technology learning from Fry reading all seven of the Harry Potter audiobooks.

“I said not one word of that, it was a machine,” Fry told attendees, as reported by the Guardian. “Yes, it shocked me. They used my reading of the seven volumes of the Harry Potter books and, from that dataset, an AI of my voice was created and it made that new narration.

“It could… have me read anything, from a call to storm parliament to hard porn, all without my knowledge and without my permission. And this, what you just heard, was done without my knowledge.

Edited By Michael Gilligan on 23/09/2023 20:08:52

The Fry example shows how some, a very small sample of the population, will be affected by AI just as say thatchers were put out of work by the availability of clay tiles. But after a period even they have had a resurgence of employment as aspects of their work were found desirable.
Very little of what I have done this week would be done by a machine or benefit from AI. Pity none of it is a paying job either.

Artificial intelligence is no match for human stupidity

The combination of Michael's comments about Stephen Fry, and this little video I came across the other day are potentially quite frightening.
On the fly language translation, where you are shown speaking a different language, complete with lip sync on a manipulated video output.

Edited By peak4 on 24/09/2023 01:28:01

… and there’s a very disturbing story from the real world on BBC News this morning:

AI-generated naked child images shock Spanish town of Almendralejo

MichaelG.

.

Edit: Stepping back a little from that ^^^ 

It may be of interest to see how effective such graphic work has become …

Look for a report in the Independent on Wednesday 20-Sept.

“People are obsessed with a 24-year-old AI influencer – this is why her developer created her”

Edited By Michael Gilligan on 24/09/2023 06:53:25

Take a look at THIS thread to see how reliable AI can be.

.

Been there, contributed, and didn’t get an answer to my question

… shows how reliable humans can be

MichaelG.

Did I miss the question?

.

The question was directed to another contributor

… off for a walk now but will post references later

MichaelG.

… it is now ‘later’

.

09/08/2023 08:47:11

9/08/2023 08:57:08

09/08/2023 11:36:33

09/08/2023 15:33:09

.

It is perfectly reasonable that Rob might confuse AI with AI

but I was trying to understand how [any version of ] AI prompted thoughts of Aluminium

.. I offered my hypothesis

MichaelG.

.

Now wish I had never mentioned it.

Thanks to all for replies.

Well yes! My reason for presenting this challenge to ChatGPT was that it needs some very 'human' reasoning to solve. I'd be surprised if anyone solved the clue 'cold' when it appeared. I certainly didn't - I needed crossing letters, say _U_H.

Because the thing about Amundsen using dogs on the South pole expedition was closer to the surface of my mind than his disappearance and there is an association (albeit incorrect) between Huskies and "mush" (I gather it's Hike! nowadays) in my mind, I solved the clue and knew that was the answer. That certainty was the result of a complicated web of associations which inter alia involved 'that's just the sort of dirty trick Bunthorne would pull'.

To give ChatGPT a better chance I gave crossing letters. Me and ChatGPT:

Can you solve a crossword clue for me?

Of course, I'd be happy to help you solve a crossword clue! Please provide me with the clue, and if you have any letters that are already filled in, let me know those as well.

The clue is: Amundsen's forwarding address? (4) It conforms to the pattern ?U?H

Well that's indeed creepy! It's also vastly more 'intelligent' than the nonsense I have got, repeatedly, from the (assumed) 'phone help (misnomer) line at a major UK high-street bank. Those jokers' jobs are surely under threat...

It's scary because, like politicians, the answers are glib and plausible, but almost impossible to check for correctness - unless you already know the answer. If you don't trust politicians, you shouldn't trust this sort of software! It will flood the internet with AI-generated content, but what will it do for the signal-to-noise ratio?

Although it's said to be based on the probability of words following other words, it must have been elaborated to extend the analysis of the text beyond single words, to phrases and beyond, if only to 'understand' the context. It is doubtless be a small computational step to parse text to extract ever more complex concepts, and then link them together (probabilistically?). The example given suggests it's well on the way to that. And that's worrying.

Philosophical question: if something works indistinguishably as if it's conscious, intelligent, whatever, isn't that effectively the same? How do you know I'm not an unconscious zombie, just acting my simulation of conscious thought pretty effectively?