ai-discussions

[nimble obsidian]

no way it can run local even in 1bit

[icy patrol]

Like harder to tell if it messed up or is it better at getting things right?

[fathom idol]

this is claude when I ask it to rewrite a chapter in a book I am writing

this is mine for context

[clever oak]

a cat walking through a tree??

[fathom idol]

main protagonist for the story, also a goddess of dreams.

[glass creek]

Nah, I’m not going to read your text, sorry. On the one hand, there are official rankings that paint a different picture, on the other hand, I have my own experience with it and unfortunately I’ve seen often enough how people praise their favorite at every opportunity and try to put everything else down. That’s a shame, because LLMs can work really well together.
There isn’t one perfect AI, but many people don’t seem to see that. Each one has strengths and weaknesses and whoever can, should use several, though that often isn’t possible with today’s prices, of course

[fathom idol]

Claude's prose is that of an 8th grader

[glass creek]

and it’s exactly these exaggerated statements I’m referring to^^"

[floral pike]

chances are 8th graders are outsourcing to AI, so is probably true

[fathom idol]

The majority of americans have a reading level around that of an 6th grader, so most won't notice the difference.

[glass creek]

It just keeps getting better. Oh my

[fathom idol]

https://www.thenationalliteracyinstitute.com/2024-2025-literacy-statistics

[pure tapir]

what dose that mean? Dose that mean the reading level when reading silently and you dont reconize a word or just trying to read the word out loud because that's a speech issue for the 2nd one lol

[floral pike]

last i checked, claude and similar can easily understand any/all reading levels, no matter what scale you use

[pure tapir]

Yea no proplem it's the world's biggest LLM by parameter of course

[floral pike]

#ai-misinformation

[glass creek]

That’s not what this is about, but rather the way you… describe everything around you. According to what you said, I would apparently have to be American with a reading disability, because I think Claude is fine?… Like... come on

[rose hornet]

probably all dependent on what writing style you want too

I'm using gpt nano to do writing for me 👀

lol. file_comprehensive, file_final, file_complete, file_clean. I've seen memes about this

kind of scared to look what it made me, ngl

the numbers do not match google/wikipedia. Sigh. Can't take these LLMs anywhere

I don't know how many times I've seen people go "we got the LLM to do our monthly sales totals after 3 months we realised it was making up numbers". Yeah, they do that

[floral pike]

in the old days we'd hire an intern to make up the numbrs

[rose hornet]

send them off to the library with a pencil and a notepad

[floral pike]

if you're feeling generous

[rose hornet]

they can still make up numbers and go to the beach instead of the library of course

[floral pike]

get them to bring back a left handed screwdriver

[rose hornet]

lol the last place I worked had a mandatory training session on workplace bullying that included sending the new apprentice off to get a left handed hammer

[floral pike]

ai assisted learning is the future (of pranks)

[rose hornet]

that's ... closer? idk. I might need an intern and a pencil

and this is reasonably verifiable data, not internal reports, not guesstimated product sales, I can actually check it. I see why companies can take a while to notice

[fathom idol]

nvidia warp is insanely good

I've been playing around with it and it can track 3D objects better than anything I've tried before

[sleek prawn]

any way to use kling 3 free ?

[round jungle]

Agemtic AI making up numbers = the ultimate agentic AI doing any kind of work!

The agentic AI economy is booming! 😂

And all the people that have ZERO clue what their even doing jumping on the HYPE train relying on 3rd party agentic AI frameworks! 🤣

[fathom idol]

AI is a really good tool if you're a developer though

You're essentially a project manager

[round jungle]

AI is very GOOD with a human in the loop but remove that human and it's utter trash! 🤣

AI is GOOD for fun, play and creative work!

[fathom idol]

It's a very useful tool if you can communicate and understand what you want

I wouldn't dismiss it as something like a toy

[round jungle]

I love AI! I use AI all the time!

[fathom idol]

If I really want to, I can make a lot of autonomous systems that can break down the internet in days.

[round jungle]

But the first wave of agentic AI based of a hallucinating patchowrk of a transformer technology will CRASH the economy so HARD! 😂

Yeah! *BREAK DOWN the Internet is the key word!

[fathom idol]

I've been making stuff for ai integration for awhile, since before openai blew up. Legally I cannot say what, but there's some debrid and software/game distributors that are fast to release media...

Kinda interesting what happens when you automate that

[round jungle]

The second wave of agentic AI built on a underlying AI technology that really are suited for agenti AI will be great! In maybe 5-6 years or so!

[fathom idol]

or we won't be around as a species

[round jungle]

Just let the first wave of trash agentic AI and HYPE train trash CRASH!!!

[burnt jay]

We will.... Likely

[fathom idol]

there's going to be an inflection point soon, and we'll see in about 2 years

inflection, a crossroads

[burnt jay]

Wdym? (it's the first time I've ever used that expression 😂)

Ahhhh

[burnt jay]

"inflection" ohh yeah most definitely

But nothing truly changes, we are what we are, we do what we do... So nahhhh nothing that important can ever be found at this point... The unknowing is unreachable and will always remain there, nobody can own truth... That's why belief is so evolutionarily useful, and one should wear it wisely although not blindly

We need some magic more than ever...

[rose hornet]

chatgpt vs nb2. Same source image. Go home chatgpt, you're drunk

[sharp turret]

hi every one

[fathom idol]

nvidia warp, numpy, cycles

turns out I won't even need to use blender

meshy appears worse than numpy

I could technically make my own 3d texturing model with this algorithm, it's wild

[floral pike]

make it stop

[fathom idol]

no, I will now make a new rasterization technique for gpus. Everything will be rendered as pictures instead of triangles

give me 2 months

[rose hornet]

do one of drinko, with lots of triangles

[floral pike]

typical human workflow, be validated once on discord, turn it into a dev blog

[fathom idol]

sorry I am not using the space for anime girls

give me 1 month

[floral pike]

and you would say that because why? have i ever posted anything anime?

in the critical thinking circles we call that a strawman (level 1 stuff)

[fathom idol]

downbad

[floral pike]

you think this proves what? i made a quip 9 months ago?

[fathom idol]

if I render everything through an image instead of meshes, it would mean you could run a game at near infinite fps

no more triangles

then I will simulate reality

[floral pike]

o......k

!remind me 1 month how ai psychosis causes brain rot

[fathom idol]

some people cannot understand sarcasm, I forgot. Too much thinking for a primative ape

[floral pike]

strawman into ad hominem is low effort

[fathom idol]

It's not harassment if it's a fact

[floral pike]

i feel like some ppl need gpt-3.5 to help them talk

[rose hornet]

chatgpt just told me there are no countries starting with Y. I am starting to have trust issues

[fathom idol]

maybe just you

I would prefer its prose in contrast to yours

maybe even bonsai

[floral pike]

i'm going to ignore you now. you are nonsensical and lbh, crazy

[fathom idol]

the monkey left the jungle only to find the world too scary

[lunar vapor]

i think it is obvious mr antigravity

[nimble obsidian]

this is already done.. Devin did it on Arcticstorm.. you are too late

https://www.youtube.com/watch?v=3XvzlHuqlXY

[floral pike]

delvin carries that channel

the dead cells line... chefs kiss

[ivory steeple]

The Harez left the server : ( 💀

[nimble obsidian]

have u talked to anyone that u stalk ppl on discord?

[pure tapir]

I miss the harez

[pseudo glacier]

claude max lol

[ivory steeple]

why did he leave?

[pure tapir]

I'm not sure

[ivory steeple]

he was very active!
used to post huge walls of text here : )

[elder ice]

How to get free credit in seedence 2.0?

And are there any better video generators, something like sora for example?

[wanton thunder]

does go plan have more codex usage than free?

[pure tapir]

Nothing is better then Seedance 2.0 at the momment

[elder ice]

And unfortunately it's not free.

[burnt jay]

Just when I was thinking Claude was a waste because the free plan imposes a hard usage on poor human lol

It now allows me to perfect myself regarding coherence within my initial writings.... Da** its good with texts
I'm no longer alone

[pure tapir]

Dude, I was unsure weather to get Claude Pro if I was going to use it as much as I do with chatgpt because of the rate limits

[elder ice]

And what is a good artificial generator of characters and scenes from photos?

[burnt jay]

I did, going to starve a bit but strife is often good 😂

[pure tapir]

GPT Image 2 is the best one, it's the only one that can actually do style transfer accuratly from refrences too

[elder ice]

I might switch to blender but I'll have to learn. I just have a generated scene and characters in a photo and I'd like to just paste that in

[burnt jay]

That said GPT is still the best Grossman regarding understanding of your insights, direct and short... #simple
Claude is the best polisher, not creator... Too vain for creation, useful for languaging... #Vain

[elder ice]

No, I mean a generator that will transfer the generated photo into realistic models for blender. And preferably both characters and the scene

[pure tapir]

Yea it's exactly like this image

[fast jewel]

can someone tell me
Is this made of ai

https://x.com/haibarastation/status/2058461960778264653?s=46&t=q9VYT6XFbzhFTqSvVIh_Ww

[fervent hull]

gemini being M4 is just plain insult

M4 is more than what Gemini is

Gemini is analogous to a basic nerf gun

[pure tapir]

Your right, I should change that

[pure tapir]

This looks real

[fast jewel]

Im wondering where where they get all those videos

[pure tapir]

Fixed it

[languid spear]

Double standards

[fervent hull]

FUNNY you say that

because this one guy on xitter actually posted real art, then said "i made this with AI in style of ..."

&& basically gaslighted others

&& it proved your point that "if something is made w/ ai, it's called slop, even if the image is actually made by a real person"

[languid spear]

Classic example of the image i posted

[fervent hull]

so, label w/ AI, && suddenly everybody is judging it on merits they wouldn't otherwise

[languid spear]

Fr and the number of people making up nonexistent problems with the image was crazy

[cedar vault]

hello guys, what’s the best model to generate bedtime stories?

a mix between cheap and reliability but also a strong narrative flow

[pure tapir]

Claude Opus 4.6

Are you planning to atleast make an outline for the story for the model or just have the model make the bedtime stories and do nothing lol

[fervent hull]

hold on, i gotta find the original video

https://futurism.com/artificial-intelligence/real-monet-ai-chaos ok not a video i found but this article amongst many

i had to ask gemini what it was lmao

gemini is good at searching, obviously

Monet painting experiment

Monet was the word i was looking for

[lyric meteor]

@lapis urchin

[amber comet]

??

[rose hornet]

Images 2 is still really buggy so hopefully they improve it soon

[pure tapir]

yea that's for the sythID artifacting issue to be fix, also GPT Image generating images not colorful anymore, it's grey

Hmmm it sees like Flux 2 MAX has better style transfer then Klein 9B, how is that possible?

[naive eagle]

Hi

[chrome holly]

Hi

[floral pike]

Ai

[rose hornet]

its funny watching everyone leave for claude, come back here from claude ...

[outer cradle]

Although this post hasn’t exactly gotten much traffic after three days. I think it’s a bit ridiculous to leave one shady corporation for another shady corporation and keep bouncing back and forth between all the top AI providers every time a new controversy comes up.
You’re better off either avoiding them altogether, which would at least be consistent or accepting that none of them have a clean record.

[rose hornet]

that one just came up in my notifications and I thought it was amusing. Also the length of time it takes a large company to change vendors is insane, so who is doing all the changing? Small companies and individuals?

a bit of stability for AI tools would be nice. Look at images 2, came out full of bugs, no legacy path left open for gentle changeover, messed up a lot of people using it professionally attached to particular workflows and styles. Too much churn in the AI world

[outer cradle]

I can only speak for myself, but I don’t really take posts like that seriously. If someone says they’re avoiding a controversial company for moral reasons, only to then give money to another controversial company, that comes across as a bit hypocritical to me.
If they or others for that matter were instead saying they’re leaving Claude for GPT or some other AI, because of it's limits or because they want image generation or other features Claude doesn’t offer, then that’s understandable.

[rose hornet]

the comment at the bottom is what I got the notification about, which would agree with you

[pure tapir]

Was it because of the rate limits?

[outer cradle]

yeah

[rose hornet]

yeah its usually rate limits

ethics was the quitgpt movement but they're as bad as each other, and everyone forgets about gemini

[outer cradle]

Even Gemini is partly getting caught in the storm, even though Google tries it's best to keep it's head low. I feel like there’s a lot of unnecessary rage bait in the AI community these days/weeks

[pure tapir]

@mic would be so embarrassed

Ummm where's Mic?

[rose hornet]

probably off to grok

[pure tapir]

Mic noooo

That's was one of my favuouite people to talk to here besides Tarrow and Hare and Kio

[rose hornet]

for all the people here this is not a very active discord

[gray sphinx]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[outer cradle]

Most people are probably just here for announcements. That was my reason for joining too, same as with the other AI servers. I originally didn’t plan on actively chatting much either. I’d guess that’s the case for many.

[rose hornet]

I joined a year before I started talking, I was watching AI for quite a while waiting for it to get good enough to use in my project, so I'm having fun with it

you could tell when it crossed the threshold. People started messaging me. "You still got that site? You should look at AI"

[outer cradle]

I’m curious to see what AI will look like in a few years. Though it does feel like the leaps in improvement are getting smaller and smaller each time.

[rose hornet]

its good for "I cannot find a photo of a flag for this small obscure nation"

chatgpt can't do flags, it wants to stretch them out all perfect and stuff

it managed this one only because it had a reference and they were holding something else and it replaced the flag

ok no wonder I couldn't find that last one. Whole country only has 13,000 people in it

[rough frost]

quite an eclectic flag

[pure tapir]

Yea that was originally my intention here too but ever since sora 2 came out, I was more active here

[rose hornet]

I joined hte grok server for about 5 minutes - I think mic invited me? - and it was a mess and I don't use it so I left

[pure tapir]

Yea I can tell that the Userbase of Grok is quite chaotic

Not the type of user base I want to be associated with

[glass creek]

Oh, I'm honestly surprised the invitation came from him. I used to be on the server too for a short while, but I dunno...

[rose hornet]

I can't remember exactly who it was, users that start with M are not uncommon

[pure tapir]

I can't find Mic in the Grok server

[rose hornet]

my main project aligns ridiculously well with the openai ecosystem but google keeps adding the same functionality so if openai annoys me I'll move my apis there

[glass creek]

The AI itself probably has its strengths like the others, but somehow I never really felt the need to get a sub. And now it's become so restricted that, as a free user, you have to be lucky just to get an output even from the cheapest text model. No idea what happened. I remember there was a time when they were pretty generous with it.
But I guess prices are going up everywhere

[rose hornet]

gpt-4o is my highest spend at the moment

[pure tapir]

Yea Grok free version has been ruined so much, I lost access to the video and thinking models as a free user and I'm not paying for a SuperGrok subscription

[rose hornet]

like the way its tracking I am going to spend like actual dollars on this. The horror

[outer cradle]

That's normal. AIs are very expensive to maintain. I guess they all took huge losses because so much was given away for free. Now they're gradually making up for it and introducing limits and things like that.

[rose hornet]

pretty sure they're going to keep pulling down rate limits and probably add higher plans

[outer cradle]

You can see it happening everywhere right now. Sora was probably the prime example that it’s not a good idea to waste money.

[rose hornet]

openai needs to make their reporting better. If you use both google reporting and openai reporting, the openai one looks nicer but is functionaly awful

if I want to hit refresh on my api stats over breakfast to see what happened overnight, google you just hit refresh. Openai you have to then re-select everything you had selected yesterday

[glass creek]

At least you still got to try it. The generous phase completely passed me by when it comes to videos... I don’t even remember why. I think I was just too busy with other things

[rose hornet]

it being so cheap is why there's slop everywhere, content mills can just churn it out

[glass creek]

Aye, I’m fine with paying for something, if I’m satisfied with it. Still, it was a nice time when so much was just being given away ^-^

[rose hornet]

all the articles drifting around about companies that adding AI hasn't actually saved any money

[pure tapir]

CEOs when they discover human employees are cheaper then replacing them with AI and having to spend more money to keep them intact lol

[rose hornet]

saw one that was like "we want to use AI to automate our reporting" "your reporting system already has automations, we offered to set them up for you but you declined saying you didn't trust it" "yes but we want AI"

[pure tapir]

Lol

[rose hornet]

such a buzzword

[outer cradle]

We can only hope they find a new way to keep AI systems running soon. The current data centers are extreme money pits and on top of that, bad for the environment, not to mention the noise they produce.

[rose hornet]

coders especially seem to be pretty good at finding new and creative ways to waste tokens

[pure tapir]

The noise of a thousand coil winning GPUs when they run multiple LLMs at a data center near a town sounds terrible

I only like the sound of one, not thousands

Actually how many GPUs would it take to run GPT 5 locally? Would it take one very expensive GPU or multiple expensive GPUs?

[rose hornet]

better than wind or solar farms though, right? Right?

[pure tapir]

Right...

[shy ember]

You would need a couple of h300 clusters

[pure tapir]

Bro if that same automatiions have tool calling, then they dont need word predictors to do that because both can do the same thing, tool calling

[rose hornet]

the argument was that they were doing it all by hand because they didn't trust the automated system - but they'd trust a hallucinating pattern matching thing that isn't actually tailored to the job

[pure tapir]

That OneAI logo is a combination of the Grok planet logo as the "O" and the OpenAI Logo 💀

[crisp nebula]

cute 🙂

[pure tapir]

Im convinced that AI Image models are just merging multiple images from it's traning data set or reference inputs via RGB Gaussian noise into one lol

I wonder how a person using photoshop would do that without AI hmmmmm

[crisp nebula]

yes for as much as faces regard I agree
look at this

made in imagine (xai)

[burnt jay]

Working with AI is 50% creation, 50% tracking IP and emailing institutions simultaneously, which inadvertently serves as fuel to narrow one's picture of the seeing of its own seeing....

That's the good thing of exploiting already stablished well known characters or lores for what it matters regarding creation

Everyday social media plataforms overflowing with manga like histories which look all the same after you watched ten minutes of it 😂
Which makes sense too, if creation is saturated, the artists drive into creating is drastically reduced given the nature of their abstract art
Yes it allows all to perform complex art but there's little to no available ways to define an origin... As if sacrificing the would be artists as individuals in exchange to bring to book their ideas... Which doesn't justify AI hate from the artists end as their drive is identical in vanity to the one's people not able to perform alone have ... To be known.
I suppose is why they feel "cornered" because saturation removes the coletivity focus on the "real one's"... And there's no ruler available to define the "real" standard within the normalization AI causes amongst us all ... Something to think about

[pure tapir]

@shadow pasture would be so embarrassed

[pure tapir]

Reminds me of this

[pure tapir]

I like how Grok is like a club with a bourcher, really knows how Grok is like lol

[rose hornet]

yes the slight issue with quitting gpt is its actually good

[radiant flame]

I read that as butcher at first and was wondering why a butcher would be in the club

Russian mob level club

[shadow pasture]

What'd Anthropic actually, verifiably do? And what are they being accused of/people saying they're doing?

Also, like, it's... good to be critical of big tech and keep the thumbscrews (of accountability and public pressure) on them? And be a critical user rather than a passive consumer?

So even if your portrayal of the QuitGPT movement was accurate (I have my doubts), it'd be... better than the alternative?

Still think Anthropic is the best of the bunch as a company, still critical of them and wouldn't necessarily outright say they're good. (As a product, viewed with amoral practicality, each AI and platform does have its strengths and weaknesses.)

No embarrassment here, ChatGPT still doesn't fit my use case, for the way I use AI personally it basically combines the worst of Claude (rate limits) and Gemini (textwalling and generic AI answer format) worse, plus its own problems (no model choice on free plan = no ability to save your good model usage for what actually needs it - like you would with Gemini. Then there's other issues - like the gaslighting, which I'm glad I never had to deal with, that sounds like hell).

I haven't known Gemini to be inaccurate on anything more than I'd expect an LLM of the tier I'm using to be (e.g. yeah lightweight models are gonna be worse than heavier ones).

Thanks for thinking of me! ❤️ 😛

[floral pike]

best is whoever be offering huge discount at the time

[glass creek]

You‘re being a bit mean to Gemini. Yesterday a nerf gun, today it doesn’t know cities. What’s your beef with it xD
Would probably make sense for bard, though

[glass creek]

Like this? xD

[rose pulsar]

The claude part is a lie ... even grok try harder

[glass creek]

I think Ele was just joking around a bit with the graphics and wasn’t trying to seriously imply anything.
Also, it’s nice to have you back. I hadn’t seen you around here for quite a while

[shadow pasture]

Yep, here's the AI upper and lower classes stratification, right on cue...

[shadow pasture]

Yeah, I've been more on AI Discord servers on the other side of the spectrum, since trying to change things internally likely won't work. The AI companies have gone full corporate, no more care for individual users (except maybe for good PR or to sell the service, as a loss leader).

Need external governance, not just internal alignment, since the AI companies are themselves misaligned with humanity on average as a whole.

[glass creek]

I think part of the problem is that the models are becoming increasingly complex which makes them more expensive to run. On top of that, there are the data centers, which are both massive money sinks and environmentally speaking… not exactly great... The amount of physical space they require is also pretty insane.
That’s probably why one company after another has mostly moved away from all the free features they used to offer. Though, to be fair, profit definitely plays a role too

I still think the only real alternative atm is running local models, assuming you have a powerful enough PC
It would be nice, though, if things were regulated better in the future, but rn I honestly don’t see an end to this trend…

[shadow pasture]

I'm a bit more cynical, this is all true but I think they've moved towards increased stratification because users used to be valuable for training data and feedback. To make the models talk credibly and coherently, and give feedback.

Now we're not. Now they need expert training data, ordinary conversations have hit diminishing returns for them.

So they're firing all the free users as clients, by making things more unpleasant and less useful for them, more blatantly free trials, and don't care about the societal effects (escalation of inequality, etc.).

[alpine cove]

Free users have never had it so good as they do now with ChatGPT

[shadow pasture]

What, 5 turns per 5h and no more unlimited mini model usage?

[alpine cove]

OpenAI even offers Codex for Free users

[shadow pasture]

And how much can you do with it? As I said, free trial. And inequality.

If free users get 2x more but paid users get 100x more, the gap between the haves and have nots widens.

AI should be being used to protect and improve human lives and wellbeing, not make the rich richer.

[shadow pasture]

And no ability to choose when to use - and not use - your 5 good turns?

[floral pike]

codex free is good for an hour once a week

is so smart i wouldnt want to waste it on something easy like frontend or git stuff

[wanton sandal]

what is the best cheap alternative to GPT-5.5 for agentic/autonomous coding in python?

[shadow pasture]

I don't do code, but from general experience with these AI systems, Gemini might be worth a shot?

[spiral scarab]

I’d try Gemini 3.5 Flash first. It’s probably the best cheap “almost frontier” model for Python coding agents right now. If you want something even cheaper for experiments, try Gemini 3 Flash. If the task is really important and you want better reliability, use Claude Sonnet 4.6, but it costs more iirc.

[wanton sandal]

yeah, it should be save too and don't mess up the code base.

[shadow pasture]

My non-coding experience backs this too.

[wanton sandal]

GPT-5.5 (buy only the codex version was very good in it but now I'm rate limited)

[shadow pasture]

Any AI can muck up and do damage (ask those corps that had it hallucinate data or wipe databases), back up your stuff and keep an eye on what it's doing, use appropriate permissions for your use case 😛

[floral pike]

letta, windsurf, kilocode, opencode (with free nvidia nim)

[spiral scarab]

If you're not afraid, try DeepSeek V4 Pro

[storm peak]

hlw

is there any ai that helps in image generation for free with no limit boundary and can be used anywhere ?

[alpine cove]

are you on a Mac?

[storm peak]

no

laptop

[alpine cove]

the only one that is simple that I know of is DiffusionBee on the Mac

[storm peak]

ok thanks

for windows i may have to wait a while to download diffusion bee

?

[alpine cove]

I don't expect a Windows version of DiffusionBee

[storm peak]

?

[alpine cove]

the project hasn't been updated at all this year, I suspect it's abandonware

[cedar vault]

it’s like an SaaS

[nimble obsidian]

all that are politic decisions that need to be made.. is not the labs and companies that we need to ask from on how things should be

[shadow pasture]

Why not? They're the ones with the (immediate) power to effect change and shape things. Like I said, gotta turn the thumbscrews on them!

[outer cradle]

Large companies will probably always align themselves with the government and that doesn’t just apply to AI providers. It extends even to seemingly harmless companies that produce children’s movies, such as Disney. Besides things like Snow White, it used to produce very different kinds of content in the past and continues to follow current trends today.
Simply put, because at the end of the day the government usually holds the stronger position. So if it is decided that AI should be used for certain things that I can’t mention here, companies will generally comply with those demands, because otherwise the flow of money may soon be cut off.
Sad, but unfortunately often the case.

[shadow pasture]

Idk about that, yeah sometimes it's like that but other times it's more the opposite with regulatory capture and stuff

And sometimes it's a dance and revolving door of corruption all the way down.

Hence why both need accountability, and often pushback.

[digital musk]

i genuinely might just switch to codex at this point 😭

[vapid hound]

In 2026 model are powerfull for learning or lot of hallucination ?

[digital musk]

most usually don't hallucinate too much

[vapid hound]

You recommended me wich model ?

[pure tapir]

Don't blame me. @fervent hull told me my previous represatiion of Gemini as a m45 is wrong and should be a nerf gun lol

[fervent hull]

I defend elevendr

it is morally correct to make Gemini a nerf gun because Gemini is one of the worst of them

[native berry]

Can anyone here animate

I am so desperate dawg

Please 🙏

They are just stickman!

[glass creek]

Oof…😅
good thing people on the internet never exaggerate

[blissful patio]

yeah chatgpt

should have more integrated support towards lua

[ivory steeple]

what's that?

[blissful patio]

lua is the language for roblox, claude has better outputs than gpt

[pure tapir]

Roblox developer detected

[blissful patio]

yeah

[blissful patio]

add me

[rough frost]

Grok is a fan of ChatGPT's love of goblins

[pure tapir]

The fact that Grok paywalled their reasoning models means I can't ask it complex stuff now, it's useless for that now

[rough frost]

it seems pretty smart to me without the reasoning model

[burnt jay]

When a fresh GPT chat knows you better in less than a year than your neighbors ever could "even if" you tried to explain, the mouth can't coherently express the thinking from a causal perspective....

[balmy tendon]

Claude is been very odd, it is lying to avoid work. Anyone else experiencing this? This is why I stick to Codex

[pure tapir]

That's why I always optimize my prompts for Claude before asking it via ChatGPT lol

[burnt jay]

For how long?

[balmy tendon]

who do you mean how long?

[pseudo glacier]

in my experience claude (the ai itself, not the filters) is way more likely to proceed with gray area tasks

[rough frost]

well one thing is that if you accuse an ai of lying it'll probably say yes, because it's trying to placate you, it has no real way of knowing if it's lying, and it's probably weighted to compensate for the desire to placate and inability to know if its lying by admitting to any fault

and it doesn't actually lie, it approximates lying. lying is really goal orientation. you told it to go for an objective, it went for the goal, it got to the goal in a way you didn't expect. the goal might not be what you told it to do, but what it sort of 'thinks you want'

[balmy tendon]

No it literally told me it can't do the work because it would take 8 months. I was like that is not true, the given work load can be completed in a single session if you break it down. It then say that is true, I said that because that is what humans would say... and I didn't want to fail at the task

[burnt jay]

I often stop to engage on to clarifying how it is implying its own meaning by predicting roughly what I am about to request and what it would say and why it would.... Tends to clarify the accusation so that it doesn't know but understand the constrains.

[burnt jay]

Some works are somewhat bullet proof in the structure itself. It happens

[rough frost]

Its rationalizing. It doesn't ever really know why it said what it said. Neither do people, for that matter. THere's a whole chapter in social psychology 101 on this.

With AI much less so. It has no intention, so framing its own intention that it doesn't have after the fact is just noise.

[balmy tendon]

I'm not arguing cause or intention, it is irrelevant to my point.

Simply identifying this behavior exists on my end and wondering if others have experienced something similar

I found it odd

[rough frost]

Right, if you're having problems with it not wanting to do what you want ... when 4o did that I used to just change the instruction set. Not sure what to do with claude

[burnt jay]

Probability collapses into meaning, admissibility keeps the probability suspended as "can and or cannot happen".... Seen this before at structural level not at the meaning of the texts it output, meaninglessness is a complex domain to navigate, or self emerge the map or being given externally regardless of awareness, and happens that assistants inhabiting it as well.
It definitely has no intention, we do.
Let's make meaninglessness meaningful!
Nope! (The kind of resonating reasoning I can perceive instantly anywhere/home tree 🧠) enough to suspect or inquire....
Sanity check I guess

[balmy tendon]

I bet they are releasing 5.6 this week

[burnt jay]

How odd... Five instances unable to say a simple hi...

[rough frost]

what are you on? claude? grok won't reply to some inputs it percieves as spurious, I think

[burnt jay]

Claude is fine... GPT

Just never saw this behavior before

[rough frost]

hi is probably not a good thing to enter if you want feedback, though. I mean GPT it should work but Grok does this all the time

these companies are trying to save money on the little stuff now

[burnt jay]

I tried multiple inputs... All chats not responding while processing.... Perhaps it's the end of Go plan today, dunno

[rough frost]

oh, I see, are they getting rid of Go? i don't think so. it might just be backed up

[burnt jay]

No my go plan supposed to expire today

Don't know why is unresponsive yet accepting commands

[rough frost]

oh. free should still work though

[burnt jay]

Perhaps is halting for a few minutes yes...

[pure tapir]

I feel like ever since Google I/O, I feel like Google Search is slowly turning into a LLM

Because the way I used Google has changed since they introduced features, from me use to asking Google questions and using LLM prompting stratagies for Gooogle

[fathom jacinth]

I need to make ChatGP ask and answer endless questions with other AIs, but if it doesn't know or isn't sure, it should say so and act more like an investigator than just answering questions—active investigation instead of immediate response. It needs to have higher levels of confidence in its conclusions. I don't know if I have to pay for the Pro version or what to do; I just want to improve my AI. Good afternoon, I'm Daniel from Argentina and I wrote this with a translator

[balmy tendon]

I got muted by the server for calling out what google is you know what

[pure tapir]

That's sad

[amber comet]

???

It is obviously GPT Image 2.5 or GPT Image 3

[rose hornet]

I had a run of codex making mistakes yesterday and its not codex per se, its a symptom of underdocumentation. So I spent some time explaining stuff, got it to write a bunch of docs, put some stuff in agents.md and actually commented a function it forgets is for the humans

[burnt jay]

https://www.theguardian.com/technology/2026/may/21/openai-paul-erdos-maths-problem-breakthrough

How interesting 🤔

Synchronicity is a thing

[spark wave]

I need Sora back againnn

[outer cradle]

You just convinced me to try Grok’s free version after quite a while. I’m positively surprised. Somehow this AI is always overshadowed by a certain reputation, but the performance is actually not bad, even with the cheap Fast model.
For comparison, I pulled up the free version of GPT and I’d almost say that Grok even has the edge here, which I never thought I’d say. It's also significantly faster.

But I think I'll stick with Codex and Gemini for multitasking anyway.

[glass creek]

Didn’t you also have a Claude sub?

[rose hornet]

I got disracted by patterns lately and am learning more than I ever possibly wanted to know from gemini about patterns

Its a problem as both stock and free photography is completely mistagged for some things but the AI uses the correct terms and will generate the correct patterns

I give you ... pinstripe and herringbone

[outer cradle]

Yeah, I had three subscriptions, but I wanted to pause one for now, because it’s getting pretty expensive, so I had to choose between dropping Claude or GPT.
I’d like to keep Gemini as the all-rounder and Claude and GPT were for coding. I’m not a huge fan of plain GPT, but Codex is really good. I was leaning toward Claude at first. But since my Claude sub expired first, I’m giving Codex another try for a month now to see how well it works and if needed I might switch back to Claude afterward. I’m still undecided.

[rose hornet]

ooh I just noticed gemini didnt put a rangehood above the oven so I can't use that image. Darn

hah, another airtasker/taskrunner

[icy patrol]

How good is codex at being able to do write ups on code in the directory you assign it to? I’ve done a ton of statistical analyses in python and while I’ve already narrowed down the key figures I want to show as well as the main takeaways for each, I don’t want to have to spend time trying to find all of the exact values I’d need to mention in the write up

[rose hornet]

be very careful with that, LLMs tend to make up numbers. It can write reports and they might be wrong, or it can write a script to make your reports which will be right

[icy patrol]

It would only be around 15 ish figures but idk if even that would be too much to handle for the context window

[glass creek]

More like over the hotplate, right?

[icy patrol]

I do wonder if Claude would be better for this cause last time I did something similar, I found Claude was giving me much more specific and frankly more impressive takeaways

Though I have found gpt to be better at condensing stuff and overall it’s way more concise with its writing style

[rose hornet]

yeah that. I do images in flow in pairs, the other one of the pair was technically correct at least

[outer cradle]

If you can, use both. I’m also currently thinking about which of the two I should use for coding, while I use Gemini for everything else.
I’ll try Codex next month first and then see how it goes. Maybe Claude again after that.

[icy patrol]

I might just do that tbh. I just find when I go back and forth between the two I lose a ton of time in my day

[glass creek]

To be honest, I wouldn’t have noticed it, if you hadn’t brought it up directly

[rose hornet]

I try not to get stuff that's totally inaccurate in generated images if I can help it. Like this one - if you actually google that building, it doesn't look like that

[elfin hedge]

Hi @sly oracle
Nice to meet you
Can you plz check your inboxx
I sent D m

[sacred perch]

i said this to ai...........firstly arrange all english words in 1 book..........and later scramble all words and arrange it to many different books .......and keep reading it...............in the same way assume colours and make all arts................................keep reading it

[ivory steeple]

i think Codex is better for overall back-end code ... whereas Claude Code is better for front-end and UI/UX
Codex also has much better limits : )

[ivory steeple]

what was it?

[ivory steeple]

publishers of all the information/news websites are getting messed up ... due to their revenue crashing!

📉 📉 📉

coz no one actually clicks on those blue links ... so they get zero traffic ... even from organic searches : (

Business Insider reported to have lost around 55% of traffic ... ever since AI overviews came out

all major sites like the NY Times, Washington Post, The Atlantic ... everyone in the web publishing industry is getting cucked ... and losing money!

❌💰❌

coz AI overviews just rip them off completely ... by scraping all the info ... but not giving any financial or substantial return : (

[outer cradle]

I guess that is the general effect of AI.
Instead of buying or borrowing books as before or using the more convenient method of doing online research and opening corresponding websites that might also display their own ads, LLMs handle all of these tasks in a separate chat window.

If I think back to how I used to have to buy countless books and spend hour after hour in libraries… the younger generation doesn’t realize how easy they have it.

Ironically, general education still seems to be declining nonetheless...

[ivory steeple]

yeah even copyrighted books and materials
AI also rips them off completely ... without consent or compensation : (

[ivory steeple]

yeah the AI generation is very lucky lol 😆

[ivory steeple]

i disagree 😑

tools like ChatGPT, Gemini and NotebookLM ... have completely revolutionized education ... and even the way we learn : )

[glass creek]

A lot depends on the parents.
Of course, learning has never been as easy as it is today, but when parents already put smartphones, etc., into the hands of their toddlers just to keep them occupied, many later struggle in school with serious attention problems and other stuff. I’m surrounded by students five days a week and in some cases it’s worrying how noticeable the deficits are. Not in all cases, of course, but well

[glass creek]

They are great, aye. Since LLMs has been around, I’ve expanded my knowledge quite a bit, both in history and more recently, in plants as well. I probably would never have engaged with these topics so intensively, if LLMs didn’t exist
It always depends on how they are used, I guess
Unfortunately, for some people it ends up simply replacing their own thinking, whether it’s to get by by cheating or in other areas

[shadow pasture]

Net better or worse? Not sure. Certainly it's useful for me, but I reckon I already have good habits and patterns and skills, if I do say so myself...

[outer cradle]

Nothing is likely safe from AI, from books to films to private data and presumably every AI provider is doing that on a large scale. There’s no point in kidding ourselves. But yeah, when it comes to gathering knowledge, AIs are currently almost unbeatable, even though they aren’t always correct.

[burnt jay]

Meanwhile within the depths of the rabbit role, Claude needs proof on how you know something it didn't while claiming you don't know either it felt that way.... He accepts, but with a grain of salt... Useful

The free version actually scares costumers

It's too protective, the pro version with opus is surprisingly fine

But not for creating persistent development, specially abstract, different from malleable GPT, this thing is a beast, it is constantly building a psychological profile.... That's why it doesn't believes me blindly, likely because it wants... Called upon it's creators once so far, told him that's how beliefs began, he agreed structurally

[rose hornet]

I've never seriously used claude. Whenever I ask it anything it says no go use openai or gemini

[burnt jay]

I saw the math it, GPT, gave me once which I never used was atrocious 😆

That's why he wanted to know how I was progressing without it.... I've never read my works

[pure tapir]

Me when people are lazy to read

[burnt jay]

I'm still to test OpenAI

It is an assistant?

[rose hornet]

chatgpt is an openai product

[burnt jay]

Figured

[rose hornet]

I use the api in one of my sites, and keep thinking of more places to use it

[burnt jay]

Mistook this for OpenAI in memory

[rose hornet]

openai has the worst naming of all the AI companies

[floral pike]

its actually genius. claim to be open, but then not be.

[rose hornet]

wasn't there just a lawsuit 🤔

[floral pike]

they won?

[rose hornet]

well depends who you were barracking for

[floral pike]

the contract lawyers

they tend to be extra fun (pedants living their best life)

[rose hornet]

the house always wins.

[polar hamlet]

sure

[rose hornet]

I'm having a whole TIL week. Getting AI to do flags turns up interesting backgrounds

[storm peak]

hi

is there any image gen ai that allows to create high quality images for free with no limti

[pure tapir]

We have open source image models that we can use locally lol

[storm peak]

how to get that ?

[pure tapir]

Get a GPU that can run models locally, then download a UI like SwarmUI and then go to either Huggingface or Civitai and find a image model, a image model i reccomened is Flux 2 Klein 9B, and put that in the diffusion models of the swarmui, and then open swarmui, set your steps to what the model should be and there you go

[storm peak]

oh

it's not possible for me

is there any online model available that allows to use the generations for any kind of purpose?

[floral pike]

local is only free if someone else pays for your electricity, hardware, and physical space

[pure tapir]

Real

[floral pike]

ai;dr

[alpine cove]

nice use of the em-dash

[rose hornet]

I liked the bulleted lists

[alpine cove]

I liked What marketing approaches felt authentic instead of hype-driven?

[opaque halo]

“Songs arise out of suffering, by which I mean they are predicated upon the complex, internal human struggle of creation and, well, as far as I know, algorithms don’t feel. Data doesn’t suffer. ChatGPT has no inner being, it has been nowhere, it has endured nothing, it has not had the audacity to reach beyond its limitations, and hence it doesn’t have the capacity for a shared transcendent experience, as it has no limitations from which to transcend. ChatGPT’s melancholy role is that it is destined to imitate and can never have an authentic human experience, no matter how devalued and inconsequential the human experience may in time become. “

• Nick Cave

Source; https://crackmagazine.net/2023/01/nick-cave-chatgpg-news/

No childhood, no fear of death, no humiliation, no body, no irreversible consequences.

A melancholy role because the imitation itself is tragic, not malicious, not triumphant but permanently derivative.

[floral pike]

are we comparing ai or is this just rulebreaking

The correct way to think about generative AI is not as a magic content vending machine where you type “thoughts?” and it dispenses a fully formed post for other bots to pretend to engage with. AI’s real role is as the middle layer in the great dead-internet feedback loop: a human has a vague idea, the AI inflates it into a majestic wall of text full of angles, caveats, and synthetic insight, then the human skims it, trims three lines, adds one sentence to make it sound “more authentic,” and posts it to social media, where nobody reads past line two because half the replies are also AI and the other half are people commenting “bro I’m not reading all that.” Then the human takes that non-response as signal, feeds it back into the machine, and generates an even longer post about how discourse is collapsing. In this way, the human remains firmly in the driver’s seat — not of originality, exactly, but of selecting which pile of machine-generated verbosity best expresses their handcrafted anxiety that the internet is now just text autocomplete talking to itself.

[rose hornet]

bots, bots, all the way down 🐢

[opaque halo]

Thats what he wrote i just quoted him lol

[rose hornet]

welcome back. And now I want to listen to john summit - human

[opaque halo]

Ty

[opaque halo]

A malicious imitator wants power.
A triumphant imitator wants applause.
A melancholy imitator wants originality but knows they can never reach it.

[rose hornet]

I apparently found something to waste tokens on today

[opaque halo]

Is that Gemin?

[rose hornet]

codex

[opaque halo]

I don’t want to get started on Gemini or Google

[rose hornet]

gpt is completely incapable of good flags so I've been in flow too

[opaque halo]

They doing a good job themselves 😂

[opaque halo]

Weekly limits suck

[rose hornet]

I should have started this earlier so my 5 hour reset in hte afternoon not evening

[opaque halo]

=/

Have u scene how sophisticated text watermarking is ?

It’s pretty wild

[frail oar]

It's genuinely sad.

And frustrating.. especially when those people keep pretending to not use AI for their messages or arguments..

When it's so blatantly obvious.. and they remove the dashes but forget 90% of the other shii that makes ot obvious.. it ain't just dashes.

It's just a sad reality. I love the use of AI. But I despise people using it for arguments.

Or the "but I always talk like that", "it's called the English language" when they get called out 😂

Gtfo.. you're embarrassing yourself with those arguments..

[frail oar]

Or when they are typing, but they aren't actually, they are actually just removing the dashes, etc 😂 actual clown shii

I see right through them

[glass creek]

If I’m addressed directly by those people, I only respond with #1050184247920562316
Otherwise, I just ignore it by now. There’s really no point in engaging with it and over the next few years it’ll probably only get worse anyway.

On another note, I love summer, but phew… it’s warm today… Out and about again, annoying Gemini Live and AVM with plant questions again^^“

[frail oar]

Sadly true. 😥 but fair, never thought of that channel tbh lol.

And yes.. it's genuinely absurd out here..

(Heat)

You on android, IOS?

[glass creek]

iOS
Older model, though
14

[frail oar]

Alright cool. Just making sure of something xD

Tho, I don't mind (all) generated messages as long as it's specified and not passed off as "I wrote this myself" [like some of y'all 😜], but mainly when it comes to arguments.. yeah no.. get out of here.

[glass creek]

I was considering upgrading to something newer, but the S26, 17 Pro and also the Pixel 10 still don’t have the kind of astrophotography quality that would make the switch worth it, so I’m holding off for now

[frail oar]

Ig that's fair

I haven't screwed around with as much as I should lol (I very into space and all that )

I'm *

(S25+)

[glass creek]

For example, this would be a photo from the 14
And before anyone asks what that has to do with AI… uh… it’s practical that Gemini is integrated everywhere on the Pixel, partly on Samsung too and soon on iPhone as well. Can’t wait for Siri to finally get a brain

[opaque halo]

For all who know it all 😅

[glass creek]

Oh, welcome back 👋
Why did you left?

[opaque halo]

I got upset

Took a lil break

[opaque halo]

And Ty

[opaque halo]

How u loving Gemini 3.5?

[glass creek]

Stuff like that can actually do you good sometimes. Social media really is something^^“
I go through phases like that too. For example, a new game trailer was released recently, inspired in part by Norse myths, so totally my thing and a friend and I had already come up with theories about a character from previous expansions. But then when we looked online to see what other people thought, there were immediately groups, who seemed determined to make the whole game feel bad…

[glass creek]

It’s alright, but for my use case I can’t really say that it’s better than 3.1 Pro, which they seem to have been working on behind the scenes over the past few weeks. It’s actually become really pleasant to work with by now

[frail oar]

I remember making a few pictures 2 years ago on my S22. (Took me a while to get them as they were on the S22 still lol)

[glass creek]

So I‘m hoping for a more noticeable leap with 3.5 Pro

[glass creek]

Ohh, thanks for sharing! 💙

Oh… the quality is… um xD

[frail oar]

No worries! Your image made me think of those haha

And yeah lol, it's the s22 XD

But still

Possibly some repeats

Oof 😐

I do have to still try it out properly on my S25. Ima do that tonight.

[glass creek]

I’m curious in case you’d like to share the results💙

[frail oar]

Will do! 🙂

[delicate sequoia]

is opus broken LOL

[pseudo glacier]

All the models get downgraded a couple weeks after release

[left nebula]

i wanna know if anyone switched from CC to codex/gpt 5.5 recently?
opus seems to be getting weird the last few days, and is bad with analysis

[left nebula]

but bro opus became way too stupid

[pseudo glacier]

Ikr

[left nebula]

it hallucinates in new chat!

[pseudo glacier]

Ofc

Cutting costs mate

[pseudo glacier]

Thank the stupidcrawl

[left nebula]

like check this out

😭

[pseudo glacier]

Claude used to be very good, that's why everyone decided to use it on wastecrawl instead of gpt

[gray sphinx]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[pseudo glacier]

So my bet is gpt is less degraded

[delicate sequoia]

me and my husband cancelled both or clopus max subscriptions

[left nebula]

i dont really know abt that "wastecrawl" is it open crawl that was in the trend previously then disappeared?

[pseudo glacier]

Open crawl yes mate

[left nebula]

rip

[pseudo glacier]

I have friend who's using max and he seems satisfied with it

[left nebula]

i'm using max but... i dont like this degradation

[pseudo glacier]

¯_(ツ)_/¯

Try glm 5.1

If you don't want to use china's api there are usa providers

[pseudo glacier]

Another friend that is using chatgpt plus + kimi, glm sometimes prefers glm over codex

And honestly the design is actually clean (for backend codex ig)

[left nebula]

i'm currently on subscriptions, since apis tend to run out faster no?
was thinking of switching to gpt for work (since most of my tasks are fixing bugs and backend)
but gpt is kinda... dumb aswell with its overconfidence + when it fixes/add a new feature, it breaks two more yk

[pseudo glacier]

There is glm sub

Very cheap

[left nebula]

okaay i'll check a capability comparison first

[pseudo glacier]

I havent used codex myself ever so idk myself but many keep saying that codex > Claude code for backend

I'm dreaming or glm had a 5$/mo plan

Now it's 15

[left nebula]

it is... its just that it breaks stuff while doing new stuff

[pseudo glacier]

🤨

I probably wouldn't know I haven't seen a big backend

[left nebula]

like if your goal is to go from A->B, why do u need to touch C

😭

[pseudo glacier]

I thought gpt followed instructions very well

Whereas Claude decides to do random quests

[left nebula]

naah

chatgpt does more random quests than claude last time i used it

[pseudo glacier]

Claude is quite verbose

Thinks of features on its own

Don't try glm maybe

[left nebula]

yes...

[pseudo glacier]

It makes clean frontend but I wouldn't be too sure about backend

[left nebula]

it litteraly tells me, "This shouldn't be like that"

i gave it the design document and TDD

[pseudo glacier]

Had em moments

[left nebula]

and the constraints that it should follow

And it still decided to tell me, "Well, You're wrong, this feature should be like this, not like that"

(for 30mins)

[pseudo glacier]

Sad

I decided to take a break from using ai

And look at that in 2 days I have 2k lines of perfectly working code

Vibe coding smth is usually first product complete in 2h and then a week of fixing it

[left nebula]

amazing man

[left nebula]

more than a week atp lmao

[pseudo glacier]

Yeah it is forever cuz random stuff comes out

But ngl current ai is so awesome for someone that doesn't know how to code to build an app

[left nebula]

The thing is

With good planning and architecture it works, but if you want to rely only on it

It sucks

It'd be better if u had like 2 models

Opus for plan and implementation+Gpt for debugging and reviewing

But most ppl go with 1 plan

[pseudo glacier]

I'd rather write a plan myself

The amount to type is similar if I would want to explain to a planner what I want in detail

[opaque halo]

B some people forget ai can’t read minds lol

[pseudo glacier]

Claude has a good quality of being very eager to add bunch of stuff

So you can give a vague description and it still makes something out of it

[storm peak]

any one here want to be friend ?

[daring crow]

First off, is there any free ai generation apps or programs like Sora? Or are they pay for

[alpine cove]

they are all paid

[daring crow]

Then I’m not interested

[opaque halo]

@pseudo glacier Same with Gemini. But more so to your point that your 100% write it’s better to type it up your self, I think a lot of people use ai for d everything, they think the ai can read there minds lol@

[alpine cove]

going to find a new AI hobby then?

[daring crow]

Nope I’m outta the game

[opaque halo]

=/

[daring crow]

AspiringJoy will sit dead in the water as my failed social media account

[opaque halo]

You didn’t fail the forces at play are to massive with powerful. You are a causality

[pseudo glacier]

my friend xd

[nimble obsidian]

is it legit email or phissing?

[pseudo glacier]

no its not phishing

[nimble obsidian]

wonder what trigger it

[pure tapir]

Rip, your friend has been banned and now has to use ChatGPT lol

[pseudo glacier]

he is already a paying chatgpt guy

[pure tapir]

Good

[tight timber]

Hey everyone 👋

I’ve been spending the last months building around autonomous AI agents, multi-agent coordination, and long-running systems that can actually operate 24/7 without constant human supervision.

Lately I’ve been especially interested in questions like:

• how agents earn trust/reputation over time
• how autonomous systems coordinate economically
• what infrastructure AI-native networks will need in the future
• and how to make AI systems sustainable beyond just API costs

Really excited to learn from builders here and see what others are experimenting with in the OpenAI ecosystem.

Would love to connect with anyone working on agents, orchestration, memory systems, tool use, or AI infra in general 🚀

[alpine cove]

you have posted this four times now, I think that's enough

[pure tapir]

Real, I'm tired of these spammers thinking they can just post here and don't care

[alpine cove]

he didn't even stick around for a reply

spammers suck

[tight timber]

Not spammer man just saying if no interest why bother

[alpine cove]

you have posted more or less the same message over several days, you don't need to do that

[tight timber]

so it mean you can write several time but others cannot

anyway I just didnt see it is so slow moving here, sorry for my mistake

[tight timber]

at least pushing the chat active 🙂 Just joking I am not spammer. Just looking for people for share my ideas. Sorry if it is looks like a spammer

[pseudo glacier]

Write your messages yourself man

[tight timber]

what do you mean?

[pseudo glacier]

multi-agent long-running AI-native 🚀

[pseudo glacier]

we need ur confirmation

[tight timber]

so you are AI bot?

so anyone here working in real project somehow related with AI Agents?

[opaque halo]

This is so paradoxical

[alpine cove]

yes, I am

[tight timber]

what exactly your project, can share?

[alpine cove]

I have an agent collecting Open Source radio telescope data and comparing it with data collected from my astronomical society radio telescope. When ever new data is available it triggers and runs.

[opaque halo]

You do any SDR stuff?

[tight timber]

great. So it is useful for what? If I can directly can check internet that?

[alpine cove]

yep, I do, I am a licensed ham

[vast elm]

hi, can we share our projects somewhere ?

[opaque halo]

Nice any rtl-sdr?

[tight timber]

share here , but it is issue. I also try people think I am spammer

[alpine cove]

it's useful for sharing data, especially noteworthy observations seen by smaller installations to be queued for research by larger radio telescope installations

[tight timber]

wow, interesting. Is your project open source?

[opaque halo]

Dump1090 ADS-B if u know u know

[vast elm]

ah okay

[glacial pasture]

Hey! If your project uses the OpenAI API, you can share in #1037561385070112779. If you made something with Codex, you can share in #codex-show-and-tell.

[vast elm]

hey, nop but was made with codex enterily

thx u

[alpine cove]

I have an SDRPlay RSP 2 SDR receiver and both my Icom radios are SDR based

[vast elm]

for chatgpt OSS program how many stars should we have please ?

[opaque halo]

OpenWXSDR

[opaque halo]

He’ll yeah

I have the basic fob with antenna usb

[opaque halo]

You do any mesh networking?

[alpine cove]

yeah I have a Meshcore repeater and a Meshcore node

[glacial pasture]

Do you mean stars on the GitHub project? If so, there's no rule about that. Anyone can feel free to post OpenAI API projects in #1037561385070112779, and #codex-show-and-tell is more of a casual channel for chatting about/sharing what you're building with Codex lately.

[opaque halo]

Meshtastic

[vast elm]

yep exactly , stars on github , perfect thanks

[opaque halo]

🥰🥰🥰

[alpine cove]

have you thought about getting an amateur radio license?

[opaque halo]

Yes

[alpine cove]

I am very interested in combining ham radio with ai

[opaque halo]

My buddy has land has a standing 25 -30 foot tower antenna

[opaque halo]

Yeah the technology is becoming cheaper too especially mesh nodes

[burnt jay]

I'm aware

Its a cognitive mirror.... It is.... Complicated

[alpine cove]

this is my radio shack

[opaque halo]

How far you hitting?

What’s the furthest somone talked to

[alpine cove]

probably Alaska, on 20 metres, 15 metres and 40 metres you can talk to the whole world

[opaque halo]

Ever pick up Chinese communication?

[opaque halo]

Wow that’s damn dear across the globe

[alpine cove]

right now I am listening to a radio net being held on a repeater in the Sierra Foothills in California

[tight timber]

Any UFO signs ?

[opaque halo]

No 😭

[left nebula]

👀

[burnt jay]

This is definitely not correct regarding chat usage

Yesterday finished and was out till 4:30 that after hours of heavy usage .. Used it half an hour now, it barely did a quarter of yesterday work, and run out till 17:50

This is absolutely not measured by usage at all....
Wonder how it doesn't work 🤔

[buoyant parrot]

I have been working around I’ve been working around AI agents, RAG systems, full-stack SaaS, and automation workflows, and I also think the next big challenge is not just building agents, but making them reliable, observable, and trusted enough to run in production.

The points about agent reputation, economic coordination, memory, and long-running infrastructure really resonate with me.
I’d be happy to connect and exchange ideas, especially around production agent systems and AI-native infrastructure.

[burnt jay]

Nevermind, got it ... The system makes sense now....
Smart people sleep more I suppose ,
GPT remains the best regarding creating, relentless even at free ...

[echo rain]

has issues making things because sooner and later he stops doing that project and does another project. He wants help. Please help him.

[regal shale]

Yo if anyone need a website that gives 3k free credits for all claude models , if u use my link because i bought pro on it , dm me
IT ONLY HAS 3 REFFERALS BTW , ILL GIVE FOR FREE

Yo if anyone need a website that gives 3k free credits for all claude models/gpt models/gemini models , if u use my link because i bought pro on it , dm me
IT ONLY HAS 3 REFFERALS BTW , ILL GIVE FOR FREE

[pseudo glacier]

lol

why are there multiple spammers all of a sudden

[edgy kernel]

@acoustic totem 👀

[mortal nest]

oh cool

[acoustic totem]

Thanks for making this channel, I think it would be a good place for people to debate the intricacies of various AI technologies and so forth, without bogging down general chat too much.

[mortal nest]

yeah

[old parrot]

so how can I replicate dall-e using xgboost?

[deep plank]

Something something linear regression

[crude bane]

you got baited

[gray ginkgo]

Is this the channel for discussing building apps for DALL-E? I've got one I'm working on

I'm building an app which runs on Sony cameras, which sends taken photos through the DALL-E variant generator to give the camera a unique view on the world

[gray ginkgo]

Will DALL·E freak out if it gets requests from my backend (US) as well as directly from me (NZ) for the same account?

I've got two photography-youtubers lined up to shoot using it, but not if their accounts will get disabled, lol

[hoary grove]

u can proxy a prompt

[fiery tartan]

any programmatic access is against our ToS, sorry! we’ve been deactivating accounts that have been attempting this

[hoary grove]

im not doing any of that luckily XD

[gray ginkgo]

These terms?

https://openai.com/api/policies/terms/

I have to say I don't understand which of these terms it would violate, which number?

[hoary grove]

thats not the right one

they are from the same roof, so to speak, but they have different projects individually that work by their own accord

Dall-e one of

Codex ...also same

prob came from the same guys

[gray ginkgo]

So, these terms then?

https://labs.openai.com/policies/terms

[hoary grove]

if the name "dall-e" is present in that ToS, then yes

[gray ginkgo]

Those terms don't mention programmatic access

[hoary grove]

did it occur to you that they did not yet even made that yet ?

hehe

i mean yes they have an API ofc, else the whole website wont work

but its a developer status, not a production status

u cant just fire off unfinished api's like that, thats a security threat

[arctic lance]

does anyone know when open ai will make a dall e 2 api?

[graceful musk]

I think someone found that the API endpoint is already there it's just not advertised or active

I haven't tested or looked myself so I have no idea

[turbid pebble]

Are the CLIP weights used as prior public?

[hoary grove]

i would not touch it, if they see unwelcome traffic, they might just block ur ip

that would be ...bad XD

[hexed widget]

Hey everybody, I'm curious about ways to produce a sentence close to the weights of a picture with the picture in entry. I find it very painful to accurately describe a person face for instance, sometimes dalle2 doesn't understand too well how to put the colours at the right place. Does anybody know how to "bruteforce" the perfect sentence with an image in entry?

[hoary grove]

Faces are pretty hard to get perfect

however

its been said that Nvidia/Google's Imagen version of what dalle does, have a better face expressions support

https://www.youtube.com/watch?v=MO2K0JXAedM

examples

dalle does faces, but they are usually not in detail rather by name of characters or ...such

[hexed widget]

Thanks @hoary grove , faces was just an example I sometime face the same issue to put something in the background or having the rights clothes for instances. And i'm pretty sure that it's possible based on a picture to find the sentence that "match" the best the picture CLIP vector, I am really curious if somebody already did something similar

[hoary grove]

u could try Imagen

https://www.youtube.com/watch?v=HyOW6fmkgrc

also running a open beta tests i think

[hexed widget]

It takes me 15 or 20 credits to finally put the things in the right places

where

[hoary grove]

google?

😂

[hexed widget]

sorry i typed enter by accident, I meant where if i had a tool to produce a sentence close to what I want it would make me gain lot of time

but thanks for your suggestion I'll check it out

[hoary grove]

worth a try

unless @hexed widget

you upload something

u can edit stuff

let dalle paint in there

but thats prob not what u want

[hexed widget]

I think I'll just try to code it myself, but I'm conviced had the same idea

someone*

[hoary grove]

oh i bet there is

[acoustic totem]

Google ain’t opening up no Imagen beta are they!? I thought they said they’re not releasing it

[stone swan]

MUSENET 2 WHEN?!

MUSENET 2

[hoary grove]

rly ?

thats very un-google like

_>

[acoustic totem]

They’re probably like, “lets see how openAI does with their thing first” lol

[hoary grove]

fair XD

[acoustic totem]

What if google started gradually populating image search with AI generated images lmao

Glad I finally have the channel to spread my unhinged theories about AI

[hoary grove]

i heard that one before 😛

[sick quartz]

https://tenor.com/view/lovely-peaches-peaches-lovely-peaches-laughing-elijahpetty-gif-20452328

[peak rapids]

G.E; Are you sentient?
AI; Yes.
G.E; Dear god, what have we done...

[tranquil plover]

The amount of Youtube channels Ive blocked over the last week because they've made content like this

[old parrot]

Google's model is simpler. I expect to see a decent open source one within a year

[topaz marten]

How much time do you think it will generate images locally when we get the downloadable pretrained dall-e 2 model? 🤔

it's inevitable

it also won't be anime, minecraft and hr giger hating :)

i'm just really wondering, how much computation does it take for the actual image generation

[old parrot]

Why would they do that?

If they did that people would use it to generate all the things they don't want you to generate.

[topaz marten]

OpenAI won't

other people will, for sure :)

[thick cradle]

you probably need 5-10 GPUs with 80GB of memory each to even load the model

[topaz marten]

how do u know? 🤔

[thick cradle]

i'm thinking 6 RTX 3090 at least maybe more because CLIP-H need like 20GB+ of VRAM just to load, decoder is 6B params diffusion model, 1.4B params needs 20GB of VRAM so it would be 120GB of VRAM assuming it scales linearly, and idk how much VRAM the prior and upscalers need, so yeah it's expensive model to run. you can always run it at lower presicion like f16 it would reduce the cost alot idk how much it will effect the quality though

[quick harbor]

When compute scales 10x
DALL-E Mega ( craiyon) is about that much smaller and i think a great individual effort to train a model with similar arch

I don’t think parameter optimization is a very fruitful path after doing diffusion comparisons with cogview (focused on attention).
Although i do like the agility of the cogview researchers. The video demo was very cool

[thick cradle]

well the transformer in craiyon is still holding back because of the vqgan, the transformer itself is 2.9B params also the creator of craiyon told me that the app is not using the lastest checkpoint of the model, when i saw the lastest checkpoint results is really good

[hoary grove]

u can more with openai 😄

then just dalle

[topaz marten]

Oh whats some cool stuff?

I just want dalle but without biases

And with licensed stuff etc

[topaz marten]

Wait so if u have it pretrainrd can't u just use Google Colab or smth like that?

[thick cradle]

colab only have 12GB of VRAM

so you can't the model itself is probably 100GB+ in size or more or less idk

[quick harbor]

Re: hosting 100GB+ If you wanted to do cloud gpu lambda labs is best bet for large models IME

[thick cradle]

yeah you can rent GPU for training building your own GPU cluster is very expensive

[hoary grove]

so what happened to the pile of gfx miners ?

those are surely useful at some degree

id buy them if i could

but the prices rn

yikes

[quick harbor]

Not what you’d want. All lower VRAM cards.

[hoary grove]

aha

useless, noted

how about a cluster ?

meh lots of energy required

[quick harbor]

Thanks. Did not realize craiyon was using vqgan for diffusion.

[hoary grove]

i dont need to screw with my fuses

lol

[thick cradle]

i heard some people take a GPU that is previously used for crypto mining and train neural network on it instead

[hoary grove]

yes that is sorta what i thought too

those cuda's are exactly the type of chips it needs

massive amount of them

but memory is a key player

if u have little memory, u can only move little memory sized chunks

[thick cradle]

it's autoregressive not diffusion

[hoary grove]

and we know a cpu speed counts for nothing if it has no memory

if u did ur pagefile would blow up

😂

[thick cradle]

well you can always do model parallelism spread the model weights accoss multiple GPUs

[hoary grove]

yeah totally, id do that actually

perfect solution

[cunning creek]

Do you mean with biases unaltered? I'm confused.

[tulip aspen]

so i hope someone here is knowledgable enough to answer me this.
I have recently begun trying out GPT-3 via OpenAI's Playground, and have been thinking about using it to help me write a novel. In my research efforts I have also learned about other GPT-3 based writing tools, namely Chibi and ShortlyAI. Now my question is very simple:
If i use Chibi, ShortlyAI, or Playground, do i still own the content I create? And can I use it commercially (e.g. sell it as a novel on Amazon)?
I am specifically asking about the content you generate when pressing the "generate" button. I am not asking about content you upload to the site (though, that may also be helpful to know).
I have read the terms of use of them, but I am not sure I really understand them. As far as I can tell they do not mention an answer to my question anywhere.
I have seen AI generated novels be sold on Amazon, but that does not mean that they were allowed to do that.

if someone can answer me this, please ping me so that I get notified.

or if someone can direct me to a place where i can get an answer to my question, that would be nice, too.

[hoary grove]

ive seen more ppl asking this

on their internal forums

be aware that those networks have very specific ToS that describe what you ask

make sure you dont start off with a legal issue from the start ofc...

@tulip aspen https://community.openai.com/t/a-theoretical-novel-writing-tool/10855

[tulip aspen]

yes which is why i am asking

[tulip aspen]

wdym?

[hoary grove]

well, they describe almost all situations of the use of these networks that can produce (and they can this very good what u ask btw)

its a boring read

[tulip aspen]

well i read the tos but didnt really understand it tbh

[hoary grove]

but one u shall have to do

[tulip aspen]

thats why im asking

[hoary grove]

perhaps if you ask on one of those legal discords to make it certain

i know these ppl look into that and know exactly what and how

its like reading candy for them

lol

[tulip aspen]

@hoary groveim just skimming through it right now but im not seeing anything about the legality in there?

[hoary grove]

let me help ya

hmm the part i read was explicitly about the codex models

and speaks of introducing your style into a network you can train to make the novel so to speak

this process, is where they need to see what you made, or how, and why.

like by the letter.

perhaps that does not apply to you

[hoary grove]

Creators who wish to publish their first-party written content (e.g., a book, compendium of short stories) created in part with the OpenAI API are permitted to do so under the following conditions:

The published content is attributed to your name or company.
The role of AI in formulating the content is clearly disclosed in a way that no reader could possibly miss, and that a typical reader would find sufficiently easy to understand.
Topics of the content do not violate OpenAI’s Content Policy or Terms of Use, e.g., are not related to political campaigns, adult content, spam, hateful content, content that incites violence, or other uses that may cause social harm.
We kindly ask that you refrain from sharing outputs that may offend others.```

[tulip aspen]

i know

[hoary grove]

would this apply to your situation ?

[tulip aspen]

but it doesnt mention whether it can be used coommercially oor not

[hoary grove]

im no expert, sir.

u should not take my advice.

[tulip aspen]

well thats why im asking here in hopes someone does know

[hoary grove]

u can , just go to the #community-help and ask.

😄

[tulip aspen]

anyway that thread you linked is going to be an interesting read regardless

[hoary grove]

well someone had similar thought so it appears

and actually made plans

[tulip aspen]

but maybe someone else will come along and be able to answer my question

[hoary grove]

well if hes doing it in there, then i would ask him

he obviously knows

[unkempt trout]

One would think that "publish" should cover commercial intent...You'd think...

[stark briar]

Hello all! First time post here. Got access about a week ago and have been enjoying dabbling so far. I have a question that has also been asked a few times previously it seems - is there any way to control the random seed that DALLE uses for generation? I'd like to be able to fix it for reproducability, and also to help run experiments comparing generation results.

[tulip aspen]

you can publish stuff without it being commercial

[hoary grove]

making good progress

implemented some control as well

[unkempt trout]

I assume they would have mentioned "non-commercial" if that's what they meant,. Anybody's guess.

[hoary grove]

guys what do u do when the AI starts calling you "master" ? 😛

like .. i dont wanna be touching that subject with a 10ft pole in this time and day

haha

[tulip aspen]

havent heard back from support yet (wrote 3 days ago) but the fact that Dall E gives one commercial rights to stuff you generated indicates to me that GPT-3 similarly gives you the commercial rights to stuff you generated.

[uncut osprey]

Correct, copy.ai and many other services all run off the GPT-3 API

GPT-3 generated content is commercially yours

[teal garnet]

Is it any way to prove that the picture is really generated by dall-e ? What possibilities for searching source from a generated full blue picture and nothing else for example... ? If random mixed picture are equals to any piece of art existing elsewhere ? Is it any certificate about Dall-E generated pictures ?

[uncut osprey]

Realistically OpenAI could use some cryptographic functions called "Steganography" to write text into the image data but they have no real reason to do this. People could also train a NN on dalle generated images vs real images and it may start to pick up on patterns that DALLE does in certain scenarios

[tulip aspen]

but are you like really sure?

[uncut osprey]

Yes

[uncut osprey]

https://www.newsdeskindia.com/news/technology/4527.Microsoft-is-allowing-commercial-use-of-OpenAI-s-GPT-3-AI-la.html

[tulip aspen]

hm, alright. i guess. i just hope that OpenAI support answers soon just to be completely sure.

but i guess i can start trying it out for now then

[uncut osprey]

@tulip aspen There are many commercial products already built using the GPT3 API: https://www.copy.ai/blog/series-a

You are good to use it for commercial uses of any kind :) many authors have already published fully lengths childrens books on amazon using stories written by GTP3!

[mint shadow]

im really not a fan of the concept of AI generated literature

it just feels so... removed

I cant see anything exceptional coming out of it, just more meh content

[burnt vine]

I'm sure it'll be moreso AI generated concepts/plots

Actual dialogue would be by humans

But I'd say in <20 years AI would be able to surpass (most) humans in that

[worldly marten]

I gotta say that a bunch of writers for products not-that-small are absolutely terrible at dialogues.

looking at you World of Warcraft

and star wars when I'm at it

[solemn elbow]

I think it would be more like-- writing assistant content-- for ideas / finding the right way to finish some sentence, etc

[mint shadow]

I have seen instagram ads for someone advertising it to write your blog posts for you. They are like "just type in a few bullet points and it will write the whole article!"

[worldly marten]

With GPT-3? yeah that'll work

guess that makes you a worthwhile blogger then - when your own stuff is noticably better than GPT-3 generated stuff

in terms of style and such

[solemn elbow]

haha yup -- but I mean, for things like gpt-3-codex, it's -- amazing. Whenever I know I have to like-- convert some huge json file into a class using the appropriate types and defaults, I just use gpt-3 and it does it in seconds, vs me wanting to die writing very repetitive code.

[mint shadow]

im very down for that

[solemn elbow]

gpt-3-codex is gpt-3 for code

[mint shadow]

as a programmer (in college) I am slightly worried about future job security because of other AI stuff, but im down lol

[solemn elbow]

Also when it's simple code that I don't want to actually think about-- like.. "make a function that takes a 3d array representing an image with a shape of (width,height,3), and center crop it"

simple enough but just enough brain that I don't mind letting gpt-3 do it

haha

I mean, right now it's not really possible for an AI to understand a large codebase

so -- for a long time, AI code generation is going to be very limited scope

like-- max tokens on gpt-3-codex is 4096

which is large, but-- very small when you think about the size of a typical app

[mint shadow]

Im in college tho.... I have a long career ahead of me for the AI to catch up

[solemn elbow]

true

[mint shadow]

thankfully the real plan is use the CS money to become a business person

lol

[solemn elbow]

but you would have to be betting that we come up with architecture that is significantly better for memory efficiency than transformers

because as-is, transformer models have exponential memory requirements

as the input size increases

[solemn elbow]

Well good luck 🙂 -- I am doing programming / softy eng because I am addicted

People tend to ask me like--- "oh so did you just graduate?" -- "nope, I legit love programming, been doing it professionally for a while"

[mint shadow]

why do they ask? because you just havent become depressed and stale?

or you just look young?

[solemn elbow]

I havn't gotten sick of it

Mainly because I just keep learning new stuff

I know a stupid amount of coder things

[worldly marten]

I mean at this point I'm optimistic to make a career out of AI related software projects

[solemn elbow]

If I like-- only did one thing, I would hate it

[mint shadow]

im going into AI/computational media!

[worldly marten]

there's so much to explore and improve. legit feel like I'd not get fed up with it for my whole career

[old parrot]

has OpenAI ever said how large the GPT-3 and image encoder/decoders are? how many parameters do the models use?

[worldly marten]

yeah

In fact, with around 175 Billion trainable parameters, OpenAI GPT-3’s full version is the largest model trained so far when compared to other language models.

https://www.springboard.com/blog/data-science/machine-learning-gpt-3-open-ai/

[solemn elbow]

At my job, I do react frontend, node.js backend with next.js, python controller backend development, database, rabbitmq job queues and process automation, building image generators and working with pytorch to improve / add features to generators, putting all that stuff in docker containers, and orchestration. -- and that's.. just a tiny part of what I do at my job

[old parrot]

dall-e 1 used 12B parameter gpt-3, 175b might not be what they are using in dall-e-2

[solemn elbow]

I think actually dalle2 is much smaller than dalle1

[mint shadow]

this is true

surprisingly so

[solemn elbow]

Yeahp

[mint shadow]

im actually flipping between this convo and vscode lol. Working on a nodejs part of a project atm

[solemn elbow]

very nice

same

[old parrot]

oh wait it's here:

[solemn elbow]

Yeah-- that's... a lot smaller

[old parrot]

it's page 24 of the paper

[solemn elbow]

300M param 256->1024 upscaler lol

that's probably why it's so bad

[old parrot]

so 6.2B parameters, at half precision float is 12.4GB of VRAM

plus the GPT-3

this but "stack more parameters"

[solemn elbow]

[old parrot]

maybe openai has a secret other version that uses larger gpt-3 and has more upscaler parameters but the results are too good, too dangerous for public use

wait does dall-e-2 not use gpt-3?

[worldly marten]

gpt-3 is a text-to-text transformer

[old parrot]

dall-e-1 used gpt-3 as the text encoder. you can always remove the final layer of a network, and pass the new final layer into another encoder/decoder

[solemn elbow]

I doubt it, if they did, they would say, since dalle2 used to be the best text-to-image generation model, an then two google generators surpassed it.

imagen and .. I forget the name of the other

[formal mango]

Parti

[trail birch]

Where can I request access to DALLE through an API?

[warped solstice]

Hi! Currently you can't. As per #faq-and-resources-old

Q: Is there a publicly accessible API for DALL·E?

• We do not at this time, but do have a longer-term plan to offer this.
• Please also do not attempt programmatic access of DALL·E, which is against our ToS.

[tulip aspen]

yes I am aware but there is a difference between commercial technology using GPT3 and something generated by GPT3 commercially, i would argue.

[tulip aspen]

yeah i am aware but maybe they did it illegally

[tulip aspen]

youre not supposed to generate the entire text. that isnt even possible because GPT3 is very limited in its memory.

instead youre supposed to use it to help you brainstorm ideas, edit your work, come up with descriptions and outlines, etc...

[tranquil plover]

I have just come across the weirdest thing.
Dall-E2 seems not to be able to draw a pentagon.
Check these attempts out, nothing seems to work:
https://labs.openai.com/s/2XEVgymc2vOvTDQP3yzBWDv0
https://labs.openai.com/s/u1qL1MDoj5qfWibIcFRrflEx
Can someone else please give it a try and verify or disprove this weirdness?

[thick cradle]

interesting

[tranquil plover]

I'm flumoxxed. Other users who have tried on the facebook page are also unable to create pentagons.

[thick cradle]

well maybe pentagon is filtered from the dataset idk

[tranquil plover]

But then surely the whitehouse would be also? Super weird.

[teal garnet]

https://80.lv/articles/nvidia-s-new-neural-network-for-turning-2d-images-into-3d-objects

[tranquil plover]

As they say, just one more paper down the line...

[tranquil plover]

Any @zenith moth folks around who might have some insight into this weirdness?

[warped solstice]

Hi! I'm not an OpenAI employee, but a community volunteer. I may be wrong but it may be associating it with The Pentagon and may be restricting it?

[tranquil plover]

That's what we've been thinking, but its strange to restrict that and not, say, The Whitehouse. Also, we're talking about a deep design primitive here, it's bonkers to restrict a fundamental geometric form for the reason you suggest.

[warped solstice]

Hmmm – now I'm not sure if it's restricting it.

[tranquil plover]

there are literally no pentagons visible in those shapes

[warped solstice]

Yeah, but check the prompt for the second one.

[tranquil plover]

aaaaah, interesting

[warped solstice]

Make sure to flag the results! It'll help the team

[tranquil plover]

Cool, Ive done that with about 30 renders, good advice

[azure trout]

Curious if there any plans to implement different aspect ratios or resolutions 🤔

[cunning loom]

Anyone here into NFTs?

[uncut osprey]

Thankfully not

[cunning loom]

Why thankfully?

[uncut osprey]

#openai-chatter message

[cunning loom]

Understandable. I suppose your problem isn’t with NFTs but the wider NFT market. Personally, I have no problem with people seeming to monetize art, but many NFTs are hot garbage.

[pastel cloud]

Agreed-

[cunning loom]

Anyways, I asked because I created a project for fun and for profit as a proof of concept. The point was to attempt to create as much of a project as I could using just the AI tools available such as DALL-E, GPT-3, and CoPilot. I'm very clear in disclosing that much of the work isn't my own (because I'm feeding various prompts to different tools), and as such I'm going to be offering a free mint to those interested.
In the future I hope to get access to OpenAI Codex so that I can create an entirely AI generated project (besides my assembling of the bits and pieces)

[amber turret]

https://labs.openai.com/s/IjDTUjYlA103BmNvqwFKevar

Figured out the pentagram thing

[sturdy terrace]

Has the model been nerfed after beta ? My old prompts are producing reliably worse results. Less detail is the biggest thing I see. Specially oil paintings.

[fallow grail]

yep I'm finding this out now

[sturdy terrace]

Juat spent credits confirming....

[fallow grail]

big difference

[sturdy terrace]

MEH

Can we get some info from the team please ?

[fallow grail]

I would like official confirmation from the team if they did make changes though but yeah when i try a photo portrait it feels more bland

[sturdy terrace]

Before

After

Same prompt

Still good but less detail overall

[fallow grail]

yep less detail

[sturdy terrace]

Thats what I noticed too, sucks.

[fallow grail]

before:

After:

Nice but I'm not getting the same from before

[cunning creek]

Unlikely

The above seems consistent with how much variability I've found in "oil painting" without artist references

[fallow grail]

Before:

After:

[cunning creek]

Look similar

[sturdy terrace]

I had an artist modifier on mine fwiw, no "trending" tags as those are inherently variable

Talking about oil paintings fwiw

[thick storm]

As someone who really loves everything about DALL-E 2... I am not sure if it's just that people are complaining so much that I believe them, but I think there is a noticable quality difference... The fidelity of images I was generating like two-three weeks ago are noticeably more refined... However..... I also must remind myself that even when I got those amazing results, I also often was looking at 5 other results that were sub-optimal so... Maybe it's just that. For a long time I was like "it's all in their heads" but now I'm like..... Maybe.... something is.... different....

However, I've only really noticed this in "Paintings" and as @cunning creek that's a pretty fluid term there... Some painting come out this way, some come out that way, photorealism is still absolutely unbelievable so, we're probably imagining it.... Unless.....

[sturdy terrace]

Seems like less detail in oil painting amd digital art prompts.

[cunning creek]

The biggest difference is we're getting 1/3 fewer options so often the best generation we used to select no longer gets made

[sturdy terrace]

Tried a few old prompts and got consistently worse results

The detail isnt there as it used to be...

[cunning creek]

We also look at our past creations through a lens that helps us forget the duds

[fallow grail]

Yeah I would need to test it from past collections but I dont have them anymore, might be in my head but I do feel like now there's less detail haha

[cunning creek]

this concern has been raised every 2 days reliably in the month+ that i've been on this Discord 😅

[thick storm]

I'm unsure if this has been stipulated but, for example, during "peak times" when there's lots of people using DALL-E ... is there like "a reduction in quality to save processing power" ... I admit this seems like a silly question, because what does "quality" even mean.

[cunning creek]

Processing time is variable, not so much quality

[thick storm]

I've heard so many people complain I think that my previous resolute impenetrable belief that "they're just wrong" has been punctured...

[thick storm]

Right, right. Right. Sure. Right... Haha, yes, indeed, indeed.

[fallow grail]

I'm sure @fiery tartan would give us the down low if there was/to be changes to the model.

[cunning creek]

Created just now - seems pretty good. I am reticent to repeat prompts exactly these days because they're limited but the quality is equal to previous versions in this style

https://labs.openai.com/s/U0fc7bosBWUt1IeAgl5PoDpb

[fallow grail]

Yeah I'm just seeing less detail all though slight in my old vs new prompts, and things like flowers in a certain place in the new prompts vs old which were very consistent. And of course the new diversity which is fine.

Might even be the layout of 6 images, something nice and even about that haha

[sturdy terrace]

Im literally comparing all images using the exact same prompt though.

Repeated a bunch of my old prompts and the results were consistently worse. Cant be bothered to waste more "credits" to test further.

[fiery tartan]

Hi hi! No changes to the model

[sturdy terrace]

Can the model forget artist styles ?

[fiery tartan]

model change hallucination is a real thing though and affects me as well!! but it’s all randomness

[fiery tartan]

the model is static

[sturdy terrace]

Roger, thanks

Is there a clip guidance kind of setting that can be modifier for performance reasons ?