File: 1737751869415197.png (1.18 MB, 1200x1658)

1.18 MB PNG

>>539066936

>>539066936
Benchmarks are a scam
Look at actual user numbers

The only place where Chinese AI is actually dominant
Is in AI video, where Seedance is way ahead of the rest
And this is probably because Hollywood has been filing lawsuits left right and center
And threatening and preventing American AI Video models from getting better

In other words, China only dominates when retards in the USA sabotage their own companies

>>539066936
That's a true statement though. Literally every single "Chinese" technological advancement in the entire history of the CCP is something they stole from someone else and made an inferior copy of. Literally everything. Communist societies canno lt innovate because they remove all incentive for anyone to do so.

>>539067280
>steal kike owned data
>improve it and make it open source
based
also deepseek research papers disprove your point faggot

File: file.png (174 KB, 760x646)

174 KB PNG

>>539067241
>look at actual user numbers

>>539067241
>In other words, China only dominates when retards in the USA sabotage their own companies
So only on days that end in Y?

>>539066936
All chink AI models are just rips of western models.

>>539067588
To the extent that data is available
Claude, OpenAI and Gemini are still the widely used AI models for enterprises

File: file.png (352 KB, 3555x1999)

352 KB PNG

>>539067745
how do you rip a model and end up with a better one retard?

>>539067280
Yeah its so unfair that someone can just come along and do everything you do at a fraction of the cost while you have to bankrupt yourself to stay competitive. Oh well, that's the free market for you.

>>539068372
do you know what limited liability means?

>>539068558
Do you know how much money is currently being spent on AI?

>>539068140
It's not better it's a distilled version of the major AI models. The major difference is how they operate in novel environment best examples are the ARC-AGI-2 benchmarks where Chinese models are scoring well below the American ones

Kimi had been distilling Claude for months now

>>539066936
chinese AI is just a pirated version of chatgpt

>>539068140
China's goal isn't to be better at AI, because they will always be behind, but crashing the business model of American AI companies, when people will start using "good enough" opensource instead of overly expensive Claude or other American models then the entire business model will break apart and with it the AI bubble

>>539068372
anthropic/oai are first movers. there is nothing stopping them from capitalizing on the demand for cheaper distilled models, they are optimally positioned to do so having the original models themselves. if they can't or don't want to fill that demand, then someone else will. the idea you are entitled to any sort of ownership over intellectual "property" is pure delusion.

>>539070309
I do not think that is viable in the long run
Somebody has to pay for the compute ultimately

So what are the options
Option 1
Pay for expensive hardware
Run these open source models locally in your companies
Retarded idea, especially if you do not end up using compute

Option 2
Pay for compute from providers who have done Option 1 and provide compute on demand to companies who need it
Smarter but you are still using inferior models

Option 3
Stay attached to the companies that provide the best and frontier models such as Claude, OpenAI Gemini etc
Use their APIs
Which lets you use their non frontier models which are less expensive
While still retaining access to their best newest releases as and when the occasion demands
Smartest of all the above

Which is why the strategy outlined wherein Chinese Models allegedly will crash the business model of American models, will not work

Tl;dr
Business HAVE to stay subscribed to the best AI models
Second bests and also rans are not really an option

>>539071280
For reference see this
https://developers.openai.com/api/docs/pricing

Notice the wide variations in prices depending on the model in question

>>539071280
Major customers that are paying the most of AI currently are the private businessses, many of them the largets companies with their own compute, for them switching to the opensource is a logical move to save money, because API tokens are super expensive and they were burning money like crazy recently

Average Joe Shmoe will continue using ChatGPT and other models widely available, but the Joe Shmoe is not a profitable customer, he's a money burnder for the AI lab

>>539072111
>Major customers that are paying the most of AI currently are the private businessses, many of them the largets companies with their own compute, for them switching to the opensource is a logical move to save money, because API tokens are super expensive and they were burning money like crazy recently

That is possible
As far as I know that has not happened yet on a large scale
It COULD happen. We will have to wait and see

>>539072235
contd

I am skeptical because companies typically tend to seek what they think is top of the line

Case in point
Microsoft Office 365 still has around 350 million paid commercial subscribers apparently
Many of these subscriptions are probably unnecessary and the tasks could have been managed with free open source alternatives
Yet here we are

>>539067241
wow the jeet is right, shocking

>>539070309
>>539072111
accurate

>>539072235
its already happening on a massive scale, everywhere outside of big tech that still has $$$$ to spend on overpriced models
its started to happen there too

>>539072447
the office analogy is 100% wrong
people use office because its essentially impossible to completely replicate the functionality of office apps, leading to incompatibilities
yes, the openoffice excel alternative might have 99% of the functionality, but that wont save you when u have to work with an excel sheet from literally the 90s that relies on several hacks that ONLY work in actual microsoft excel
plus, the sub is very cheap, so there's no compelling reason to go through the pain of the transition either.
adobe products are similar, there's tons of foss ebook readers, exactly none of which are a 100% match to acrobat.
windows as a whole operates similarly too. linux is objectively better, but good luck trying to run that one win98 app that your business relies on. or, you know, games, lol.

LLMs are not like that. switching costs are minimal, and any advantage in prices without a critical loss in performance is reason enough to switch

>>539074659
>people use office because its essentially impossible to completely replicate the functionality of office apps, leading to incompatibilities
okay but does that apply in each and every case though.
I doubt it

>>539072235
API costs are still heavily subsidized.
Running on local hardware at scale is also not trivial. Its not just the cost of the hardware, but you need a full team to manage it.

>>539074985
>API costs are still heavily subsidized.
Yeah I have read about this too
So far I haven't come across an authoritative source on how subsidized they are
Different articles roll out different percentages

But yeah they are subsidized

>>539075235
Contd

I also have heard that monthly subscriptions are even more subsidized
And that of they are calculated in terms of API costs, those subscribers would have been charged even more

>>539072447
AI is payment computation adjusted. Issue is, western models are 10 times heavier but produce similar results. You'd have to pay x10 the compute to gain the same results, plus the cost of actually training the model is x100. American AI bubble is supported by trillions of fiat money being printed like there is no tomorrow, this is not grounded in reality, this will never be solved, because there are no real economy that can reimburse such investments. Chinese AIs are sustainable from the get go and they are open-source, you won't magically find your setup not working because they did some update to their policies or mounted another model altogether. If Chinese open-source is similar to western proprietary, you have no usage for the later.

>>539075522
As I was saying earlier, this is theoretically possible
But it has not yet happened in the real world
People and companies are still using OpenAI, Gemini, Claude etc.
Chinese models as of now are still not that popular compare to the above, outside of China

Of course this can change. Only time will tell

>>539074791
The biggest problems with LLMs is that you can train models on prompts and output from other models. If Claude is better at ChatGPT at coding specific software, if OpenAI has access to Claude they can get the output and train ChatGPT on that. Its the same situation with the Chinese distilled models. So any model being able to do anything means that other models will soon be able to do it as well. To some degree companies aren't even really competing or racing, they're in a weird collaboration. its like if there weren't any patents for cars and people were just inventing new things that are copied by their competitors whose designs are copied by everyone else. The only time anyone has an edge over anyone else is in the short time window before someone else has copied anyone else and it results in an incredibly fast race towards generalization.

>>539075909
I dont disagree with you.
Again, it is just that there is not evidence YET
That all of these dynamics have combined to break the business models of Western AI companies such as Open AI or Claude or Gemini

It may be the case that they are all actually bleeding money and may implode eventually
But so far it has not happened

>>539075781
They are not popular yet. Same as everything Chinese. They first copy, then start selling you for for the fraction of the cost, then their products become better due to scale advantage and low costs. It's like with everything Chinese. They are the only place where real economy is concentrated and high skilled workforce is both present and is adequately priced. Western AI get ahead ONLY because US is killing their economy and stealing their people wealth to funnel into that multi-trillion blackhole. China doesn't commit financial suicide for their models to be good, they do it the normal way, via optimisation and open-source community support.

>>539067241
Chinese AIs are the best for degenerate roleplay which is the only valid and value generating use for LLMs.

>>539075522
>western models are 10 times heavier
gross overstatement, but yes, they are heavier
most frontier models are in the 1-1.5 trillion params range. that's a 50% swing, not a 10x

>plus the cost of actually training the model is x100
again, grossly overstated, but yes, they are more expensive

>Chinese AIs are sustainable from the get go
nope. they too are subsidized, albeit to a lesser extent

>and they are open-source, you won't magically find your setup not working because they did some update to their policies or mounted another model altogether
correct.
and perhaps most important of all, if you self host, you can be 100% certain that you are running the full fat model properly, not some 8/4/2bit quant because the provider is seeing too much traffic right now and they'd rather serve a model that is 90% as good to all their customers instead of rate limiting and losing on api income
and if you ever do try to run a full fat 1T model in prod yourself, you will quickly discover just how subsidized they are

china has 2 big advantages over the us
1) no PC/woke bullshit. an inordinate amount of effort, training, params, etc goes into making sure western models remain oblivious to and/or silent about certain facts. and towards the systems that are layered on top of them, to make sure they are subsequently censored in the very frequent event that the baked in wokeness training doesn't prevent them from reaching certain obvious conclusions or repeating certain easily googlable facts, etc etc.
2) trumps gpu ban, albeit far from 100% effective, did force chink AI labs to try to make do with less. they absolutely could get their hands on the primo nvidia shit, but even so, prices were higher, and the incentive to be lean and mean was way higher. plus, see above, they didn't have to worry about their models being oblivious to and/or silent about quite as much as western models.

>>539066936
Didn't they just recently uncover that those benchmarks are mostly fake because the AIs game the systems?

File: 1487809930836.jpg (45 KB, 462x505)

45 KB JPG

>>539066936
The real question is if Kimi K3 will jannie out and have a meltdown if you dare use it as a biologist.

>>539076212
>most frontier models are in the 1-1.5 trillion params range
That's the issue, you are living in chatGPT 4 era. SoA western models are 5t, fable is rumored to be up to 10t. Do you understand why they require so much investment and infrastructure. Specifically because their strategy was always to bloat the shit out of instead of what Chinese did. Chinese models of 1t are of similar performance as wester 5t ones. This is the core issue.
Only Musk's grok wasn't bloated to heaven and hell yet, but that model isn't that smart.

File: 13535135.jpg (99 KB, 579x949)

99 KB JPG

>>539066936
That's old news.

Four hours ago they just released a new model that shot up to number 1 ranking

>>539076692
Which is why they have started offering Lite and ultra Lite versions
Have a look at this
https://developers.openai.com/api/docs/pricing

Their latest 5.6 model comes in three versions - luna terra and sol with different pricing

File: HKQMxV4akAAmoar.jpg (135 KB, 1298x1184)

135 KB JPG

>>539076348
Forget benchmarks, the free market has already decided

>>539077087
Open Router works like an aggregator for AI APis I think.
However, I think it only shows part of the market
And I am not sure what its marketshare is

Chinese Axis are the most used AIs in the West already moron
Go take a look at token used

>>539070309
Cope

>>539076692
>SoA western models are 5t, fable is rumored to be up to 10t.
American devs are bloatmaxxing to keep the data center musical chairs going. The only ones that actually make breakthroughs in optimisations outside of China are Google.

>>539077266
Contd

Got the Answer via Google Ai
OpenRouter serves over 5 million developers and end-users globally.
So it covers only a fraction

>>539076692
>you are living in chatGPT 4 era
more like kimi 2.5 era, actually, but yea, you're right, they've grown a bit since
kimi 3 is reportedly in the 2-3trillion range, we'll know when they release the weights
as for western models, they don't report those numbers anymore

>SoA western models are 5t, fable is rumored to be up to 10t
i doubt it. 5t maybe, 10 seems absolutely absurd

>Do you understand why they require so much investment and infrastructure
they don't, its a ponzi scheme
there's a big ass circular trade between nvidia, the AI labs, the big tech companies and datacenter builders
everyone is borrowing money and shoving it to each other as fast as they can, while the bubble still holds
the only company making actual profit is nvidia, ofc

>Chinese models of 1t are of similar performance as wester 5t ones
again, we don't actually know how big western models are anymore. last reported numbers were ~1.5t, but that is outdated
and no, chinese models are not quite there yet, but very close indeed. kimi is an outlier, mind you. most others are a bit further behind. still very good, still great value, but behind.
in particular, while im quite happy with chink models when it comes to coding help, im also very disappointed in their tendency to hallucinate heavily. the free chatgpt you get simply by visiting the website is miles more reliable than glm5.1, which will happily make up shit all day long.

>grok
that part is weird to me
i mean, obviously he has the money to make it a top tier one, if he wants to
i can only assume there's some strategy to run as lean as possible in a bid to reach profitability asap
mayhaps he's the only person smart enough to realize that chasing "replace all knowledge workers" is a fools errand, and he's trying to settle for "good enough to run your tesla and replace google", which is imo a very good strategy.

>>539077356
Only true for Openrouter which is a fraction of the market. Almost everyone uses the OpenAI/Claude/Xitter websites or apps because they're normgroids and first movement advantage matters.

american jews care more about lining their pockets than delivering the best cheapest product

if china leads to cheaper more accessible services then i welcome the competition. scam altman and that claude kike can go to hell

File: 175729835262.jpg (1.34 MB, 1824x2166)

1.34 MB JPG

>>539077087
Former OpenAI's CTO prefers Chinese AI models

>>539077452
It's used by developers, enterprises, comanies and starts up. They are the ones who integrate AI into software and businesses that end up being used by the masses. It's like AWS but for LLM, so it's a big fucking deal when their foundational models are skewing towards China's.

>>539077486
>first movement advantage matters.
its "first mover advantage"
and also no, it doesn't, not in this area
first mover advantage is huge in social media that relies on network effects
eg facebook
everyone is on facebook because everyone else is on facebook. that's its essential value, that everyone is on there
and everyone else is on facebook because it was the first one to break through and become mainstream
and if some1 else makes a better version of facebook, it wont matter, because the value of facebook is that everyone is on there, not that its website is a bit faster than competitors (it isn't) or that it has a better privacy policy (ahhahahahahhahaha) or fewer ads or whatever.
youtube is similar. content creators are all on youtube because all the viewers are there. and all the viewers are there because all the content creators are there.

notice how neither one of those apps ever lost users to a clone? even FOSS clones without any censorship of privacy violations?
however, both of them lost viewers to different apps that supported different use cases
in the case of facebook, snap and insta
in the case of youtube, twitch and tiktok

anyway, LLMs are NOT like this. an LLM is just a machine that answers questions (well, more accurately, fills in text, but whatever)
it doesn't matter which machine you were using yesterday, or which one everyone else uses
the switching cost is $0. all you have to do is change the API url and key, and tadah, now you're using some other LLM

>>539077897
Probably but they still represent only a small proportion of the market
Chinese models can only declare victory once they have captured the broader market

Until then, the Western Models carry on without any threat to their business model

>>539067179
>weebs are watching donghua
They aren’t. They aren’t watching gookshit either.

>>539066936
>western
you mean kiked. there are no western models, only jews trying to set up an automated panopticon surveillance state. its why they want data centers in space. angry mobs cant burn them down.

>>539077087
This isn’t an indicator of the “open market” as it does not factor in enterprise, just individual users (who often aren’t paying)

>>539067241
>when retards in the USA sabotage their own companies by importing endless jeets to run them into the ground

>>539078021
openrouter represents the market of users who actually have to pay fair(ish) prices for their use, aka API users
everyone else is on a "plan" (read: hugely subsidized costs) and/or has unlimited token budget because the big tech company they work for has infinity money to burn. for now.

the current western model is NOT a long term viable model
nobody makes money by offering a $5 service for $1, which is what chatgpt/anthropic/etc are currently doing
you can get a fuckton of users and keep them, but then what?

see >>539077960 for the discussion of first mover advantage, and how it doesn't exist here
this is why ai companies are fucked. it doesn't matter how much they subsidize this shit now, there's no way to lock in customers to LLMs
the literal microsecond that chatgpt and anthropic and google start charging proper prices for their AIs, they will lose all their customers to chink models that offer 95% of the perf for 20% of the cost

File: dumbjak.jpg (67 KB, 919x1024)

67 KB JPG

>>539078256

>>539078354
>the current western model is NOT a long term viable model
Maybe
But as I stated earlier, these theories have not yet materialized into reality
The Western Models are still chugging along merrily

>>539078354
There is a way, if you model is uniquely proficient in doing something the other can't. But since they all are pretty locked in on coding and agentic stuff there is no room for niches, so only price/value matters in the end.

>>539078256
yea, this is going to become a HUGE problem eventually
most of the big tech companies rely on some sort of incredibly effective lock in effect
eg microsoft with users forced to run windows and/or office for compatibility with older apps/docs reasons
or the big social media companies with network effects
companies like that can survive a long time while still offering an inferior product, but not forever

and eventually the jeetification of the tech workforce will result not just in inferior products, and not just in hugely higher operational costs, but plain non-functioning products, outages, etc

companies have tried the jeet thing in the past, with outsourcing, and it was 100% a failure
simply changing the location of the jeets to the us or europe will not change anything, it will still be 100% a failure. sooner too, because now the costs are higher as well.

>>539078647
>There is a way, if you model is uniquely proficient in doing something the other can't
nothing like that exists in the realm of LLMs
they are programs that fill in text, nothing more
none of them do anything unique beyond that
the closest we ever got was when deepseek were the first to do "reasoning" with r1, and even that was trivial to replicate, we had reasoning distills of it within a couple days

most of the competition in the space is mostly about chasing efficiencies. in training, in inference, in quantization, in sparseness
the rest of it is all about data sources, real or synthetic. and china has a huge advantage here, because they have no reason to fear breaking western copyright laws to obtain as much data as they want

>>539066936
>Americans spend trillions of dollars, while Chinks then copy them and develop models almost just as good for a fraction of the cost
Yes, which is why the notion that China would want to stop US data centers is completely ridiculous. It's like stopping someone from doing homework that you then copy. I'm sure China is absolutely fine letting Americans spend enormous amounts of capital so they don't have to.

>>539078529
>Maybe
there's no maybe about it
offering a $5 service for $1 is not viable, period
you can do for a while, for as long as investors are happy to cover your operating losses
but eventually, you have to turn a profit and show those investors you can pay all their money back, and extra profit on top

>>539066936
Gemma 4 31B is a fucking marvel.
You can run it on a mid-upper mid gaming PC. The uncensored version is really good at roleplaying too.

>>539079007
I prefer to wait until these 'enlightened' predictions become actual reality
Until then I choose to be skeptical of such theories

>>539078866
>nothing like that exists in the realm of LLMs
That is bad faith simplification. LLMs are what their datasets are. They are pattern recognition engines that do what they have 'read' the most.
Feed them obscure porn, they'd be the best porn companion. Feed them social interaction, they'd be best partner. Focus on multi-modality and make them proficient in text encoding media generation models, they'll become artists and composers. Of course is you feed them JSONs of tool calling and github repositories and then severely penalize anything else, they'd be that one same thing everybody aims at creating.

File: AI22-26.png (83 KB, 1111x668)

83 KB PNG

>>539066936
>includes a picture that disproves neither idea
>>539076692
>fable is rumored to be up to 10t
retard
>their strategy was always to bloat
The strategy has always been innovation