[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/pol/ - Politically Incorrect


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1755010184553.jpg (371 KB, 2400x2400)
371 KB JPG
China won the AI war
>>
>>540208986
They did a textbook nationalist pragmatism.

They decentralized ai from american centralized control.

They decomodified Americas ai models.

They're going to delegitimize the american ai models.

Its nice to see Americas financialized centrally controlled (among gays and Jews) get fucking rekt.

I'm buying champagne when open ai, anthropic, from and the other ai companies go belly up.

These faggots are playing stupid games for even stupider prizes.
>>
>>540210456
This
I hate american Ai more than the most ravinous leftist freak.
BUT I FUCKING LOVE OPEN SOURCE AI.
Unironically good job China!
>>
>>540208986
But they haven't lost the battle
>>
>>540208986
We will all be losing the AI War.
AI is the Beast.
>>
how can you pretend the west are the good guys when all our AI is extremely resource intensive and proprietary (not to mention all the major AI corps are owned and operated by jews), and china's AI is lightweight and open source?
>>
>>540208986
DeepSeek has gone downhill. It used to be decent - better than Grok but worse than ChatGPT. But now it's started translating all of my prompts into Chinese, and I think the double-translation is affecting its accuracy.
>>
>>540217033
I think it is because they moved away form model distillation to real data training and most of it's training is done in Chinese and very little in English. So when it taps into mostly new data or starts sourcing Chinese web pages, it may randomly switch to Chinese.
>>
>>540217622
That could actually be an advantage then. I usually use it to cross-check ChatGPT, so having a significantly different training data set will help keep their answers independent from one another.
>>
>>540210456
The biggest cost for AI is not developing the LLM, it's sustaining the data center and that requires a lot of computer hardware and electricity

Open source AI isn't useful to most people because we don't have the equipment necessary to even benefit from it.
>>
>>540219373
Look a little further into the future. Right now, everyone is building out fab capacity as hard as they can. At some point there is going to be enough dram to fill the channel and then some; and then consumer devices will begin getting bigger again. We should have been on track for 128gb laptops being common by now, but everything has become sidetracked. I think in a couple years that progress resumes, and the end result is that by 2033 it is super common to run local open weights models. Specialized accelerators actually designed for AI like some of that new risc-V gear will become more common and eventually people will just expect local inference.

Frontier models on supercomputers will have their place but for your day to day tasks you just don’t need it. As it is, local models on a decently big home machine right now do a very good job, albeit not as good at tool calling as commercial ones. We are finally seeing interesting progress in computing again after a long stifling period of reskinning the OS out of boredom… I can’t see how people aren’t more excited about this. We are going to give stupid people something to help them think, finally…. It might actually even result in them bettering their lives. Who knows. Amazing time to be alive.
>>
>>540208986
Deep seek is what I call fucking ur mom not only because I have to find her under the ocean at lung crushing pressures but also because her vagina is so massive I can't find the womb. How about that buddy?
>>
>>540217033
this. it's unusable
almost every single prompt now comes in chinese and i need to keep reminding the damn thing to translate over and over and over
>>
>>540208986
AI is already slop i can't imsgine how fucking shit chinkAI is
>>
>>540208986
I personally help accelerate this by having multiple openAI accounts burning tokens

Codex max using gpt 5.6 sol on max reasoning costs OAI approximately 86,000 US per month if you use 100% of the tokens on a $200 plan.

I'm using those tokens to pump out SaaS products that I use to undercut American businesses. Effectively cloning their applications and selling it at 1/2 the price with 1/50th the workforce behind it.

Fuck burgerland. Fuck your entire judeo country. You fucking kike lovers have fucked the planet and given it all away just so you can get another drop of small hat long beak cum down your throat.
>>
>>540221259
The worst part is you faggots choose this. You have the mechanisms and the culture that enable you, the people of burgerland, to seize your country from foreign interests overnight. America could be American again in a matter of hours if you decided to order66 your masters.

But you don't. And you won't. Instead you elect another wall kissing pedophile to represent you as a people.
>>
>>540217033
So I take it you use it a lot to translate Chinese? I really don't use AI but can't you clear the previous interactions so that it stops assuming you want shit translated?
>>
>>540208986

Shut up!
You make these gay post every day for years and not because your being payed by the ccp but because you have a tiny penis and you didn’t go to Harvard to compensate
Now fuck off back to plebbit
>>
>>540217033
That seems to be a recent bug with it. If you have Search button enabled, it's searching the web with Baidu (Chinese Google) and getting lots of Chinese language results, which is filling up the context window with Chinese and the model then answers your question Chinese.

If you don't need search, you can toggle it off. It's annoying to have to reply it to write its answer in English but I've never had the issue with the search turned off. Hopefully they fix it. I think they might not have noticed this since a lot of their users are Chinese. I think all that's needed is to tweak the search tool.
>>
>>540208986
DeepSeek SuckDeep
>>
>>540210456
Based
>>
>>540220131
You're never getting cheap consumer memory or compute ever again.
>>
>>540214431
they distill inferior models from our ai lol lmao
>>
>>540219373
the data center is the thing that trains the llm
the tokens it takes to generate scarjo in a bikini are already half of what they were a year ago
This is how deepseek can continue to produce the highest end model from 6 months ago at a quarter of the cost and electricity
>>
>>540223087
this. macbook is the only worth it product now
>>
>>540223087
you will, just not till the end of the decade
this kind of thing has happened before, just not in anyone born after 1965's adult living memory
>>
>>540217033
Deepseek 4 Flash just released a few days ago, this is a new model, not the old one
>>
File: kektr.gif (2.1 MB, 360x360)
2.1 MB GIF
>>540208986
Distilled chat GPT and gave it away for free as open source as a fuck you kek to them, for that reason alone I use it
>>
>>540210456
Yep, they decentralized it, took all the regulations off, open sourced everything, they used the Ancap model and it crushed America, really pathetic that America has become a centrally planned shithole, now fucking China of all places has to show them how to do capitalism.
>>
>>540223341
The US is $16tn in debt collectively over the AI boom and now China is giving it away for free. They're going to end up collapsing the US economy kek
>>
>needs 8x A100 80GB to run the 634GB VRAM REQUIRED model
costs 12 dollars an hour on a server provider
>>
>>540223467
FP8 quant is 284 GB, the FP4 quant requires 146 GB to 158 GB
>>
>>540212414
You're not even a real christcuck, so cram your LARP up your ass, fag.
>>
it's dogshit, I used it
claude and GPT far better
>>
>>540219373
All those data centres are worth dogshit junk if you don't have the training and the models, local models are the future, if China can beat America in local decentralized models then it's over for America, it saddens me what America has become, the once great nation that invented free unrestricted markets turns it's back on it's own heritage only to get whooped by another semi-communist country that adopts the principles it threw in the trash, and rightfully so, what a great shame to America and good on China for having the balls to open up and succeed to the benefit of everyone, despite all their problems.
>>
>>540223422
Yep and I'm here for it, the jewish AI scammers need to be put out of business and this is exactly the way to do it, also deepseek is insanely cost effective when you run the numbers, its not even going to be that expensive to fuck them over. If they fix the language problem everyone is evetually going to be using deepseek
>>
File: 1756827477432659.jpg (62 KB, 934x960)
62 KB JPG
>>540223647
It ranked first on the independent VulcanBench, surpassing Claude Fable 5 and its free
>>
File: HO2p4_VaMAAhFjp.jpg (51 KB, 1080x572)
51 KB JPG
>>540223749
Deepseek 4 Flash is so cheap, it doesn't even show up on the cost graph. America's frontier model Fable is 105x more expensive.

Plus when the hardware catches up, you will be able to run these models locally, so no dependency on datacentres at all. NVIDIA is releasing RTX Spark computers in the next year that will allow local running of these models. You already can run smaller models on Apple hardware right now.
>>
>>540221259
based
>>
>>540223821
>RTX Spark
you're retarded spark has max 128gb or max 64gb
>>
>>540210456
Both fight to be the lead of the next "big tech" thingy.
Chinas advantage is, that burgers have a rotten, corrupt, zombie economy. Keeping up the fake funny lines with fake money from a fake hegemony will bring the burgers to their knees. For each dollar China can spend on research/dev, burgers have to spend multiple Million dollars for HR, DEI, smearing and putting into their own pockets, before they even get an output, let alone meet the same output.
How long can burgers keep the "virtual investor capital" up and flipping it with "itself"?
How much of burger BIP is "creative bookkeeping"?
When the burger terrorarmy buys a screw for $600, does this really create "$600 of value"?
Yes I know, you are too stupid to think about such things, burger education system demands it tolls. Call it tech-dept if you'd like.
The pimp is naked, and the whores are not getting paid to tell him.
>>
File: 1781870012889877.gif (639 KB, 540x357)
639 KB GIF
What AI war? China releases everything as open source and advocates for AI alliance. Fucking american capitalist pig jews.
>>
>>540223422
All the US had to do is open source their research, but no the US patent office wants to keep getting that sweet cash and control, honestly pisses me off, they are handing the world to China, at this point I wouldn't be surprised if they've all been bribed by China to shut down US AI, the US is so unbelievably corrupt now you can't even imagine. Fucking failed nation.
>>
>>540223876
There will eventually be a 122b NVFP4 quant model once the Spark comes out
>>
>>540217033
Learn Chinese. It's the Chinese century
>>
>>540223905
I would prefer the US win, but not the Trump version of the US. If it ends up being China's century, then so be it.
>>
>>540223972
>ignoring the obvious reproducibility / replicability crisis
you really are retarded
>>
>>540223821
The API costs 0.14 per million input tokens and 0.28 per million output tokens (dollars), basically it means, that if it prints two lord of the rings trilogies worth of text/code/whatever it costs 0,30
>so no dependency on datacentres at all
This would be insanely hilarious considering the literal trillion dollar data center building projects in America which nobody wants
>You already can run smaller models on Apple hardware right now.
Already am, not on apple tho
>>
File: BA3.png (1.01 MB, 1000x500)
1.01 MB PNG
>>540221259
>>540221329
Utterly fucking glorious
>>
>>540223904
America is fighting a war against itself at this point.
>>
>>540208986
Are their dancing kitten videos better than American dancing kitten videos?
>>
>>540224066
This is the Nature index, the most reputable journal on the planet, replication crisis applies here far less if at all. And also replication crisis applies to US and Chinese research equally, so it is irrelevant, China is still winning on volume of papers accepted into reputable journals
>>
>>540224082
>This would be insanely hilarious considering the literal trillion dollar data center building projects in America which nobody wants
They're building them here too, it's going to be a fucking mess when local models reach a point where everyone has the hardware to run them
>>
>>540223972
Maybe China will become the new America and take a more Dengist approach as they improve. God knows what the future of America is, democrats will win every single election in the future and they are like rabid dogs, unless America reforms like the USSR they have a totally doomed future.
>>
>>540224239
I have a fucking huge datacenter in my small town and they are expanding it too, easily the largest building in the entire town, Europes most powerful AI supercomputer is supposed to be in the same building, nobody was asked if they want it here and what are the potential downsides
>>
>>540219373
The best models i can run locally are all chinese. And my PC is an absolute shitbox.
>>
Oh cool another open source larp thread
>>
>>540224183
https://www.science.org/content/article/misconduct-sleuth-china-swiftly-gains-acclaim-calling-out-questionable-papers
>>
>>540224419
Get a bunch of trustworthy guys in your area and set it on fire.
>>
>>540224567
AHAHAHAHA
Laugh at the wumads
>>
>>540223821
Luna is v4 flash but better in every aspect including price. Data center independence is truly the only good thing, the rest is delusion.
>>
File: HOddDurWcAEpikL.jpg (149 KB, 1100x1252)
149 KB JPG
>>540208986
>Chinks release a new update on their Deepseek V4 model
>Real ball knowers have been vibecoding with opencode + deepseek flash for peanuts running 24/7
>Sometimes you can even get a free API key from NVIDIA's page (I know if its free you are the product) and it just werks
>Shill new benchmark scores on xitter
>Jeets start using it because benchmark number is bigger than before
>Error 503: Service Queue is full

Fuck this gay world
>>
>>540224678
>China flag
Please bro make my country a Chinese province already, China can't stop winning and I want some of the glory
>>
>>540224082
the datacenters are meant to replace masses of workers, and to hunt down the billions of unwanted people. The (((rich))) think they finally have the key to get rid of the peons.
>>
File: lumiaifactory.jpg (136 KB, 1200x630)
136 KB JPG
>>540224632
Where to begin?
>>
>>540212414
>matrix multiplication is the beast
what did he mean by this?
>>
>>540221259
Kek
I tell Claude to write out some massive slop with "make no mistakes" etc. using the free web chat. I don't even read the output. Costs me nothing and cost them something.
>>
>>540223087
the memory scarcity will move new players into the space, no one is going to let SK Hynix monopolise the market for long
>>
>>540210728
>Joogle Gemini
>upload a pic of a anime character and black military side cap with a German eagle on it
>ask it to fit the cap on the character
>"Sorry, I cannot create unsafe images"
There was never going to be any "war with China", just as the kikes have made abundantly clear since October 2023 onwards. It was always going to be a controlled demolition, as evince by the rats scurrying from them golem's hide to Argentina et al..
>>
File: SHIP OOPSIE!.webm (597 KB, 626x360)
597 KB
597 KB WEBM
>>540224783
You're a fucking nong that fled to the philippines
>>
>>540224800
Checked, yeah datacenters store shit tons of information in them and obviously in modern times they have pretty much everyones online footprint stored somewhere, its cyberpunk dystopia but gayer and more jewish
>>
>>540224183
>replication crisis applies here far less if at all.
thats not how that works
the replication crisis has only been getting worse, with no solutions proposed and no end in sight
>>
>>540224836
does "make no mistakes" actually work?
>>
>>540219373
I can run local LLMs on my ten year old PC. Pretty sure businesses can run much better open models on their own servers instead of paying for tokens from anthropic
>>
>>540225560
Literally yes but not in the way you think.

By writing no mistakes you're causing the LLM to focus attention on the word mistakes, and what it could mean. By writing no mistakes, you're instructing it to validate the responses after they're generated, rather than telling it to just get the right answer the first time.

All LLMs are for all intents and purposes ambiguity reducer functions. The intent is achieved even if you've misunderstood the mechanism.

The problem is now you have an ambiguous definition of what constitutes a mistake. Nor have you provided a workflow for what to do next.

For example, one of my prompts includes:
"After creating the pr, use the MCP testing documentation founds in /docs/ to run the application, and take screenshots of each UI change, storing these screenshots in /artifacts/{release number}/. If any page causes an uncaught exception, or the line coverage is less than 100%, the PR will not be approved."


No instructions = max ambiguity
Some instruction (no mistakes) = better
Autism instructions = ideal

T. Many years software industry. I was dead serious about saying I use AI to mass produce SaaS.
>>
>>540226258
I've prices one, a GLM5.2FP8 on A100s costs about $550k AU for the hardware. I'm likely going to buy one if it tips 300k. Shove it in my basement and whip it like a little piggy for 1000tok/sec at about gpt 5.4 reasoning levels.

The problem is that is nothing compared to the giga models and those will forever cost billions, and have giant data centres. Every business went cloud for a reason, AI will be no different.
>>
File: HF1nkifasAEh_eq.jpg (75 KB, 720x647)
75 KB JPG
>>540221259
exchange "the juden" with "capitalists" and you are a fine fellah in my book.
The juden is a scapegoat by rightwingers so they dont have to admit the fuckups by the capitalistic system destroying our society everywhere. Are jews controlling and influencing shit in the world? Yes they do. But they are just a intersection. Not all of the Epstein Class are Jewish. Not all of the corrupt elites are jewish. Realize who your enemy is. Dont shoot on the wrong goal.
>>
>>540208986
Nobody in this thread knows how to vibe code or has made anything worthwhile.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.