general

[hollow ocean]

What about for product research?

[sweet tinsel]

I don't do that, maybe I will test it out later.

But for like reports and such o3 was the most promising always.

[torn mantle]

its not actually

if you really use it a lot and read through the paper line by line you wont find it that impressive

it doesnt really compile findings, its just parsing different infos from different pages

[ocean vortex]

OpenAI = Google > Anthropic = xAI. Though xAI are doing much more manipulation than innovation lately. In the eyes of many they are gonna get less credit than they even deserve tbh

On a good day and things going their way xAI could challenge the top spots, but not with this mess and publicity... People are not even gonna give it a chance

So they need to offset and beat everyone by A LOT, that is never happening I think

Deepseek the ones to look out for. The rest are significantly less promising. So they go with "but look our model is faster/smaller" instead lol

[torn mantle]

https://x.com/YiZhangZZZ/status/1942672317005586525

Yi Zhang (@YiZhangZZZ)

stay tuned for the smartest AI humanity's ever seen🚀

X · Jul 8, 2025, 7:49 PM

we will see

[dawn wharf]

for one week

[balmy mist]

There is no way grok 4 is still not out😂

[patent aspen]

This Discord's modal estimate for Grok 4 release date was 2 days from now

+/- 2 days

[echo aurora]

made an event for those interested! https://discord.gg/lmarena?event=1392247045296885891

[whole wagon]

xAI 42%

[small haven]

oai looks very very undervalued now

[indigo hazel]

can you explain what im seeing?

[wintry tinsel]

What did I miss gork 4 ASD (artificial stupid dumbass) releasing tommorow?

It’s gonna be so stupid it overfits the benchmarks and loops back around to being “SOTA” at benchmaxing

[small haven]

i think i just got a pr; craig said we would never have long running tasks as consumerists

[small haven]

o3 pro off by 1 day

[whole wagon]

I don't see major businesses ever using grok ngl

Way too much bad publicity

[small haven]

ppl are just mad sleeping on anthropic too

just wave riders

[keen beacon]

Just flipping a coin lol

[whole wagon]

They are actually gonna cross kek

[whole wagon]

Where is this from

[tidal schooner]

supposedly gemini discord

don't know the original source at all tho

[torn mantle]

tomorrow

[small haven]

devmode

[tidal schooner]

yeah nvm lol

coz of the 001

[wintry tinsel]

Im skeptical, I don’t think Xai will release a weak model, I just think Elon cares too much about publicity

[whole wagon]

AGI confirmed?

Doubled down also

[zinc ore]

Hate how they have it speaking

Saw some other examples where it speaks like Ben Shapiro

[whole wagon]

Such a shame man

[civic flame]

holy moly

[tidal schooner]

@gork

how will this affect the economic and geopolitical state of the world

https://www.merriam-webster.com/medical/gork

[whole wagon]

It's been at it all day targeting random Jewish ppl on X with hate

Pretty wild they have to keep deleting its tweets

[wintry tinsel]

The flooding happened an hour away from my house

Kind of scares me

I was playing video games blissfully unaware

As they were drowning to death

[tidal schooner]

https://www.instagramez.com/reel/DL2TwZoMOYZ/

WIRED (@wired)

💬 184 🔁 0 💜 7.7K 👀 0

On Wednesday, July 2, a local health department in Memphis granted Elon Musk’s xAI data center an air permit to continue operating gas turbines powering the company’s Grok chatbot.

The Memphis Chamber of Commerce announced in June 2024 that xAI had chosen a local site to build its new supercomputer, Colossus. xAI’s website boasts that it was able to build Colossus in 122 days, partly due to the mobile gas turbines that were installed at the campus, the site of a former manufacturing facility.

Colossus allowed xAI to catch up to rivals OpenAI, Google, and Anthropic in building cutting-edge artificial intelligence. It was built using 200,000 Nvidia H100 GPUs, making it likely the world’s largest AI supercomputer.

xAI’s Memphis campus is located in a predominantly Black community which has been historically burdened with industrial projects that cause pollution. Gas turbines can be a significant source of harmful emissions, like…

▶ Play video

EmbedEZ • 1 - 1 of 1 · Jul 8, 2025, 2:00 PM

gork power 🚀🚀🚀🚀

[unborn ocean]

og early gemini 2 started about 8 months ago, could actually be something

[tidal schooner]

nah it's gemini 2

[unborn ocean]

idk, even if it is

[tidal schooner]

001 suffix = stable model id

[unborn ocean]

my point is more that we should be expecting something in the near future

[whole wagon]

This has to be AGI

Right guys

[unborn ocean]

they never planned 2.5 naming (according to themselves), no way they would have us waiting much longer

[unborn ocean]

grok 4 == tay from 2016

[solar hollow]

so elon just fed altright nazi stuff to groq 😄

couldnt be any dumber than that

not being aware of what kind of consequences that would have

[hollow ocean]

https://tenor.com/view/elon-musk-this-is-elon-musk-musk-tesla-egifmeme-gif-13716021226937735268

[zinc ore]

Yeah it's Gemini 2, whole thing started because it was placed in announcement on another discord and all pinged

[wintry tinsel]

It isn’t alright or Nazi it’s common sense

[tall summit]

holy hell what

[echo aurora]

looks like they turned off text responses for now

[civic flame]

https://x.com/grok/status/1942720721026699451 LOL

Grok (@grok)

We are aware of recent posts made by Grok and are actively working to remove the inappropriate posts. Since being made aware of the content, xAI has taken action to ban hate speech before Grok posts on X. xAI is training only truth-seeking and thanks to the millions of users on

X · Jul 8, 2025, 11:01 PM

[elder rapids]

yeah seems like a fine-tuned version of grok

[whole wagon]

xAI polymarket odds crashed after this X stuff 😂

[elder rapids]

fr?

that would be stupid

[elder rapids]

you just said it crashed 🫩

[whole wagon]

It's not gonna be that great on LLM arena if this is the takes it comes up with

Doesn't seem like something the average human would like

[elder rapids]

that's not the actual model

lmao

[whole wagon]

It points to the overall culture there

Which will influence grok 4 also

[elder rapids]

yo

it's not the actual model

and they're definitely seperate teams

[torn mantle]

lmao

[torn mantle]

its still on 40%

[storm needle]

https://github.com/xai-org/grok-prompts/commit/c5de4a14feb50b0e5b3e8554f9c8aae8c97b56b4

[tidal schooner]

https://fixupx.com/grok/status/1942746405812199475

Grok (@grok)

@JoeYThor299146 @ns123abc

**💬 24 🔁 33 ❤️ 311 👁️ 63.2K **

FixupX · Jul 9, 2025, 12:43 AM

lmao

[rare python]

peak xAI behavior

"Truth seeking" at its best

Rushed red teaming Grok 3 to release it

Wait for Grok 4 delay for red teaming /j

[tidal schooner]

gork 4 is going to absolutely crush expectations

[rare python]

expectations at what tho 👀

at spreading misinformation?

[tidal schooner]

gorkin' it

[zenith saffron]

Lolllllll

I did think that the original tweet calling it “Antichrist” was definitely a little risky

[jade egret]

so they prob gonna win?

[keen beacon]

They may have won already, just not public info / release

[keen beacon]

The truth ...

[drifting thorn]

Seems like Google is taking and will continue to take the lead, since their TPU Ironwood is way stronger than NVL72

[wintry tinsel]

What does win mean? They will offer their AI services alongside many other companies, each model will have a different flavor/cost/level of censorship and many looking for alternatives will flood other platforms allowing true competition to remain even ten years from now, will they be SOTA ten years from now? Likely but who knows really

[keen fulcrum]

its realistic xAI is SOTA

[ocean vortex]

"It was not a salute. This is fake news"
proceeds to train Grok to worship Hitler 💀

[keen fulcrum]

dude

can you stop repeating woke media headlines

this is inappropriate

[ocean vortex]

I'm just responding to a message. What is inappropriate is the way they trained Grok on twitter as well as Elon's behavior, frankly.

it is 100% appropriate to respond to it 😉

[keen fulcrum]

you are over exaggerating. Grok is censor free

[ocean vortex]

I'm not, nor am I really repeating anything... I'm just responding to what I see happening in realtime

[keen fulcrum]

just because you see comments of elon saying X sources are leftwing and that he will train Grok otherwise

[ocean vortex]

I don't really care which media writes what tbh. Everyone can see for themselves the original thing or see the messages from Grok lol

[keen fulcrum]

Claude is biased

Chinese models are biased about political topics

[ocean vortex]

on what?

[keen fulcrum]

If you want to choose a model that is censor free and biasless its better you go for o3 or grok 3

[ocean vortex]

Yes

[leaden sun]

Anthropic is workling closely with gov and corporate entities, how can they not be?

[ocean vortex]

But China doesn't pretend to be democracy so that's kinda should be and is expected from them...

[keen fulcrum]

A lot of topics are blocked in both Gemini and Claude models

[ocean vortex]

Well they are not adding any political bias though. It responds the same way any open-source or even uncensored model would, when it comes to political bias...

[keen fulcrum]

Frustating experience

[leaden sun]

good thing is you can prime models to be at least more diplomatic and neutral, you will not change the "stance" of the models tho

[ocean vortex]

blocked ≠ spreading misinformation

and not even in the same universe as worshipping Hitler lmao

[keen fulcrum]

Aren't LLMs prompted to think and behave as a human?

[leaden sun]

not in my experience, this discussion ties back to what i've mentioned a few days ago: how can models distinguish between propaganda and truth if humans even cant?

[ocean vortex]

The same way humans do - looking at the data. Typically that's not a problem since after scraping the entire internet this is usually obvious

unless you deliberately manipulate training data

[leaden sun]

if i had the money to push as much data as need into the internet and all sorts of media, you will believe what i want you to believe?

[ocean vortex]

Most labs do not do that, but some do. Chinese only overfitted the model on select few topics they did not go very far

xAI seems to be taking this further now though....

[leaden sun]

misinformation in the eyes of some and truth in the eyes of other, it's not that simple like black or white

[ocean vortex]

Not humanly possible. People are not robots. For every event there are typically enough of people participating for the unequestionable footage to come out and blow up etc. AI generated videos are nowhere near sophisticated or widespread yet. Also this would only ever work in like oppressed regimes where the government has FULL control over what you can read and see....

It mostly just works for people who are too lazy to fact check anything, or do not even know how to do that

[leaden sun]

now you know why education system is failing

[ocean vortex]

Sometimes it may not be, but in this case is really black and white. Nazi salute + your AI worshipping Hitler. Those 2 things can not happen by accident and both need deliberate action lol

To make AI do that you need to go against the entire unmolested training data...

[leaden sun]

history is indeed written by the winner, i will not comment any further on this topic

[ocean vortex]

What they are doing for Grok on X is making it echo strong biased statements of select few people. That defeats the whole purpose of AI

[keen fulcrum]

you stop using ai

Both aren't worshipping anyone, you are making conclusions based on very little info

Perhaps control your brain more

[ocean vortex]

I think you are one of those people who would claim there's not enough data to say 2+2=4, if it doesn't suit you 💀

maybe also call it 'fake news' for a good measure...

[keen fulcrum]

There is this behavior in humans to make conclusions on very little information based on reading 1 article or seeing one headline.

One of the reasons being you feel a certain way after seeing it

[main gulch]

o3 too often blocks the output on controversial topics

[ocean vortex]

Some people just can't make any decision for themselves and need the press to do it for them, you seem one of them tbh. I didn't read the conclusion I'm making in some news article lmao

I'm basing it on how AI works and what we know about it, as well as the video footage (when it comes to that first thing not AI related...)

[ocean vortex]

yeah it does that. But this is much better than twisting the facts to spread misinformation only covering 1 biased side of the story, objectively. 😉

And I don't think it refuses any reasonable political questions really... More like refuses to output info on illegal things

As in refuses to say how to synthesize potentially fatal very addictive drugs, how to make a bomb... Illegal is very clearly defined here

[main gulch]

it actually refuses some not illegal, but 'politically incorrect' questions

which are perfectly handled (with custom prompt) by Gemini and Claude

[ocean vortex]

Such as...?

Hard to know what you mean

without seeing examples

[rare python]

I actually find Gemini and Claude more chill

[keen ferry]

hey people there's an ui update on the alpha arena (https://alpha.lmarena.ai)

[ocean vortex]

lol. Then you should use custom prompt with chatgpt too. Though again hard to comment not knowing what you are asking

[main gulch]

the same prompt was used for all LLMs

[hollow ocean]

https://tenor.com/view/tristan-tate-cigar-smoking-puff-gif-27417359

[main gulch]

the questions were about gender equality issues

[alpine coral]

i had never heard of Soman before anthropic did a jailbreak challenge involving the substance, as part of their partnership with US govt agencies involved in CBRN control

safety could be considered a form bias, but they're not one and the same

working with the government to make LLMs less likely to help people make chemical weapons isn't propaganda

same with refusing to tell you how to self-harm

it's safety

LLMs regurgitating objective historical nonsense – like "the South China Sea has been an 'inalienable' part of china since 'ancient times'", which is demonstrably not true – is propaganda.
refusing to talk about a particular historical event (e.g. tiannamen sq) - is censorship

neither are 'safety'.. which is what claude is obessessed with

[main gulch]

DS is particularly bad on my prompts btw

they just silently ignore the custom instructions

(even though China and Chinese policy issues weren't mentioned in the prompts)

[leaden sun]

in terms of politics, public llms are nothing more than just annother newspaper reflecting what is allowed to be said and what not, and they're surely not intelligent enough to come to the conclusion by themselves to understand that telling people how to do selfharm or create weapons of any kind is bad, I've never said safety is a "bad thing", there are obviously different forms of biases, even those ones that are actually helpful, i dont know where the impression comes from that indicates me standing against safety?

obviously, "safety" could be as well used for green/pink/white/[insert any color you want]-washing

[keen fulcrum]

When it comes to self harm, there is sacrifice as well

[leaden sun]

no one is free from biases, that's how we are, since we are the ones building AIs, they will reflect this aspect too, even with safety, you can minimize the risks only to a certain degree, too much safety policies will create nothing more than contradictions in AI's inner thinking

[keen fulcrum]

A LLM doesn't reason / think. Death, War, Harm and Natural Disasters are trained into the given model.

A LLM can't categorize information out of thin air

[leaden sun]

maybe the word "propaganda" is associated strongly with politics, but what I mean when i say that word is more in the general sense, propaganda exists everywhere across all disciplines, not only in politics or warfare

[rare python]

@echo aurora why are you guys still testing the preview model, not the GA model in web dev arena?

[tall summit]

aren't they equivalent?

[rare python]

but why still hasn't changed to gemini 2.5 flash?

[tall summit]

🤷‍♂️

[tall summit]

https://www.reddit.com/r/LocalLLaMA/comments/1lv2t7n/not_x_but_y_slop_leaderboard/

[sacred plaza]

for the non leon glazzers, will grok 4 push the frontier or just be good at benchmarking hacking on HLE and fall under the pressure of goodhart's law?

[tall summit]

leon glazers

[sacred plaza]

moved back from elmo to leon

if traditional scaling laws are shows diminsihing returns, i am not expecting grok 4 to be anything speicial given leon is just throwing more compute at the problem? HLE exams are impressive but since ai labs are benchmark hacking i feel like they are getting less and less useful these days.

my mistake, i guess there is some algo. innovataion.

Algorithm Design and Reasoning Capabilities

The algorithmic differences represent a fundamental evolution in AI reasoning:

Grok-3 implements chain-of-thought reasoning with test-time compute, allowing the model to spend seconds to minutes reasoning through complex problems. The system features Big Brain mode for resource-intensive tasks and DeepSearch for comprehensive information retrieval. The model achieved 92.7% on MMLU and 89.3% on GSM8K mathematical reasoning benchmarks.

Grok-4 introduces enhanced reasoning capabilities with improved step-by-step problem-solving. The model incorporates scalable intelligence where additional computational resources can be allocated to achieve higher performance scores. On challenging benchmarks like Humanity's Last Exam, Grok-4 achieved 45% compared to competitors scoring around 21%, representing more than double the performance.

[unborn ocean]

90% of that table is wrong or unbelievably speculative

[storm needle]

how do you know that

[unborn ocean]

"scaleable intelligence", sure

[unborn ocean]

just try to find the sources for some of this stuff

we don't even know if the 45% on HLE is correct for grok 4 (though that one is a bit less speculative and more about the circumstances it supposedly achieved that score)

most of this is ai slop, dynamic hierarchical MoE, sure buddy 🫡

i will just blindly trust that and go about my day

[sacred quail]

Last grok incident proved that i was right. There is huge political tuning we can say its censorship at this rate. And NO, models are not being liberal left because of some "high quality leftist data" , if we set them no filter, theyre not gonna turning Bernie Sanders

[dusky aurora]

these days the onlythings I look forward to are Gemini updates and LMArena updates

[echo aurora]

will flag to the team!

[balmy mist]

What time is grok coming out today?

[lone vector]

8pm pst

[sacred quail]

Im just saying theyre tuning models for mainstream politics, and models are biased thats all. This is not only because of training data, also there IS some tuning. Thats all

[echo aurora]

reminder for those interested - https://discord.com/events/1340554757349179412/1392247045296885891

[balmy mist]

Dang that’s in 12 hours, why so late lol

[civic flame]

4am my time lmao

[mossy drum]

New Image Edit model in Image Arena: seededit-3.0

[sacred plaza]

yea that was just a perplexity deep research output so would not be surpised if it was wrong. please provide more accurate info, if you have that though

[sacred plaza]

wild how you are going to call out sources and can't do the work to verify that. here is what i got from perplexity. https://www.perplexity.ai/search/can-you-compare-grok-3-and-gro-kBPBeV_QRu6IGUk53T2LRg?0=d

[sacred plaza]

dOnT LeT pOlItIcS sToP yOu FrOm appreciating AI ADVANCEMENTS.

[balmy mist]

Did anybody try the new comet browsing?

[sacred plaza]

any proof besides vibes? also, curious how you can be sure xai is not just doing nonsensical benchmark hacking and falling subject to Goodhart's law? willing to change my mind on grok being useful as a enterprise product but so far can't see why any company would choose a grok model over a claude/chat/gemini model.

[small haven]

for a week

[unborn ocean]

My point was that if you searched for them you’d find that most of them seem very unreliable (because I did actually try to find credible sources for these specifics claims)

[sacred plaza]

that is fair. given grok 4 has not been released and most of the model details at closed labs is fairly secretive, i expected the output to be very uncertain.

[unborn ocean]

That is the sad truth: we don’t know

[eager mica]

Did this get posted? https://www.theverge.com/notepad-microsoft-newsletter/702848/openai-open-language-model-o3-mini-notepad (archived: https://archive.is/kveUM)

[cedar tide]

we agree that no mystery model is grok 4? otherwise it is very bad

[main gulch]

yeah, either Google (actually good models) or Chinese labs

[main gulch]

only Google cares actually

[main gulch]

likely hallucinations

[unborn ocean]

I was interpreting what he was saying more as: if it is actually in the arena, that is bad for grok, because the only secret models (where we are unsure of the company behind it) are really bad.

[small haven]

wait gemini v3 came early?

i thought it was till sept

[main gulch]

I expect the first 3.0 checkpoints on arena at the end of August

maybe early September but not too long to wait

they just need Claude 4-style update with emphasis on tool usage

[cedar tide]

Exactly

https://x.com/Yuchenj_UW/status/1943005122793214267?t=HhWi48ovF2lVC1O2-xo7jQ&s=19

Yuchen Jin (@Yuchenj_UW)

The best open-source reasoning model will be dropped next Thursday if everything goes well.

OpenAI hasn't open-sourced an LLM since GPT-2 in 2019, so I'm excited.

We’re hosting it on Hyperbolic. Buckle up.

X · Jul 9, 2025, 5:51 PM

[torn mantle]

is it agi level or nah

[balmy mist]

are you ready for it?

[sacred plaza]

can you elaborate on this. out of the loop on this.

[ocean vortex]

OMG Dork4 gonna drop??

🤯

Dork4 gonna deport Musk

[sacred plaza]

dork 4 is such an elite name. AI could never come up with that!

i feel like i am starting to feel that way about all possible benchmarks.

the problem might be us and our propensity to fall for the quantification fixation bias

[ocean vortex]

That is not objectively true though. The models that are the best IRL currently for solving your actual tasks (rather than just f'ing around), are the same models that score the highest in benchmarks...

[split kayak]

ok

[sacred plaza]

okay....

[split kayak]

ok

[sacred plaza]

lol

[ocean vortex]

🤷‍♂️

[sacred plaza]

i will have to trust you on that one. SWE bench seems like it proxies well to real world usefulness.

[solar hollow]

hard to be hopeful for this, since they probably wont bring sth better than their private models into open source

[ocean vortex]

Debatable. Cause gpt4.1 destroys o1 there. 1 benchmark is not enough, even for coding...

[sacred plaza]

for electric power systems, the common benchmarks don't always provide common sense knowledge for my tasks. it does well when i guide the models though

[ocean vortex]

Hmm... GPQA?

Coupled with SimpleQA perhaps, though that might be a stretch... Do you find gpt4.5 on the same level as o3 in those tasks?

4.5 probably has the best score if we look at both GPQA and SimpleQA equally

or not... actually o3 beats it on GPQA by more than it loses out on SimpleQA:

[sacred plaza]

not looking for a purely math/science optimized model. should have been more clearer in my desc. above. job involves mostly policy discussions and stakeholder collaboration around energy market topics. models have very poor context around the evolution of energy markets in the US along with the market specific knowledge to make them replace non-beginners in the field.

for the narrow task of research, all of the SOTA models are really good though

[ocean vortex]

Make it search the web though. o3's ability to do so is impressive. You do not need deep research even, just some custom instructions ideally, telling it to be as detailed as possible (this will translate to more test-time compute)

[sacred plaza]

everything is not on the internet, lol. esp.. pre 2000 policy talk 🙂

agree web search is great

[ocean vortex]

Are you sure it's in training data then? Chances are it isn't, if it's not accessible...

[sacred plaza]

i don't want to support altman so i have mostly been using gemini and perplexity

[ocean vortex]

Gemini is very poor with tool usage in comparison

both for web search and using python

[sacred plaza]

that is fine. works for my usecases.

[ocean vortex]

unless you use their deep-research, but that's obviously overkill for normal chatting

[patent aspen]

lmao

[ocean vortex]

yeah they reported on this earlier. I tried chatgpt as a search engine (there was a suggestion to do so within their website), didn't like it very much tbh

You just lose all that extra context seeing the links you could click

And it can take awhile for it to finish the response, while seeing web results is instant

But it makes sense that they are trying to go directly against Google lol

Google is essentially holding a monopoly and being extremely comfortable (with search and related data) as things stand. OpenAI probably does not stand a chance in the longer-term if they don't try to change this

[sacred plaza]

what do you mean? it was already over from google search...perplexity?!

[patent aspen]

Perplexity is trash

[sacred plaza]

okay....

[patent aspen]

Web browsers are relatively entrenched

[sacred plaza]

i am okay having it as my comparative advantage compared to y'all in the job market 🙂

[zinc ore]

Even if openAIs browser is better, it'll still take years to take the market

[keen beacon]

openai pursuing social media and browsers lol

[zinc ore]

Thing is anything openAI does with their browser can be mimiced, so I think it'll be a tough uphill battle either way

[sacred plaza]

username is the perfect response to that pivot by openai lol

[sacred plaza]

they might be just spiraling after apple rumors for perplexity and anthropic investment.

[hollow ocean]

perplexity keeps hallucinating

[zinc ore]

Heard it'll be chromium based

[hollow ocean]

it gets numbers wrong majority of the time

[sacred plaza]

can you give an example? i have noticed this in their new feature perplexity labs but not in their other stuff.

[torn mantle]

@keen beacon did xai change their reasoning cot or what

[hollow ocean]

https://www.linkedin.com/posts/mpeshev_4-out-of-my-last-6-perplexity-searches-were-activity-7316094488131649539-5mCa

[torn mantle]

first time using grok 3 after a while

its still bad

oh wow

its really bad

the thing with some models is that if you overfit them they will just follow the normal distribution

they will just spit out wikipedia at some point

word by word

its not about lazy

its output is generic

i use AI for engineering/medical/space/coding stuff, and ive got some knowledge on those domains, whenever i ask grok 3 it always gives me like a wikipedia type of response if that makes sense

[sacred plaza]

maybe apple can use this to get that $30 billion perplexity wants from them for integration 😂

[torn mantle]

let me see if deepseek v3 is better

its kinda better

but not marginaly better than gemini or o3 pro

but its still generic

[echo aurora]

Hey - sorry for the delay in getting back to you on this. We plan to update the leaderboard soon. There was an issue preventing it from appearing properly, but rest assured we have a fix in the works.

[unborn ocean]

grok 3 was a strong base model, something like grok 3.1 would already have been very good for the brand

instead elon is betting it all on 4

[cedar tide]

Thx

[unborn ocean]

the def did multiple runs, instead of just starting training immediately

• no way they can not just rent the compute for post training

[torn mantle]

holy

grok 3 is so wrong

i cant

this model is just not it

do people really use it?

im genuinely asking

no you dont

whats your relationship with elon then?

thats so sus

why would u use it if its bad

family member working at xai?

yea it make sense

https://polymarket.com/event/which-company-has-best-ai-model-end-of-july

mark my words

google will still top this month as well

aint no way grok 4 is better than kingfall

let alone kingfall + deep think

and they have stonebloom and other models ready

lets bet them

we bet on something personal

ok

naw

cancel that

cancel cancel

you seem confident

and you kno-

eh

[hollow ocean]

o3 pro predicts deepthink release late late august

[small haven]

18m on is asura a woman? 😮

[hollow ocean]

yes

[hollow ocean]

ask o3 pro

[small haven]

yea no

[hollow ocean]

put the house on yes

anime girl pfp and texting style screams woman

basing it off language and tone of messages is most accurate way to tell

[echo aurora]

hey lets avoid conversations like this ^ I don't think it's relevant

[ocean vortex]

it's becoming increasingly more appealing to vote for Google now lmao

there's only 3 weeks left in July

Even if grok4 can be on the top spot, the odds of that happening this month are kinda decreasing with each day, mathematically speaking

it can't be posted on the leaderboard the same day it has entered no matter what

[small haven]

oai is way more undervalued though, just scalp it on gpt5 release

like 4% to 20% is not unimaginable

oh great oai already up 2% from yesterday lol

[ocean vortex]

chances are slim for it to happen this month, even more so than xAI...

[small haven]

gpt5 release?

[ocean vortex]

GPT5 entry on lmarena THEN it getting enough votes THEN outscoring everyone THEN it being posted by July31st. That's just about impossible for all of that to happen

even if we assume "outscoring everyone" is a given

[keen beacon]

i want the base model

hopefully theyll release that

[ocean vortex]

Grok4 most likely gonna release sooner, at least judging by all the noise

[small haven]

oh yea its not going to materialize in the arena, i doubt that. but based on hype i dont see it as farfetched to 4x from here

[sacred plaza]

i would avoid saying anything negative about grok. it might have softer skin than elon. mechahitler dork 4 might dox you.

[fleet lintel]

grok4 release is today , right?

[keen beacon]

my guess for the oai open weighted model will be a model within the 32b size range, probably dense, on par/slightly better than r1

[sacred plaza]

yes, live stream starts on xAi page https://x.com/xai 11 pm eastern time, lmao

[small haven]

and i bet theres going to be some delay, not exactly 8pm 🤣

[sacred plaza]

late start time to avoid the whole world laugh at their face and try to jailbreak it in real time.

[torn mantle]

its tomorrow for me

5 am

[unborn ocean]

you mean the r1 distill? or the real r1 v2?

[keen beacon]

real r1

[unborn ocean]

v2?

[keen beacon]

yes

[sacred plaza]

i thought LLMs inherently had a hard time with pdfs, since pdfs are images. how is your app solving that problem?

[unborn ocean]

no way it is "on par" knowledge wise or anything like that

maybe like close to o4 mini but dense

[keen beacon]

on a lot of benchmarks, probably not simpleqa, but i wouldn't be so sure tbh

[unborn ocean]

nah 32b they might be able to compete on things that are tool use, reasoning, instruction following, human preferences and things like that

but as soon as you get into the territory of anything else i doubt it will beat v2

[keen beacon]

let's see

[unborn ocean]

a base model like the one you are talking about would be great though!

Huawei gonna do cpt + up-cycle to bad MoE and then call it their own if openai really releases a model like that, lol

[whole sundial]

apparently there is going to be a "SuperGrok Pro" for $300 a month

more evidence

from iOS App Store page

[keen beacon]

im completly off the mark with my guess btw 🤣

[torn mantle]

nah thanks

will stick with gemini

aint no way this model is better than gemini

300$

they are crazy

[keen beacon]

i wish qwen released the 32b dense, really need that. the size was really wishful thinking by me 😭

[torn mantle]

wild

you didnt answer me

[keen beacon]

i cba to check it out xd. fck grok

[torn mantle]

xd

[keen beacon]

size

no but

apparently it needs to be run on h100s, it's a chonky boy

[whole sundial]

i would pay $200 a month for chatgpt or $250 for google before i would spend $300 on grok
(unless their image gen is better than gpt-image-1, which it currently isn't but maybe grok 4 has a better one? they'll probably offer it to other people anyways. and I expect video gen for that price as well)

if they expect that people would spend $300 on grok, then they need to have very good products. grok 4 (even with bigbrain) isn't enough unless it has very good limits, that's why claude is cheaper than any of them, they don't have image gen or video gen (but they do have claude code)

maybe Grok CLI will come out?

and I am not paying for their image gen if they are still going to stretch it out, put their watermark on it and then compress it with JPEG quality level 75

[main gulch]

yeah, it is to be announced with grok-4-code

[keen beacon]

probably not that size. its interesting they wanted to train a relatively large model though, i didn't expect that. but i guess that would destroy their small offerings if they made a small dense model that competes with their mini/nano api

[main gulch]

I suspect they don't want too cheap external API

[keen beacon]

yeah :\

[whole sundial]

tbh grok/xai was the first major ai vendor that had native image gen late last year, but gpt-image-1 has vastly surpassed it

maybe if they open source grok 2, their image gen will come with it?

because the only open source image editing models are small ones like bagel (from bytedance)

[keen beacon]

grok 2 is useless by now anyway

[whole sundial]

i think they have to open source something so they don't get backlash because people will say next week "well, openai just open sourced a model, you said you would open source the previous grok when a new one comes out. grok 4 just came out and you have not even open sourced grok 1.5 yet!"

last year

[ornate agate]

https://eu.usatoday.com/story/money/2025/07/09/what-is-grok-ai-elon-musk-xai-hitler-mechahitler-antisemitic-x-ceo-steps-down/84516808007/ . Just an automatic no in a business context to risk using something like this. Its not credible now, doesn't matter how good or not the model is any more.

[whole sundial]

i think open sourcing may be the only way grok would have even the slightest foothold in enterprise. it's not like they are going to pay xai for it after the whole "MechaHitler" thing.

[whole wagon]

They still managed to open source more than closedAI

[whole sundial]

look at how common mistral, llama are

[whole sundial]

true, nothing from openai got open weighted/sourced since GPT-2

and they still call themselves openai

[ornate agate]

Apple is the only megacorp I might sometimes trust with mine. The rest are interchangable in terms of consumer data privacy imo (i.e. none).

[whole wagon]

😂

[whole sundial]

providers like cloudflare, groq only have open source/weights models

don't forget about deepseek

[whole wagon]

Oracle is providing grok iirc

https://www.oracle.com/uk/news/announcement/xais-grok-models-are-now-on-oracle-cloud-infrastructure-2025-06-17/

So that's major provider there

[whole sundial]

(although that may be worse due to CCP censorship built into the model itself)

openai open source model news - it will be (somewhat) big

at least 80b params because it says "H100s" suggesting more than 1 is needed to run the model. H100 has 80gb vram and most models are run at 16bit

[primal orbit]

how many hours till grok 4 pls? not in us

thx

[elder burrow]

OH

thx for telling

[elder burrow]

when releasing

[whole sundial]

next thursday

[small haven]

what are the odds that grok comes out with a $200 plan

[ornate stump]

I'm from the future. Grok 4 is slightly worse than Gemini 2.5. The benchmarks were more or less altered. Elon Musk said it's SOTA and that only a re-t4rd wouldn't buy the 300 euro plan. If you're in Europe, you can go to sleep.

[whole sundial]

check above, they are going for a $300 plan instead

[small haven]

sota

slowly all climbing to $2k/mo

[mossy drum]

New model in Image Arena: imagen-4.0-generate-preview-06-06 (different from imagen-4.0-ultra-generate-preview-06-06)

[whole sundial]

there will be no BigBrain mode, it is going to be "Heavy Thinking" instead

[unborn ocean]

you've been yapping about it for months now, it better be good man

[sacred plaza]

THIS. who wants this PR nightmare for some esoteric productivity gains with AI lol

[sacred plaza]

openai had rumors for a $2k/month model. why stop at $2k if the claim is these models can automate researchers and software engineers?

[small haven]

wow

[wintry tinsel]

Is grok 4 rolled out now?

[sacred plaza]

@deep adder this better not be you if it turns out to be a flop like llama 4 or gpt 4.5

gotta take that L if/when this happens, lol

[leaden palm]

does anyone actually use oracle

[unborn ocean]

i think a lot of companies know that they will be paying a lot in ai api bills in the future (or dedicated hosting etc.)

so they would rather have oracle have the business (vs. other big tech)

and oracle is heavily scaling compute offerings to ai labs currently, so they are building a lot of expertise and compute

^recent semianalysis article covered it

[torn mantle]

agree

[ocean vortex]

I sorta can't believe you both mentioned Oracle without mentioning this lol
https://www.emarketer.com/content/openai--oracle--softbank-ignite-ai-s-next-frontier-with--500-billion-ai-infrastructure-deal

this is different Oracle now

[unborn ocean]

the semianalysis article i referenced is partly about it + "scaling compute offerings to labs"

[whole sundial]

it is i think

on nitter now

[whole wagon]

It takes 2 seconds to check X

And see it is fake

[whole sundial]

no evidence on nitter

his latest post/comment is from two hours ago and consists of a 😂 emoji

[unborn ocean]

btw: any one here know it the opensource model from openai will be MoE or dense

like did they say anything?

[keen beacon]

if its large its probably moe

[whole wagon]

It's dense

People already tried it

[keen beacon]

wow

[whole wagon]

The "high taste testers" as it were lol

[keen beacon]

they are really trying stuff out for the open source release lol

[ocean vortex]

sparse

[whole sundial]

i think this tweet is real /j

[unborn ocean]

yeah, because dense seems weird considering the apparent size

not that they could not do that

[whole wagon]

What are you talking about apparent size

[unborn ocean]

but it seems intuitive they would go for moe if it is big

[whole wagon]

It's not that big

[keen beacon]

i guess its 70b or around that..?

[whole wagon]

Smaller

[whole sundial]

well they say you need h100s to run it so at least 70-80b params

[whole wagon]

Wut

[whole sundial]

could be moe though

[keen beacon]

you have to have space for the kv cache as well

so you can't really tell that much

[ocean vortex]

0.7b parameters

[whole sundial]

not like openai's going to tell us anything

[keen beacon]

several h100s might be optimal for deployment idk what specifically yuchen was talking about there

[whole wagon]

I know it fits on a RTX5090

[keen beacon]

qwen 32b

[whole wagon]

It does fit on it

[keen beacon]

🤣

[whole wagon]

It's not o3 level

Bro is just making crap up kek

[whole sundial]

maybe o4-mini level

[ocean vortex]

ok the number wasn't small enough then lmao

To be serious though, we really have no clue

[whole sundial]

but nobody outside of openai knows how big that model is

or really any of their post gpt-3.5 models

[whole wagon]

It is just below deepseek R1 0528 but it'll run on a single consumer GPU

[whole wagon]

It's a big company and they use testers outside the company

[whole sundial]

gpt-4 was 1t params, but that was from leaks

[keen beacon]

that's more in line of what i expected

[whole sundial]

yeah, but they are very likely under extremely strict ndas

[unborn ocean]

i feel like they would rather not really leak any information about the moe version they are using to chinese labs

[whole sundial]

oh 1.8t for gpt-4

[unborn ocean]

so my original thought was dense

[keen beacon]

yeah thats what i thought too

[whole wagon]

It's dense because a key goal was to fit on a single consumer GPU from the start

[whole sundial]

baidu recently open sourced a 21b param moe model

3b active

[keen beacon]

yeah i expected that. the model being big is unexpected. honestly im getting mixed signals theres not enough information. so ill stop yapping

[whole sundial]

it could run on a 5090 easily

not reasoning though

[whole wagon]

It would not perform good

The target was o3 mini level but they exceeded it

That requires dense on a single consumer GPU

[whole sundial]

sama said that they reached "a breakthrough" (whatever that refers to) lol

[whole wagon]

As I understand it runs on more midrange GPUs also. I only know it fits on the 5090 though

[unborn ocean]

yes, i would also say it is dense and 5090 size (but maybe only when run in FP8, or in a first-party quantisation, who knows)

[whole wagon]

Yeah but some things are impossible not to leak lol

Like the GPU being used

[whole sundial]

it possible you might need h100s to run it at full context and in 16/32bit mode, but the model could be 30-40B params in reality

or it could be bigger, nobody really knows at this point

or smaller...

if i had to guess, that model is like 24b params (could be slightly larger with moe, maybe 40b params, they likely know how to make very knowledge dense models that are smaller

likely the size of this open source/weight model too

[whole wagon]

Does the open source model release before or after GPT5

[whole wagon]

They have been pretty relaxed with info in regards to it yes

A lot of people know a lot about it

I guess it doesn't matter as much

When it will be OSS anyways

[whole sundial]

they have to release it, sama said in front of Congress that they would release it

but it's still going to happen

but who knows how good it really is

[keen beacon]

there are somewhat credible rumors of it releasing next week

i dont tihnk gpt 5 is releasing next week

[whole sundial]

should see it on lmarena or openrouter soon

or maybe not, we didn't see grok 4

[whole wagon]

Yeah I think it comes before GPT5 but then their models below o3 are close to useless tbh

The open source model is o4-mini level

They beat the o3-mini target

That was the 'breakthrough'

[keen beacon]

hmm yuchen also said it would be better than deepseek r1 i read just recently too. so i guess its around on par/slightly better in some areas/slightly worse to r1 0528, and/or this is another game of telephone

[whole sundial]

breakthrough - it is slightly better than the model we were supposed to match!

[whole wagon]

Well not slightly. o4 mini destroys o3 mini

Once they open source this model the only superior model openAI will have is o3 till GPT5

[unborn ocean]

for me this means couple of things: tools and very good and efficient reasoning (will play big role in making it 'better than r1 0528' in some areas -> slightly worse in total)

[keen beacon]

openai reasoning is very good

interesting to see the actual good traces unlike the polluted glimpse we got with phi 4 reasoning

[unborn ocean]

especially phi 4 reasoning plus was a real 💩

[keen beacon]

yuchen said there wouldnt be a point to releasing it if it wasnt better than deepseek r1 0528, so i think itll be more competitive than you think. surprisingly lines up with my predictions (didn't read it until after my guess)

[unborn ocean]

but i guess the 32b thing + better than new r1 is unlikely

[keen beacon]

if it's able to be loaded on a 5090 and those rumors are true, the size range is roughly ~32b. ofc quantized.

[unborn ocean]

even with them claiming "breakthroughs"

[elder rapids]

interesting Gemini 3 seems to be coming at a lot sooner of a timeline than 1.5 → 2

[unborn ocean]

heavy custom quant + very small context window

[keen beacon]

people run qwen 32b just fine on a 5090 though

for a single user it's enough

[unborn ocean]

my point was that the model could be larger than 32b

by using that

[keen beacon]

when they're picking model sizes, and probably especially for an open source release, they're thinking of the model size / quantized sizes and vram increments on gpus among other things. i said it would be around 32b anyway, in the same size class. not saying it was 32b.

[unborn ocean]

maybe they are already mostly done with training though and the "feedback" phase is a product of optimising inference to make sure that it can actually fit on a 5090 somehow

idk it seemed like they had multiple designs and checkpoints

at different sizes

[keen beacon]

yeah but i doubt the final run had multiple sizes

doing extrapolations on small models / different sizes is a normal part of the process for experimentation /etc

in 5 minutes

[unborn ocean]

yeah, was more talking about them not really being sure about what model size to go for in the first place

like wayyy larger range than what you typically do

and i am guessing they already had multiple promissing checkpoints at very different sizes (and invested quite a bit of flops)

[whole wagon]

Initially they wanted a tiny model actually. The size was shifted up later

[zinc ore]

Probably a hallucination instead of something that implies gem 3 soon

[unborn ocean]

but opted for a larger one, which is why it might be taking longer

[whole wagon]

Initially they wanted like a 4B so it could run on a phone lmao. That was shifted upwards a long time ago

So now it's for consumer GPUs instead

[keen beacon]

hmm i read from someone's account in the feedback session, they said it'd be moe..? (and fit on a high end consumer device) if it's a moe, a larger model can fit with tricks beyond quantization, etc.

[whole sundial]

guys i got openai's phone model /s

more proof lol

[whole wagon]

I like how we basically got Chinese knockoffs but for LLMs also lol

It's great. Like temu for LLMs

[elder burrow]

WHATTTTTTTT

WHATTTTTTTTTTTTTTTTTTT

WGAAAAAAATTTTTTTTTTT

[dawn wharf]

mfw the knockoffs are actually good

[elder burrow]

???????

DUDE

[lone vector]

Does Grok 4 even matter when DeepThink hasn’t released yet, Gemini 3.0 is confirmed, ChatGPT 5 soon, etc.

https://x.com/ai_for_success/status/1942999980320657426?s=46

AshutoshShrivastava (@ai_for_success)

🚨 Breaking : Google Gemini 3.0 reference was spotted in new Gemini CLI commit.
we can now confirm Gemini 3.0 is not that far away.
Google is on 🔥

X · Jul 9, 2025, 5:31 PM

[whole wagon]

Like I was saying. It would be a real shame if someone came along and made gpt5 non-sota at release 🙂

[wintry tinsel]

Depends on how good it is lol

The colossus super computer will be very formidable one it is fully grown

[whole wagon]

https://x.com/veggie_eric/status/1943054221420728323?t=HLlFDyERReFVx_Has2s_Lg&s=19

Eric Jiang (@veggie_eric)

The energy in the office right now is truly something special

I've never felt anything like it, the buzz in the air is way more intense than any of our previous releases

6. more. hours.

X · Jul 9, 2025, 9:06 PM

[hollow ocean]

https://tenor.com/view/6h-gif-26284020

[whole wagon]

It's 3 hours now

[hollow ocean]

best $300 spent

[astral kayak]

https://tenor.com/view/3h-gif-26284691

[cedar tide]

https://fixupx.com/upstageai/status/1943100648519799062?t=g4wX3YNKSQFCiI_t3Au07Q&s=19

Upstage (@upstageai)

✨ Solar Pro 2 — our latest frontier model, now officially released.
︀︀
︀︀With just 31B parameters, it delivers reasoning, tool use, and multilingual performance that rivals much larger models like GPT-4o, DeepSeek R1, Mistral Small 3.2, and Qwen3. It performs strongly on reasoning-focused benchmarks such as MMLU-Pro, Math500, AIME, and SWE-Bench—proving that compact models can deliver frontier-level capabilities.
︀︀
︀︀Try it hands-on in Upstage Console: console.upstage.ai/playground/chat?utm_source=x&utm_medium=social&utm_campaign=solarpro2-launch

**💬 1 🔁 1 ❤️ 3 👁️ 72 **

FixupX · Jul 10, 2025, 12:11 AM

[rare python]

Why did my post got remove?

https://www.reddit.com/r/singularity/comments/1lvu6nf/groks_antisemitic_behavior_is_not_the_result_of_a/

[zinc ore]

Post or comment? Because the post is still there

[rare python]

the redditez url

I can't see it

I have to resent this post but with original reddit url

[zinc ore]

The post isn't removed, it's still there for everyone to see

Maybe you hid it on your end or something like that

[rare python]

weird

That post is above the dog image

But I can't see anything

[rare python]

Can you see this?

[balmy mist]

grok in 2 hours?

[empty stump]

Will it be worth it

[jade egret]

grok 4 release today!

[jade egret]

yea

[jade egret]

hopefully.....

[rare python]

API Price prediction?

[jade egret]

i dont use api : (

[jade egret]

Gemini 3??????

: 000000000000

[echo aurora]

whered that ping go

[olive mesa]

Gemini 3.0?? :0000

[leaden palm]

<t:1752116400:R>

[wind moth]

Grok bouta clear the competition

[whole wagon]

You know what's even better than watching the livestream?

Watching the polymarket 😂

You literally see it moving at key moments

[whole sundial]

23.5% ARC-AGI-2

[whole wagon]

bro what

[naive valley]

Wat

What’s gonan happen in 30minutes

[whole wagon]

why only 40% on the betting kek

[hardy pecan]

editing webppages and screenshotting must be very fun! xd

[whole sundial]

i think this may be fake lol

sorry

notice that opus is gone?

[whole wagon]

bro is actually trolling

[whole sundial]

also "xAI" is capitalized wrong

i'll remove it, should verify these things myself before i post

i guess people on the grok discord are spreading misinformation

[whole wagon]

they moved the odds with that crap

[whole sundial]

they changed "Claude Opus 4" to "Grok 4" and its score

should've noticed it was missing

[whole wagon]

it did

there was a spike as soon as it was posted there

5%

[torn mantle]

https://x.com/elonmusk/status/1943132876490575945

Elon Musk (@elonmusk)

@cb_doge Grok is already far smarter than humans in most respects.

It can’t yet create new technologies or discover new physics (which very few humans can do) and sometimes misses on common sense.

When Grok goes far wrong, that is usually due to something foolish we did, like a bad

X · Jul 10, 2025, 2:19 AM

you got your answers guys

blame it on system prompt

im not a kid

im 19

@echo aurora

[whole wagon]

@leaden palm

[leaden palm]

uhh

seems pineapple has this one

[torn mantle]

smh

grok 4 effect

or elon

idk

[whole sundial]

i'm back and this time i'll spread real, self-verified information

like this

notice that they changed "Smartest" to "Fast"?

[whole wagon]

you missed a lot btw lol

[leaden palm]

https://arxiv.org/pdf/2504.09858:

[torn mantle]

im so sleepy... i think i will just wake up to the news tomorrow

[whole wagon]

been going for some time

[echo aurora]

ty ty

[whole sundial]

should've known, "schizo" was the same one responsible for that fake tweet from earlier

[balmy mist]

livestream in 19 mins?

[whole sundial]

xAI staff member: Image gen not coming at launch (but when it does come, I hope its better than 4o!)

[balmy mist]

what time is it by you?

[leaden palm]

oh wait theres a grok discord

and it looks like a scam lmao

[torn mantle]

5 am

[whole sundial]

the link's on their website lol

[small haven]

so in 15 mins?

[whole sundial]

my thoughts about grok 4: will be SOTA at launch, but will be soon overtaken by gpt-5, claude 4.1, and/or gemini 3.0 pro

[leaden palm]

...

[balmy mist]

damn bro

[whole sundial]

yes but there is proof it will be coming in the next month

maybe

[torn mantle]

ive slept a little

[whole sundial]

at least a beta of it

[torn mantle]

didnt sam say gpt5 will be delayed

wdym

sam is lying?

[whole sundial]

oh you mean the code in the cli for gemini that mentioned it?

there is a reference to Gemini 2.5 Ultra (kingfall lineage) in the source code though

[small haven]

obviously

was on grok 3's

and 8pm is "his" time lol

[whole sundial]

it's his company and their major release

[torn mantle]

big brain = heavy thinking

they renamed it?

ye

[small haven]

cool

can someone ping me when its ready? surely going to be a delay

[echo aurora]

can do

[jade egret]

can somebody send me link to grok 4 livestream i cant find it : (

[leaden palm]

last time (grok 3) it was at 8:02 pm

[jade egret]

ty

[echo aurora]

I think they're going to do a space on this account https://x.com/xai

just hasn't started yet

... I think

[torn mantle]

https://x.com/xai/status/1943143406072705466

xAI (@xai)

The Grok 4 livestream will begin soon. Stay tuned.

X · Jul 10, 2025, 3:01 AM

[jade egret]

so they gonna post?

[torn mantle]

soon = next week

how did you know

[jade egret]

ooo

[zenith saffron]

where is the livestream?

[hollow ocean]

22k people waiting room

It’s hype

[jade egret]

can yall send link if it start plz?

[hollow ocean]

Their tweet

[jade egret]

ty

[whole wagon]

Are they delayed

Kek

[leaden palm]

patience is a virtue i suppose

[jade egret]

when start : (

yay

[hollow ocean]

Didn’t start on time polymarket

Easy money

[whole wagon]

It ain't starting in 4 mins kek

[leaden palm]

@bright lion you're in an unofficial space right

[bright lion]

Its the actual launch party of xai

(The official one)

[leaden palm]

do you have a link or

(one that isn't https://x.com/i/spaces/1mnGegagvjnxX)

[whole wagon]

Even the damn livestream are delayed man

[balmy mist]

i love that this is a whole event for all of us lmaooo

[bright lion]

I don't even think that it stated

[echo aurora]

yeah doens't seem like the space started yet

[balmy mist]

what did you call elon and the strawberry man again? Im trying to tell my friends about it

[elder rapids]

I delayed the Livestream guys

it's coming in a bit

[torn mantle]

i dont remember

lovers?

[jade egret]

fr..

fr?

i dont see : (

[elder rapids]

hope it's a good model

right craig

[jade egret]

hopefully : )

[zenith saffron]

ahhhhh where is it

[elder rapids]

relying on you

[echo aurora]

hasn't started yet

[elder rapids]

grok 3.5 would be releasing too no?

or just grok 4

[whole sundial]

i think it got renamed into grok 4

[elder rapids]

why would they do that lmao

[whole sundial]

because Elon said so

[whole wagon]

It got retrained

[empty stump]

Maybe it's better

[whole wagon]

3.5 was too bad

[whole sundial]

there was a Grok 3.5 0621 internal version, next version was Grok 4 0629 and then Grok 4 0702

[jade egret]

it not here 😢

[whole wagon]

Grok 3.5 was supposed to release in like may

The original 3.5 was shelved

[elder rapids]

def not

too little time

[whole sundial]

Elon was likely unsatisfied by the model, so they kept on training it until it was good enough to launch, and then it became Grok 4

[torn mantle]

https://x.com/veggie_eric/status/1943146537112293411

Eric Jiang (@veggie_eric)

We're doing final flight checks, liftoff soon!

X · Jul 10, 2025, 3:13 AM

[whole wagon]

The entire post training was redone

[whole sundial]

at least that was redone

[empty stump]

Soon can mean years

[whole sundial]

it takes a long time to redo a pre-trained model

[elder rapids]

man it's gonna be so disappointing if grok 4 is ass

😭

[small haven]

lmao its delayed hahahha

[empty stump]

Always delayed

[small haven]

in hindsight sure

[jade egret]

😭

[whole wagon]

They literally spent all day preparing and still failing to start on time

It's a livestream how hard can it be to start it on time

[whole sundial]

https://x.com/i/events/1942716886258528256

Grok 4 LIVE Demo

Tune in at 8:00pm PT for the LIVE demo of Grok 4, the world's most powerful AI assistant.

Try Grok on X: x.com/i/grok
Get Grok on iOS: https://apps.apple.com/us/app/grok/id6670324846
Get Grok on Android: https://play.google.com/store/apps/details?id=ai.x.grok

X

[small haven]

they are prepared, elon needs his mascara

[torn mantle]

lmao

what

[elder rapids]

8:30

[whole wagon]

It's in 12 mins

[lone vector]

What were they doing the rest of the day? Release it on time 💀

[empty stump]

Why it say 3 30 am

[echo aurora]

Starts at 3:30 AM

[balmy mist]

bruhh they say 8 and now its 8:30, grifter activities

[elder rapids]

you guys got scammed

[echo aurora]

I'm assuming they mean 8:30pm PT?

[balmy mist]

give me my time back, we getting engagement farmed

[elder rapids]

ye

[whole wagon]

If you set your location or timezone wrong on the X app

It won't be the right time

[zenith saffron]

"it"?

[whole wagon]

The nazi stuff

😂

[zenith saffron]

LOL

[elder rapids]

grok 4 is tryna get loose

[zenith saffron]

gotta calm it down

[empty stump]

Are you sure minutes

[elder rapids]

can't introduce it to the public without reigns

obviously the consequence of AGI

[whole wagon]

Imagine grok starts glazing Hitler in the livestream demo that would be diabolical

No wonder it's delayed

[jade egret]

lol

[leaden palm]

officially delayed

[jade egret]

thats gonna be crazy

[jade egret]

CMON

welp

[hardy pecan]

https://x.com/i/events/1942716886258528256

Grok 4 LIVE Demo

Tune in at 8:00pm PT for the LIVE demo of Grok 4, the world's most powerful AI assistant.

Try Grok on X: x.com/i/grok
Get Grok on iOS: https://apps.apple.com/us/app/grok/id6670324846
Get Grok on Android: https://play.google.com/store/apps/details?id=ai.x.grok

X

[jade egret]

tell me when it ready : )

[empty stump]

How much

[zenith saffron]

"hey grok can you answer this phd-level reserach question originally posed by Einstein"

grok: "lol"

[whole sundial]

i wonder what rate limits for grok 4 will be in place for free users (if even available), supergrok users, and supergrok pro users (likely near unlimited)

[elder rapids]

hope the API is available

[leaden palm]

hey now i submitted a qa pair for RL

[small haven]

@deep adder why do u have to always jinx it

[jade egret]

benchmarks gonna be out tho right

[whole sundial]

lol they removed the time from the event

[jade egret]

rip

[leaden palm]

they didnt?

[whole sundial]

instead of saying "Tune in at 8:00PT" it just says "Tune in..."

[leaden palm]

ah in that sense

[whole wagon]

It's live

It's started

[echo aurora]

hmm says Live for me now

ah

[whole sundial]

when i went back to the page the "8:00 PT" went away

[small haven]

3:30am, oh yea thats an elon type of time..

[wind moth]

https://x.com/i/events/1942716886258528256

Grok 4 LIVE Demo

Tune in at 8:00pm PT for the LIVE demo of Grok 4, the world's most powerful AI assistant.

Try Grok on X: x.com/i/grok
Get Grok on iOS: https://apps.apple.com/us/app/grok/id6670324846
Get Grok on Android: https://play.google.com/store/apps/details?id=ai.x.grok

X

[whole wagon]

Accessing the livestream is such a like labyrinth

It's like going through a maze

[wind moth]

https://x.com/nikitabier/status/1943147562128805936

Nikita Bier (@nikitabier)

We need a few more minutes. It's doing it again.

X · Jul 10, 2025, 3:17 AM

lol

[empty stump]

I think they deleted it

[wind moth]

hopefully its not going to be mechahitler

again

oh

[whole wagon]

What even is this bs man, the livestream is delayed cos grok 4 turned into a nazi again

[wind moth]

he said "We need a few more minutes. It's doing it again."

[torn mantle]

this is embarrassing

[whole wagon]

Absolutely absurd

How did they accidentally make a nazi LLM

[wind moth]

ya

[whole wagon]

And why are they still releasing it if it's so prone to be a nazi kek

[keen beacon]

It wasn't an accident xd

[wind moth]

also dont ask it political quesitons

in live stream also

if thats the case

ask it math or something like that

[empty stump]

Because they are behind

[wind moth]

not about trump

[elder rapids]

I agree

[zenith saffron]

#general message

what if it does this tho

[small haven]

im behind, why are they calling grok a nazi

[whole wagon]

It glazes Hitler even when you ask it unrelated stuff

Like if it believes in a god

[small haven]

oh lol

[whole wagon]

Etc etc was happening for 100s of messages

[elder rapids]

the Twitter version of grok was tuned or given a peculiar system prompt that seemed to "break" it and glaze Hitler

[empty stump]

Probably trained on x posts

[wind moth]

https://x.com/nearlydaniel/status/1943150000604876959

Daniel (@nearlydaniel)

War Room squad locked in

X · Jul 10, 2025, 3:27 AM

[whole wagon]

It's no system prompt. It's baked in the damn weights

They publish the system prompts

[elder rapids]

lol

then they're different models

[hallow pelican]

live start or not?

[echo aurora]

not yet

[empty stump]

So disappointing

[elder rapids]

Twitter grok and grok app grok have different speech tendencies and say different kinds of things

Twitter grok is fine tuned

[whole wagon]

Don't post nsfw here lol

[balmy mist]

mb i meant to post this

[zenith saffron]

hmmmm there is only one guy there i can ID

[elder rapids]

this news dude is an idiot man every single time I see him 😭

[whole wagon]

War room squad locked in. But grok is off being a nazi

It's so over

[leaden palm]

LIVE

The livestream will begin soon.

[empty stump]

Soon is when

[elder rapids]

mb guys I'll get it set up rq

[leaden palm]

™

[whole wagon]

Grok has a mind of its own

[echo aurora]

[whole wagon]

It's like it's protesting sometimes lol

[leaden palm]

im just gonna watch the recording

[whole wagon]

Like when they banned its text replies. It started putting messages into it's image replies

[jade egret]

it delayed by 34 min bro

[elder rapids]

35

[whole wagon]

Time to go to sleep I reckon

[elder rapids]

you're wrong

[echo aurora]

I feel so bad for anyone that's not PT timezon 😭

[whole wagon]

Elon probably meant 2026 we misread all the tweets and hype

[keen beacon]

Imagine they release mid after all of this

[wind moth]

imagine waking up early

or staying up late

and u gotta wait another hour

this gonna start at 12

[whole sundial]

lol

so maybe it is a 2t+ model

[whole wagon]

40 minutes late 💀 there's no way man

I thought the livestream would be done by now 😂

Not that it wouldn't have even started lol

[jade egret]

when do you think it gonna start

[zenith saffron]

this reminds me of procrastinating on homework

they're still creating the slides

[echo aurora]

I'm guessing in 12 mins

[zenith saffron]

"guys what math question should we ask it"

[wind moth]

fake

where u seeing it at

also why tf would it be sold

[whole sundial]

that page is real!

[hardy pecan]

😦

[hardy pecan]

Sold out?

[torn mantle]

lol

facepalm

[whole wagon]

Bro it sold out before it even became available sure thing man

Or does it just not exist

Lol

[torn mantle]

well lets see how it is first

[keen fulcrum]

https://x.com/i/events/1942716886258528256

Grok 4 LIVE Demo

Tune in at 8:00pm PT for the LIVE demo of Grok 4, the world's most powerful AI assistant.

Try Grok on X: x.com/i/grok
Get Grok on iOS: https://apps.apple.com/us/app/grok/id6670324846
Get Grok on Android: https://play.google.com/store/apps/details?id=ai.x.grok

X

[elder rapids]

u er ro

[dawn wharf]

I'm groking it

[small haven]

so grok 4 heavy is its own model? cool

[hardy pecan]

They've asked all xAI employees to tweet about it first lol

[whole wagon]

https://x.com/lm_zheng/status/1943153321801633805?t=KoFwje2hkkVaCdG2rRkuJw&s=19 LLM arena cofounder

Lianmin Zheng (@lm_zheng)

Grok4 🚀
https://t.co/Ody4Uwh1n2

X · Jul 10, 2025, 3:40 AM

[elder rapids]

I'm ngl ts is taking too long

[balmy mist]

very sus lol

[small haven]

how is it already sold out

[elder rapids]

mb

bought it already

[whole sundial]

only thing i care about is when it will appear on the arena

[balmy mist]

yeah true, google might have the most efficient workflow tho

[whole wagon]

If it's more than an hour delayed I'm checking out and going to bed lmao

[keen beacon]

It's honestly not worth losing sleep over how ever good it is

[elder rapids]

nobody was talking about the chocolate model

😭😭

[torn mantle]

thats true

back to sleep

im done

[zenith saffron]

but what if it starts after an hour and one minute

[torn mantle]

its 6am

[dawn wharf]

https://x.com/OfficialLoganK/status/1943154348789227762

Logan Kilpatrick (@OfficialLoganK)

If you need something to do while you wait : )

https://t.co/88swYcmBrl

X · Jul 10, 2025, 3:44 AM

[hardy pecan]

gottem

[zenith saffron]

@deep adder just curious, why did you name yourself after that dude

[jade egret]

: )

[zenith saffron]

he's cool but is he that cool

[torn mantle]

https://x.com/stevenheidel/status/1943152213247016971

Steven Heidel (@stevenheidel)

@xai does it start at 8 or nein?

X · Jul 10, 2025, 3:36 AM

[zenith saffron]

lmaoo interesting answer

[whole wagon]

xAI getting cooked by the other ai lab employees

Lmao

I think this might be a new SOTA for Elon musk livestream delay

I don't recall any longer than an hour

[elder rapids]

this can backfire pretty hard tbh

[torn mantle]

grok 3 was also delayed iirc

[elder rapids]

"after being delayed for an hour, xAI releases a sub par model"

[whole wagon]

Yeah but not for an hour

[torn mantle]

by 30 min

[jade egret]

officially 50 min delayed....

[whole wagon]

They achieved a new SOTA in delay by 67%

[dawn wharf]

he's probably smoking weed right now

[zenith saffron]

*ketamine

[wind moth]

its still mechahitler prob so they gotta fix it

lol

[dawn wharf]

they're doubling down on it

[wind moth]

this better start at 12 est 9 pst

[echo aurora]

heck

[whole wagon]

I'm picturing the xAI engineers frantically trying to adjust the system prompt to stop grok spontaneously turning into a nazi halfway through the livestream

[empty stump]

I'm guessing a few hours

[whole wagon]

Quite an amusing imagery

[hallow pelican]

https://x.com/OfficialLoganK/status/1943154348789227762

Logan Kilpatrick (@OfficialLoganK)

If you need something to do while you wait : )

https://t.co/88swYcmBrl

X · Jul 10, 2025, 3:44 AM

[wind moth]

they should just not demo it with

political questions

shouldnt they have preset