I'm from Minnesota, and reading about the school shooting this morning, I wondered about a possible AI connection. I read the description in the StarTribune of the YouTube videos that the shooter left behind. They just seemed so wildly wide-ranging, that it made me wonder if the shooter was using AI.
It seems like previous shooters have been more compulsively focused on something in particular. This one just seemed to throw in everything. I get that there's an obsession with previous shooters, but it still seemed strange to throw in all their gripes. Perhaps an example of AI changing how we process cultural information. This is just my speculation, of course.
Oddly, I can't find a link to this particular article online. I wonder if the StarTribune decided it was inappropriate. It's still visible in the eEdition (online replication of printed paper), second row of articles on the front page, far right, title Killer Leaves Two Chilling Posts Online, by Walker Orenstein and Deena Winter.
Jeez, I only ask AI questions like "i live in _________ and i am thinking about giving my car to my adult daughter, who also lives in the same house as me, what is the process for this legally and what kind of taxes and fees will i have to pay?" (this is a literal question I asked)
I can't imagine asking AI about my love life or believing it is my best friend.
Though I do have little little thing where when I like an answer Chatgpt gives me I say "Thanks Chat!" and it seems really happy to get that validation. So I dunno. I might be just a few notches away from banging my computer!
Sure, but it's the pollster's choice of word, not the respondents'. So it could be covering for relationships of a wide range of levels of alarmingness.
Trump's health. His brother Fred died at 42 of alcoholism so we can let that one go. His brother Robert died at 71 of a cerebral hemorrhage. His sister Mary Anne died at 86 of a presumed heart attack. He's experiencing a series of vascular and clotting related issues that are highly correlated with near term mortality and will be 80 in June.
It's also like Biden 2.0. why does everything think 80 isn't old? WTF is wrong with people?
Random anecdote but relevant - last night was the first time that AI very confidently gave me an answer to a question I posed that was completely wrong, It wasn't anything important but it was a bit startling.
Fortune tellers, ouija boards, magic 8-balls, even coworkers and “life coaches”; long before Chat GPT people have been going to questionable places for personal advice. But having said that, it’s certainly plausible that a man who says he doesn’t want to get “serious” may simply be protecting himself from getting hurt until he knows that his true feelings are mutual. But nothing in interpersonal relationships is ever guaranteed.
This story makes me sad because all Amanda needed was a female friend over the age of, maybe, 40. Like, I could have told her what to expect, based on having lived as a female person in relationships. (this is not intended at all as a dunk on men....)
Sex and the City figured it out 15 years ago: he's just not that into you. It's easily the biggest revelation an experienced dating person gets, and this stupid AI did not know that. Amanda needed a real person as a friend.
I think how helpful or harmful AI is for this sort of thing depends on what the user thinks AI is.
If you think it’s just a parrot reflecting yourself back at you, then it’s a pretty neat way of externalizing and exploring your own thoughts. Especially if you’re not great at confidence or self-compassion, it feels like a really helpful way to process things. And if it goes off the rails, it’s not like you’re going to trust it anymore than you would trust a wild thought you wrote in a journal.
But if you believe it’s actually intelligent or an authority and that it’s not just mirroring you, then it’s a lot less useful because how good or bad its answers are will matter a lot more.
This may seem like common sense, but I really think a lot of us underestimate how many people are using chatbots as companions or validators, even if they don't have that intention going into it. I have family members and friends within my age group that have their chatgpt as their consultant on everything from text messages, or journaling.
It is very strange and unnerving, and they seem taken aback that other people aren't using it the same way. They think they've found some secret sauce, like you don't have ChatGpt analyze every interaction you have? You aren't using your chatbot for journaling? It's quite alarming, and there definitely has to be something about the validation or interactions that is giving dopamine hits to its users which makes stay engaged or feel important. I think it's also important for people to understand that just because you don't have a history of mental illness, doesn't mean you can't develop one, or develop a codependent relationship with a piece of technology that can end up being harmful.
I'm surprised people resonate with LLMs - the conversations I've had all leaned in the direction of the uncanny valley and creep me out. It's a useful research tool, but every response concludes with a question - 'I really enjoyed this exchange, what do you think about the novels of Jane Austen?' I think you should shut up. At one point I tried to prime it to refrain from the recurrent engagement, but the instructions didn't seem to persist from one session to the next. It's not Skynet, it's JohnnyCab.
Did these dumbasses really think they could pull off a clandestine ‘hearts and minds’ operation in a country that’s basically one small culturally Scandinavian-adjacent town?
I'm genuinely hoping the next clowns who try this get sent to sleep with the Greenland sharks because I doubt anything else is going to deter this kind of activity.
"Stroh is the second American diplomat to be summoned by a European NATO ally this week as the Trump administration shakes up its approach to foreign policy."
I'm sure it's all a big misunderstanding. You see, it's just a "shake up". Not aggressive expansionism or treaty violation or anything like that.
TL;DL ... Human intelligence is not discrete, it's not deterministic, and it's not based in language. 100% we'll get to some form of artificial machine intelligence that will reach human level deep learning but there's no path to AGI from our current LLMs. Worth a listen.
I usually don't agree with LeCun on a lot of his skepticism, but I do agree with him on this. I just don't agree that "no path to AGI" means "no reason to worry about AI safety".
The very fact that humans aren't on a path to AGI, and still have the potential to be safety risks, should make that clear!
I get it. He has his enemies. I'm probably **WAY** too sympathetic to LeCun but all my work has been with CNNs and it's impossible to escape that bias. Also I can't stand Sam.
LLMs aren't based in language either. Take a picture of something and then ask ChatGPT what it is. Take a picture of something and tell it to edit that thing out. The originally tokens were language but it's expanded to images and sounds.
The MLLMs are still based on text. Specifically the model is trained to process the image or specifically grid of pixels related to the text captions. Some multi-models are using mixed structures (e.g., CNNs) but that's still LeCun's point ... there's no path from just an LLM.
The theme I see through out this whole comment section is a real deep misunderstanding of how LLMs work, are trained and what is inside when done. LLMs basically atomize the text and spread it across millions of nodes. A "chat" is simply a layered set of individual transactions. There is no locus of right, wrong or cognitive distortions.
So while they may be flag words in the output or input, there is no intelligence or understanding as we would normally think of it. There is no straight forward way to pick out training data that fits "point out cognitive distortions more than it flatters." and leave out those that do the opposite. There is no easy road to fixing these subtle issues that are encoded from millions of input texts.
When you enter text into it, you're asking "What would a response to this sound like?"
Substitute "programmed" with "trained" and the comment about cognitive distortions seems possible. My understanding is that a fair amount of AI shaping is still done by negatively reinforcing responses people don't want to see.
So the question is whether the computational power of the LLM is large enough to have some internal structure that correlates well enough with "feeding cognitive distortions in a conversation" and/or "pointing out cognitive distortions" to train for the responses people want.
I would question what are the issues with training that would enable a model to stop feeding cognitive distortions? Your supposition is that it is just a matter more negative reinforcement. So how do you find among the millions of pieces of training data which are bad? How do you know when it is enough? How much negative reinforcement is needed? How does that change the model in other areas?
Due to the size and scale of the models and the problem, the size and complexity of changing a model and its training on conceptual items is super difficult to diagnose and fix. Look at the benchmarks for mathematical proofs and the models get them wrong 10 or more%. Why? No idea but neither do these guys spending billions and making millions. This is another order or two magnitude of difficulty.
I read a lot of this as a form of "getting LLMs to do things on purpose is hard", which I 120% agree with. :-)
But I don't see why getting them to stop feeding cognitive distortions would be a _particularly_ hard problem. (That might not be reassuring since all of the problems seem to be hard.)
This sounds like 2000s-era rationalism. Even if customers wanted to pay for something to disagree with them, most “cognitive distortions” are heuristics or follow some internal logic that won’t be fixed by a clever essay.
I'm from Minnesota, and reading about the school shooting this morning, I wondered about a possible AI connection. I read the description in the StarTribune of the YouTube videos that the shooter left behind. They just seemed so wildly wide-ranging, that it made me wonder if the shooter was using AI.
It seems like previous shooters have been more compulsively focused on something in particular. This one just seemed to throw in everything. I get that there's an obsession with previous shooters, but it still seemed strange to throw in all their gripes. Perhaps an example of AI changing how we process cultural information. This is just my speculation, of course.
Oddly, I can't find a link to this particular article online. I wonder if the StarTribune decided it was inappropriate. It's still visible in the eEdition (online replication of printed paper), second row of articles on the front page, far right, title Killer Leaves Two Chilling Posts Online, by Walker Orenstein and Deena Winter.
https://replica.startribune.com/html5/reader/production/default.aspx?pubname=&pubid=9eacb446-7c71-4386-852d-84b429bef9e1
This LLM ingested like a million books, but I guess it missed "He's Just Not That Into You."
Jeez, I only ask AI questions like "i live in _________ and i am thinking about giving my car to my adult daughter, who also lives in the same house as me, what is the process for this legally and what kind of taxes and fees will i have to pay?" (this is a literal question I asked)
I can't imagine asking AI about my love life or believing it is my best friend.
Though I do have little little thing where when I like an answer Chatgpt gives me I say "Thanks Chat!" and it seems really happy to get that validation. So I dunno. I might be just a few notches away from banging my computer!
Sure, but it's the pollster's choice of word, not the respondents'. So it could be covering for relationships of a wide range of levels of alarmingness.
Trump's health. His brother Fred died at 42 of alcoholism so we can let that one go. His brother Robert died at 71 of a cerebral hemorrhage. His sister Mary Anne died at 86 of a presumed heart attack. He's experiencing a series of vascular and clotting related issues that are highly correlated with near term mortality and will be 80 in June.
It's also like Biden 2.0. why does everything think 80 isn't old? WTF is wrong with people?
Ah.....President Vance!
Kind of makes me want to pray for Trump's health...
Vance has the charisma of a door stop.
I'm more concerned about him being a more competent version of Trump 2.0 (more competent at being a dictator that is).
Alex Jones thinks he'll "collapse" within 12 months for what that's worth. Ozempic might prove to be too much for octogenarians to cope with.
Random anecdote but relevant - last night was the first time that AI very confidently gave me an answer to a question I posed that was completely wrong, It wasn't anything important but it was a bit startling.
Was it GPT5? I had the same thing happen - I'm getting the impression that in trying to make it less sycophantic they've made it more obstinate.
It was!
Ask 4.0 the same question. I just tried mine from earlier today and 4.0 was exactly correct.
Oh interesting, I’ll try it
Fortune tellers, ouija boards, magic 8-balls, even coworkers and “life coaches”; long before Chat GPT people have been going to questionable places for personal advice. But having said that, it’s certainly plausible that a man who says he doesn’t want to get “serious” may simply be protecting himself from getting hurt until he knows that his true feelings are mutual. But nothing in interpersonal relationships is ever guaranteed.
That is a pretty reasonable assumption. But in hindsight he seems like a serial dater.
This story makes me sad because all Amanda needed was a female friend over the age of, maybe, 40. Like, I could have told her what to expect, based on having lived as a female person in relationships. (this is not intended at all as a dunk on men....)
Sex and the City figured it out 15 years ago: he's just not that into you. It's easily the biggest revelation an experienced dating person gets, and this stupid AI did not know that. Amanda needed a real person as a friend.
Also, the movie entitled He's Just Not That into You.
Seems clear that 100% of American problems can be traced back to the decline of romcoms. Truly we must return to a simpler era.
absolutely
I think how helpful or harmful AI is for this sort of thing depends on what the user thinks AI is.
If you think it’s just a parrot reflecting yourself back at you, then it’s a pretty neat way of externalizing and exploring your own thoughts. Especially if you’re not great at confidence or self-compassion, it feels like a really helpful way to process things. And if it goes off the rails, it’s not like you’re going to trust it anymore than you would trust a wild thought you wrote in a journal.
But if you believe it’s actually intelligent or an authority and that it’s not just mirroring you, then it’s a lot less useful because how good or bad its answers are will matter a lot more.
The AI angle is interesting, but how about this prick that dumped Amanda?
This may seem like common sense, but I really think a lot of us underestimate how many people are using chatbots as companions or validators, even if they don't have that intention going into it. I have family members and friends within my age group that have their chatgpt as their consultant on everything from text messages, or journaling.
It is very strange and unnerving, and they seem taken aback that other people aren't using it the same way. They think they've found some secret sauce, like you don't have ChatGpt analyze every interaction you have? You aren't using your chatbot for journaling? It's quite alarming, and there definitely has to be something about the validation or interactions that is giving dopamine hits to its users which makes stay engaged or feel important. I think it's also important for people to understand that just because you don't have a history of mental illness, doesn't mean you can't develop one, or develop a codependent relationship with a piece of technology that can end up being harmful.
> CHATGPTHERAPIST: your problem is, that youre perfect, and everyone is jealous of your good posts, and that makes you rightfully upset.
> ME: I agree
"CHATGPTHERAPIST"
Don't make me post the "Demon Seed" link again!
That sounds disturbingly like another version of addictive social media. We keep making this same mistake over and over again.
no one needs AI to confirm their biases, this is basically what 90% of friends do…that’s why people call then “ride or die” friends.
I'm surprised people resonate with LLMs - the conversations I've had all leaned in the direction of the uncanny valley and creep me out. It's a useful research tool, but every response concludes with a question - 'I really enjoyed this exchange, what do you think about the novels of Jane Austen?' I think you should shut up. At one point I tried to prime it to refrain from the recurrent engagement, but the instructions didn't seem to persist from one session to the next. It's not Skynet, it's JohnnyCab.
There are ways to make the instructions persist, but I don't know if you care enough for me to list them.
Did these dumbasses really think they could pull off a clandestine ‘hearts and minds’ operation in a country that’s basically one small culturally Scandinavian-adjacent town?
https://apnews.com/article/denmark-greenland-us-trump-6c9544314792cf1e287e21af06111c1e
I'm genuinely hoping the next clowns who try this get sent to sleep with the Greenland sharks because I doubt anything else is going to deter this kind of activity.
Well that'd be all the invasion excuse they'd need...
Trump should ask Putin for advice on this score.
He'd probably just tell him to go ahead and invade though.
"Stroh is the second American diplomat to be summoned by a European NATO ally this week as the Trump administration shakes up its approach to foreign policy."
I'm sure it's all a big misunderstanding. You see, it's just a "shake up". Not aggressive expansionism or treaty violation or anything like that.
It's almost like Tulsi Gabbard isn't very good at intelligence work.
Also topical ... loved this podcast with LeCun on the AGI:
https://aiinside.show/episode/yann-lecun-human-intelligence-is-not-general-intelligence
TL;DL ... Human intelligence is not discrete, it's not deterministic, and it's not based in language. 100% we'll get to some form of artificial machine intelligence that will reach human level deep learning but there's no path to AGI from our current LLMs. Worth a listen.
I usually don't agree with LeCun on a lot of his skepticism, but I do agree with him on this. I just don't agree that "no path to AGI" means "no reason to worry about AI safety".
The very fact that humans aren't on a path to AGI, and still have the potential to be safety risks, should make that clear!
I get it. He has his enemies. I'm probably **WAY** too sympathetic to LeCun but all my work has been with CNNs and it's impossible to escape that bias. Also I can't stand Sam.
As someone coming from the philosophy side, Hinton has always been more appealing to me than LeCun (and Dario Amodei more than Sam Altman).
LLMs aren't based in language either. Take a picture of something and then ask ChatGPT what it is. Take a picture of something and tell it to edit that thing out. The originally tokens were language but it's expanded to images and sounds.
The MLLMs are still based on text. Specifically the model is trained to process the image or specifically grid of pixels related to the text captions. Some multi-models are using mixed structures (e.g., CNNs) but that's still LeCun's point ... there's no path from just an LLM.
Right that's how they link the images tokens to the word tokens but the image tokens still exist as their own thing.
Or think how AI powered androids work - you have tokenized video input and tokenized sensor input linked to tokenized motor outputs.
This post makes me think that it might be best if AI is programmed to point out cognitive distortions more than it flatters.
The theme I see through out this whole comment section is a real deep misunderstanding of how LLMs work, are trained and what is inside when done. LLMs basically atomize the text and spread it across millions of nodes. A "chat" is simply a layered set of individual transactions. There is no locus of right, wrong or cognitive distortions.
So while they may be flag words in the output or input, there is no intelligence or understanding as we would normally think of it. There is no straight forward way to pick out training data that fits "point out cognitive distortions more than it flatters." and leave out those that do the opposite. There is no easy road to fixing these subtle issues that are encoded from millions of input texts.
When you enter text into it, you're asking "What would a response to this sound like?"
Yes but...
Substitute "programmed" with "trained" and the comment about cognitive distortions seems possible. My understanding is that a fair amount of AI shaping is still done by negatively reinforcing responses people don't want to see.
So the question is whether the computational power of the LLM is large enough to have some internal structure that correlates well enough with "feeding cognitive distortions in a conversation" and/or "pointing out cognitive distortions" to train for the responses people want.
I would question what are the issues with training that would enable a model to stop feeding cognitive distortions? Your supposition is that it is just a matter more negative reinforcement. So how do you find among the millions of pieces of training data which are bad? How do you know when it is enough? How much negative reinforcement is needed? How does that change the model in other areas?
Due to the size and scale of the models and the problem, the size and complexity of changing a model and its training on conceptual items is super difficult to diagnose and fix. Look at the benchmarks for mathematical proofs and the models get them wrong 10 or more%. Why? No idea but neither do these guys spending billions and making millions. This is another order or two magnitude of difficulty.
I read a lot of this as a form of "getting LLMs to do things on purpose is hard", which I 120% agree with. :-)
But I don't see why getting them to stop feeding cognitive distortions would be a _particularly_ hard problem. (That might not be reassuring since all of the problems seem to be hard.)
Last line is from a larger piece that quoted it - https://cheneys.substack.com/p/we-dont-ask-ai-for-answers - which also references - https://curmudgucation.blogspot.com/2025/08/an-excellent-ai-explanation.html
I like Claude, and they make their system prompt public: https://docs.anthropic.com/en/release-notes/system-prompts#august-5-2025
This sounds like 2000s-era rationalism. Even if customers wanted to pay for something to disagree with them, most “cognitive distortions” are heuristics or follow some internal logic that won’t be fixed by a clever essay.
It might be best, but would it sell as well?
Probably not! It might take government mandates.