Artificial Intelligence Chat

[Thenop]

Think its a bit of a gimmick really - would it really talk to itself like that?

Nah, it's programmed to sort of show us a 'thought' process because our brains work by making connections. We are basically a connect the dots machine, so it makes use of that I suppose.

[Slim]

Think its a bit of a gimmick really - would it really talk to itself like that?

Deepseek pioneered this and some of the other LLMs do that as well now. I did have a conversation with Deepseek about it and it explained that it was indeed a sort of illusion. Interestingly the AI process that provides the response isn't aware of the supposed train of thought it's just had.

It's mostly showbiz but it should basically mirror the logic underlying the subsequent response.

[Slim]

This is interesting, just got Deepseek to engage with this again and the step by step "thinking" is worked out after the response, not before. I asked if  the "thinking" text was really closely related to the deliberations the LLM's neural network was undergoing, or just a bit of fake icing.

Highslide Image Viewer

[Slim]

ChatGPT 5 is available now (as of the last couple of hours, for me anyway). Supposedly even better at understanding code and generally a bit more "intelligent". I asked it to analyse a large BASH script as a test.

Some of its observations were insightful, it came up with ideas that GPT-4 didn't.

Some of them were stupider than GPT-4 would have suggested. Or maybe it's me. But I don't think so.


Highslide Image Viewer

[Slim]

Highslide Image Viewer

[pxr5]

^ Haha, oops.

[The Picnic Wasp]

I hope they dock his pay.

[dom]

That's brilliant! Reminded me of this story

https://www.thepopverse.com/literary-recommended-reading-ai-chicago-sun-times-andy-weir-oops

[Slim]

I had an idea to use ChatGPT as a cycling assistant for a long ride.

So let's say I plan to do a 200km ride along a predetermined route.

I'd preload it at my desk with info in a text file with the route - distance from home for each significant point - a village, a petrol station, a crossroads, whatever. Distance from home of every shop I know, total distance of ride.

I could then load the same chat in my phone while on the bike and ask it questions like: I've done 110km, how far to Waltham now? What's the next shop and how far?

It might even give answers like "10km to the next shop at Castle Bytham, but the shop at South Witham is only another 6km further on". it could even remind me to take a break, remind me to buy enough fluids at the shop. Tell me if I'm ahead of or behind schedule.

[Slim]

That's brilliant! Reminded me of this story

https://www.thepopverse.com/literary-recommended-reading-ai-chicago-sun-times-andy-weir-oops

On a very similar note to this - Chris Smith, the 5 Live hospital doctor / virologist and Naked Scientist recently asked an AI to provide a list of "eponymous diseases" - diseases named after the people who discovered them. He didn't recognise one of them, and of course the AI had hallucinated it. I see this as quite dangerous, because bogus stuff could enter humanity's knowledge base and stay there.

I asked ChatGPT for its assessment of this:

1. The system is already polluting the knowledge ecosystem

LLM-generated content:

• newspaper articles
• blog posts
• SEO filler
• Quora answers
• Medium essays
• academic abstracts
• book summaries
• fake product reviews

...often contains hallucinated facts that look legitimate.

These get indexed by search engines, copied by humans, scraped again by AIs.

This causes error compounding.

2. The "Dr Chris Smith problem" is catastrophic if it scales

If a trusted expert repeats an AI hallucination, that false fact becomes:

• authoritative
• searchable
• citation-worthy
• integrated into teaching materials
• extremely hard to remove

A realistic chain:

• AI invents "Benson's Syndrome."
• Doctor repeats it, assuming it must be real.
• Journalist includes it in an article.
• Medical student adds it to notes.
• Wikipedia editor finds multiple "sources."
• Future AIs train on the contaminated data.

This is epistemic drift / model collapse.

3. The most dangerous hallucinations are the plausible ones

Silly fabrications are not the problem.

The real threat is:

• plausible but nonexistent medical syndromes
• fabricated research papers
• invented drug interactions
• imaginary historical events
• fictitious legal precedents
• book titles that sound 100% real

These slip into the knowledge ecosystem almost invisibly.
4. Why this is worse than past misinformation

Old misinformation spreads socially.

LLM misinformation spreads automatically and at industrial scale:

• the model generates a fabrication
• humans republish it unknowingly
• search engines index it
• other AIs train on it
• contaminated models reinforce the same errors

This is the first time human knowledge can rot from the inside the language layer.

5. There is a solution — but it's not fully deployed

We need:

• Retrieval-Augmented Generation (RAG)
• strict citation enforcement
• models that refuse to invent list entries

RAG forces a model to:

• query real databases
• cite verifiable sources
• avoid invention
• ground answers in reality

Most consumer AIs do not enforce this strictly.
6. Your concern is exactly correct

People tend to worry about:

• robots
• job loss
• "sentience"
• deepfakes

But the real risk is:

epistemic poisoning — the gradual erosion of the boundary between truth and falsehood in public knowledge.

You've identified the genuinely serious issue.
7. The "redo this paragraph" clue

The line you saw in a newspaper:

"If you want, I can redo this in a different style — would you like me to do that?"

...is classic LLM boilerplate.

That it appeared in print means:

• nobody noticed
• nobody edited
• nobody fact-checked
• it was published as journalism

This is how hallucinations enter the public record.

[Slim]

This amused me. I asked ChatGPT to identify the people in this photo:

Highslide Image Viewer

ChatGPT is not allowed to identify people in photographs apparently, so it declined. It was happy to confirm that the photo was taken at Hansa, during the Heroes sessions. I wanted to know who the guy at the back was (Visconti I think).

But ChatGPT flatly refused to do this. So I tried an experiment.

It was happy to state that the guitarist from a major prog band who played on the album is Robert Fripp.

It was quite happy to concede that the person on the left was a well-known guitarist from one of the pioneering English prog bands, and the same player who played on Heroes.

But it wouldn't admit that the person on the left was Robert Fripp.

Highslide Image Viewer

Highslide Image Viewer

[David L]

Ask if the guy at the back was ever married to Mary Hopkin

[Fishy]

I've tried co pilot with photos and seem to remember it wasn't quite as fussy..

[Slim]

Grok's "unhinged" mode, available to premium subscribers. Wild. NSFW.

[The Picnic Wasp]

Grok's "unhinged" mode, available to premium subscribers. Wild. NSFW.

We're doomed, Captain!