AI General Thread

[the-pi-guy]

Actually another super annoying thing that AI does:

"That's not what I mean, I mean _____"


"You're right, let's lock in to the real heart of the issue." (somehow still not talking about ____ )

[Legend]

Actually another super annoying thing that AI does:

"That's not what I mean, I mean _____"


"You're right, let's lock in to the real heart of the issue." (somehow still not talking about ____ )

"Hostile reframing."

So freaking lame. But a great reminder that llms still are clueless about basic convos.

[darkknightkryta]

Neural nets are going to do what neural nets do.  The faster the business men in charge realize this is just another data structure the better.  Like, they were better off making industry specific neural nets instead of the behemoth general "I know everything" nets which gets its data points crossed.

[Legend]

Neural nets are going to do what neural nets do.  The faster the business men in charge realize this is just another data structure the better.  Like, they were better off making industry specific neural nets instead of the behemoth general "I know everything" nets which gets its data points crossed.

The behemoth nets can do cross domain learning however which is very very useful. You might not have enough data to train a net for your specific industry in isolation but a behemoth net can be fine tuned on it just fine.

[darkknightkryta]

The behemoth nets can do cross domain learning however which is very very useful. You might not have enough data to train a net for your specific industry in isolation but a behemoth net can be fine tuned on it just fine.

Not worth the money or data centers.  Hell I stopped using Gemini since it cross contaminates conversations referencing irrelevant things.  I spend more time trying to get it to stay on task than anything.  That's the problem with large neural nets.

[Legend]

Lol. I tried having the new gemini 3.5 help me with a programming thing.

"I could use wave function collapse, but is there something better for this specific thing?"

Gemini: "yes use this algo, a million times better!"

"Ok it's fast but it doesn't make sense, these two values are the same."

Gemini: "yes it can't handle stuff like that. Why, was that important?"

[the-pi-guy]

Sorry for the silly post, but I feel like this is the most reasonable thread I've seen on Era about AI, probably ever:


https://www.resetera.com/threads/wsj-mind-blowing-growth-is-about-to-propel-anthropic-into-its-first-profitable-quarter.1526530/

[Legend]

Sorry for the silly post, but I feel like this is the most reasonable thread I've seen on Era about AI, probably ever:


https://www.resetera.com/threads/wsj-mind-blowing-growth-is-about-to-propel-anthropic-into-its-first-profitable-quarter.1526530/

Yup there is real money now.

And Anthropic is paying SpaceX $1.5 billion a month for compute. I think we're about to see the ai hype investments get dwarfed by the sustained steamroller of paid customers.

[the-pi-guy]

Yup there is real money now.

And Anthropic is paying SpaceX $1.5 billion a month for compute. I think we're about to see the ai hype investments get dwarfed by the sustained steamroller of paid customers.

One of the mods like a month ago, made a post where they compared using AI to finding out your food was made out of ground orphans.

A lot of people in there were openly talking about using AI at work. 

[Legend]

One of the mods like a month ago, made a post where they compared using AI to finding out your food was made out of ground orphans.

A lot of people in there were openly talking about using AI at work. 

People are pyscho with ai. I really do not get it.

Have issues with slop? Sure I fully get it.
Have worries about what ai will mean for people's livelihood? Sure I fully get it.

Everything else is just made up nonsense or general fears/issues.

Like the Ghosts TV show last night was all about their house being turned into an ai datacenter. Parts of it made sense, the show focused on how bad it would be for the ghosts living there vs a mansion, but it's still silly that this became something the mainstream thinks about.

Will help spacex a lot though. Move datacenters to space.

[the-pi-guy]

Does anyone else ever just find it surprising what isn't out there? 

Like right now, I would really like to have an LLM+image generator pipeline that would create contextual images that go with the LLM response. 

That doesn't feel crazy hard or anything, yet I'm not aware of any free solutions.  

[Legend]

Does anyone else ever just find it surprising what isn't out there?

Like right now, I would really like to have an LLM+image generator pipeline that would create contextual images that go with the LLM response.

That doesn't feel crazy hard or anything, yet I'm not aware of any free solutions. 

It is definitely a valid observation! While separate, free APIs for text and image generation exist, seamlessly linking them into a single, automated, conversational flow—where one model dictates the precise context for the other without extra prompt engineering—is a powerful combination that isn't readily available as a simple, free product yet.
The technological pieces are largely there, and paid, proprietary systems are starting to integrate these steps, but you are right that a free, open pipeline isn't a prominent 'ready-to-use' solution.
I have generated an image showing exactly what this requested pipeline looks like.





Gemini can kinda do it. I just told it to respond to you and include a generated image. One prompt.

Edt: did a second test. Yeah gemini can fully include multiple generated pics seemlessly within its generated response. I tested it with three pics.

[nnodley]

I'm not a fan of AI still, but i've been using Codex to put all my thoughts for my big game into documents. Much easier.

I've also started working on a smaller parkour heist game that I've been having Codex give me blueprint guides to set up nodes and stuff. So far it's been doing a great job. Though you still have to have a pretty significant knowledge of blueprints or you'd be lost

[the-pi-guy]

Random collection of AI thoughts - some I've talked about before, and some I don't think I have:

- LLMs obviously don't have 3D space. 
And this is a common failure.  Weird example that I haven't actually done, but this is easy to illustrate. If I were to tell an LLM to write a story where Elena is sitting in front of Nathan looking out the front window, and she will kiss him. A lot of times they'll say something like "Elena leans forward to kiss Nathan".  A lot of mostly good models will fail to realize that something like this is impossible. Elena needs to turn around.

- a lot of models will start to fail at Dungeons and Dragons type roleplaying. Where they will start to forget which character they're supposed to be playing as. Or they won't realize that a specific character doesn't have access to certain information. 

- It's a little interesting seeing biases in LLMs. This happened a while ago. 
One of the weirder ones that I've seen. I gave it a prompt, and it would talk normally. But when I gender swapped the roles, it would get flirty. Then I recall trying to go back to the other prompt to see if I could get it to flirt, and had a hard time getting it to do that.  

It's also funny how different models like different names. 

This is an easy one: 
Write a 3 paragraph story about a man going to the grocery store

Copilot really likes the name Daniel. 
GPT-OSS also does.
GTP-5 avoided giving a name.  
Llama likes the name John for this story.
Gemini likes Arthur I guess  
Claude Haiku 4.5 apparently likes the name Marcus.  

[Legend]

"The View from the Glass
The rain was streaking across the wide pane of the front window, blurring the streetlights into soft, amber smudges against the gathering dark. Elena sat on the cushioned bench of the bay window, her back resting against Nathan's chest. His legs were bracketed around hers, his arms loosely draped over her waist—a steady, warm anchor against the chill leaking through the glass.
They had been watching the storm for an hour in silence. It was the kind of quiet that didn't require filling, built from a long-standing comfort where every habit was familiar. Nathan's chin rested lightly on her shoulder, his even breathing a constant rhythm against her neck.
Elena traced the path of a single raindrop as it raced down the outside pane.
"The wind is dying down," she said, her voice quiet in the dim room.
"A bit," Nathan murmured, his chest vibrating against her back. "Still coming down hard enough that I'm glad we're inside."
Elena watched his reflection in the glass. In the dark window, his features were silhouetted by the dashboard-like glow of the stereo across the room. He wasn't looking at the street anymore; he was looking at her reflection.
Turning around on the narrow bench, she shifted her weight until she was facing him completely, knee to knee. Nathan blinked, his hands moving to her hips to steady her as she moved. The question in his eyes never made it to his lips.
Elena reached up, her fingers brushing the cool skin of his jawline, and leaned forward to close the remaining distance.
The kiss was deliberate, warm, and familiar—a quiet punctuation mark to the end of the day. Nathan responded instantly, pulling her just a fraction closer as the rain continued to beat a steady, rhythmic retreat against the glass behind them."

Hot.

There was a benchmark like this with a glass of water and a dishwasher. Most early models would get it wrong, but currently I think it is solved with that test.

You are fully correct though that it is still an issue in general.

Models are ghosts. They are echoes of what they experienced during training, with poor generalization. Hence why a few weeks ago you could ask "How many ps are in google?"

https://g.co/gemini/share/5e3c30e3be70

My brother complains all the time about llms used for d&d.

I got arthur with Gemini too!