The Lemon Fool

servodude wrote:

BullDog wrote:I have asked ChatGPT some technical questions I would ask of an engineer in an interview. The answers are fine. They're learned from articles on the internet. I think there's a world of difference recognising patterns of words and giving a copied text book response and actually being intelligent.

It also understandably makes the same mistakes that are common among the average of the stuff on the internet
- or it did when I gave it a couple of interview questions I like to use to weed out the bluffers
Interesting it could eventually be convinced of its error and would remain trained for the rest of the session (unlike your average bluffer who tends to die on the hill)

Yes, in an interview with me, the engineer would be asked to give me examples of where he came across various issues related to his answer. To probe breadth and depth of a subject. That's the bit where ChatGPT is a lot less convincing in it's answers. Context, nuance, lateral thinking in problem solving, it's not there, yet.

I am convinced there's a role for AI in monitoring and optimising the performance of industrial processes. After all, most of the decisions an operator or technician makes is based on recognition of patterns or outcomes from the past. But we're still a long way off rolling it out for real.

bungeejumper wrote:

BullDog wrote:I have asked ChatGPT some technical questions I would ask of an engineer in an interview. The answers are fine. They're learned from articles on the internet. I think there's a world of difference recognising patterns of words and giving a copied text book response and actually being intelligent.

I understand that chatbots also have a habit of making stuff up! Which would seem to underline the need for caution WRT their outpourings, especially in legal or reference contexts. And especially not in science.

I'm not sufficiently well up on this topic to have a worthwhile level of understanding, but I could imagine that these bots are designed to look for "bridges" whenever they find two quoted facts that don't quite meet in the middle. If we, as humans, found such a gap, we might be tempted to theorise about such a "bridge". Indeed, you couldn't read an Agatha Christie without doing exactly that!

It's called "hallucination" and you can find a good example at viewtopic.php?p=552760#p552760 and the posts that follow, where I asked ChatGPT what's the fastest flying mammal and it came back with the peregrine falcon and, after a lot of further grilling, invented references to a (non existent) paper in a (real) scientific journal to bolster its claim that the peregrine falcon is both a bird and a mammal!

There's an interesting article explaining the problem at: https://arstechnica.com/information-technology/2023/04/why-ai-chatbots-are-the-ultimate-bs-machines-and-how-people-hope-to-fix-them/

The important thing to understand about Large Language Models like ChatGPT is that actually all they are is a humongous set of statistical data on associations between words. As the article says, "The only thing they know how to do is to pick the next best word based on statistical probability against their training set", and, referring to Bing Chat, "It tries to produce the highest probability continuation of the string using all the data at its disposal".

Mike4 wrote:

csearle wrote: I still feel they should identify themselves as AIs though rather than masquerade as humans.

Here, you put your finger on my unease. I feel there is a deceit happening when a bot posts masquerading as a human, and it could and should be called out.

If a bot identifies itself as a bot when posting, I see no problem. For now anyway!

The EU has just (yesterday) passed an act (comes into force in May)

https://therecord.media/eu-parliament-p ... regulation

The article doesn’t say so but I have heard that AI sytems (chatbots, artbots & the like) will have to identify themselves and their output as such. Oh, but wait, we’re not in the EU.

GrahamPlatt wrote:The article doesn’t say so but I have heard that AI sytems (chatbots, artbots & the like) will have to identify themselves and their output as such.

I hope they've carefully considered exactly where such rules apply. In a context like lemonfool it would make reasonable sense - though perhaps the site owners might deserve a say over allowing a bot that doesn't comply. But in an academic research context it could be unduly onerous.

GrahamPlatt wrote:The EU has just (yesterday) passed an act (comes into force in May)

https://therecord.media/eu-parliament-p ... regulation

The article doesn’t say so but I have heard that AI sytems (chatbots, artbots & the like) will have to identify themselves and their output as such. Oh, but wait, we’re not in the EU.

That's going to be a tough one to enforce.

"Stop getting at me, Ursula, I'm not a bot. You must be thinking of my identical twin, who I haven't seen in twenty years. Same name as me, but that's your problem, not mine. And by the way, I'm a US citizen, and I've got an expensive lawyer, and Elon Musk is my cousin, and he says you can't restrict my constitutional right to say what I want, where I want, in privacy.

Go ahead, make my day. And don't complain if I re-register in Tehran or Moscow or Pyongyang, or somewhere else where you can't do much about me. Love, Z."

BJ

GrahamPlatt wrote:

Mike4 wrote:

Here, you put your finger on my unease. I feel there is a deceit happening when a bot posts masquerading as a human, and it could and should be called out.

If a bot identifies itself as a bot when posting, I see no problem. For now anyway!

The EU has just (yesterday) passed an act (comes into force in May)

https://therecord.media/eu-parliament-p ... regulation

The article doesn’t say so but I have heard that AI sytems (chatbots, artbots & the like) will have to identify themselves and their output as such. Oh, but wait, we’re not in the EU.

We will have a chance to enact it properly then. C.