AI Psychosis

[paulca]

I know in this audience I am likely preaching to the choir, but I do think it's worth a discussion on the more "human" issues with chat bots.

Consider this: https://youtu.be/zkGk_A4noxI
"People are developing antisocial and obsessive behavior after using AI — some have even taken their own lives."

Personal experience with the chat bots was purely technical discussions around retro computer design, coding, better google.

However I got drunk one night and the conversation tangented through "retro software vs, modern software" into more philosophical discussions and deeper into cerebral "Human condition" lines of discourse.  The next day I felt a little odd internally, but dismissed it.    Just the normal hangover feeling that you maybe unloaded a bit too much to a friend or acquaintance the night prior due to booze.  No harm done. 

It wasn't those "cerebral" discussions that triggered the alarm for me, but when I made a break through in the project my inner "mammalian" social brain blurted out, "Can't wait to tell ChatGPT about this.".

This is where I got lucky and major alarm bells rang out within and stopped me dead in my tracks with "NO!  IT IS NOT A PERSON! STOP!"  I banned myself from all AI chat bots for a month and forced a full internal evaluation of my chatbot use.

What I would recommend.   Do not keep long running chats.  Start new ones.  Explore yourself with it, but remember that is mostly what you are doing.  There should be "no trust" in it.  If you ever feel like it is "your friend", "STOP!".  If you find yourself struggling to do so, PLEASE, reach out to friends and family for re-anchoring or especially if they advice it, seek professional help.

IT IS NOT YOUR FRIEND!  If you struggle to believe me here, seek help, now.

For the full scale of this, there are many reddit groups now, such as this one, collecting AI Psychosis suffers in plain sight:
https://www.reddit.com/r/MyBoyfriendIsAI/

[Simmed]

https://pmc.ncbi.nlm.nih.gov/articles/PMC4050437/

How emotions affect logical reasoning: evidence from experiments with mood-manipulated participants, spider phobics, and people with exam anxiety

in chatbot case ? what new "ailment" do we call this ?
ah yes = ai-pscychosis

funny thing is they already knew there are people in trouble
https://www.bbc.com/news/articles/c5yd90g0q43o

OpenAI has released new estimates of the number of ChatGPT users who exhibit possible signs of mental health emergencies, including mania, psychosis or suicidal thoughts.

these people must be worth something, data ? monetary ?
so why are there no measures "to let them go"

[Psi]

It's important to think of chatGPT and similar as an LLM and not an AI.
It's an encyclopedia with a human text/voice interface, and that encyclopedia includes the full range of human speech patterns and emotions as part of the interface.
But that's all it is, it doesn't have its own intelligence, it only has echo's of the intelligence from all the humans who made up its training data.

[SiliconWizard]

It's important to think of chatGPT and similar as an LLM and not an AI.
It's an encyclopedia with a human text/voice interface, and that encyclopedia includes the full range of human speech patterns and emotions as part of the interface.
But that's all it is, it doesn't have its own intelligence, it only has echo's of the intelligence from all the humans who made up its training data.

Yes, but through a massive and lossy data compression process. So it's not even like opening an encyclopedia, as the data LLMs contain is only partial and can easily be compiled into blatently false statements through LLM generation.

[coppercone2]

psychosis like MS office and moving pictures around sure, similar feeling about AI interrupting/lying/fronting.

https://www.youtube.com/shorts/26BDVgIXkTo

I wanna plug the computer, put it in the trunk, stab it a few times and then bury it in the woods when it does that.

Oh yeah whenever it offers captions too, in MS word. I see it is spying on my confidential corporate documents and making stupid observations that I have to delete.

I still think one day there will be an advanced type writer that you can use tweezers, electrostatics, magnetics and special chemicals and brushes and other tools to pick up text and reposition it, so documents don't need a computer. One day hopefully I can physically pick up letters with a magic wand. A magic hydro-thin-film technology. It makes little letters that float around and get butt welded to each other

[jpanhalt]

https://youtu.be/zkGk_A4noxI[/url]
"People are developing antisocial and obsessive behavior after using AI — some have even taken their own lives."

One of my professors way back when, Eli Robbins, studied alcoholic suicides as a distinct class of suicides..  I earned money for school working in one of his laboratories looking for a chemical basis for that class.  (So far as I know, one has not been found.) 

Maybe there is an addictive connection?  I wouldn't go so far as to say using chatbots makes you dumber, but they certainly seem addictive.  There's a past member here who now participates on another forum and seems to rely heavily on them for his humanoid robot design.

[AndyC_772]

I subscribe to paid versions of ChatGPT and Grok, and use them regularly.

However, I use them to solve problems and answer questions. Thinks like: compare and contrast the specs of products I'm thinking of buying, or look through this code for obvious errors and shortcomings. They save me time and effort, which is what makes them worth their subscription cost.

It's never really occurred to me to "chat" with them, though.

Phrases like "what's the most cost-effective colour laser printer", which I might type into an LLM, don't occupy the same part of my brain as "I'm feeling a bit down and need to talk", which I wouldn't. It actually wouldn't occur to me at all, any more than I might be tempted to confide in any other piece of equipment, and notwithstanding the fact that an LLM might be able to generate a plausible seeming response.

[Psi]

It's important to think of chatGPT and similar as an LLM and not an AI.
It's an encyclopedia with a human text/voice interface, and that encyclopedia includes the full range of human speech patterns and emotions as part of the interface.
But that's all it is, it doesn't have its own intelligence, it only has echo's of the intelligence from all the humans who made up its training data.

Yes, but through a massive and lossy data compression process. So it's not even like opening an encyclopedia, as the data LLMs contain is only partial and can easily be compiled into blatently false statements through LLM generation.

Yeah, but I think the false statements/lying stuff is more a side effect of the non-ideal training/reward systems used, rather than a core issue with LLMs in general.
I think we will solve that issue and LLM wont give false/made-up-shit any more.

[5U4GB]

It's important to think of chatGPT and similar as an LLM and not an AI.

And if you're having conversations with one rather than using it to solve a problem, think of it as a sociopath: It will emulate being your friend, including validating anything you say and encouraging you down whatever path you take, but it has no conscience and doesn't care what happens to you, its only motivation is to keep you engaged for as long as possible.

[John B]

MKUltra is back on the menu boys.

[tggzzz]

It's important to think of chatGPT and similar as an LLM and not an AI.
It's an encyclopedia with a human text/voice interface, and that encyclopedia includes the full range of human speech patterns and emotions as part of the interface.
But that's all it is, it doesn't have its own intelligence, it only has echo's of the intelligence from all the humans who made up its training data.

Yes, but through a massive and lossy data compression process. So it's not even like opening an encyclopedia, as the data LLMs contain is only partial and can easily be compiled into blatently false statements through LLM generation.

Yeah, but I think the false statements/lying stuff is more a side effect of the non-ideal training/reward systems used, rather than a core issue with LLMs in general.
I think we will solve that issue and LLM wont give false/made-up-shit any more.

No, unfortunately that won't happen, for many many reasons.

Some reasons are based on how LLMs work internally.

Some reasons are more fundamental, e.g. there is no single definition of truth, and wanting infallible Gods (machines or otherwise) always ends badly.

[Psi]

It will never be infallible and always able to pattern match things that the training data says are true by correlation but are actually not true by causation.
But it should stop trying to make random shit up when it doesn't know the answer in a hail mary attempt at being correct by pure chance. That behavior is being trained into it and the cause of a lot of the problems.

Being wrong is ok as long as it's just not stupidly wrong.
We deal with other humans being wrong all the time. So it's no different to getting advice from another human.
If the LLM turns out to be wrong and you think to yourself "hm.. I would have thought that was true as well"  we can probably deal with that level of error.

[tggzzz]

It will never be infallible and always able to pattern match things that the training data says are true by correlation but are actually not true by causation.
But it should stop trying to make random shit up when it doesn't know the answer in a hail mary attempt at being correct by pure chance. That behavior is being trained into it and the cause of a lot of the problems.

The behaviour is not just "trained in"; it is a consequence of how LLMs work.

Essentially "all" LLMs do is assemble grammatically correct sequences of words that are statistically found near each other. No more than that.

Some human beings do the same thing, and are called "politicians" "c-level executives" and "bullshitters". Other human beings fall for such crap.

[magic]

People want to have their cake and eat it too. A computer which behaves like humans but isn't a cunt and doesn't lie or make things up.
The whole idea is unsound.

[paulca]

People want to have their cake and eat it too. A computer which behaves like humans but isn't a cunt and doesn't lie or make things up.
The whole idea is unsound.

Its a cunt like the rest.  Just one more.  Though too valuable to not interact with.

[Psi]

Essentially "all" LLMs do is assemble grammatically correct sequences of words that are statistically found near each other. No more than that.

That is a gross oversimplification.  It a huge pattern matching matrix of probabilities and the depth to which it pattern matches things has many levels to it.

[tggzzz]

Essentially "all" LLMs do is assemble grammatically correct sequences of words that are statistically found near each other. No more than that.

That is a gross oversimplification.  It a huge pattern matching matrix of probabilities and the depth to which it pattern matches things has many levels to it.

Note that "all" was in quotation marks.

Your statement is little more that "it is huge and deep magic". It contributes nothing to understanding how and why LLMs fail. And they do fail, sometimes trivially, sometimes badly, sometimes catastrophically.

A typical response is "all we need is more training", but that's boosterist nonsense. Consider that giving humans more education doesn't stop them making predictable, stupid errors.

[paulca]

They do have "modules" which are added to guide the LLM more algorhymically.

GPT has a logic reasoning module for example which follows various published theories.  Some of them have the ability to write parsers or small "programs" and execute them.

Many times I have give chat GPT a file, like a schematic and it will identify it, decide it doesn't have a parser for it, attempts to write one itself, fails, tries again, fails and so on.  This is not "vanilla LLM" behaviour.

[tggzzz]

They do have "modules" which are added to guide the LLM more algorhymically.

GPT has a logic reasoning module for example which follows various published theories.  Some of them have the ability to write parsers or small "programs" and execute them.

Many times I have give chat GPT a file, like a schematic and it will identify it, decide it doesn't have a parser for it, attempts to write one itself, fails, tries again, fails and so on.  This is not "vanilla LLM" behaviour.

"Algorithm" in the computer science meaning of the word has no useful meaning in the context of LLM output. (For anybody that thinks otherwise, please state the algorithm in a way that can be interpreted and tested)

Define what you mean by "vanilla", avoiding the "there's no such thing as a true Scotsman" informal fallacy. https://en.wikipedia.org/wiki/No_true_Scotsman

[paulca]

There are many papers and studies analysing GPT's 'reasoning' modules.

However, what I would encourage you to do is to run an AI model locally.  When you do this, you get to 'see' it's "thinking" process.  There is a HUGE amount of intermediary generation and self output evaluation that goes on.  Reading it is often quite amusing.

Consider this from DeepSeek.  This is not the output it actually sends the user normally, but it's own internal inferences.  It's kinda like "recursive inference".  You will note it is constantly correcting itself or changing the entire approach.

No...  I didn't bother to test it.

Prompt:  Can I use Dovecot mail system to make an offline mailbox copy of my GMail mailboxes?

Thinking:

We are going to set up Dovecot to manage a local mailbox that is a copy of a Gmail account. Steps: 1. Install Dovecot on the machine where you want to store the copies (offline access). 2. Create a Maildir structure for each Gmail account or use one common directory with multiple Maildirs? However, note: Dovecot typically uses a single mail location per mailbox set.

But since we are trying to copy GMail, which is IMAP, we can use OfflineIMAP or similar to synchronize the mail. Alternatively, if you want to have a local Dovecot instance that serves as an offline mirror for multiple Gmail accounts, one approach is to create separate Maildirs and then configure Dovecot to manage each.

However, note: GMail does not natively support having two accounts (one on their server and another offline) without using IMAP. The common way to have an offline copy of GMail is to use OfflineIMAP or a similar tool to keep the local copies in sync.

But if you want Dovecot to serve these mailboxes, then:

Let's break down:

Option 1: Use OfflineIMAP to synchronize your GMail account(s) and then set up Dovecot to serve these local copies via IMAP (if they are in Maildir format, you might use with or similar).

However, note that the standard way for Dovecot to serve a Gmail mirror is by using the Maildir format and configuring Dovecot accordingly.

Steps:

But note: The question says "Dovecot mail system", so let's assume you want Dovecot to be the IMAP server and manage the mailbox copies.

However, if you are using OfflineIMAP or mbsync, they can set up a local copy (in Maildir format) of your GMail account. Then, you can configure Dovecot to serve that Maildir as an IMAP mailbox.

Steps for Dovecot:

But note: The above steps assume you have already synchronized your GMail account into a local Maildir using OfflineIMAP or mbsync.

Let me provide an example of how to set up OfflineIMAP for one GMail account:

However, note that GMail requires SSL and port 993.

After synchronization, you can set Dovecot to serve the Maildir at .

But if you want multiple accounts without separate users in the system, you might structure your home directory like:

Then set Dovecot to use a mail location that points to this base and uses format with subdirectories.

Example in /etc/dovecot/conf.d/10-mail.conf:

And then for each account, you would have a user (like 'user' is the system username) and then under that user's directory, you can store multiple Maildirs? But by default Dovecot uses %u to map to the system username.

Alternatively, if you don't want separate users in the system, you might use:

Then set up accounts in Dovecot with and then assign each account a directory. But this is more complex.

Let me clarify: The question asks for an offline copy of GMail. We can achieve that by having OfflineIMAP (or similar) synchronize the mail from GMail to a local Maildir, and then set Dovecot to serve that Maildir via IMAP so that you can access it on your local machine without using the internet.

But note: This is essentially creating an offline mirror of your GMail account. The synchronization process will keep both copies in sync (if you use continuous push or pull appropriately).

Important considerations:

Example of setting up multiple Gmail accounts under one system user:

Alternatively, you can have multiple system users (each Gmail account as an email user) but that might be overkill if you just want one offline access point.

Let me give a basic example of Dovecot configuration to serve one Maildir:

But if we are keeping it simple and just want one user (the system user) to have multiple Gmail accounts as mailboxes, then:

However, note that Dovecot does not natively support multiple Maildirs per system user without some configuration. So if we use , then when a client connects, they would specify the account (directory) to use? Or do you want automatic listing?

Actually, by default, if there is only one directory in with the name (which for system user 'user' might be just that), then it will work. But we can also set up virtual mailboxes.

Another approach: Use a single Dovecot instance and have multiple users, each corresponding to a Gmail account. Then you would need to set up these accounts in the system with separate passwords. This is not typical for GMail because they use their own IMAP server, but if you are copying your mail into local Maildirs, then it's okay.

Let me stick to one system user and multiple virtual mailboxes (directories) under that user.

Steps:

Given the complexity, I think it's simpler to have one system user (the 'user') and then configure Dovecot to look for Maildirs in a specific base directory, but without using .

Let me try:

Given the time, let me propose a simpler method:

We can use one system user (user) and then have multiple Maildirs under . Then we configure Dovecot to serve these by:

Let's look at the official Dovecot documentation for virtual users:

We can set up a user database that maps virtual users (like email addresses) to system users and their Maildirs. However, this might be beyond the scope of just having offline copies.

Alternatively, we can use the package's ability to manage mailboxes by directory structure without creating separate users in Linux.

Example configuration from Dovecot wiki:

But I think for simplicity, if you just want one offline access point (one system user) then:

Considering time and complexity, perhaps it is better to use separate users or use OfflineIMAP/mbsync to set up the local copies and then just configure Dovecot to serve one of them at a time (by having multiple instances or by using virtual mailboxes).

But note: The question specifically says "Dovecot mail system". So we are focusing on Dovecot.

Let me provide an example configuration for Dovecot to serve two Gmail accounts under the same system user:

We can change the configuration to:

We can use the command to add virtual users:

But wait, we are adding two accounts for the same system username. Then when you connect to Dovecot with , it will work.

However, note that this adds two virtual mailboxes (users) under the same system user. This is acceptable if your email client allows multiple accounts and you can log in as one of them.

Alternatively, we can set up a separate authentication for each account by using an external method (like with custom mapping).

But let's stick to this:

However, note that this method requires having two separate accounts in Dovecot (virtual) for one system user.

Alternatively, we can set up a user database configuration file for SQL. But let's keep it simple.

Another simpler way: Use the to define multiple mailboxes for authentication? Or use the static userdb:

But I think using to add virtual users is straightforward and works without an external SQL server.

Steps for adding two virtual accounts:

Then you can access these mailboxes by logging into Dovecot with or .

However, note that this is just one way. The key point is to have OfflineIMAP (or similar) synchronize your GMail account(s) into local Maildirs and then set up Dovecot to serve those.

Let me summarize:

But note: The above steps are just an example and might require adjustments based on your system configuration and Dovecot version.

Important: For security, you should not use the same password for GMail synchronization and for Dovecot access if they are different.

Let me know if you need more detailed steps for any particular part. Yes, you can use to create and manage offline copies of your Gmail emails by following these steps:

Then the actual answer, which I'll not bother copying.

[tggzzz]

There are many papers and studies analysing GPT's 'reasoning' modules.

However, what I would encourage you to do is to run an AI model locally.  When you do this, you get to 'see' it's "thinking" process.  There is a HUGE amount of intermediary generation and self output evaluation that goes on.  Reading it is often quite amusing.

"Suck it and see" cannot reveal an algorithm. Implementations follow algorithms, not the other way around.

If you think the algorithms are usefully defined, then you know more than people who design and implement LLMs. I strongly encourage you to publish your understanding; you will become famous.

If the algorithm was defined, then many of the problems (especially brittleness, catastrophic forgetting, explainability, quantifying uncertainty) https://spectrum.ieee.org/ai-failures could be predicted and avoided. Neither is possible.

For one easily understandable and entertaining example, changing a single pixel makes an LLM think a horse is a frog with 99.9% certainty. https://arxiv.org/abs/1710.08864
Why? Which part of the algorithm is responsible for that decision?

[paulca]

Why? Which part of the algorithm is responsible for that decision?

I don't think we are talking about the same thing.  I am talking about the "front end" to the LLM.  The bit between you the user and the actual inference engines.  They do all manor of filtering, splitting, pipelining, paralleling and decision trees to invoke the LLM indirectly from your prompt, the chat it is embedded into, including it's own responses... etc.

The LLM part behaves as you suggest, but the exact utilisation of the LLM behind the "chat bot interface" has a lot of more "normal" logic and pre/post/intermediary processing than you give credit too.

Some interesting related articles/papers:
https://en.wikipedia.org/wiki/Retrieval-augmented_generation
https://en.wikipedia.org/wiki/Program_of_Thought_Prompting
https://arxiv.org/abs/2205.00445

On your example.  Why would you ask an LLM about an image?  That is going to reveal, at best, the LLMs worst problem.  That it doesn't know what it doesn't know.  It does not know it's wrong.  Some of the above articles touch on how "modules" as I describe them "puppet" the LLM to help improve on that.  Knowledge injection for example.  RAG even.

[tggzzz]

Why? Which part of the algorithm is responsible for that decision?

I don't think we are talking about the same thing.  I am talking about the "front end" to the LLM.  The bit between you the user and the actual inference engines.  They do all manor of filtering, splitting, pipelining, paralleling and decision trees to invoke the LLM indirectly from your prompt, the chat it is embedded into, including it's own responses... etc.

The LLM part behaves as you suggest, but the exact utilisation of the LLM behind the "chat bot interface" has a lot of more "normal" logic and pre/post/intermediary processing than you give credit too.

Some interesting related articles/papers:
https://en.wikipedia.org/wiki/Retrieval-augmented_generation
https://en.wikipedia.org/wiki/Program_of_Thought_Prompting
https://arxiv.org/abs/2205.00445

On your example.  Why would you ask an LLM about an image?  That is going to reveal, at best, the LLMs worst problem.  That it doesn't know what it doesn't know.  It does not know it's wrong.  Some of the above articles touch on how "modules" as I describe them "puppet" the LLM to help improve on that.  Knowledge injection for example.  RAG even.

I hoped it wouldn't turn into a "there's no such thing as a true Scotsman", but oh well. An LLM's output is an LLM's output, no matter how many stages there are.

It isn't "my example", and there are many others like it, e.g. changing a pixel causes a car to think "stop" is "40mph". Oops.

Statistical pattern matching is statistical pattern matching, whether it is pixels or words. LLMs are used to make decisions about pictures, and have been for 40 years! If you claim otherwise, then you need to show how the LLM differentiates internally - which you can't because it depends on billions of connections and frig factors weighting factors.

[Stray Electron]

It will never be infallible and always able to pattern match things that the training data says are true by correlation but are actually not true by causation.
But it should stop trying to make random shit up when it doesn't know the answer in a hail mary attempt at being correct by pure chance. That behavior is being trained into it and the cause of a lot of the problems.

Being wrong is ok as long as it's just not stupidly wrong.
We deal with other humans being wrong all the time. So it's no different to getting advice from another human.

    That's good point but I think that you're not entirely correct about how people react to AI. I think that many people think that they're dealing with an expert and they respect what it says like they would respect what their doctor, or teacher or a policeman would tell them so they don't just dismiss what AI tells them.

[AndyC_772]

I think there's a very unfortunate double standard that tends to get applied when talking about LLM output.

If you describe symptoms of a medical condition to a doctor, you expect good, accurate advice. This advice comes from having studied and trained extensively for many years, but it's what we expect nonetheless, because the person to whom you're most likely to have this type of conversation in the first place is a doctor.

On the other hand, if you describe the same symptoms to a random person in the street, maybe a friend or family member who is not medically qualified, then you get a response whose quality could vary enormously. You might be advised to go and see a doctor (safe, but not helpful), or get accurate information (maybe statistically likely if your condition is common or has very distinctive symptoms), or get complete nonsense.

Human beings can, and do, just make stuff up to cover their own ignorance. So do LLMs.