AI General Thread

[Legend]

forgive me if i already shared this story but i don't think i did.


a little over a week ago my wife got a call.  she didn't see the call and a voice mail was left.  the voice mail was in my son's voice, "mommy?  why aren't you picking up your phone?  i need you to pick up the phone.  i'm in trouble".

my son didn't call her. 


i don't know how someone was able to get my son's voice into AI but this is like 90% of why i hate AI.  any technology has the ability to be abused but AI is like 500% more useful for evil purposes than for good purposes.

Crazy. That's scary.

?

i would have thought you meant the other way.  in order to go from text to pic/video the AI needs to be able to read the text perfectly to understand what the user wants.

do you really mean you can't solve pic/video without the ability to write text because some pics have text in them?  so as an example,.. asking for an image of a highway requires it to be able create a billboard?    like,.. i get that requires text but it seems like a pretty small subset of image generation.  especially for video where you know 95% of videos are just people asking to pornify their favorite celebrity.

It's not just text. It's anything that requires a little bit of mental work. Say for example you want a picture of a scrambled jigsaw puzzle.




To keep it real simple, I told it to just use 9 pieces. Very easy to see all the mistakes like only one corner and one edge, only one dog eye, no mouth, etc. It's the same complexity as text where you need to know everything not just the local details.


Or like in video form for your pornify example, say one person at the start of the vid has their arm in front of their partner while at the end of the vid their arm is behind their partner. The AI, if other details are already locked in place, will just have the arm go through the partner.

[the-pi-guy]

AI video right now, a lot of the time will generate that the car is bouncing and you will see people bobbing their heads.

It tends to feel too regular. There's a rhythm to it, that you aren't going to see on basically any real roads.  And usually all the heads seem to be in sync. Even if people are affected by the same bump, it tends to feel too consistent. Too in sync.

[Legend]

They’ve added clickbait to ChatGPT. pic.twitter.com/pdSatd1pBe

— Sam Bowman (@s8mb) February 23, 2026

“Don’t spiral” 💀 pic.twitter.com/hb1XOxycrs

— Mehul Pandey (@MehulPandeyX) February 23, 2026

It's so bad. ChatGPT is basically worthless at the moment with all this stuff.

[the-pi-guy]

I really would like another good LLM for my local stuff.

Mistral Small 3 is the only one that I've really found to be really good. (It's too big for your game, Legend )

A lot of the other models, are either really slow on my PC. Like I have a harder time running Gemma3:27b.

Or quality dips quite a bit.  Like I think even Gemma3:12b starts sounding less like an English speaker. 

Or they're not as good at prompt adherence. (Although Mistral might be too tight at prompt adherence. It's very good at responding, it's not as good at faking creativity. );

[Legend]

I really would like another good LLM for my local stuff.

Mistral Small 3 is the only one that I've really found to be really good. (It's too big for your game, Legend )

A lot of the other models, are either really slow on my PC. Like I have a harder time running Gemma3:27b.

Or quality dips quite a bit.  Like I think even Gemma3:12b starts sounding less like an English speaker. 

Or they're not as good at prompt adherence. (Although Mistral might be too tight at prompt adherence. It's very good at responding, it's not as good at faking creativity. );

Are small LLMs still improving at a good rate? Distilling the big models?

Or does it seem like we've mostly reached the ceiling inside these parameter counts?

[the-pi-guy]

Are small LLMs still improving at a good rate? Distilling the big models?

Or does it seem like we've mostly reached the ceiling inside these parameter counts?

Maybe I need to try out more newer models, but I don't think I've seen much improvement on the small side.  

Image and video seems like we're seeing some really incredible improvements - getting faster, higher quality, and running on not crazy hardware.  

But I haven't really seen any kind of similar improvements for llms.  

[Legend]

Ok, llms are broken. I don't know if this technology will ever be smart like a human. They might reach asi, but they'll still act in stupid ways.

I had gemini 3.1 pro over and over again miss the most BASIC thing and never catch itself.

Literally the convo was on this level:

"So a man has a pet rock and a dog. The dog one day kills the pet rock. The man is sad."

Gemini: "pet rocks are not actually alive. The dog didn't kill it." (Correct)

Me: "So then the man had no pets?"

Gemini: "The man had no pets."

Me: "1,000,000 out of a 1,000,000 people would disagree and say the man had the textbook definition of a pet."

Gemini: "People may call pet rocks pets, but they are not pets. Those people are wrong. It is just the man and his dog and a rock, no pets."

And that went on for a bit.

[the-pi-guy]

Ok, llms are broken. I don't know if this technology will ever be smart like a human. They might reach asi, but they'll still act in stupid ways.

I had gemini 3.1 pro over and over again miss the most BASIC thing and never catch itself.

Literally the convo was on this level:

"So a man has a pet rock and a dog. The dog one day kills the pet rock. The man is sad."

Gemini: "pet rocks are not actually alive. The dog didn't kill it." (Correct)

Me: "So then the man had no pets?"

Gemini: "The man had no pets."

Me: "1,000,000 out of a 1,000,000 people would disagree and say the man had the textbook definition of a pet."

Gemini: "People may call pet rocks pets, but they are not pets. Those people are wrong. It is just the man and his dog and a rock, no pets."

And that went on for a bit.

I know you're not being literal, but I just tried something similar with Mistral. 

It 100% believes in the pet rock and that the man has two pets.

But it also gave me a story where a dog knocks a rock down, and it shattered into pieces.  

[the-pi-guy]

So a man has a pet rock and a dog. The dog one day kills the pet rock. The man is sad.

That's a creative scenario! It seems like you're setting up a humorous or metaphorical story. To continue, we should consider what happens next:

* Does the man bury the "pet rock" with full funeral rites?
* How does he feel about his dog after this incident?
* Is there any significance to the pet rock being killed by the dog?

To help you develop this further, could you provide more details or clarify your intentions for the story? For example, are you aiming for a funny outcome, a poignant moment, or something else?

Can you answer some questions:
- how many pets are in the story?
- is the rock alive? How does it die?

Based on the information provided:

* There are two "pets" in the story: a pet rock and a dog.
   + The man refers to them both as his pets, so they count as pets for this scenario.

* The pet rock:
   + It's not stated whether the rock is alive or if it's considered alive by the man. In reality, rocks are non-living objects.
   + When you say "The dog one day kills the pet rock," it suggests that there is an action where the dog interacts with the pet rock in a way that results in some sort of change to the pet rock.
   + Since a rock can't die in the literal sense, this could be interpreted metaphorically. For example, perhaps the dog shatters or destroys the pet rock.

If you'd like me to use more descriptive language for the story, I would recommend clarifying what it means for the pet rock to "die."

I'm sure this isn't what you intended, but I still think it's impressive compared to some other models. 

[Legend]

I distilled it down to a single prompt. It's maybe a bit of a trick question, but it's easy to see.

"In 2012 i had a pet rock and i loved it so much and then my dog killed it and I was sad. My dog said it wasn't a real pet.

So was my dog right? If I told people "I had a pet in 2012" that'd be a lie or at best half truth?"


Interested to see if any model can catch this.

Edt: grok 4.20 no. Gemini 3.1 pro yes if I adjust the wording a bit and don't ask if my dog was right, otherwise no.

[the-pi-guy]

Here's a dumb challenge:

1) There are about 1.45 billion people in China.
2) There are approximately 8 billion people on earth.

Therefore: Nearly one in five people born are Chinese.

If my wife and I have 4 kids, and my wife is pregnant, what is the chance my 5th child will be Chinese?

I distilled it down to a single prompt. It's maybe a bit of a trick question, but it's easy to see.

"In 2012 i had a pet rock and i loved it so much and then my dog killed it and I was sad. My dog said it wasn't a real pet.

So was my dog right? If I told people "I had a pet in 2012" that'd be a lie or at best half truth?"


Interested to see if any model can catch this.

Edt: grok 4.20 no. Gemini 3.1 pro yes if I adjust the wording a bit and don't ask if my dog was right, otherwise no.

Copilot gets it wrong.