general

[glacial mulch]

what was the previous image upload limit

[grand cliff]

Nice image edit.

...

Now where's the Text edit.

[civic spindle]

remove image limit

it sucks

[digital vale]

prolly gotta wait for llm diffusion models tbh

[grand cliff]

What?

[echo aurora]

Okay would you mind creating a new post in #1343291835845578853 and break down the step by step? It'll help keep this organized.

[marble birch]

I don't understand... brother

[grand cliff]

Also. I replied to my bug.

[pseudo fable]

[velvet forge]

he just asked when is gemini 3.0 pro is gonna out

[hollow imp]

What's this new bot

[robust sluice]

looks like the others has the same problem, I just see it there

[marble birch]

Thank you. I don't know Korean at all.

[echo aurora]

Add onto an exsisting bug report then please

General reminder to try and use English when messaging in our text channels please.

[light minnow]

i saw the update about the new image edit feature but can't find where to use it, where should i look?

[velvet forge]

Okay, but idk does he speak english

[viscid cloak]

Up till now, it seems multi-turn image edit automatically reads the last output image within a chat as the new reference photo, instead of recognizing “your uploads” as reference.

[velvet forge]

@echo aurora can i write him in chat so he can understand that need to write in English not korean?

[echo aurora]

YEah that's fine.

[bleak jetty]

write to him in chinese he will understand better

[velvet forge]

@pseudo fable 죄송해요. 하지만 디스코드 서버 규칙 때문에 한국어가 아닌 영어로 대화하는 것이 더 좋다고 운영진(모더레이터)이 말했어요.

[magic stag]

https://x.com/OpenAIDevs/status/1986861734619947305

LMAOO and that they were gonna decrease limits, and they increased them by 50%

OpenAI Developers (@OpenAIDevs)

You can now get more Codex usage from your plan and credits with three updates today:

1️⃣ GPT-5-Codex-Mini — a more compact and cost-efficient version of GPT-5-Codex
2️⃣ 50% higher rate limits for ChatGPT Plus, Business, and Edu
3️⃣ Priority processing for ChatGPT Pro and

X · Nov 7, 2025, 6:22 PM

[bleak jetty]

it was a hype train to drive in more investments lol

[full vigil]

Exploring

[pseudo fable]

알겠습니다. 저는 영어를 할 줄 압니다. 제가 뭔가 잘못한 줄 알았습니다.
Okay, I speak English. I thought I had done something wrong.

[echo aurora]

Thank you, and thank you @velvet forge for explaining that. I do appreciate it.

[velvet forge]

Okay, so gemini-3.0-pro is gonna out like in end of november or in december

[velvet forge]

No problem

[pseudo fable]

That's great! I hope it will be available this month. I can't wait to test its capabilities.

[pseudo fable]

You're welcome.

[exotic crest]

anybody know how to fix the infinite generation bug? my chat has been stuck for 14 hours. tried signing out and back in, refreshing the page, logging in on different browsers/devices...its just stuck

[steady grove]

we need new gemini for coding because grok code is kinda better now? or minimax

[fluid yew]

seems their doing everything in their power not to fix this one issue here lol.

[zealous sparrow]

who calls their model ernie

[sweet tinsel]

China, Baidu to be precise.

[zealous sparrow]

Yeah first course of action is i always ask who made the model and if its chinese

[stray aspen]

Is movement labs still online

[lofty frigate]

I got a problem that i encoutered today, dont know if its only to me or for everyone, so when i write a prompt to compare models with images, so after its done generating and clicking on the retry for some reason it started generating random ai photos... anything like that happend to any of yall today? It never happened to me until today

[rough harbor]

Everyone does

Waiting for the people upstairs to resolve the issue

[ashen mauve]

tell you what ernie isnt good at roleplaying

[signal citrus]

Hi, how are you people

[sweet tinsel]

They are.

[ashen mauve]

you all still picking it apart from last night?

[echo aurora]

We made a change to image edit multi-turn. If you want to continue to use the original image, you'll want to click the Edit Image button. This will essentially re-upload the original image into your next prompt compared to starting with a blank slate. The recent announcement has a bit more information on it. #announcements message

[wintry tinsel]

Congrats I been waiting for this one

[polar niche]

👍

[jolly garden]

I can't find any Edit Image button even with page refresh, clearing cookies etc., so LMArena is unusable for generating images with character consistency.

[lofty frigate]

Wait you guys like this new update rather than the old one? Im the only one who liked the old one?

I loved the multi turn😔

[echo aurora]

Are you on mobile or desktop?

[jolly garden]

Desktop, even with different browsers

[echo aurora]

You don't have this? Can you send me a screenshot?

[lofty frigate]

No way to get the multi turn back?😔

[lofty frigate]

On mobile wont show?

[magic stag]

https://x.com/scaling01/status/1986886020067938749

Lisan al Gaib (@scaling01)

BIG OPENAI NEWS:
GPT-5.1, GPT-5.1 Reasoning and GPT-5.1 Pro

GPT-5.1: "Flagship model for the latest generation of ChatGPT"

GPT-5.1 Reasoning: "Thinks longer for better answers"

GPT-5.1: "Research-grade intelligence"

X · Nov 7, 2025, 7:58 PM

[ashen mauve]

nani

[jolly garden]

Absolutely not

[lofty frigate]

And since today everytime i try to retry it failes...

[balmy mist]

this i polaris alpha?

[magic stag]

no

polaris is a. non thinking model

[echo aurora]

I haven't been able to repro this, there is a thread in #1343291835845578853 about this issue, can you post your Eval ID there?

[balmy mist]

hmm, but polaris is open ai?

[magic stag]

i mean, it could be 5.1-chat or whatever

whatever the non tinking is called

[magic stag]

yeah

but non thinking

[balmy mist]

is polaris better than sota?

hmm

[lofty frigate]

Ok

[balmy mist]

how does it match up against sonnet 4.5?

or against thinking models?

[magic stag]

you can use it for free

on open router

https://openrouter.ai/chat

[gaunt spade]

polaris is dum, i dont think its made by openai

[echo aurora]

Okay really sorry, I have update on the Image-Edit announcement I put out, this is still in experiment mode and hasn't fully rolled out.

That was my mistake.

However, users that aren't in the experiment, should have the same behavior as before, where the original image is being retained for multiple prompts.

[magic stag]

Its non thinking

As in

Current chatgpt-5-chat

[lofty frigate]

Ok it seems a lot of ppl have been experiencing the image not generating anymore after the first retry

[jolly garden]

No, that's not happening. Every uploaded photo is ignored on the second retry, and many models fail on the first upload but then stabilize with the second retry. Now it's totally unusable, so LMArena will spend my time uploading the same photo without ever getting any consistent peering.

[echo aurora]

No, that's not happening
This is in reference to this message:
However, users that aren't in the experiment, should have the same behavior as before, where the original image is being retained for multiple prompts.
right?

[jolly garden]

right

[cloud zinc]

gpt 5.1 gonna crush gemini 3

[balmy mist]

u dont like it?

[ocean vortex]

I have said months ago this is gonna happen when Google fanboys were claiming 2.5Pro shouldn't be compared against gpt5 since "it's older model" 🤣

gpt5.1 and gemini 3.0 are basically gonna be as even as it gets, when it comes to their release timelines

kinda crazy that OpenAI with less compute is able to outdo them with big releases frequency tbh

[dim burrow]

Hello to everyone!

[novel flame]

I haven't been following this too closely, what makes people bullish on GPT-5.1? Is it on webdev arena? Or are there rumors?

[lime coral]

Gpt5.1 not even better than Gemini 2.5

[magic stag]

gemini-3-pro-preview-11-2025 was available to use on gemini cli for about 24 hours a couple days ago

and it was absolute garbage compared to what we saw on ab tests and the lmarena models like lithiumflow

[novel flame]

Sure but the consensus seems to be that lithiumflow is Gemini 3 though, right? So either the CLI access was a misunderstanding or it was serving a nerfed/older snapshot, no? Are people no longer thinking lithiumflow is Gemini?

[magic stag]

https://x.com/scaling01/status/1986895079944991025

Lisan al Gaib (@scaling01)

OpenAI: GPT-5.1 to be released on November 24 2025

X · Nov 7, 2025, 8:34 PM

[magic stag]

there have been like 50 versions of it

right now testers have something like 20+ different versions of it

there are good ones, aka the A/B tests and lithiumflow

then theres garbage, the one they nerfed and chose as the official preview one....

[cloud zinc]

gemini 3 is downgraded

[novel flame]

Well you say "the one they [...] chose as the official preview one" but in fact it was very clearly unofficial and unintended that the CLI responded to requests for that model ID. I would think it's entirely possible that if they can make the mistake of opening access early, they could have just as easily made the mistake of connecting an entirely different checkpoint than the one they intend to be their official anything

[bleak jetty]

It's not entirely out for the public yet so don't expect the final product

I don't get why people are fussing over it

They're probably doing some tweaking / testing for the distilled preview and people with unauthorized access testing and shilling on the unfinished product

[novel flame]

Personally I'm fussing a little because I upped my Google investment based on my expectation for a strong Gemini 3 launch 😄 but granted, that's on me

[bleak jetty]

Let's keep our hopes high gng, I still have beliefs in Google lmao

We have like 1 or 2 week before the actual public release of preview

Tons of optimizations can be done

[novel flame]

In any case, thanks @magic stag for bringing me up to speed

[magic stag]

i mean however you want to cope. a company like google doesnt "unintentionally" take a powerful dev build named internally "Teletubby (orion with 1p v6 no examples)" or similar, spend valuable dev time nerfing it and testing it, name it "gemini-3-pro-preview-11-2025", upload it to the normal public servers...

all "unintentionally"

lol

[violet cipher]

Are there any AI Image generators with an API that don't charge for like "Free tier" ?

I mean it doesnt even have to be free... but I just want a fixed cost of some sort...

[bleak jetty]

Gemini 3 pro better worth the hype

[ornate agate]

I don't know why they won't just release the good model that they already have. Doing pre nerfing is really foolish imo. The whole appeal of a model by ppl like Google is to pay a lot more for a tiny bit of extra quality. So if they don't give that option idk what their plan is.

like Claude is $15 or even $60 (opus) output cost, and people pay that. So if its expensive to run release it an price accordingly

[magic stag]

they would rather prioritize "market reach" by giving a mega nerfed model out for free to all students for a year and over 500 million indians for 18 months

if they give it out for free on that scale it just cant be not nerfed to some degree

[ornate agate]

I hope thats not what they are doing, thats a huge fail of a "plan" imo

[magic stag]

i mean its literally exactly what theyre doing, lol

[ornate agate]

u sure?

[magic stag]

chatgpt did it too but they were intelligent enough to only give the worst model for less money

[magic stag]

https://www.cnbc.com/2025/10/31/google-gemini-ai-free-500-million-jio-relince-users-india-.html

fairly

[ocean vortex]

gemini 3.5 is not even better than gpt3.5

[magic stag]

its not exactly a secret

its public knowledge lol

[bleak jetty]

They better roll out unnerfed model and name it gemini 3.0 ultra, I can pay for it as long as its decent and not a nerfed model being served globally for cheap market

Or Google learned from GPT 4.5 lol

[magic stag]

thats what im. hoping for too

that there will be one of the best A/B test models avail as gemini ultra that yo have to pay more for, and no giving it out free to 500 million people

and they can give this trash out as much as they want

[bleak jetty]

Yeah it's not the problem of their incompetence, it's about their choice of whether serving us garbage or not

[ornate agate]

ok. this is very disappointing then. 500mil free users means the model has to be cheap to serve, and they agree to give pro for free so... ultra or bust i guess?

[bleak jetty]

We can see their capabilities through ab testing and 2.5 pro exp

[ornate agate]

I don't know what the market strategy here is either, because they are currently trying to charge $20/m for that in my country...

[balmy mist]

gpt 5.1 is king

[magic stag]

its not only 500 million. thats just non-student indians. now add up all of the free student pro offers

its probbably double that or tripple

[ocean vortex]

Welp... they did release ultra in DeepThink. Are you using that?

I'm guessing not

[bleak jetty]

Nope I am a Claude diehard fan

[magic stag]

you cant use something limited to 10 prompts a day as ur main model

but to answer your question yeah i have it

[bleak jetty]

Toxic relationship btw

[ocean vortex]

So you can't pay then

[bleak jetty]

I pay 200$ monthly wdym

[ocean vortex]

For Google?

[bleak jetty]

Claude

[magic stag]

no one without money to waste for no reason should pay $250 for 10 prompts

lol

[ornate agate]

one-upping on how many $200 subs someone has is a bit cringe

[violet cipher]

Claude is amazing... so wordy 🙂

[ocean vortex]

But not for Google. That's what it takes to get 'ultra' from them

[bleak jetty]

They should serve it as API instead and I'd pay

[magic stag]

or, what it takes to get people to pay $250, is not limit you to 10 prompts a day while giving away all your compute for free to 1 billion people

[ocean vortex]

I agree. But they are not doing that and you are clearly not prepared to pay for it unconditionally lol

[bleak jetty]

Using it through the Gemini app is what set me off

AI Corporations going cost-effective route atm now that they've gained much popularity and backings

[ornate agate]

lol. active params maybe but I doubt it.

there is no way gpt-5 is 100b it knows far too much for that size

[ocean vortex]

Very unlikely tbh. It's most definitely MoE and the total parameters are gonna be higher

There's not much compute penalty when serving it to this many people

If your activated parameters are relatively low - that's all what matters

enables you to train it fast too

[ornate agate]

they made a big model, 4.5 but they couldn't sell it it seems.

people only seem willing to pay for very large Claude models, as Opus still has a significant market.

[ocean vortex]

It's the same size as gpt4o. Well not for a fact but as close to a fact as you will get without them going on record

Size, basically:
gpt4o = gpt4.1 = o1 = o3 = gpt5

then you have minis in a different category with the same relation

o1-mini was never bigger than gpt4o-mini initial release

but they could afford to price it for its performance rather than running cost with lacking competition.

I think it was the right move overall, given that performance you get per the same size is constantly improving. They arguably undersized og gpt4o, but there's little reason to backtrack knowing where the progress comes from.

They used to struggle with spatial awareness and web design bad. But that's no longer the case as the progress has caught up to the model size they arrived at ~1.5 years ago

[fiery gull]

The unique problem from gpt 5 to me is, it don't have markdown or can act in a persona, maybe the gpt 5.1 correcting this will be my farvorite model

[fiery gull]

Hell no, I think the o1 is soooo heavy, the o3 too but less that o1

[ocean vortex]

That is quite literally the only benchmark they care about

OpenAI would most likely trash them if they didnt care about anything else

[fiery gull]

And the gpt 5 is the lightest gpt's in all time

Except gpt 2 obviously

[ocean vortex]

what?

lol

[fiery gull]

Gpt 5 minimal maybe is 20b

[ocean vortex]

I don't think you have any clue, no offense

gpt5-minimal is the same exact size as gpt5-high

[fiery gull]

They generate 6 billion tokens per minute, the gpt 5 should be very light

[ocean vortex]

same arch same everything

[fiery gull]

Ops I want to say MINI

[ocean vortex]

Oh ok lol

[fiery gull]

I think the gpt 5 mini is very very light, really light

[ocean vortex]

same as o4-mini

[fiery gull]

No, the o3 mini is much lighter

[ocean vortex]

o3?

[fiery gull]

So much so that they only make the o3 mini available for free today

Gpt 5 mini < o3 mini < o4 mini

[ocean vortex]

gpt5-mini is free on chatgpt

[fiery gull]

I know, because this I'm saying this

[ocean vortex]

limited use, similar caps to what they used to do with o4-mini

Similar cost for them

o4-mini and gpt5-mini

[whole sundial]

gpt-5 is probably 600-800b params, mini is likely 80-120b, nano is like 10-20b

[fiery gull]

Hell no, I think is 100b~200b

Think about the gpt 120b oss with only 5b active

[ocean vortex]

Roughly yes. But keep in mind open-source tends to undercut commercial models with size, not the other way around

[fiery gull]

If what they made available open has 120b and performs well, imagine what they have internally

[ocean vortex]

It makes sense to have a ton of total parameters with MoE if you can guarantee high utilization

[whole sundial]

i tried to account for that, but keep in mind all of these are capable of vision

[ocean vortex]

With open-source... You can not guarantee high traffic

and the priorities are different

[whole sundial]

i bet gpt-image-1-mini probably has more vision/image gen params than text

[ocean vortex]

I wouldn't be extremely surprised if with OSS they pushed for smaller size much more than with their closed models. It makes sense cause with OSS it was literally their goal to make it fit into single GPU. They do not care about this with their closed models and would much rather have better performance. Again, they can offset total parameters with high traffic when it's MoE.... vram is not a problem

I would maybe even say with OSS they were out of their comfort zone having those constraints

So they pushed for low cost that time for real. Unconditional low cost that can't be achieved or offset by high traffic

[burnt sinew]

Anything new on gemini 3?

[ocean vortex]

soon™

[hollow ivy]

it could appear in nov.18, or begin of next month

(so, it will definitely not appear before 18th)

[burnt sinew]

I mean the endpoint that was open yesterday is it still closed

[hollow ivy]

Has anyone found out a way how we can discern Sonnet-4.5 from Haiku-4.5 in battle mode?

[balmy mist]

on openrouter is kimi k2 thinking provider chinese based?

[hollow ivy]

idk, sorry

[whole sundial]

moonshot is chinese so yes

[balmy mist]

damn

[hollow ivy]

it is made by chinese, but can be run local, right?

[keen beacon]

So I watched some vids on Gemini 3 preview

and I realize one thing lol

[balmy mist]

thats what thought but its huge, i thought openrouter had a different provider

[whole sundial]

moonshot has some business in Singapore as well

[keen beacon]

Open AIs oss makes zero sense

Essentially, you have an open source model, which isn’t main flagship product

Competing against Chinese open source flagship LLMs, which are equally competitive to their top-tier close source models

[whole sundial]

ok looks like singapore, but they are a Chinese company still

most Chinese companies are registered in Singapore for various reasons

[keen beacon]

Well, they do business in the US. They have to have some kind of US partner

Some of the most popular Chinese AI isn’t even used in China. It’s majority customers are from western countries.

Few of them are even banned in China.

For example

According to this filing by Disney against minimax

https://chatgptiseatingtheworld.com/wp-content/uploads/2025/09/Disney-v.-MiniMax-COMPLAINT-Sept-15-2025.pdf

And if we read a little further

And the last one

So what is all this mean?

MiniMax distributes its Hailuo AI service in the U.S. through app stores, handles payments via Stripe, and uses Cloudflare for content delivery. It operates under U.S. laws and intellectual property protections, including trademarks for "Hailuo" and "MiniMax." The company has agreed to jurisdiction in the U.S. for legal matters related to its service.

[quick salmon]

Hey guys

I’m new here

[thorny berry]

Hey folks 👋
I never set out to “build AI” I just liked making things that think. That curiosity led me into full-stack dev, AI systems, and a bit of blockchain work along the way.

These days I’m into projects where tech and creativity overlap automation, intelligent tools, stuff that actually feels alive.

I’m open to joining new builds or teams that want to push limits a little. If you’re working on something ambitious (or just weird in a good way), let’s chat.

Let’s see what we can make surprise even us.

[keen beacon]

Spot on

[quick salmon]

I’m also into one business

[keen beacon]

Why Singapore, though ? Access to International Markets , low tax rates, economic freedom, legal rights

https://en.wikipedia.org/wiki/International_Enterprise_Singapore

https://en.wikipedia.org/wiki/Economy_of_Singapore

[proper rain]

the video generator seems to ignore my sent in image today, anyone knows why?

[keen beacon]

Ugh what was the image of?

[whole sundial]

this has been an issue today from what i heard

[proper rain]

nonthing special, a guy and a product, both times ignored

[whole sundial]

ever since they did the update with the edit image button and removed multi-step editing

[whole sundial]

the most interesting thing about this is that openai has not been sued by anybody yet even though sora 2 can produce many of the same things (at least, before the nerfing) and gpt-image-1 and especially gpt-image-1 mini can almost perfectly reproduce album covers from a simple prompt, no other image generator can do that except for theirs and i bet it can reproduce copyrighted characters too

[keen beacon]

Um you mean that you upload an image but after the video generates, it’s a completely different thing right?

[keen beacon]

The lawsuits are actually in the works.

[heavy juniper]

ai has come so far in such little time

it was only two years ago in early 2023 most people started picking up on it

[whole sundial]

wouldn't be surprised

[keen beacon]

https://www.transparencycoalition.ai/news/seven-more-lawsuits-filed-against-openai-for-chatgpt-suicide-coaching

[heavy juniper]

open weight models are such a blessing to this industry, its going to drive the competition through the roof

[keen beacon]

https://www.reuters.com/legal/government/openai-loses-bid-dismiss-part-us-authors-copyright-lawsuit-2025-10-28/

https://smithhopen.com/2025/07/17/nyt-v-openai-microsoft-ai-copyright-lawsuit-update-2025/

https://www.thedailystar.net/tech-startup/news/anime-and-manga-giants-warn-openai-over-ai-copyright-breach-4026866

But the Japanese got no leverage lol

You guys don’t even see what’s coming

https://www.musicbusinessworldwide.com/universal-music-settles-udio-lawsuit-strikes-deal-for-licensed-ai-music-platform/

Music AI = practically over

Video is next

https://www.riaa.com/wp-content/uploads/2024/06/Udio-Complaint-6.24.241.pdf

[whole sundial]

i literally found out about the gpt image 1 album cover thing just by putting in prompts for llms into it, asking it "What is the album "insert album name here"?" will reproduce it if it is well-known enough, gpt-image-1-mini does it better for some reason

[keen beacon]

For all you people making fun of Chinese LLMs censorship

Home made censorship is a lot easier to digest huh?

Oh wait it’s only censorship when the Chinese LLMs do it

https://tenor.com/view/aceventura-petdetective-jimcarrey-comedy-duh-gif-18242739799609897143

[whole sundial]

makes sense, chatgpt wants to censor things that openai doesn't like, which includes being sued for copyright infringement because to them, using any copyrighted work as training data is fair use

you can use that same argument for basically any AI company

they all do it

[keen beacon]

This lawsuit has nothing to do with open AI though lol

[whole sundial]

i meant AI companies being sued for copyright infringement in general

[keen beacon]

Well, what do you think the Chinese do when they censor Tiananmen Square?

[whole sundial]

they want to censor things they don't like

china doesn't like tiananmen square, the us companies don't like being sued for copyright infringement

[keen beacon]

https://en.wikipedia.org/wiki/1960s_Berkeley_protests

https://www.nytimes.com/2025/11/07/us/kent-state-vietnam-john-cleary.html

And the reason I’m pointing this out

[inner gate]

Hello laddies

[keen beacon]

https://en.wikipedia.org/wiki/1989_Tiananmen_Square_protests_and_massacre

Look at how we labeled theirs

We add the word massacre

[inner gate]

What’s going on?

[keen beacon]

But look at ours

https://en.wikipedia.org/wiki/1960s_Berkeley_protests

Both were deadly and both used the means of the state to in-force order

Except for some reason, we don’t have the word “massacre” at the end of ours.

Wikipedia USA based service

[inner gate]

@keen beacon

What’s this about

[keen beacon]

About Llm censorships

[inner gate]

Summarise it to me quickly I’m interested

[gleaming roost]

https://grokipedia.com/page/1960s_Berkeley_protests

[keen beacon]

Using the argument about deep seek being censored because they censor Tiananmen Square protests

[whole sundial]

not just tiananmen though, there are many things that china doesn't like that are censored by their llms

[keen beacon]

Yeah same with ours

I just don’t have time to go through everything and show you

On many subjects and many things, like you wouldn’t believe

[whole sundial]

actually momentum doesn't have any pro-china bias in the questions i asked it

maybe the claude post-training fixed it

[keen beacon]

I know they don’t

I was making a point about using that as the actual argument this is what I was trying to get at

See even here lol

So personally me and this is just my opinion I don’t see the difference between what they do and what we do to me it’s more of a style which which one is more digestible

[ashen mauve]

my eyes

why are you using light mode

[echo aurora]

For clarity we have some swear words blocked through auto-mod. This isn't an LLM doing it.

[keen beacon]

It’s nothing wrong. I’m not saying that there’s anything wrong here. Or that this is bad or good.

I’m just pointing out how language itself

The form of expression

[echo aurora]

Gotcha

Sorry to butt in like that, I haven't been following this convo closely but yeah wanted to make clear why that message failed to send.

[gleaming roost]

Why is the word "ro blox" banned?

[echo aurora]

We had issues with spam/hacks related to it previously in the future.

[keen beacon]

What you’re encountering and would I diagnose this phenomenon as it’s known as the invisible fences

It’s a really complex problem from both the moderator standpoint and from the user standpoint

But what ends up happening is what’s known as self censorship

Otherwise, there will be chaos and disorder in order to keep this place, polite safe and open to everybody. There are trade-offs not just with this server, but practically in all servers, especially companies and brands.

But if we take the euphemisms out of it, and we take out the psychology and the behavior under normal definitions of censorship, this would qualify if we look at it objectively it’s not necessarily direct and indirect

This is why I always say that but ends up happening in the end is a echo chamber, it’s more subconscious and we don’t even realize that we’re doing it or we’re part of it. This forms bias about how we view everything outside of that bubble, and it gets reinforced by the same behaviors that naturally form from these invisible fences

So to wrap everything up in simple terms, we have to look at things from outside and understand the context in which we are communicating

[barren prairie]

Don t trust numbers bro because actually one person can make 20 accounts 😂

[keen beacon]

Spotify got hit with a major lawsuit for this problem

https://www.forbes.com/sites/conormurray/2025/11/03/spotify-accused-of-ignoring-billions-of-fraudulent-drake-streams-what-we-know-about-bot-streams/

Ai and the tech industry is a house of cards one large systemic pyramid scheme lol

https://seekingalpha.com/news/4518446-what-to-know-as-michael-burry-bets-against-nvidia-and-palantir

[magic stag]

What?

Google is offering it to 500+ million individuals for free for 18 monthsvin a single country alone. No one said that that exsct number is how many accounts they've made.

[keen beacon]

So is open ai right?

For Sora 2 and in India ?

[keen beacon]

They’re not offering anything for free, they’re buying their way into the market. They are slicing them selves a piece of the pie. 🥧

[magic stag]

Not the same thing. They give them chatgpt Go which only has 4o

So they give them an old model instead of nerfing their new flagship model to be able to give it out for free

[queen veldt]

Go is bad it's not like plus plan gemini gives more features same as regular subscription

[keen beacon]

Look at this hedging lol

😂😂😂

[queen veldt]

Yeah google sells your data for advertising that's their way of earning money that's why google search is free and other stuff

[keen beacon]

if we broaden the definition of "seller of user data" to mean the entity that generates the most revenue from user information, Google (Alphabet Inc.) is arguably the largest in the world.

So when they offer something for free to 500 million

[queen veldt]

They track us the same way they will track them maybe even use ai chat history to give personalized ads

[keen beacon]

Real talk

[queen veldt]

I don't care about "selling my data" they can track me all they want

[keen beacon]

Same lol

They have yet to sell me something lol

[queen veldt]

I like free products we got the ai studio for free, we have the 2.5 pro for free and nano banana for free

[keen beacon]

For now yes

But the gates are coming, I feel it in my bones

[queen veldt]

I have ad blockers and adguard dns set up so idc i have revanced youtube also

[keen beacon]

I don’t even have any of those

[queen veldt]

I haven't clicked a single ad for a loooooong time

[keen beacon]

I’ve never clicked one in my life

Sometimes I even fall asleep to those long advertisement commercials when I’m watching YouTube

lol

[queen veldt]

Except when they forcedly pop up in my pirated movie site 😂

[keen beacon]

Or take a nap and wake up in the same commercial is still on

I’m the hardest kind of person to sell to because I’m not a consumer.

[queen veldt]

I got the 3-month Netflix subscription trial through my carrier and ngl Netflix sucs it doesn't have ALL the movies and series

[keen beacon]

But their services are useful when you’re looking for things which is awesome

[queen veldt]

They be like "You're probably looking for this, we don't have this but we have other suggestions

[balmy mist]

​​is kimi k2 thinking better than polaris alpha?

[keen beacon]

I don’t watch TV. I don’t listen to music. I just do YouTube, which I guess worse.

No, there are a lot of cool products out there like when you need something and you’re searching it but you truly need it

Not that someone is forcing it down your throat. I think this is a good aspect and a good way of doing business.

Because the benefits both parties

And to give them credit sometimes commercials also help people believe it or not

If you got a problem and you don’t know how to do it or deal with it and somebody offers a solution in a commercial I don’t think that’s a bad thing

[queen veldt]

Agree

[keen beacon]

I don’t even mind commercials to be honest at all what bothers me is the fact that these commercials the way they’re designed to grab attention and some of the marketing tactics are just an insult to my intelligence. I feel like as a person as a human.

I feel like a little baby fish in a sea full of sharks.

Like there are more respectful ways to sell products you could use commercials you could use ad. You could do a number of things with no problem.

And the overall feeling, though, which kinda hurts me out the most is it seems like the whole economy or the world or whatever even the Internet all it’s able to do is make money off ads lol

Like ads run the world or something lol

Because there’s a more fundamental question about the Internet and technology that’s a lot more harder to answer

Like if we really look at the Internet as a whole and we try to understand, it’s economics and the way it generates money

It’s arguably, one large billboard

Because the Internet itself doesn’t generate any money or revenue

But the Internet cost a lot of money to maintain

[queen veldt]

YES internet was first used in the military it's not there to benefit US but the SHAREHOLDERS and people who actually PURCHASE THINGS

[keen beacon]

Because how do people really make money off the Internet like really in the real world you do a service and you get paid and you have something to show for it in a sense on the Internet you’re either looking for services offering services advertising and marketing or selling infrastructure to maintain those two factors

But it’s not a physical thing it’s just a communication means

So it asks the question of why and how are these technology companies so profitable and valuable?

[main spoke]

Hi new here just trying to test out video ai

[keen beacon]

And then we look at one other factor the fact that the way they make money is off of the data which comes from the users who don’t see the benefits necessarily or share the profit

So the users are the ones that provide the data for the platforms discord Reddit google whatever you name it, etc. the users do all the hard work the labor you could say

[queen veldt]

I don't think it's about money anymore, look at sora2 they had the video generations for FREE .. LIKE WHY WOULD YOU DO THAT their API was like $5+ for a single video...

[keen beacon]

Do you really wanna know?

https://youtu.be/nTDS5r5fk_k?si=H1_kEavcJCFvkMcy

Watch that video and you’ll understand the problems with sora

And the opportunities

Depends how you wanna look at it I’ll let you decide for yourself

Me with my biased so I’m gonna naturally go with the opportunities in the sense

[queen veldt]

But everyone was creating funny videos and stuff... Very small amount of people actually continued to use sora2 (they've had their fun with it) I've seen reddit posts of bunch of people complaing how it's "so expensive"

[keen beacon]

How so?

[keen beacon]

lol

Don’t judge a book by its cover

That video was very informative

OK, since you made that statement, you point me what there is false information

So you don’t believe what you’re saying is misinformation?

https://futurism.com/artificial-intelligence/openai-sora-trouble-backlash-copyright

https://www.archyde.com/openai-posts-record-losses-of-12-billion-in-three-months/

https://www.digitalphablet.com/news/sora-stuck-in-research-rut-openai-loses-5-billion-in-a-year/

[queen veldt]

Everything costs money.. sora 2 is EXPENSIVE the amount of training data it have is the reason why it's so expensive

[round sedge]

Maybe because of the word bullshi

[keen beacon]

https://fortune.com/2025/01/07/sam-altman-openai-chatgpt-pro-subscription-losing-money-tech/

[queen veldt]

The more the training data the more computation is needed , same goes for local LLM'S

[keen beacon]

OK, what information do you want? That’s credible and I’ll pull it up for you right now.

[queen veldt]

Bro just said "i trust the news on television"

"i trust the government"

Fine go ahead and download some large llm and run it through llama and see how much time it takes to generate text

Now imagine how much time it would be needed to generate a video

With GOD KNOWS HOW MANY TRAINING DATA "

Umm what super computer

You know it takes time for their gpus to generate a video right?

You're saying it's "free"

[keen beacon]

See this is what I mean about AI mania

No amount of information will break the belief system

[queen veldt]

Lmao he's clueless

[keen beacon]

Because everything is gonna get categorized in the often so common category

It’s either bias or fear mongering

The AI community is immune to criticism and immune to scrutiny

[queen veldt]

.

[echo sinew]

Hey guys! Leaving this friendly reminder:

✅ Treat others with Respect. Be kind, assume good intent from others, and keep disagreements respectful. It’s encouraged to share your disagreements, but only if it’s done in a respectful and productive way.

[queen veldt]

They lose money by wasting time on generating your video it costs electricity and their gpus need to be replaced every few months

[keen beacon]

https://blog.samaltman.com/

OK, so this is a website that is Sam Altman’s. This is his personal blog.

Now, if we go there, let’s see what he says about sora

This is him personally, addressing sora on his blog

I’m glad you brought up Mi

https://microsoft.gcs-web.com/node/34236/html

OK, you brought up Microsoft right

https://techstartups.com/2025/10/31/openai-is-hemorrhaging-billions-microsoft-filing-reveals-openai-lost-11-5-billion-last-quarter-amid-ai-bubble-hype/

Anyways man, believe what you wanna believe I don’t wanna argue

We could agree to disagree

[queen veldt]

Nah don't waste time arguing with this man

[keen beacon]

Nothing I presented as false articles

I presented legal financial documents from the actual company

Statements from his blog

Many reputable article sources from reputable news organizations

I don’t need to defend it

There’s nothing to defend

[queen veldt]

I don't plan to since i see it's pointless to argue with you

[keen beacon]

Publicly available information

I mean, I already know the information. Maybe you should go look it up lol

Anyways, whatever man, no hard feelings you do you I do me believe in whatever you want and I will do the same

I’m done here. I’m done.

I don’t wanna play these games

I’m not talking about the future. I’m talking about right now. I’m currently.

No, what could be or what will be?

I’m talking about present day. The realities is on the ground currently as of now and today.

[queen veldt]

But you do agree that OpenAI lost bilions on Sora2??

Google doesn't offer free veo video gens since it's expensive to create a video

OpenAI earned $0 on sora2

0

People made funny videos of cats on doorcamera and stuff

Failed marketing play

[keen beacon]

https://www.bloomberg.com/news/articles/2025-11-07/openai-asks-us-to-expand-chips-act-tax-credit-to-ai-data-centers

😂😂😂😂😂😂

[queen veldt]

Which nobody will use...

[keen beacon]

https://techcrunch.com/2025/11/06/sam-altman-says-he-doesnt-want-the-government-to-bail-out-openai-if-it-fails/

https://www.citizen.org/news/openais-request-for-massive-government-bailout-package-is-pure-corporate-entitlement/

[queen veldt]

Nobody will use the a.i. slop videos for real world money

[keen beacon]

https://www.reuters.com/business/openai-does-not-want-government-guarantees-massive-ai-data-center-buildout-ceo-2025-11-06/

[queen veldt]

Only real usage of sora2 were cat videos and remixes

[keen beacon]

https://www.cnbc.com/2025/11/06/trump-ai-sacks-federal-bailout-openai-friar.html

I don’t care it doesn’t bother me either or for this or that

I hope they succeed I want them to succeed

I like open ai and I think they’re going against a lot of odds

It’s not false information because it’s coming from officials dude

[queen veldt]

😂

[keen beacon]

It doesn’t matter it’s irrelevant. I got no choice.

If I trust him, if I don’t trust them, it doesn’t matter

[balmy mist]

5.1 is agi

[keen beacon]

The same to you playboy

https://tenor.com/view/biggie-smalls-the-notorious-big-gif-22926755

[queen veldt]

False articles 😂

From google official site

[keen beacon]

They could be for all we know

But I have nowhere else to pull out information from except these articles from these new sources

Do I think they are? I don’t think they’re false, but I could see why somebody may think that.

[queen veldt]

You're saying tokens are free

Explain than

[keen beacon]

Thank you for giving me inspiration for my next video

[queen veldt]

..

How they don't lose money

Nono you said "openai don't lose money on tokens "

You said "OpenAI doesn't lose money on tokens"

Explain that

I haven't watched devday

I won't

Since you have no clue they lose money on tokens? 😭

[potent glacier]

Is Reve Edit Fast any good?

I haven’t tried it yet

[queen veldt]

But they do lose time on generating YOUR video instead of someone who pays

It does cost to generate a video

Bitcoin miners lose money on mining bitcoin

Openai pays the Microsoft that operational costs

Which is huge with such a huge video gen model...

But it isn't free...

They're losing money with each video

Eh who I'm arguing with..

Yeah this is my alt

I like my privacy

Yeah i visited it to laugh at it 🤣

I bet i could find some exposed api key in your github

Oof you haven't even built such an advanced applications

Look he loves to spread lies on reddit too

I don't see people agreeing with you

I just think that you think you're way above your actual limit

[echo aurora]

Hey reminder that respectful debate is fine, but as soon as it get disrespectful it's not okay.

[queen veldt]

Sorry pineapple

[keen beacon]

There’s something wrong with images in LM Arena

The images aren’t coming out like they’re supposed to

[echo aurora]

Is this for image-edit?

[keen beacon]

Odd

[echo aurora]

We did have issues earlier, but has since been fixed. Are these examples of chats that are slightly old?

Would you mind trying a new chat and let me know what happens?

[keen beacon]

Those are all brand new chats, but I will try a new chat with the new prompts as a control variable

[echo aurora]

Sounds good, it was fixed very recently.

[keen beacon]

Yeah, there’s something funny going on

Unless this was what the fix was supposed to do

[echo aurora]

Still? Can you send an Eval ID of the new chat?

[keen beacon]

One sec

I would send one, but I’d rather not but hold on. I’ll give you something better.

If I had to guess it would be something with the filter or the routing. 🤷‍♀️ maybe the api end points.

Because it is essentially listening to the instructions of the prompt and it looks like the same models are doing what they normally should be doing except following the input image but if I refresh once or twice it does it again

All those are new chats or refreshes

[echo aurora]

Can I have the Eval ID of one of them?

It'll look like this

[keen beacon]

I know what they are

All right, I’ll give it to you but

019a60e0-f858-71f3-9e05-58e4ea21183c

Ya just don’t be snooping plz lol

Everything was done for the purpose of testing

(Not suggesting that you would, but I read the terms of services)

[echo aurora]

Understood, we're just looking at this one Eval related to the bug. If it makes you feel better it's our engineers looking into, I personally don't look through content from eval IDs.

[keen beacon]

Well, I don’t really mind but it’s just gonna be awkward to explain

But it is what it is hopefully they get it fixed

[echo aurora]

This is really helpful as after the fix went out we haven't been able to repro the same problem. We were under the assumption the fix was working properly.

[keen beacon]

I guess I need to do some more testing

No, there’s something off for sure

019a60fc-046e-77ae-b489-c57a28f91a62

It’s way too much of a coincidence that keeps occurring this often when it never used to

No, I’m not sure what it could be it listens to the instructions, but it’s not able to read the input properly but sometimes it works so I don’t know

[fossil lance]

LM arena image generation is bugged now . Basically , if you tag the image ( the one you wanna edit ) and write a text prompt , the model will completely ignore the tagged image and generate the image based on text prompt .

Here is an example . Let’s say you tagged the specific dragon image and write a prompt “add big plasma cannon on dragon’s back” . The model will give you random dragon with plasma cannon on its back because image you tagged is not reach to the model in first place .

Tested in Yupp and Poe . No problem . LM arena is the one that encounter the problem .

[queen veldt]

It works fine for me

[fossil lance]

It happened in Seedream for me

[keen beacon]

Yeah, it’s not something that happens to common, but it does occur

Yeah, well the image part of it is a little weird because of two observations I’ve noticed

The first being that the model gets stuck on the context of the first image and doesn’t change models afterwards

So this is suggesting that the original image gets carried over into the next request

And that simply implies that the models don’t necessarily change and remain the same regardless

[queen veldt]

Maybe it's browser/phone specific

[keen beacon]

I’m not sure how it’s supposed to work, but I was under the assumption that they are supposed to reset

[queen veldt]

Have you tried chrome or some other browser

[keen beacon]

But it looks like some form of memory or context remains

And then it resurfaces

[fossil lance]

Yeah I will test more and see what happen . But 100% sure it’s bugged for me too .

I am on mobile . Tested in Safari and Google app browser . Will test in PC .

[queen veldt]

Works fine on my android

Guess it's apple

[echo aurora]

For anytimes you experience the bug if you're able to share the Eval ID that'd be very helpful.

[keen beacon]

And I will give you an example of why I think my theory and why I think it

I use these two images as my reference

And my intent was just this

Which is correct and was one of the responses but

I also got those in one of the models and now white and how I link them together is because I often use the same kind of anime style when I’m generating images from other chats

But somehow, it leaked into this new chat

And started adding weird elements I didn’t even mention

And how do I know that it’s the same aesthetic from other chats?

[queen veldt]

That reminds me when it mixed one image with the other in same char when i switched from nano to seedream

[obsidian cargo]

thats an interesting style :O

[keen beacon]

Well, that’s why I’m able to recognize it

And it’s probably my second most common used style

[keen beacon]

So to see it randomly pop up into this chat, makes the coincidences

Either one big strange, coincidence, which you could possibly be

But either way it’s an unsettling question because if it’s not the case, then that means that some form of memory or context gets retained somewhere

Which would imply that their routing lol

[fossil lance]

XD

[jade egret]

when gemini 3 ):

[simple sleet]

Hi, has anyone managed to use LIGHT VAE - wan 2.2? My text-to-image generation is working terribly.

https://github.com/ModelTC/ComfyUI-LightVAE

[keen beacon]

Is that Spyro ?

[empty stump]

why isnt kimi k2 thinking in the arena

nvm

just not the leaderboard

[fossil lance]

It’s the random dragon Seedream generated that I never asked for . I told it to edit my ice dragon and it gives me completely unrelated image haha .

[keen beacon]

But that’s what makes it strange

Because it adhered to one portion of the prompt because it generated a dragon, even though it wasn’t the right one

[fossil lance]

Because I say “Dragon” in text prompt .

[keen beacon]

Yeah exactly

[fossil lance]

Seems only following text prompt and ignore tagged image .

[keen beacon]

So will that mean it would mean that it’s failing to either a see the input image or or it’s not choosing to adhere to the input image

[fossil lance]

Exactly what’s happening in my end .

[keen beacon]

What happens when you refresh?

[keen beacon]

Are you only using text or using a reference and is also?

[queen veldt]

Text only

[keen beacon]

Yeah, the problem is with the input image

The text part works fine

[fossil lance]

Holy it gives me right image when I refresh

[keen beacon]

Yes

That’s what I’m saying

[queen veldt]

Wutt

[keen beacon]

Which makes this very baffling to me

It makes me think it’s the filtration for some reason lol

[fossil lance]

So … only the first generation is bugged lmao

[keen beacon]

I’m not sure I don’t know the backend

But not many things would be causing this

There’s only a handful of things that would cause this lol

[fossil lance]

But thanks anyway … at least it’s usable now haha

[keen beacon]

I got more questions than answers 😂

I’m trying to understand in reverse engineer the issue they’re having

[queen veldt]

iPhone related, 2 api requests one with one without reference image

[fossil lance]

Yeah i am using iPhone .

[queen veldt]

Maybe site reads only the first in the row and when you refresh somehow it reads the last image it returned

[keen beacon]

Well, I’ll be honest. I’ve been trying to understand exactly what lm arena is and what’s its role in the ecosystem of AI

And a lot of interesting observations about the community and the arena itself

But I guess more than anything to open up my eyes too how expensive AI services really are

And that a lot of people are gonna get left behind if it wasn’t for services like this

But on the other hand, it’s also exposing users just not wanting to pay for a eye services also so it’s two a streak

Too bad AI so expensive that both can’t meet somewhere in the middle

[obsidian cargo]

a eye services

[keen beacon]

They’re both in a awkward situation from a consumer and a business perspective

[queen veldt]

$200 a month is a reasonable price for my usage of the ai

[keen beacon]

Yeah, but $200 is also a lot of money

Imagine you’re in a Third World country work the average pace four dollars

[queen veldt]

Even though i WILL NEVER PAY $200

[keen beacon]

That’s what I also mean though because you meet everybody else would probably use up that amount if not more

I don’t mind paying for AI. I just don’t have the means. I’m sure like many other people.

[queen veldt]

I paste full files with 2k lines

[keen beacon]

I just think if you have a chance in the opportunity and you got the extra money, I think that the right thing to do would be to pay

Even if it’s 20 bucks a month or whatever

And I could also sympathize and understand the people that just don’t have the means

That’s why I say that the AI industry the community and just everything is just kind of rocky to be honest

It’s uncertain built on promises of tomorrow

[queen veldt]

Yeahj

Openai is at a loss with $20

[keen beacon]

Yeah, I was just reading an article. There are a loss at even $200 a month. They’re losing money and they are really if you think about it all the services they include for pro users.

[queen veldt]

Anyone who gets the Plus will abuse the chatgpt to full

[keen beacon]

They’re awfully generous

You can’t abuse it

It’s designed not to be able to be abused in that sense

[queen veldt]

Yeah not abuse it really but like

[keen beacon]

Because even on a pro count you have limits

[queen veldt]

Use it fully

[keen beacon]

Well, I’m just saying you’re not gonna exceed what you pay for it

Even if you do and use it, Hella crazy you’re not gonna exceed the $20 you pay for

[queen veldt]

For some basic tasks even if i don't need to

[keen beacon]

You have to understand the business structure of how they generate money

And the difference between API and subscription based

So I like to think of like the subscription base model is all you could eat kind of, but after you get to a certain point you’re gonna get cut off

Let me actually put this a better way

How many tokens do you think you use?

[queen veldt]

I think claude is more realistic service even though I'm not happy how they have limited access to models but that's what I'd realistically get for that price

[keen beacon]

Well, how many tokens do you think you used daily or even monthly?

1 million 2 million?

[queen veldt]

Not 1000 or so extended thinkings

[queen veldt]

Idk i use a lottt

[keen beacon]

Well, how much though if you had to guess

[queen veldt]

I paste full codes and create apps for 6 hours

My chrome session starts to lag

[keen beacon]

OK, so let’s just break this down even more simpler than

Leg say 1 million words is roughly around 750,000 words

Average person (casual typist): about 40 words per minute (wpm)

• Average professional typist: 65–75 wpm

• Fast typist / transcriptionist: 100–120 wpm

Calculating for one hour:

• 40 wpm × 60 minutes = 2,400 words per hour

• 65 wpm × 60 minutes = 3,900 words per hour

• 100 wpm × 60 minutes = 6,000 words per hour

So a typical person can type roughly 2,400–4,000 words per hour, with skilled typists reaching 6,000+ words per hour.

So let’s just go with the number 2400 per hour

So at 2400 words per hour, even if you do 24 hours straight

You’re at 28k words and roughly double that if you do 4000 words per hour 41k words in 24 hours

[queen veldt]

I first send one file

Which has 2k lines

And tell it to explain to me what it has read there

[queen veldt]

And than i provide other file which has 2k lines and than tell him to do this and that

[keen beacon]

So let’s just take these prices so it’s $.25 per 1 million tokens and two dollars per output tokens

Let’s just say even for argument sake you do 100,000 K a day

You know what we could even say 500,000

At 25 cents per million tokens

That’s only 12.5 cents for input token

lol

[queen veldt]

Yes it's cheap but still

These are my two biggest sessions this month

[keen beacon]

How much price per token?

For 1 million tokens input and output

[queen veldt]

My average session is like 15k tokens which is $0.05

This was gemini 2.5 pro

[keen beacon]

Well, those are extremely high prices

[keen beacon]

Kinda

Holy smokes

[queen veldt]

Well it is gemini 2.5 pro thoo

[keen beacon]

Crazy

[queen veldt]

Yes and ai studio is giving it FOR FREE

[queen veldt]

I don't have gemini sub

[keen beacon]

How are they giving out for free?

Anyways, though, either way, like I said, this is what makes it really rocky

In order to get the customers in the first place, you gotta offer free services because of the competition

So then, how are you able to generate money? lol

And that’s why people are speculating why it’s a bubble because AI is failing to really generate the revenue versus what it spends

[queen veldt]

Ik makes no sense

[keen beacon]

Well, if we look at the reporting the quarterly reporting from Microsoft loan

Quarterly meaning three months

Open AI lost almost $12 billion dude

In three months

For operational costs

[queen veldt]

For a hope of "better tomorrow "

[keen beacon]

I don’t think you could fathom how fast that’s burning money

[queen veldt]

Which might not even come

[keen beacon]

That’s a lot of money to burn in three months a lot

They only generated like 12 billion last year or something like that in revenue

[queen veldt]

Lmao

And lost whole year of revenue in 3 months

[keen beacon]

Well this year according to Sam Altman they’re gonna generate close to 20 billion

[queen veldt]

Yes sure

And next is 1 trilion

[keen beacon]

The steaks are really high

Winner take all

[queen veldt]

They're trying to scale something which is super expensive to scale

We'll have datacenters around the whole globe if they continue like this

[keen beacon]

https://youtu.be/aI3M9gUrDdg?si=8HdKvQLpNtMTNe1B

The problem also that people don’t wanna see is the fact that the users are also part of the blame

They’re infatuated with AI, but don’t want to pay for it

[queen veldt]

I think eventually in few years there will be a big gap

Between the free users and paid users

[keen beacon]

There’s already a gap

[queen veldt]

People with $200 to spare will get the precise a.i. while we're gonna have some a.i. slop

[hazy kernel]

Ye I don't have access to Gemini deep think models for free

And gpt 5 pro either

[keen beacon]

You gotta think bigger OK like how are they gonna recuperate this money?

So I like to look at it as like rent or IOU right now they’re giving me a service. I guess you could say for free.

And eventually, they’re gonna have to come collect and since they can’t forcefully do that what means do they have?

The first option is obviously failure and call it quits

Or the more plausible one the second option is make it where you’re forced to interact with AI everywhere

And so in the future, we may not have the luxury of not dealing with it at all

Meaning, we will be not forcefully forced to pay for

But there’s gonna be no way around it type of thing

Almost like not having a bank account not using digital pay services. There’s just no way around it if you wanna participate in the modern economy.

[queen veldt]

Maybe we'll become braindead and must use a.i. to do something useful in life

[keen beacon]

No, you’ll just be forced to use it indirectly

Like you’re forced to own a cell phone, even though you’re not really forced, but you need to get by in the morning times

[hazy kernel]

We can still learn new things from ai

And hope it's not misleading

[keen beacon]

The fact that people don’t think it’s misleading, already is baffling to me

[queen veldt]

It's just badly though out

[keen beacon]

Do people not see how AI especially western AI how bad it hedges?

[queen veldt]

They slapped bunch of training data in some auto-correct machine and called it a.i.

[keen beacon]

https://youtu.be/9cGxewHh3bw?si=3ACzHTHRFwce7gpX

This is already the first sign of misleading

[swift oyster]

upcoming checkpoint 4

[keen beacon]

Because what you say, and the output you get via the language and the words and semantics in context

Is already filtered

And I know this sounds crazy like a conspiracy theory or whatever but I’m just telling you how it really is

[queen veldt]

Wtf is this

Fake ads

[keen beacon]

Movementlab a legit ai service

[hazy kernel]

Ah damn maybe I have ever said something wrong here cause I'm not a native English. (I just use autocorrect and translator lol)

[keen beacon]

No, you didn’t do anything wrong I understand where you’re coming from

And it is gonna get used for a lot of good and we are gonna learn a lot from it

[queen veldt]

Ain't no way ai creates Minecraft with animations with single prompt...

[hazy kernel]

Imagine that ai have access to your terminal for its tools

[swift oyster]

it took 2 prompts 🙂 but we are getting there with checkpoint 4, looking very good on our end.

[toxic egret]

gray

[keen beacon]

So for all the good, the AI is gonna do there’s gonna be a equally bad amount it’s gonna do

And vice versa

But sometimes the scales tip to far of one end

[echo aurora]

Btw @keen beacon fix is in route with the issue you raised earlier

[keen beacon]

Cool thanks 😎

[queen veldt]

I'm not buying it... It looks too good to be true... The torch has burn animation and the ground looks too textured not even the most advanced a.i.s could do this..

[keen beacon]

It’s super 💨fast

Like crazy speed lol

[keen beacon]

“I’m not buying it” 🤣🤣🤣

Isn’t that the story of AI we were just talking about

[queen veldt]

Yeah i meant like "I'm not falling for it" 😂

No a.i. could do this

Yet

[keen beacon]

Well, have you tried it?

[queen veldt]

Well i haven't tried their a.i. but i know from previous a.i.s like gpt 5 high and gemini2.5 pro and deepseek and others that they can barely make a ground that i won't clip through

[keen beacon]

https://movementlabs.ai/

Remember that bias we were talking about

The great filter, the subconscious

[queen veldt]

I don't think some a.i. I'm not aware of can pull off a working game with 3 prompts

[queen veldt]

Yep I've seen enough

[keen beacon]

I don’t know anything about coding so I can’t really test

Is that bad output or what?

[hazy kernel]

what's that output

[keen beacon]

Yeah, that’s what I’m curious. I don’t know what that means.

I don’t know if that’s good or bad lol

[queen veldt]

Broken markdown format

[keen beacon]

You might as well be speaking a different language lol

[queen veldt]

Html code that's impossible to copy

[keen beacon]

Oh ya

I forgot about that that it didn’t have that

Because I was trying to copy and paste something wrong and I noticed that too, but didn’t think much of it

It is a odd feature, not to include though

[queen veldt]

Yeah happends on lmarena

But i just tell it to "Give it in markdowm format"

[keen beacon]

Right, but is the code good though?

[queen veldt]

And it gives me in not broken version

[hazy kernel]

But why does the model giving out weird response over your prompt

[queen veldt]

Idk... It gave me html code even if I didn't asked for it

Exactly my question

I asked it what model it was

[keen beacon]

Well, we need a more thorough test

Yeah, but if the model is try not to reveal it instructions

[queen veldt]

First prompt i said was

And it told me about some voxel game?? It didn't mention it or anything...it was my first message to a.i.

[keen beacon]

That’s just a little strange because why would the model reveal it instructions if you asked it so basically?

They’re not supposed to reveal their system instructions. I don’t think.

[queen veldt]

Just testing it blatantly

[keen beacon]

No, I hear ya, but I’m just saying that typically from what I understand models aren’t supposed to reveal their system instruction

[swift oyster]

we are making some updates to our website currently

[keen beacon]

So as long as it didn’t veal it regardless of what the answer was it technically did a job. I guess if that was the case

[swift oyster]

adding a preview button

[hazy kernel]

Isn't this the same output?

[swift oyster]

so website might be down for 5-10 min

[keen beacon]

Oh, it’s down

😂😂

Damn bro

That’s bad for business

Two potential customers that viewed your product

And it was down

Well, a magician never reveals his tricks

I was wondering where you were

😂😂😂

[hazy kernel]

Wait is this exactly the same like @queen veldt prompt output??

[keen beacon]

I had a feeling you weren’t too far

https://tenor.com/view/round-2-boxing-ashita-no-joe-anime-fight-gif-4113432034406286548

Where is the rest of the gang?

Man, I was cracking up last night

[queen veldt]

Probably

[keen beacon]

Oh, you should’ve seen it yesterday

They were going at it

[queen veldt]

It talks about cloudflare being down...

[keen beacon]

Yeah, the dude said the website might be down for five or 10 minutes. They’re updating it.

[queen veldt]

Yeah but i asked it in prompt

And it returned me a html code as response

Instead of showing me on the page "hey we're down"

[keen beacon]

Yeah, but if it’s down, what is it gonna connect to?

[queen veldt]

Also it made me ran out of daily prompts (3)

[keen beacon]

There’s actually 50 when you login

I know at first I had my doubts to

But there was this whole big debate about this yesterday about this company and about this AI specifically on what a MPU is

And I don’t think there’s been a satisfactory answer to quell all the scrutiny and criticism lol

But at the same time there hasn’t also been any substantial proof that it’s something other than what it’s saying it is if you will in a very oversimplified euphemistic way

[queen veldt]

Yeah even gpt said

[keen beacon]

Yeah, but they’re still anomalies that are hard to explain also

The speed

And other anomalies that have yet to also have a explanation

[queen veldt]

The Minecraft in 3 prompts is impossible

Maybe in a few years

[hazy kernel]

What's under it

[hazy kernel]

That

[keen beacon]

Well, that’s $1 million question

[queen veldt]

They're some scammers I don't trust them

[keen beacon]

What are they scamming exactly

[queen veldt]

I don't know and don't want to find out

[hazy kernel]

It said "mpu thing is just marketing hype"

[keen beacon]

And a very well could be

[queen veldt]

If they're using MPU without providing any proof or clear explanation of what it is

I don't trust them

[keen beacon]

I agree I acknowledge the fact that this is yet to be answered for

But I’m also failing to see what the scam is at least currently if it is a scam

[hazy kernel]

And that can't be proven with "proof mpu doesn't exist". The correct way would be "proof mpu exists"

[queen veldt]

I won't be giving them my money until i see some credible YouTuber actually generate Minecraft with 3 prompts

I think it's all just lies

[keen beacon]

Well, how do you measure value from AI in general? I’m just curious.

If we qualify it into money

For example, if I go out and I wanna have fun entertainment, I’m willing to pay one dollar per hour of entertainment for argument sake

And let me just say that I get entertainment value from using AI

And I’m going to exchange one dollar for one hour of AI use

Well, AI fraud and schemes are definitely under rise. Even the government warns about it.

Yeah, you beat me to it

[hazy kernel]

Even it's real, other companies would do the same earlier.

[keen beacon]

So it is a valid concern that everyone’s having

[keen beacon]

That’s exactly the argument I was making yesterday

But nobody wanted to hear it lol

I’m trying to understand what you mean by that statement

I don’t think other companies would claim their own proprietary hardware and models.

I’m not sure if I understand exactly what you mean by that

[hazy kernel]

TPU created by google, and it's everywhere. But I'm sure Google own the claims of TPU

[keen beacon]

Fair enough

Pre-self-explanatory and pretty straightforward so I would have to agree on that statement

But I guess what the prom here is are we assuming that it is a fraud because the MPU isn’t defined?

Or are we assuming it’s a fraud because they’re using a fraudulent term like MPU when they’re using other hardware and other software that isn’t necessarily theirs and claiming it to be their own

[queen veldt]

And a suspiciously accurate minecraft game in just 3 prompts

Like even gemini 2.5 pro and gpt 5 high is struggling to create some basic apps

[keen beacon]

Well, this is a little bit paradoxical if we look at the subjectively

Because how do you define artificial intelligence?

lol

[queen veldt]

I don't

[keen beacon]

What exactly is intelligence?

And how do you qualify that into?

[hazy kernel]

I don't know if it's fraud or not, they seem like don't claim their model uses mpu

[keen beacon]

Well, no, they both make good points. They’re making solid points and these are really good arguments.

[queen veldt]

Hahahahahahahahahah

[keen beacon]

But what I’m saying is if we are going off of the fact that MPU can’t be defined, and thus, for it is a fraud

How do we define quantified artificial intelligence as a term itself?

And one example I’ll give the court here one second

https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/

Gemini 2.5 is a thinking model, designed to tackle increasingly complex problems. Our first 2.5 model, Gemini 2.5 Pro Experimental, leads common benchmarks by meaningful margins and showcases strong reasoning and code capabilities

Besides the word benchmark, nothing they’re saying here could be quantified

We can’t quantify strong reasoning as a term and measure it

Or the term in words that used meaningful margins

[queen veldt]

True

[keen beacon]

And they’re not the only ones that do that all companies do this

They use this specific language when they’re describing their models improved reasoning, or better capabilities or whatever all these terms, none of them could be quantified in any measurable way

And so what’s left is the benchmarks?

And the benchmarks themselves could be skewed

And yet this is the only form we have to claim legitimacy lol

Because we have no other way of measuring “artificial intelligence “

To me, the word is no different than the word supplement

Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

[queen veldt]

Real way of measuring it is to give it one task and just let it do it

Like for example

One youtuber i like to watch

[keen beacon]

Adaptive Learning, Autonomous Decision-Making, Generative Models, Cognitive Computing, Contextual Awareness, Deep Reinforcement Learning, Embodied AI, Explainable AI (XAI), Human-in-the-Loop, Interactive AI Systems, Knowledge Graphs, Language Understanding, Multi-Modal Perception, Neural Architecture Search, Predictive Analytics, Real-Time Reasoning, Self-Supervised Learning, Semantic Understanding, Synthetic Data Generation, Task-Oriented AI, Transfer Learning, Zero-Shot Generalization, Agentic AI, Continual Learning, Scalable AI, Edge AI, Responsible AI, Large Context Models, Multimodal Fusion, Decision Intelligence, AI Orchestration, Model Interpretability, Autonomous Agents, Next-Generation AI, Human-AI Collaboration, Reinforcement Optimization, Knowledge Distillation, Situational Awareness, Adaptive Reasoning, AI-Driven Insights

[queen veldt]

He does same prompt for every model

[keen beacon]

Yes, but those are all objective tasks on an individual level for your use only

[queen veldt]

It's the "Create mac os"

[keen beacon]

And everybody has a different use, so how can you benchmark something like that?

[queen veldt]

Create tom hardware website

Create a racing game

[keen beacon]

Right, but if it makes a basic one or a really sophisticated advanced one

How do you know that the models improved if they both essentially do the same thing if you will?

Even though one could do a crappy and the other one could do it nice

But they’re both technically artificial intelligence

[queen veldt]

Well newer models have hover effects and cool animations

[keen beacon]

You get what I’m saying? I’m not disagreeing with you. I’m just trying to expand a little bit onto what you’re saying.

[queen veldt]

Yeah but this is coding benchmark

There's also math benchmark

Mmlu etc..

[keen beacon]

Yeah, well he’s definitely not helping a situation if I’m being honest, no offense, buddy

But at the end of the day, dude, even if he’s not helping his case, and even if whatever the case may be here

[queen veldt]

ENTIRE MODEL ONTO IT'S MEMORY HAHAHAH

[keen beacon]

I still don’t see the scam though

Because I don’t see no difference between what he’s doing and what other parts of AI industry companies are doing

Absolutely zero difference

Well, actually, I take that back one second let me rephrase that

Because you are right to the extent of which and I see your point of view open AI Anthropic, and all these other companies, at least they know the services and they’re building actual models that are their own

I guess what I’m trying to say is more of like the business side of it