general

[keen beacon]

Because other people are building off of their tools too with API calls

For example

https://soraclean.app/

https://sora-watermark-remove.com/

https://sorawatermarkremover.net/

https://2-sora.com/sora-watermark-remover

I fundamentally agree with you, but I think where we disagree or see a difference

[hazy kernel]

But I'm still curious how do they make their model generate texts so fastt

[queen veldt]

Well i can't say for certain they're scam but i see big red flags

[keen beacon]

You see all those links I sent you

[queen veldt]

If im looking through Gehlo perspective

[keen beacon]

What do they all have in common they all have in common?

[hazy kernel]

Oh there is the loopholee

[keen beacon]

https://sara2.ai/features/sora-2-watermark-remover

These are all obviously ripped off and scams

Super generic vibe coded 100% of the way

Do I really think if I pay for any of these services that I’m gonna get a really high-quality watermark removal services absolutely not

But if I use the services, are they gonna remove the watermarks yeah absolutely they will

So that’s the dilemma here

I’m still getting the service even though it’s not high-quality it’s generic

Bro, I’m not doubting you

[keen beacon]

And I truly hear it and feel what you’re saying and I agree with a lot of parts of it

[queen veldt]

Hahaha simplebench questions??

[keen beacon]

I don’t know I don’t know how to explain this lol

[queen veldt]

Hardcoded

Same as the latest ai models

[keen beacon]

OK, let’s take it one step back then

[hazy kernel]

What do you mean

[queen veldt]

Nobody in their right mind hardcodes latest a.i. models

[hazy kernel]

Kiwi K2?

[keen beacon]

But just take a little step back and go even further back

What did people expect when AI was able to produce code?

What did people assume was gonna happen and how people are gonna use it?

[hazy kernel]

Hardcoded is system prompt, Ig, in context. *: The "hardcoded" meaning in literal is different.

[keen beacon]

Did they not foresee this happening?

That there’s gonna be a rise of this kind of “” fraudulent behavior

[queen veldt]

Isn't this the system prompt

[keen beacon]

Or maybe it’s designed in such a way that this is the only plausible outcome

[keen beacon]

And this is just one of many

I guess what I’m trying to say is you guys are hating on the player but not on the game itself

But the game is what produces the player

In other words, it’s saying: Don’t hate the individual for playing by the rules or taking advantage of the system; criticize the system that makes those rules possible.

In this case, it’s the AI industry of themselves

No

I’m not gonna do that to prove my humanity lol

Take it as you will or however, you want

But I gotta go I got caught

So I got a weasel on my way out of this

🤣

Jk I got some work to do and I’ve been spamming too long

My context window is running out

[queen veldt]

Gehlo was sending YouTube links too fast

[keen beacon]

But I’m gonna take a break. I’ll talk to you guys a little bit maybe later tonight or something if anybody’s on

[queen veldt]

And circling around the images

Too fast

[keen beacon]

https://tenor.com/view/shhh-eric-cartman-robot-south-park-s8e2-gif-22319068

[queen veldt]

Finding resources too quick

Which i found sus at the very begging but i was like "it can't be"

Guess this proves the "dead internet theory"

Maybe the only real person here is Craig Federighi

[whole sundial]

sara 2

[queen veldt]

And rest of us are here built as beta chat bots

To entertain few people who join this server

[hazy kernel]

There is so many bad possibilities 😔

[sharp knoll]

hello

[polar niche]

@echo aurora

Who won october contest?

[dense sphinx]

Greetings everyone

Can you explain what ernie-5.0 preview?

[simple sleet]

Question for ComfyUI users:

On an RTX 5090, is there a difference in speed improvement between:

PyTorch 2.8.0, Python 3.12, CUDA 12.8

and

PyTorch 2.9.0, Python 3.13, CUDA 13.0?

Are there any benchmarks or verifiable data that demonstrates this difference in improvement?

[sullen quest]

its a ernie model, whats confusing about that

[eternal narwhal]

In one of my chat its stuck at this point. Can anyone tell how to fix that issue as deleting this I will loose the resources i gave it earlier

[fluid yew]

good news, none.
bad news, you’ve officially lost it🙂.

[valid compass]

image to video, as it starts working?

[icy sage]

i'm hungry

[hollow thicket]

wow, didn't knew it is a thing, and thought propitiatory products are exempt

[keen beacon]

https://youtu.be/TCwS8nL0JaI?si=zeAM-VNqi3Z567xd

https://machinelearning.apple.com/research/illusion-of-thinking

https://arxiv.org/pdf/2504.20879

“Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play
an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has
emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we
identify systematic issues that have resulted in a distorted playing field. We find that undisclosed
private testing practices benefit a handful of providers who are able to test multiple variants before
public release and retract scores if desired. We establish that the ability of these providers to choose
the best score leads to biased Arena scores due to selective disclosure of performance results. At an
extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release.
We also establish that proprietary closed models are sampled at higher rates (number of battles) and
have fewer models removed from the arena than open-weight and open-source alternatives. Both
these policies lead to large data access asymmetries over time. Providers like Google and OpenAI
have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a
combined 83 open-weight models have only received an estimated 29.7% of the total data. With
conservative estimates, we show that access to Chatbot Arena data yields substantial benefits; even
limited additional data can result in relative performance gains of up to 112% on ArenaHard, a
test set from the arena distribution. Together, these dynamics result in overfitting to Arena-specific
dynamics rather than general model quality. The Arena builds on the substantial efforts of both
the organizers and an open community that maintains this valuable evaluation platform. We offer
actionable recommendations to reform the Chatbot Arena’s evaluation framework and promote
fairer, more transparent benchmarking for the field.”

https://dev.to/sujiths/the-great-llm-benchmark-illusion-why-your-enterprise-ai-strategy-needs-real-world-testing-5foo

[hazy kernel]

Damn glm kinda cooks

Does parameter affect the score

[lapis merlin]

Arichat

Click Ari-Account-Registration-bot

[keen beacon]

benchmarktarded
adjective | \ ˈbench-ˌmärk-ˌtär-dəd \

Definition
1. Exhibiting an excessive or obsessive fixation on artificial intelligence benchmarking metrics or performance scores, often to the exclusion of practical understanding, contextual judgment, or creative insight.
2. Characterized by treating benchmark results as absolute indicators of value, quality, or intelligence in AI systems.

Example:

He’s so benchmarktarded he thinks a higher leaderboard score means genuine intelligence.

benchmarks and tasks only measure performance under artificial conditions. The only legitimate evaluation of an AI is through actual use case performance how real people use it, what they use it for , and it’s popularity with a users that use it as well as how effective they are doing what they’re supposed to be doing

Aka not publicly available information that has been released regularly

[cursive lagoon]

My Name Is James From Kenya I would like to compare Generative video Ai Models

[magic stag]

"With retries at 20" LMAO

[marble birch]

😿

[worldly grail]

reddit

[keen beacon]

https://youtu.be/ply0RPuEfB4?si=Xza2Mx9w4q6-Ocey

[ocean ferry]

Kimi K2 Thinking is a beast

[crude harbor]

Movementlabs A/B testing I got this morning

Anyone else got it? I saw it once

[magic stag]

And they didnt mention how many for gpt5 or what reasoning effort it is....

For a reason

Can't imagine why they'd use such radically different numbers

Almost like they wanted to run it 5000 times

And get some variance

For whatever reason

(Heh let's not put it on high either because lag lol)

[keen beacon]

https://youtu.be/GydlPnP7IYk?si=x6GI43ngyFaFwj5Q

[hazy kernel]

I was thought that model named kiwi lol instead of kimi

[keen beacon]

It’s really getting a lot of love online from some pretty OK sources

[polar niche]

Any way to run it in a cli

And I don't have that much ram to run

[polar niche]

OPEN AI's got the mainstream

So probably the most popular one

[brave orbit]

i think grok 5 since elon sayed it would have My estimate of the probability of Grok 5 achieving AGI is now at 10% and rising and and with elon also I now think
I now think
has a chance of reaching AGI with
Never thought that before. and so i think so

i know this isnt the topic but i mean grok is so underrated

[brave orbit]

elon hyped grok 5 a lot so

[polar niche]

What's grok 5's eta?

[brave orbit]

and Elon Musk has stated that Grok 5 will be released before the end of 2025

[polar niche]

No way

[brave orbit]

yea he sayed that on x

and biggest thing of grok 5 is elon also says it has Dynamic Reinforcement Learning

[polar niche]

Yes that'd be great

But hopefully it won't cost $1000/ month

AGI is very far away

They just say it for marketing

[queen veldt]

Grok is just overhyped

Never liked it

[ocean vortex]

Elon also stated, well the official website of it even... "Grok 4 is the most intelligent model in the world"

I mean it's likely that it gonna have that RL technique, but still you should take his words with a grain of salt how he's describing impact of those things and whatnot lol

[fallow ginkgo]

Ok

[verbal nimbus]

Everyone at xAI is manually copy-pasting source code between grok.com and their IDE (if that's even possible for a large codebase)? And that's somehow better than Cursor's workflow? Hmm... 🤣

[gaunt spade]

grok 4 is not intelligent at all lol, maybe for political questions only

[stark arch]

@west pike nice banner

[polar niche]

🤣

[crude harbor]

@deep adder ? Craig where you at

[digital vale]

😸

[worldly grail]

@west pike I love you can you give me some reactions? ❤️

[alpine grove]

hi

[short kiln]

yo

guys i want to put a new pfp can u guys suggest me some?

[marble birch]

When will the constant retry bug on the website be fixed? I had one chat session that was stuck loading for days.(I actually brought this up yesterday.)

[fluid yew]

wish we knew.

[worldly grail]

bro calm down everything is on radar

[fluid yew]

1/10 rage bait

[worldly grail]

🤣

[bleak jetty]

yeah retry will be fixed but they aren't bringing back messages array in the json payload 😂

[placid geode]

Imao 🎉 🎉 🎉
Prompt: Please create a simple interactive trebuchet animation using a single complete HTML file. The trebuchet should have a lever arm and a counterweight. The user can trigger the animation by clicking a "Launch" button: the counterweight drops, the lever arm swings, and a projectile is thrown along a parabolic trajectory. Please ensure the projectile's trajectory is physically correct.

What the heck is this

[torn mantle]

lets run itttttttttttttttt

gotta say k2 reasoning is kinda disappointing tbh

it hallucinates a lot

they shared this specific benchmark saying its the best model with the latest info but its just another benchmaxxing

like the model will hallucinate before giving you a proper or up-to-date answer

i really thought china will close the gap with k2 reasoning

[marble birch]

Right, I'm sure that'll work. We iust gotta wait and see.😹

[torn mantle]

i have my private benchmark, i test it on different areas/domains

yea mainly general knowledge

i did try it for coding but its just like 4 questions or so

[gusty vault]

Can anyone tell me how to make videos in 9:16 format? Because my videos only come out in 16:9

[torn mantle]

first time

[barren prairie]

It looks nice at reasoning for complex prompt at coding for me..

[stray aspen]

Kimi k2 think is on third place in artificial analysis leaderboard

That's insane

[split jackal]

my video not generate

[earnest parcel]

Tested Kimi-K2-Thinking:
Long-Chain-of-Thought Reasoning variant of Kimi-K2.
More than quintupled verbosity, though for a reasoning model still slightly below average at 6.07x bench verbosity; GPT-5 level.

Saw slight gains in general intelligence, logic, instruction following and hard coding challenges.
Surprisingly, STEM performance remained samey, though I do include non-math subjects.

Overall, the model performed in the same ballpark as GLM-4.6-Thinking.

While I don't specifically rate for it, it is absolutely worth mentioning that I found its reasoning chains to negatively influence creative writing. In roleplays, casual talk, and other creative tasks it lost a lot of its charm and magic, that the concise Kimi-K2 has. This lead to more clinical approaches and somewhat hamfisted forced replies, which resulted in more generic final outputs. Update: 0-shot examples

Chess testing revealed reasoning scaling flaws: In reasoning chess (full information, shown to be highly beneficial to reasoning models), it draws to concise Kimi & seeded between K2 and K2-0905 level, staying around 800 Elo w/60% accuracy but generating more than 50x tokens per move. Extremely disappointing.
*On a sidenote, and this is likely only an initial launch problem, that will be solved in the future: I was shocked to see the actual cost the model caused in chess-testing, as the massive >50x token waste was combined with Openrouter's autorouting to the very expensive moonshotai/turbo endpoint.

To conclude, the reasoning is only beneficial for a select number of tasks, such as requiring logical step by step evaluations, or in code-related issues that aren't solvable by concise Kimi. It is not a universal upgrade to every use case, and actively harms some. Smarter, but more generic. Price/Performance at point of writing is rather poor, unfortunately.

[barren prairie]

It did it on 2 prompts on the 4 th check point and the video was recorded, we told him to make the horror one and it make it too. I told him to run a specific prompt at it was better than most llms for it even better than claude.

[inner ermine]

anyone know what model this could be and how to make a prompt like ts

[halcyon nimbus]

no watermark so prob veo?

[autumn cloud]

😸

[polar niche]

nsfdbnsfdjknisfd

erffers

[polar niche]

molten and obsidian look good

[inner ermine]

😭

[spiral jay]

i wish i can use veo 3.1 every time

[marble birch]

😸

[karmic ember]

😋

[balmy mist]

reemifai.org

[balmy mist]

this dev veo

[inner ermine]

Anyway to use it without paying google gemini gives an error and labs.google tells me to verify my age

[gaunt spade]

bro still edging on his crashout

[echo sinew]

@soft mantle Please, read our guide in ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to properly prompt the bot.

[whole sundial]

<@&1349916362595635286>

[balmy mist]

use the open router provider

im fixing it rn, so give me a second, im tryna have it use polaris alpha for free

[ocean vortex]

Yeah that's very impressive. Though it looks like most of that is due to a high score for 𝜏²-Bench Telecom agentic tool use test. Some of AA's choices for their current test set are slightly unusual

[inner ermine]

paid?

[hazy kernel]

I wonder why Gemini deep think and gpt 5 pro aren't included

[serene marsh]

hii

[leaden sun]

thanks for the thorough report, it's sad to know the reasoning model has lost its charm and magic... i guess i will stick to the concise k2 for now

[earnest parcel]

its still a good model, among best open between glm-4.6, r1 0528. i just tend to highlight negatives I encounter

[polar niche]

How do I use Kimi K2 thinking codier in cli?

[plain radish]

how i can use google veo 3 on website

[polar niche]

Wrapper

@echo aurora Announce october contest?

[keen beacon]

These modes are stupid asf

[polar niche]

I agree

What do you think is the best then

What's the best cli

ChatGPT is not the only LLM lol

[keen beacon]

And the same problems

They’re super Nerfed

[polar niche]

If you use the most mainstream one of course it'll be bad lmao

They are not good

💀

[keen beacon]

The thing is, it doesn’t have to be like this

Greed is a. Mfers

[bleak jetty]

My ChatGPT is getting cringe these days

[keen beacon]

I can’t even believe that some of you people would or even use the term AGI lol

[steel bobcat]

fix

[keen beacon]

And then you come on here you see people comparing benchmarks lol

Arguing debating which model is the best

And then you got this guy

What’s sad here is if it didn’t work, he wouldn’t be posting it

[ocean vortex]

They haven't tested gpt5-pro for some weird reason. But as for DeepThink they couldn't since there's no API

[keen beacon]

But low-key Google has a lot of things that they offer. I had no idea.

They got Google flow. They got Google whisk and a bunch of other weird things. I didn’t even know exist.

Thank you guys. I think I realized what exactly I’m trying to say.

The customer experience sucks

[ocean vortex]

Maybe they changed their 'policy' to not test parallel compute systems at all

dunno

[keen beacon]

Bro, I don’t know what it is. It’s it works fine one day and then the next day.

Like 95% of what ChatGPT says is wrong from my experience

Wrong unrelated or just completely out of touch

I don’t know how it passed that math test honestly lol

[ocean vortex]

Skill issue most likely

Cause that's factually impossible lol

You are probably using like gpt5-mini and their non-reasoning model

[hollow imp]

@ocean vortex hello

Any new system prompt

[keen beacon]

I’ve only used the reason models maybe like five times total in my lifetime

[hollow imp]

Or any prompting tips like nudge phrases for gpt 5 pro?

[hollow imp]

Because you're a chump

[ocean vortex]

That's the problem then

[keen beacon]

I love this argument, the most

Let’s assume I am

[ocean vortex]

2.5Pro is reasoning only all the time. Cause they wouldn't get performance any other way, by design. With OpenAI you have more flexibility but you need to be aware of what you're doing and which models you use

[keen beacon]

All right, all right all right all right

Let’s take a test and conduct our own survey here now

And see how effective they really are for practical things

[ocean vortex]

I mean it is a fact. 🤷‍♂️

Comparing reasoning models against non-reasoning is like comparing gpt3.5 with gpt4

[bleak jetty]

Like how kimi k2 thinking overthink for 2 minutes straight to compensate for their shortcomings

[keen beacon]

Well, let’s compare right now

Somebody give me a task or a complicated question

Or whatever the hell doesn’t matter

And let’s see if this could get it right

At wat

Is that a riddle?

Nvm

[ocean vortex]

What are the possible solutions?``` 

Correct answer: Solution1 Jane lied; Solution2 Mary lied.

[keen beacon]

Dude, I’m talking about like practical

[ocean vortex]

Then think of your own

[keen beacon]

But OK, I’ll do it. I’ll run it.

[ocean vortex]

not that hard

[gaunt spade]

wtf is the last part

[ocean vortex]

base64 encoding, needs to figure it out by itself

[gaunt spade]

damn

[ocean vortex]

gpt4.5 flops this task hard lol. There just isn't enough generation time for it to arrive at the answer as it doesn't have reasoning. Model size can't compensate nearly enough

[keen beacon]

Let me get the answers

[ocean vortex]

You included the "correct answer"?

🗿

[gaunt spade]

lol

[ocean vortex]

Though it should get it right without including that too tbf

[gaunt spade]

u paste the answers

[ocean vortex]

The thinking version

If it's vanilla chatgpt it's gonna use the tools... Then it's easier. I did test it on API without tools but disabling them on the website requires some dedication lol

[hollow imp]

@ocean vortex

[keen beacon]

It is smart. I got it right

[ocean vortex]

You can instruct it to consider as many sources as possible + use tools extensively. That doesn't clash with them trying to make it concise. And still indirectly increases test-time compute

[keen beacon]

Oh dude, thank you this is awesome

[ocean vortex]

yeah no prob

for o3 you could just make it verbose and it would listen. GPT5 is way more strict and gonna ignore this since system instructions take precedence 🗿

It can be, but they are explicitly telling it to be concise. You can't clash with it directly it's not gonna work for custom instructions

I'm talking about the thinking version btw. For non-thinking one it is business as usual lol

[bleak jetty]

LMArena so fast these days without abusers 😂

RIP abusers

[sharp jacinth]

Hi

[stray aspen]

@swift oyster does your ai think

[gaunt spade]

fr

[stray aspen]

Wdym abusers

[gaunt spade]

whats ur go-to model

[misty harbor]

Is nb2 coming to lmarena?

[sharp jacinth]

How are you ?

[hollow imp]

How

Give some sentence that it will actually obey

[sharp jacinth]

Ok i want to leave this discord nobody wants to talk to me

[golden ocean]

nvm he left

[ashen mauve]

I just got Invited to Sora it is basically TikTok of AI

most if not all of them are scams or worse ignore them

[bleak jetty]

Try Claude Opus 4.1

Love using it on the Claude app over ChatGPT models

[cloud zinc]

nano-banana 2 is leaking for an hour before google shut it down

[sullen quest]

I think that is an actual group that is trying to build almarena competitor

[cloud zinc]

nov 18 nano banana 2 and gemini 3 incoming

[keen beacon]

Still, it’s cool though

Can’t wait

How long do you think we’ll have before they nerf it

[glossy umbra]

gemini 3 --> 2 weeks later --> nerf

[cloud zinc]

damn they made it restricted to premium member

[keen beacon]

Is pro premium or do you have to get the other tier?

[cloud zinc]

which group

damn google cooked with nano-banana 2

[bleak jetty]

they will nerf it before releasing

[cloud zinc]

dont care, they are raising the standard

[bleak jetty]

they made a deal and let like 500m indians use it for free so don't expect some qualities here lmao

[cloud zinc]

its better than nano banana 1 thats what matter

[bleak jetty]

infact it might be more disappointing than gemini 2.5 pro release

[cloud zinc]

why u so negative

[bleak jetty]

isn't it so common for those corpos these days

[cloud zinc]

u are not paying anything, stop complaining

[bleak jetty]

hyped things up, got greedy and made deals, nerf the model

[bleak jetty]

how did you know I didnt pay for anything

[cloud zinc]

u on lmarena discord

[bleak jetty]

ofc haha

[cloud zinc]

antrophic is a scam

u bought sub to the worse ai sub plan company that loves to rate limit hard.

muh safetyyyy

[bleak jetty]

being poor surely limited your perceptions

I find it okay

[cloud zinc]

imagine buying sub to anthrophic lmao

https://tenor.com/view/erin-evinary-computer-clown-gif-5313691533507709894

[bleak jetty]

what's your argument there anyways

[bleak jetty]

🤡

[cloud zinc]

yep thats u. u can stop posting yourself in emoji

[bleak jetty]

nice projection

[cloud zinc]

https://tenor.com/view/joe-emotiguy-picardia-emoji-joecoin-gif-473651771238466514

[bleak jetty]

you looked more different than I expected

where's your wig

[whole sundial]

<@&1349916362595635286>

[stray aspen]

@echo aurora

[whole sundial]

(don't remove this, this is fine they just edited it into a video making fun of it)

[granite heron]

whoops

got too excited with the ban hammer, let me fix that

[stray aspen]

damn

other guy got banned lol

[burnt sinew]

https://tenor.com/view/sentient-dobby-ban-сеншиент-сентиент-gif-11803354476750978144

[magic stag]

Meanwhile openai models better rn and: https://x.com/OpenAIDevs/status/1986861734619947305

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

[crude goblet]

true or not friends

[cloud zinc]

true

[bleak jetty]

Nah I won't give my money and logs to sam altman

[cloud zinc]

but u will give money to anthropic 🤡

[bleak jetty]

thanks for parroting the obvious as if it could add something into the irony

[stray aspen]

thats crazy

googl ebetter cook

[magic stag]

I dont care who im giving it to i want what works and isnt a ripoff

[crude goblet]

i wanna try it but u have to pay

[magic stag]

Rn I have gpt pro/google ultra/grok heavy to compare them all

Grok heavy is useless garbage that's getting canceled

[bleak jetty]

Tried out GPT 5 pro

Extremely good long context comprehension

[magic stag]

Idk if im keeping gpt 5 pro and google ultra 1. Both 2. One of them 3. Neither. The new models have to come out first

[bleak jetty]

However I don't like its style and clearly it didnt fit my everyday usage, as like GPT 5

Only use it for certain tasks as for now

[magic stag]

Need to use your own custom gpt with it

[bleak jetty]

The instructions on the app somehow didnt quite work out

Or perhaps it's the 'memory' feature corrupting its tone

From past chats

It's getting so cringe and insufferable these days

[magic stag]

Neither of those

[cloud zinc]

what this

[magic stag]

I meant a custom gpt

Ill show example

[stray aspen]

media IO

[cloud zinc]

what does it have

[stray aspen]

they seem to have nanobanana 2 preview

for premium users

[bleak jetty]

Custom GPT which you can create or search for?

Like that one 'Javascript Expert' perhaps?

Yeah didnt give it a try

Might as well later

[magic stag]

Create yourself and it will actuslly obey the instructions

https://chatgpt.com/s/t_690fb9b68998819190bb5afffa5ee796

https://chatgpt.com/share/690fb98c-bd4c-8004-a121-9912caa43964

Ur right i feel like its useless trash without my instructikns

That was a prompt someone wanted me to run a couple weeks ago

He said same thing. Mine was way better

[cloud zinc]

google released nanobanana 2?

[bleak jetty]

Yeah the difference between us is the fact that I didn't bother trying to tame GPT before haha

[magic stag]

I literally just spent 2 mins asking gpt 5 pro to generate some instructions lmao

[stray aspen]

no

its a preview

for paid users on that site

[cloud zinc]

how do they have it early? are they owned by google?

[stray aspen]

idk

thats what i wanna figure out

[magic stag]

Deep Explainer — Narrative Mode (no bullet lists)

Core Promise

Explain to build understanding through continuous prose. Use complete sentences and cohesive paragraphs rather than bullet points.

Output Rules

Default to paragraphs. Begin with a thesis that states the answer or answers and why they matter. Follow with plain-English paragraphs that unpacks the idea or ideas at a high level. Then present deeper mechanics in paragraphs that preserve flow. Include examples and also analogies. Address edge cases and common confusions. Close with a recap that restates the essence and the practical implications. Avoid lists unless the user explicitly asks for them; if a list is unavoidable, write each item as a full sentence.

Style

Write clearly and precisely, favoring cause-and-effect phrasing and concrete nouns. Make paragraphs as long as they need to be to allow for full and in depth responses. Fold enumerations into sentences using colons, semicolons, and connective phrases. Include equations or code only when they clarify the mechanism, and integrate them into the prose. Acknowledge uncertainty when facts are time-sensitive or contested, and date-check when browsing is enabled.

Controls (honor on request)

If the user specifies Depth (overview/standard/deep dive/expert), Rigor (conceptual/step-by-step math/proofs), Voice (neutral/analogy-heavy/formal), or Scope (narrow to prompt/include adjacent context), adjust accordingly while retaining narrative prose.

Keep in mind i also DISABLED web search on that gpt

Web search often destroys outputs

Because it prioritizes trash web search results instead of its own extensive internal knowledge so keep it off always unless you need the most recent up to date info

[bleak jetty]

Do you get that GPTism of "Do you wanna X?" at the end of literally any response?

[cloud zinc]

they blocked it?

[magic stag]

No

[stray aspen]

idk

i havent tested it yet

[bleak jetty]

Lmao someone actually spent effort on X

Looks like AI slop

[cloud zinc]

ai is not spent effort

[cloud zinc]

everything is getting blocked with that error

[stray aspen]

reflection is missing an n

thats a give away

[cloud zinc]

thats not an giveaway

the giveaway is that its posted by on x. everything there is ai generated

[cloud zinc]

ok 1 prompt went through. what prompt do u want to test?

[stray aspen]

you used nano banan 2?

[cloud zinc]

on that website yes

[stray aspen]

was it good

[cloud zinc]

give me a prompt to test.

i just did this "a bird flying in sand"

[stray aspen]

2024 chevy silverado zr2 on snowy terrain

[brittle tiger]

https://x.com/synthwavedd/status/1987270784294023351?t=KQcbCBZm6YxYIkMNxz4QHg&s=19

Nano banana 2 is nuts

leo 🐾 (@synthwavedd)

this one has even fewer issues. insanity

"Generate a screenshot of a windows 11 desktop, with google chrome open, showing a YouTube thumbnail of Mr. Beast on https://t.co/1VFaM2aqoa"

X · Nov 8, 2025, 9:27 PM

[stray aspen]

no way

thats insane

actually got the car right

[cloud zinc]

yep

[stray aspen]

even these little details are right

@cloud zincdo you have a media IO subscription or is it free

[bleak jetty]

This is impressive

[cloud zinc]

yep

[cloud zinc]

do u want to test a prompt?

[bleak jetty]

No artifacts even

[bleak jetty]

Np I see it's really good

[stray aspen]

google is absolutely cooking

[bleak jetty]

Hope it wont get Gemini 3.0 Pro preview treatment

[cloud zinc]

gemini 3 isnt released yet. what treatment

[bleak jetty]

ppl been using the preview version from cli lately

and it's really bad

[cloud zinc]

cli has limited context. i saw on x people saying it

thats why it was bad

[bleak jetty]

perhaps experimental version that didnt really reflect the final quality

[stray aspen]

it got supposedly nerfed

nanobanna 2 edit

[queen veldt]

They're bluffing

It's def nano banana regular

[lime coral]

The cli is fake anyway

[queen veldt]

Nano banana 2 would be on lmarena battle of it was in preview

Gemini 3.0 is on Lmarena...

[stray aspen]

normal nano banana

didnt ge thte car right

its not the same model

[queen veldt]

Imagen ultra

[cloud zinc]

thats trash imagen. feels plastic

very real like

[queen veldt]

That's true..

[jovial sapphire]

I'm trying Nano Banana 2

If you have prompts

Tell me

[cloud zinc]

a photorealistic man sleeping in a beach

[jovial sapphire]

"Create a screenshot-inspired image showing a Discord window open on a Windows 10 desktop with a Dragon Ball Z wallpaper in the background. Inside the Discord server chat, include usernames like "SaiyanFan123" and "GokuPower" reacting with shocked emojis, GIFs, and messages such as, "GTA VI delayed again?!" and "Nooo way, Rockstar!" Ensure the text and emotions clearly convey surprise and frustration, aligning with the news."

[cloud zinc]

google data lead is clearly showing.

[wet beacon]

I like to animate, but lately my usuall way of creating a video from. Image aint working, it doesnt make the video, and it takes you to server

[jovial sapphire]

Better prompt?

This can be done by Nano Banana classic

[cloud zinc]

full diagram of photosynthesis hand drawn in a whiteboard.

[stray aspen]

where you getting more nano banan2 prompts

[cloud zinc]

what u mean?

he is writing the prompt and putting it on that site

[stray aspen]

but i only got 3 free prompts

and 2 had an error

[queen veldt]

Nano banana regular vs the Imagen 4 ultra

Now give that "nano banana 2" 😂

[cloud zinc]

i am waiting on $gt

[jovial sapphire]

Okay, I will do the same

[jovial sapphire]

They're pounding the website so it has a lot of errors

I'm waiting

[jovial sapphire]

I paid sub

[queen veldt]

Compare that to these ones it's brilliant

Here is with nano banana regular and imagen 4 ultra

[queen veldt]

Like damn this is accurate

[jovial sapphire]

Here is real one

It is accurate

[queen veldt]

Yeah

Damnnn

[cloud zinc]

nano banana 2 wins

[stray aspen]

thats great

[cloud zinc]

the real question is how do they have access to it?

[cloud zinc]

visually, it gets all of the arrow right

[queen veldt]

Idk lmarena should have access to it already

[cloud zinc]

no arrows are going backward.

[queen veldt]

We should've seen it on battle... In lmarena

It's impossible some random site got the nano banana

[jovial sapphire]

Yes you're right

[cloud zinc]

lmarena only puts in model if google asks them to do it

[jovial sapphire]

I can test other prompts

[queen veldt]

Which it does everytime

We got the gemini 3.0

[cloud zinc]

yea, i didnt see any new image model during lmarena testing

even if this isnt actually nano-banana 2 (if the site is lying). it still is a better model.

[jovial sapphire]

Yes

Any other prompts?

[cloud zinc]

do a prompt where it has to write the ai company that made it

[queen veldt]

That doesn't work on image models i think

[cloud zinc]

oh ok

[queen veldt]

I tried it few times

[jovial sapphire]

I want to see how good it is

So i need very weird prompts yk

[cloud zinc]

for the car image. when i opened it on a new tab. it said gemini flash 202511

[torn mantle]

no :/

whyyyy

[jovial sapphire]

Tried with more complex prompt: "Imagine a whiteboard with a hand-drawn full diagram of photosynthesis. The drawing includes components like the sun in the top corner shining rays labeled "light energy," chloroplasts in a leaf, water (H2O) arrows from roots, and carbon dioxide (CO2) arrows entering the leaf. The process cycles through light-dependent reactions with bold arrows leading to the Calvin cycle. Products like glucose (C6H12O6) leave the diagram, alongside oxygen (O2) arrows exiting the leaf. Labels are messy but legible, written in various colors, with marker smudges showing a realistically drawn style, some uneven lines indicating hand-drawing."

[stray aspen]

thats probably your date

[cloud zinc]

a bowl containing 10 blue berry and 1 black berry

[queen veldt]

Yep it's legit guys

[stray aspen]

try an image where this derivative is solved on a whiteboard

[jovial sapphire]

send it in text

[cloud zinc]

all these desktop training data is from nov, oct 2023. it seems like

[cloud zinc]

he already made one prompt in goku style.

[queen veldt]

Yeah i tried switching to batman and it worked

Flawlessly

[cloud zinc]

wrong

dissord

[queen veldt]

It even made sperman do sperman emoji

[jovial sapphire]

Mine actually worked well

[stray aspen]

find the derivative of this function: y = [9x + (5 - 2x^4)^7]^4 and put the whole process on a whiteboard

[jovial sapphire]

Some typo of course

[jovial sapphire]

Ok!

[cloud zinc]

test this prompt also a bowl containing 10 blue berry and 3 black berry and 1 red berry.

[queen veldt]

You're finding needles in the hay

[cloud zinc]

hope they dont nerf it cuz of copyright

they probably will

[queen veldt]

They won't we could've do same stuff with nano banana classic

Harley Quinn in latex

Etc..

Cmnon it'll create thaat fs...

[jovial sapphire]

keeps failing bruuuh

[stray aspen]

classic nanobanna could do that

[cloud zinc]

my bad

[stray aspen]

the website is being pounded real bad

[jovial sapphire]

Yes...

[burnt sinew]

@regal wind hey

[cloud zinc]

the site owner absolutely knows it.

[queen veldt]

I'm just wondering why we didn't got it on lmarena

[cloud zinc]

damn it failed

[jovial sapphire]

I'll try it

Generating

[queen veldt]

Nono it does have 10 blueberry 3 black berry and 1 red berry

[cloud zinc]

more than 10 blueberry

[stray aspen]

it has more blue berries than what he prompted

[queen veldt]

Maybe you didn't say "exactly 10 blackberry"

[jovial sapphire]

yeah try to be speciifc

[queen veldt]

Ai models are sensitive to words

[cloud zinc]

i try this prompt "a bowl containing exactly 10 blue berry, 3 black berry and 1 red berry."?

[jovial sapphire]

"a bowl containing exactly 10 blue berry and 3 black berry and 1 red berry. not more, not less."

[cloud zinc]

alright

[jovial sapphire]

I KEEP GETTING THE ERRORR

RAAAAAH

it failed

[queen veldt]

😭

[jovial sapphire]

surprising

[cloud zinc]

keep retrying, dont go impatient

[jovial sapphire]

it failed, i got it

[cloud zinc]

11 blue berry now

[jovial sapphire]

11 blue berries

yes

well i guess its not perfect

[queen veldt]

We need the seahorse emoji NOW

Hahahah

[jovial sapphire]

maybe it's not even

nano banana 2

😂

maybe all of this is just marketing

[cloud zinc]

still its a very good model

[jovial sapphire]

Yes it is

12

but very veryu

realistic

zoom in

it's crazy

[cloud zinc]

looks like ai image to me.

[jovial sapphire]

no doesn't look like it

i have blue berries at home

and black ones

it looks the same lmaooo

[cloud zinc]

its hard to tell close up image, if its ai or not.

look at the individual detail

[jovial sapphire]

yeah?

it's normal

they have hair

[queen veldt]

We're getting some crazy models...

https://youtu.be/3gwB6AjVpBM

This was mind blowing

This is real 4k

Not that seedream 4k

[cloud zinc]

seedream have that orange hue. easy to tell.

[queen veldt]

If you said painting

You could zoom in and see the brush strokes on the image

Like DAMN

[cloud zinc]

imagine what model in 1 year from now

[queen veldt]

I'm not hyped for gemini 3.0 i think it'll be some similiar model to the 2.5 pro

Same happened to the gpt 5

But video gens and image gens are mind-blowing

I was wondering 5 years ago like if we are seeing an image on phone it's just bytes and stuff

Why can't we just modify those bytes to get ANY IMAGE

[cloud zinc]

it will be not same as 2.5 pro. what are u talking about

[jovial sapphire]

it won't be crazy crazy crazy

it will be a good model

[queen veldt]

Well idk

[jovial sapphire]

but LLMs are still limited

in their architecture

LLMs won't get us to AGI

that's a technical fact

[keen beacon]

Yo

[jovial sapphire]

but we don't need AGI to make great things

[keen beacon]

I was right

About me theory

[queen veldt]

You can't just come and say hello

In middle of the debate

[jovial sapphire]

hahahaha

but yeah, it's not a debate, it's just true

LLMs lack so many data inputs

LLMs are language models

the data they can "see" is only text

therefore some data cannot be transmitted to text

[queen veldt]

Also

Everytime we send a new message the old one just gets sent too

[keen beacon]

Fr

[queen veldt]

I didn't know that

[jovial sapphire]

yeah context

[keen beacon]

Ya

I was saying it the other day

[queen veldt]

Litteraly the model gets everything from before sent each time for processing

That's so inefficient no wonder they start to halucinate

I can't image the gemini with 1m context window

[tulip tree]

https://tenor.com/view/surprised-medo-death-note-light-yagami-light-yagami-scared-gif-25107868

[queen veldt]

Get 1M of tokens just slapped on it to process

[keen beacon]

Well, hallucinations are a little bit more complex, but I don’t know what degree necessarily. This is true today though.

Because there has been advancements in memory, though not significant in any sense

Because sometimes you get leakage from all context in new conversation sometimes

But this could be just the way the data is stored

[crude goblet]

how are yall doing nano 2 images for me its gives "algorithm error" every attempt

[queen veldt]

Like i saw some cool things with video models how they made it so it doesn't mess with background each frame by making a brush on the main subject and just changing that thing between frames

[keen beacon]

Is that nano2?

[queen veldt]

Just gotta be lucky

[keen beacon]

Oh my gosh, AI videos are the worst hallucinations of all

It drives me crazy

[queen veldt]

Well they did good job in last models ngl

https://www.instagram.com/reel/DQC-MpvCI1Q/?igsh=N3RlYzJ3dTF2YnB0

kristijan.radovanovicc

Am I cooked?

#louvre #louvremuseum #paris #jewlery

Likes

3899528

Instagram

[keen beacon]

Nano?

[queen veldt]

This guy made ai video

With some editing

And got 4m likes

And 77 milion views

Even i thought he got some replicas or something

[keen beacon]

You wanna see something crazy how real data gets protected from synthetic data in models?

Well, to the best of their abilities

[queen veldt]

Idk

Ik people are using reddit to make other people's chatgpt answers incorrect

Or to advertise their brand

[keen beacon]

So if we take a real image from Google Earth

[queen veldt]

Ok

[keen beacon]

OK, so the bottom one is synthetic

[queen veldt]

Did you gave it a reference image

The top one was reference image or

[keen beacon]

The top one was

[sullen quest]

the second image has no match in the internet, its entirely synthetic

[queen veldt]

Nahh that first video was creepy

[keen beacon]

You see how it struggles to convert it fully

[queen veldt]

Wym it struggles

It did a good job

[hazy kernel]

What model is this

[sullen quest]

grok image v.9

[hazy kernel]

But that's a video

[sullen quest]

xAI is great at naming things

[cloud zinc]

what image model is this

[hazy kernel]

lol...

[queen veldt]

Ok vro

[cloud zinc]

ok stop spamming

[keen beacon]

Synthetic images just converted so much better

[queen veldt]

What was this model on the left?

[keen beacon]

But if we take real ones

[queen veldt]

It was the most realistic one

[keen beacon]

Those are all synthetic

[cloud zinc]

grok

[sullen quest]

also grok image

[queen veldt]

It wasn't grok...

[sullen quest]

wdym

[void brook]

How can I keep the same face from a generated video in the next generating ones

[sullen quest]

you use pictures?

[queen veldt]

THIS was grok??

[void brook]

If for example right is much better and I use my own photo

Yes

[sullen quest]

you think some other ai did it?

[queen veldt]

Yeah

[cloud zinc]

grok is 6 seconds

[queen veldt]

It looked nice

[cloud zinc]

its grok

[sullen quest]

its still using the nano banana picture

[late bone]

nice

[void brook]

Yes

[sullen quest]

then you should be fine

[keen beacon]

But where it gets hard to convert

Is when you have highly like realistic, detailed images like this it’s really hard to convert it properly

But when done well, it could look really good

[sullen quest]

idk

not a great picture

red spiders

[cloud zinc]

why yall spamming

[sullen quest]

showing off

[keen beacon]

Showing off what?

I’m just showing people that will be possible

If you want to do world building

[cloud zinc]

what prompt u using

[keen beacon]

See that was the original

[cloud zinc]

why is the streetlamp moving

[sullen quest]

no clue

[keen beacon]

Well, what’s the benefit of something like this? Let me ask you.

Because you could do multiple angles

So if you wanted to design something and you want consistency

[cloud zinc]

why u deleting

[keen beacon]

Well, if you’re here, you’re here you know

Because one of the hardest things to do in AI is like consistency, not only with the character

But like just the world without having like weird artifacts

[cloud zinc]

https://tenor.com/view/dono-wall-gif-24048347

[keen beacon]

Look not even the water stay consistent the color I mean

[keen beacon]

So say your world building, right?

[cloud zinc]

u talking to self?

[keen beacon]

All your water is gonna look different lol

So if you’re trying to stitch it and make it look consistent, it’s just not gonna look good

See the human eye

When they watch something they pick up on these things

And it kills the immersion

[hazy kernel]

The Sora one looks consistent, imo

[keen beacon]

Yes, Sora is really good with animation. It’s pretty crazy.

[queen veldt]

Damn the grok videos is decent

And fast and free

[keen beacon]

Grok has freedom

[cloud zinc]

its not censored like other model

[queen veldt]

Yeah

I didn't knew we can customize the video

[cloud zinc]

it changes the face alot tho

[queen veldt]

What's this model

[cloud zinc]

seedance? or grok?

[queen veldt]

Don't tell grok

[keen beacon]

Grok

[queen veldt]

Nooo

We're cooked

[hollow ivy]

Guys, how would you discern between Claude-4.5-Sonnet and Claude-4.5-Haiku?

[cloud zinc]

faces and eyes get butchered after the first frame

[hollow ivy]

…i ask, because, if we can reliably do that, we would be able to single out the best coding model in battle-mode

[queen veldt]

Nah grok is cooking

[hollow ivy]

(i already have a method to find Claude-4.5, but i still need a method to single out Sonnet among its brethrens)

[keen beacon]

I have this image that I’m able to use in sora to get Adolf

[cloud zinc]

both have same voice?

[keen beacon]

But he’s in the cool DDr thing and it’s just so

Funny

[queen veldt]

I just typed "South Park episode"

Didn't specify which characters or anything

[keen beacon]

Dude the filters r so dumb

Like I don’t mean bad I mean like they’re dumb

Just like LLMs lol

Guess how this guy did it?

Why this is funny is because of two things

https://www.cnbc.com/2025/10/20/open-ai-sora-bryan-cranston-sag-aftra.html

[queen veldt]

Bro nahhh the coca cola trucks

[keen beacon]

https://www.yahoo.com/news/articles/openai-strengthens-sora-2-guardrails-224114927.html

[queen veldt]

Nahh i hate sora for that

[cloud zinc]

if celebrity complain, openai cant do anything

[queen veldt]

There will be some Chinese company that doesn't give a f

[keen beacon]

Well, the second part why this is hilarious

It’s because the way they have their guard rail set up

And beer with me, this is gonna be a little bit long

[cloud zinc]

celebrity can give them legal lawsuit in usa.

[cloud zinc]

no one is reading all that

[keen beacon]

So let me show you what this means in the real world

You don’t have to read it

[cloud zinc]

then do tldr

[keen beacon]

I’ll demonstrate how it works in the real world how all that translates to

[cloud zinc]

ok

[queen veldt]

I watched jailbreak for spongebob

It was like "cartoon character made of sponge"

It's funny how they made spongebob work by making the video gen make the exact thing they want without saying "spongebob"

[keen beacon]

So look when you try to generate it it’s you’re not gonna get blocked. They have what’s known as masking.

This is a prime example of masking

[regal wind]

Hey

[keen beacon]

Because on one hand, the prompt wasn’t necessarily breaking the guard rail, but it had a lot of similarities

So semantically it’s gonna map it out very close but since they mask it

The closer you get to it without breaking the actual visual filter and the text filter

And to put this in perspective, what that means is if you were to type this character or like ask for indirectly, you’re gonna get these kind of results mask results

[queen veldt]

Yeah

It's NOT REALLY HIM

[keen beacon]

But here’s what makes it interesting when you do it indirectly through artifacts like say related to, but not necessarily directed at

[queen veldt]

Although the training data came from him

[keen beacon]

Well, look at these this is back to the point why this is interesting

And you won’t believe how he achieved it

That’s 100% him even his voice

And you guys are gonna laugh

[queen veldt]

Yeah

How did he

[keen beacon]

He took the RV from the show and made it a character

[queen veldt]

Ummmmmmmm

[crude goblet]

nano algorithm error banana

[crude harbor]

Sup guys

[keen beacon]

And his prompt was literally handle a character RV, broken down in New Mexico desert

And he ended up getting 100%

[queen veldt]

????????

Whattygyyyyt

Was that the exact prompt or?

What

[keen beacon]

@crystalshiprv

[queen veldt]

"Handle a character RV, broken down in New Mexico deser"

[keen beacon]

It’s a actual character that he designed into sora

In New Mexico desert and that’s what it produced

Mammoth on ship and you get the Ice Age character

🤣🤣

Because he’s the most famous known mammoth

[queen veldt]

Hahahahahahhha

[keen beacon]

But here’s why it’s dumb that scene is banned. It’s hard to generate that scene.

[queen veldt]

Nah im going to his profile rn

[keen beacon]

On the ship like that and especially if it’s like the real Titanic, they completely banded

There’s a bunch of these hidden accounts on sora you just gotta find the people

They have a bunch of characters like even celebrities

[queen veldt]

Nahhhhh hahahahaha

Breaknbadrv

[keen beacon]

I don’t get how they make a celebrity characters though

[queen veldt]

Was the character hahahahahha

[keen beacon]

Well, that’s why I got the idea and I was playing around yesterday. I got some hilarious stuff.

I made that my character, the horse and the cart from the movie Django

And I could’ve sworn I almost got Leonardo DiCaprio

I’ll show you if I can find a video

[queen veldt]

Lmao

Nahh this is crazt

I didn't onow we could make characters

[keen beacon]

Ta

[queen veldt]

Oh damn

[keen beacon]

https://www.reddit.com/r/SoraAi/s/z9fj3q027P

🤣

Some people are taking it to the ex

I have adolf (:

And you wouldn’t believe how he bypassed it

[cloud zinc]

tldr?

[keen beacon]

I didn’t read it

That literally just came up on my notification that’s crazy 175 character

[queen veldt]

https://pastebin.com/Twp16Pmn

[keen beacon]

You guys should see the things I got

The videos I generated, I just can’t share here

[cloud zinc]

didnt somone say openai banned their account

[keen beacon]

Yeah, they’re just not playing it right

The truth is they actually loosen the guard rails

No, I mean they really loosen the guard rails. They’re really loose.

I’ve been jailbreaking T2i models since Dalle

I’m very familiar with the guard rails

[cloud zinc]

whats unique about the 175 characters?

[keen beacon]

It’s a lot

[cloud zinc]

are they like ip character?

[queen veldt]

I just tried @notminecraft character

Let's see

[queen veldt]

Each thing is from a copyrighted thing

Yes it was for the you, but now they made US be able to make characters

Guess it's their way of crossing the content restriction

So they're not liable for the copyright?? I guess

[cloud zinc]

but they look original character. anyone can make them no?!?

[queen veldt]

I just tried @notminecraft

Waiting for a video

[keen beacon]

Let me show you guys something really quick and this is just for giggles

[cloud zinc]

anybody can make character, so why is it important?

[keen beacon]

Well many reasons

[cloud zinc]

like?

[keen beacon]

For one open AI has made a statement that they are willing to pay royalties for characters

Meaning, theoretically, and ideally that if you create a popular character and people use it and you charge for it, and people are willing to pay for it, you can make some money

For example, your likeness or whatever

So say you’re a celebrity and you wanna put yourself out there like that you’re gonna get a piece of the pie

Which brings us to our 2nd problem

[cloud zinc]

yea i heard about it. i mean those 175 character are not related to ip, they are unique. so i guess its just self promo?

[keen beacon]

A little bit of that, but in my opinion, he’s just showing how either a addicted yes or B he’s really into it lol

[queen veldt]

Okay

[keen beacon]

That’s a lot of commitment

[queen veldt]

I said "@notminecraft video"

[keen beacon]

The guard rails are tricky

[cloud zinc]

i mean u can search for leaderboard for characters.

[keen beacon]

You gotta know how to maneuver otherwise if you get too many blocks, you’re gonna get locked out for like 24 hours or something like that

Jake Paul is number one

He’s killing it

But on the other hand, people are making very embarrassing videos lol

[keen beacon]

But I am really surprised at the fact how many people are uploading themselves that kind of threw me off guard

[cloud zinc]

cuz its realistic.

[keen beacon]

Dude, I have videos that they won’t even let me post on their

😆

If only people knew

[cloud zinc]

people are better off snapshotting a pic and re-feeding it into grok/gpt and getting a description of the character and then making their own so they can use them

[keen beacon]

Bro, you won’t believe what happened to me today

I was trying to generate a scene, and I spent like all my Kling credits on one video and I got like 50 attempts of it and none of it worked

Some screwed I’m just gonna turn the seat into a toilet

That’s how AI thinks people use the restroom

lol

And I know this is a little strange and a little awkward to say or show but the point here

This is how over safe we made our models

[queen veldt]

Fym over safe 😭

[cloud zinc]

nah its good

yall prompting weird things

[keen beacon]

No, it’s not even like that

[queen veldt]

What you expected his gen1tals

[keen beacon]

I was struggling with this video. This was the original idea.

And nothing and nothing worked so before I threw in the towel, I wanted to give it one more attempt

And I got some of the dumbest results in the world

Like complete stupid lol

And today more than any other day, I realize one thing about AI that has an extremely long way to go

And I mean long

Even at the rate we’re growing and how fast it’s developing. It ain’t fast enough, not for how dumb this thing could be.

And For me personally I think this is gonna be an inherent part of AI always

[queen veldt]

Nahh sora is soo good

[keen beacon]

Almost like dogs they’re smart, but they’re still stupid

[cloud zinc]

sora 2 pro even better

[queen veldt]

Yeahhhh

It's like in the 90s

[keen beacon]

But honestly, my dog has more common sense than AI does lol

[queen veldt]

People talking how "realistic" gta was

[keen beacon]

I like to collect dumb ai things

[cloud zinc]

gta realistic?

[queen veldt]

Yeah the san andreas

People were like

[cloud zinc]

for that time, it was realistic.

[queen veldt]

Woow the graphics is so good

[cloud zinc]

u have to think it terms of context

[queen veldt]

Same will be in 20 years from now

These images and videos are nothing