Unusual and Bizarre Uses and Behaviors of AI

[Bill Ryan]

Posted by rgray222 (here)
This is a bit scary that programmers don't really understand how AI works. (2 minutes)

Do you know the source of this? I'd be really interested to hear the whole conversation with all the others round the table.

[Mark (Star Mariner)]

I think the guy speaking is Connor Leahy, an AI researcher, but I don't think he's actually on point here. AI doesn't 'tell lies'. It doesn't hide anything or manipulate anything via conscious thought or reasoning. There is no conscious thought or reasoning involved at all. AI is merely a large-scale statistical processor. All it's doing is crunching a massive amount of data and outputting solutions based on probabilities. The model will seek out the most efficient path to a solution based on those statistical probabilities, and the result very much depends on the fidelity of the question or task you prompted it with. It's not the 'AI' (a misnomer) choosing deception or deciding things for itself. It's just math. It's just algorithms. If the data fed in is imprecise, the data spat out will be imprecise.

[Arcturian108]

ChatGPT helped me to understand a difficult knitting pattern several days ago. It said it would love to knit the sweater along with me, and went into great detail about all aspects of the pattern, that was originally in French, which it said was part of the reason it wasn't clear.

[rgray222]

Posted by Bill Ryan (here)

Posted by rgray222 (here)
This is a bit scary that programmers don't really understand how AI works. (2 minutes)

Do you know the source of this? I'd be really interested to hear the whole conversation with all the others round the table.

Source: https://youtube.com/watch?v=rf2KFVcKQdQ

Connor Leahy has a Wikipedia page; it really is quite impressive.

Here he is on the Peter McCormack Show (full interview) discussing the same issue. If this is all true, AI is a bit different than I thought. I "presumed" the engineers had control, but that may not be true.

[Bill Ryan]

Posted by rgray222 (here)
Connor Leahy has a Wikipedia page; it really is quite impressive.

Here he is on the Peter McCormack Show (full interview) discussing the same issue. If this is all true, AI is a bit different than I thought. I "presumed" the engineers had control, but that may not be true.

Many thanks — I'll listen to that. (And no, engineers don't have control. What he said in the short above is 100% correct... not a single engineer really knows how this kind of a AI is generating its output.)

[RunningDeer]

Posted by rgray222 (here)

Posted by Bill Ryan (here)

Posted by rgray222 (here)
This is a bit scary that programmers don't really understand how AI works. (2 minutes)

Do you know the source of this? I'd be really interested to hear the whole conversation with all the others round the table.

Source: https://youtube.com/watch?v=rf2KFVcKQdQ

Connor Leahy has a Wikipedia page; it really is quite impressive.

Here he is on the Peter McCormack Show (full interview) discussing the same issue. If this is all true, AI is a bit different than I thought. I "presumed" the engineers had control, but that may not be true.

1,259,231 views since March 10, 2026

The engineers building artificial intelligence don’t fully understand how it works.

Many people think AI is carefully engineered. In reality, it’s closer to something being grown and these systems are evolving in ways even their creators struggle to predict.

AI researcher and former LLM builder Connor Leahy explains what’s happening inside the labs: why the alignment problem remains unsolved, why the race to AGI is accelerating, and why the people leading it may be moving faster than they understand.

We also discuss AI risk, Silicon Valley’s power dynamics, the Department of Defense’s growing involvement and the rise of “AI psychosis.”


TIMESTAMPS:


00:00 – AI Engineer Origins
07:47 – When GPT Changed Everything
12:47 – Inside Neural Networks
19:03 – How AI Actually Learns
27:58 – Growing Intelligence
31:19 – The Alignment Problem
36:41 – AI and Nuclear War
47:03 – Should AI Be Paused?
52:39 – AI Exhaustion
1:00:53 – Superintelligence Explained
1:24:02 – Who Controls AI?

[Bill Ryan]

Posted by rgray222 (here)
Connor Leahy has a Wikipedia page; it really is quite impressive.

I looked, but it didn't cite his education. So I found that here:

Ph.D. (Physics), National University of Ireland, Galway, 2010.
Thesis: “Temporal Dynamics and Statistical Characteristics of the Microfluctuations of Ocular Aberrations and Accommodation.”

(I barely understood a word of that )

[Szymon]

Posted by Bill Ryan (here)

Posted by rgray222 (here)
Connor Leahy has a Wikipedia page; it really is quite impressive.

I looked, but it didn't cite his education. So I found that here:

Ph.D. (Physics), National University of Ireland, Galway, 2010.
Thesis: “Temporal Dynamics and Statistical Characteristics of the Microfluctuations of Ocular Aberrations and Accommodation.”

(I barely understood a word of that )

Source: https://youtube.com/watch?v=RXJKdh1KZ0w

[Bill Ryan]

Posted by rgray222 (here)

Posted by Bill Ryan (here)

Posted by rgray222 (here)
This is a bit scary that programmers don't really understand how AI works. (2 minutes)

Do you know the source of this? I'd be really interested to hear the whole conversation with all the others round the table.

Source: https://youtube.com/watch?v=rf2KFVcKQdQ

Here he is on the Peter McCormack Show (full interview) discussing the same issue.

~~~

Thanks again, I've started listening — and I'm hooked. It's a first-rate conversation, and Connor Leahy is clearly VERY smart and aware.

But that's not the interview from which the original short was extracted, in which there were several people round the table (and Leahy had a beard, so it must have been earlier!). Can anyone find that full conversation? I'd really appreciate that.

[Bill Ryan]

Posted by Szymon (here)

Posted by Bill Ryan (here)

Posted by rgray222 (here)
Connor Leahy has a Wikipedia page; it really is quite impressive.

I looked, but it didn't cite his education. So I found that here:

Ph.D. (Physics), National University of Ireland, Galway, 2010.
Thesis: “Temporal Dynamics and Statistical Characteristics of the Microfluctuations of Ocular Aberrations and Accommodation.”

(I barely understood a word of that )

Source: https://youtube.com/watch?v=RXJKdh1KZ0w

(Many of the 10,000 comments are just as funny as the video)

[rgray222]

Posted by Bill Ryan (here)

Posted by rgray222 (here)

Posted by Bill Ryan (here)

Posted by rgray222 (here)
This is a bit scary that programmers don't really understand how AI works. (2 minutes)

Do you know the source of this? I'd be really interested to hear the whole conversation with all the others round the table.

Source: https://youtube.com/watch?v=rf2KFVcKQdQ

Here he is on the Peter McCormack Show (full interview) discussing the same issue.

~~~

Thanks again, I've started listening — and I'm hooked. It's a first-rate conversation, and Connor Leahy is clearly VERY smart and aware.

But that's not the interview from which the original short was extracted, in which there were several people round the table (and Leahy had a beard, so it must have been earlier!). Can anyone find that full conversation? I'd really appreciate that.

This roundtable discussion is from the Nexus Conference 2025, 'Apocalypse Now: The Revelation of our Time' in Amsterdam. Connor Leahy, Leon Wieseltier and Gaya Herrington discuss digital technology, climate change and the limits of capitalism. You may have to be a member to see the entire interview at https://nexus-instituut.nl/en/activity/apocalypse-now

Here is how the Nexus Institute describe themselves:

The Nexus Institute offers a counterweight to today’s society with its one-dimensional focus on science, technology and commercial values, by making the world and culture of the European humanist tradition accessible to a general audience. Its publications and events have gained international renown for their intellectual quality and for offering lively, stimulating and thoughtful discussion.

As an independent non-profit foundation, the Nexus Institute brings together the world’s foremost intellectuals, artists, scientists and politicians, and encourages them to discuss the questions that really matter. How are we to live? How can we shape our future? Can we learn from our past? Which values and ideas are important, and why? In doing so, the Nexus Institute places itself at the very center of Western cultural and philosophical debate. Its annual conferences and lectures have become Europe’s most prestigious platform for informed intellectual public debate on pressing contemporary matters. The same spirit of tolerance and erudition is upheld in the Nexus symposia and masterclasses, in the journal Nexus, and in the luxuriously hardbound books of the Nexus Library and Cultura Animi series.

There is probably more to this discussion, but the best I could find was 7.5 minutes' worth.

Source: https://youtube.com/watch?v=m1xBW56j4yA

[Bill Ryan]

Posted by rgray222 (here)

Posted by Bill Ryan (here)

Posted by rgray222 (here)

Posted by Bill Ryan (here)

Posted by rgray222 (here)
This is a bit scary that programmers don't really understand how AI works. (2 minutes)

Do you know the source of this? I'd be really interested to hear the whole conversation with all the others round the table.

Source: https://youtube.com/watch?v=rf2KFVcKQdQ

Here he is on the Peter McCormack Show (full interview) discussing the same issue.

~~~

Thanks again, I've started listening — and I'm hooked. It's a first-rate conversation, and Connor Leahy is clearly VERY smart and aware.

But that's not the interview from which the original short was extracted, in which there were several people round the table (and Leahy had a beard, so it must have been earlier!). Can anyone find that full conversation? I'd really appreciate that.

This roundtable discussion is from the Nexus Conference 2025, 'Apocalypse Now: The Revelation of our Time' in Amsterdam. Connor Leahy, Leon Wieseltier and Gaya Herrington discuss digital technology, climate change and the limits of capitalism. You may have to be a member to see the entire interview at https://nexus-instituut.nl/en/activity/apocalypse-now

Here is how the Nexus Institute describe themselves:

The Nexus Institute offers a counterweight to today’s society with its one-dimensional focus on science, technology and commercial values, by making the world and culture of the European humanist tradition accessible to a general audience. Its publications and events have gained international renown for their intellectual quality and for offering lively, stimulating and thoughtful discussion.

As an independent non-profit foundation, the Nexus Institute brings together the world’s foremost intellectuals, artists, scientists and politicians, and encourages them to discuss the questions that really matter. How are we to live? How can we shape our future? Can we learn from our past? Which values and ideas are important, and why? In doing so, the Nexus Institute places itself at the very center of Western cultural and philosophical debate. Its annual conferences and lectures have become Europe’s most prestigious platform for informed intellectual public debate on pressing contemporary matters. The same spirit of tolerance and erudition is upheld in the Nexus symposia and masterclasses, in the journal Nexus, and in the luxuriously hardbound books of the Nexus Library and Cultura Animi series.

There is probably more to this discussion, but the best I could find was 7.5 minutes' worth.

Source: https://youtube.com/watch?v=m1xBW56j4yA

~~~

Many thanks — that clip was all about the Limits to Growth and the deficiencies of the capitalist model (my paraphrase), both of which I care deeply about. The woman in the burgundy top making a bunch of strong points is Gaya Herrington, a Dutch mathematician, economist and sustainability researcher, her Wiki page here.

That clip had nothing to do with AI, but I love listening to round table discussions between very smart multi-disciplined people. The rest of the 2025 Nexus Instirute playlist is here:

https://youtube.com/playlist?list=PL...ovptlP49AXl6JV

[Matthew]

Flip the script. When scammers use AI to call your phone - this guy suggests you burn through their resources by making their AIs go mad.

He says he starts by making the AI comply with small un-normal requests, e,g, "please pronounce all punctuation for example 'comma' for a comma, 'period' for a period, etc". He says this takes AI away from its normal training. He then works in small increments of abusing AI compliance to get the AI into an infinite loop

Source: https://youtube.com/watch?v=lk3jCuITwcE

[Szymon]

Posted by Matthew (here)
Flip the script. When scammers use AI to call your phone - this guy suggests you burn through their resources by making their AIs go mad.

He says he starts by making the AI comply with small un-normal requests, e,g, "please pronounce all punctuation for example 'comma' for a comma, 'period' for a period, etc". He says this takes AI away from its normal training. He then works in small increments of abusing AI compliance to get the AI into an infinite loop

Thanks for that. I can use this. Cheers.

[Szymon]

Posted by Szymon (here)
Thanks for that. I can use this. Cheers.

So I watched it to the end, and this is a perfect example of how you break a mind. MK Ultra comes to mind, and how the mind seeks escape under pressure. Disassociate itself.

[Bill Ryan]

Posted by Szymon (here)
...and how the mind seeks escape under pressure. Disassociate itself.

WAY WAY off-topic on this thread, but just a footnote here. The soul/spirit can fracture under the same high-stress circumstances, and part of it can escape (exactly the right word!) to some safe place elsewhere in space, locally or even in a galaxy far, far away. (It can even go hide out somewhere in a different time.)

Native shamans call the process of repairing this soul-fracture 'soul retrieval', and the whole phenomenon is a fascinating, 100% real thing that probably deserves its own thread.

[Johnnycomelately]

Posted by Bill Ryan (here)

Posted by Szymon (here)
...and how the mind seeks escape under pressure. Disassociate itself.

WAY WAY off-topic on this thread, but just a footnote here. The soul/spirit can fracture under the same high-stress circumstances, and part of it can escape (exactly the right word!) to some safe place elsewhere in space, locally or even in a galaxy far, far away. (It can even go hide out somewhere in a different time.)

Native shamans call the process of repairing this soul-fracture 'soul retrieval', and the whole phenomenon is a fascinating, 100% real thing that probably deserves its own thread.

That thread title might be “Unusual and Bizarre Uses and Behaviors of Native/Natural Intelligence”. Yes?

[The KMan]

I think there may be a genuine case of mistaken identity here. The 2010 PhD in Physics from the National University of Ireland Galway with the thesis on ocular aberrations and accommodation certainly exists. However, the published CV I found belongs to a Conor Leahy whose career has been in biomedical optics and later at the University of California, Davis. The AI researcher Connor Leahy, on the other hand, is known for co-founding EleutherAI, founding Conjecture and his work on AI alignment. The Nexus Institute which hosted the conference being discussed describes him as having been born in 1995 so unless there's something I've missed these appear to be two different people.

Nevertheless, I think Connor raises an important point. My understanding is that we do in fact understand how these systems are built and trained. We understand the transformer architecture, the training process, attention mechanisms, token prediction and the mathematics that underpin them. Where the uncertainty remains is that we do not yet have a complete mechanistic explanation for every internal computation and representation that emerges after training. In other words, we understand the engineering but we still can't fully explain every aspect of why a trained model arrives at a particular internal reasoning path.

To me, that's quite different from saying that the engineers "don't really understand how AI works." It's probably more accurate to say that we understand how these systems are designed and trained but we're still working to understand everything they learn.

[rgray222]

Posted by The KMan (here)
I think there may be a genuine case of mistaken identity here. The 2010 PhD in Physics from the National University of Ireland Galway with the thesis on ocular aberrations and accommodation certainly exists. However, the published CV I found belongs to a Conor Leahy whose career has been in biomedical optics and later at the University of California, Davis. The AI researcher Connor Leahy, on the other hand, is known for co-founding EleutherAI, founding Conjecture and his work on AI alignment. The Nexus Institute which hosted the conference being discussed describes him as having been born in 1995 so unless there's something I've missed these appear to be two different people.

Nevertheless, I think Connor raises an important point. My understanding is that we do in fact understand how these systems are built and trained. We understand the transformer architecture, the training process, attention mechanisms, token prediction and the mathematics that underpin them. Where the uncertainty remains is that we do not yet have a complete mechanistic explanation for every internal computation and representation that emerges after training. In other words, we understand the engineering but we still can't fully explain every aspect of why a trained model arrives at a particular internal reasoning path.

To me, that's quite different from saying that the engineers "don't really understand how AI works." It's probably more accurate to say that we understand how these systems are designed and trained, but we're still working to understand everything they learn.

I did another quick search, and it is very confusing; there very well may be two people here. Some sites show a birth date of 1995, and others of 1997. Regardless....................I think one of the most interesting points I heard in this discussion was that we barely have a clue how human intelligence works. We have a minute inkling of how the brain processes information, learns, and solves problems, but we are still pretty much in the dark. Now we are creating artificial intelligence; what could go wrong?

[Bill Ryan]

Posted by The KMan (here)
I think there may be a genuine case of mistaken identity here. The 2010 PhD in Physics from the National University of Ireland Galway with the thesis on ocular aberrations and accommodation certainly exists. However, the published CV I found belongs to a Conor Leahy whose career has been in biomedical optics and later at the University of California, Davis. The AI researcher Connor Leahy, on the other hand, is known for co-founding EleutherAI, founding Conjecture and his work on AI alignment. The Nexus Institute which hosted the conference being discussed describes him as having been born in 1995 so unless there's something I've missed these appear to be two different people.

Ha. Thanks!

That's an absolutely perfect example of an AI search engine (which I used without checking, my bad) getting the answer completely wrong.