• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Chinese A.I. Models Close the Gap With Anthropic and OpenAI

P

philb2

2[H]4U
Joined
May 26, 2021
Messages
3,471
https://www.nytimes.com/2026/06/25/...telligence-models.html?smid=nytcore-ios-share

Z.ai is on the cutting edge of a wave of powerful but inexpensive A.I. from China that is challenging the lock that OpenAI, Anthropic and Google have had on the industry. Six of the models now on the A.I. leaderboard were developed in China.

Z.ai’s new model, GLM-5.2, arrived just as U.S. businesses realized that they had to find ways to cut down on how much they were spending on A.I. It also landed when executives in Silicon Valley were becoming worried that the Trump administration was leaning toward regulating the technology.

Some software developers are reluctant to use the A.I. system that Z.ai offers from computers in China, because they worry about sharing data with the company or with the Chinese government. They are also wary of China’s efforts to censor its A.I. systems or running afoul of U.S. export restrictions.
 
Chinese AI models have surpassed US AI models in global call volume. I recently saw a comparison image showing usage by model in 2024, when China had minimal presence, vs now in 2026, when China is leading. I suspect that growth trend for China is far from over.

The founder of Z.ai replies to Musk's prediction that China will reach the level of Anthropic's most advanced model, Claude Fable 5, by Q1 2027, saying it won't take that long. The US govt restricting access to its top models won't do much good and will only hurt US AI companies if China is making even better ones available to everyone.

1782629551888.png
 
apparently trained on Ascend 910 series type of hardware (Huawei accelerators made in SMIC fabs), which if true would be the big thing here (and showing out of place banning H200 to be sold in China was)

Close in part the gap is of course true, but can be overstated here, about opus 4.7 medium on:
https://artificialanalysis.ai/agents/coding-agents?coding-agents-performance-chart=index

while taking a bit more than fable 5 max time and costing more than GPT 5.5. xHIgh because of how long and large the token use is, i.e. not much cheaper for many things, that only if you look at per tokens cost not real world task cost.
 
GotNoRice said:
What makes you think that whatever bandwagon "AI" bullshit DuckDuckGo uses has anything to do with leading-edge AI models?
Click to expand...
Because they utilize a number of AI models from OpenAI, Anthropic, and Mistral AI. They tell you right on their webpage on duck.ai. They act as a gateway to the other AI's and supposedly keeping it anonymous. If it's shit then that's because Anthropic and OpenAI is also shit.
 
The one thing I have to give Chinese models credit for is releasing them open source and ideally open weights/training data. Like DeepSeek's models before, Z's GLM-5.2 appears (from the article) to be FOSS as well and it takes a little poking around to find the github/huggingface where the Z related open source projects are available. It seems there's both a BF16 and FP8 precision model, and are at the same size of "744B-A40B" . Off the top of my head I'm unsure what degree of hardware it would take to effectively run something like this locally, but for the hell of it asking DuckAI (which pulled from GPT5.4 for this) seemed to suggest that it requires an absolute fuckton of hardware (8x H200 class 80GB GPUs) , and that even a heavily quantized 2-bit variant would need like 250GB unified RAM (which I assume it counts as VRAM + system RAM), so yeah..not likely to be plopping that on my new home server. Still, I'm glad that something like this is out there for those that can run it locally. Of course, thanks to quantizing and distillation as well as increasingly performant models there ARE variants of sufficient capability that can be locally hosted on common hardware. Still, AI certainly makes one feel like having even a 5090 + 64GB RAM system feel low end, but I digress...

The one issue people always cite with Chinese models is the concern over censorship, and I don';t know much about Z but I remember DeepSeek R1 's officially hosted instances did shy away from certain steretoypically "sensitive to the Chinese gov't" issues. Then again, US based models are often "censored" or limited and while concern for getting the current administration's disfavor is a recent issue, the longer standing motivation is more about marketing, avoidance of liability/bad PR, and avoiding easy justifications for regulation (as with the recent suicides and subsequent lawsuits). However, when it comes to the Chinese models (at least DeepSeek when I checked) self hosted versions did not have the censorship hard-coded into them as it was instead something added to the governance of the model's output after the fact - a good thing. You could ask DeepSeek's various local models about Tiananmen Square or Taiwan's independence without it refusing to discuss or spouting the ideological approved view unlike iif you asked the officially hosted variety. If Z is at least like this, then somuch the better.

The question over AI development "arms race" between the US and China is not going to be helped if the vast majority of high performance models don't just require expensive hardware, but are basically framed as a proprietary, software-as-a-service subscriptions to access. With a few exceptions (I rarely like to give Meta credit given so much they've done wrong, but some of their FOSS models and related software like the Llama series is greatly appreciated) , the open source offerings from Western companies are much lower performance compared to what is ofifically hosted and proprietary. There's also the issues about copyright and training models in the US, surrepticiously or otherwise ; China is unlikely to have to worry about this the same way, but at least is up front about it. I think we'd be in a much better situation in all regards - including those concerned about power and other resource use - if companies were all collaborating on improving FOSS models, but sadly we're seeing the handful of big tech all striving in hopes to be rulers of the new digital fiefdom.

Ultimately, if China (or anyone else for that matter - I hope those in Japan, Korea, and Europe will also be able to contribute) release top-graded performance models under FOSS licenses, for all manner of usage (agentic/chat/LLM, image/video creation, audio/voice creation etc), that are not encumbered by sanitized / censored restrictions, we'll all at least have a chance of the benefits of AI to be more widely distributed and accessible, rather than it being in the hands of a few who dole it out drip by drip to the many at increasing cost - especially once its use and interest becomes apparant.
 
And on those exact topics:

Coinbase's CEO details that switching to self-hosted Chinese AI models like GLM 5.2 and Kimi 2.7 has cut Coinbase's AI spending in half while its token usage continues to grow.

Anthropic's CEO cries about open-source AI that it can't control, and basically pleads the US govt to ban its competitors. I expect US intel agencies ask people like him to make such choreographed arguments during hearings, so they can use them as pretext to create regulations for narrative and capability control. I don't think it will work, though, except in commercial usage. Everyone else will just download the best open-source AI from wherever it's available and use it, anyway.
 
RanceJustice said:
The one thing I have to give Chinese models credit for is releasing them open source and ideally open weights/training data. Like DeepSeek's models before, Z's GLM-5.2 appears (from the article) to be FOSS as well and it takes a little poking around to find the github/huggingface where the Z related open source projects are available. It seems there's both a BF16 and FP8 precision model, and are at the same size of "744B-A40B" . Off the top of my head I'm unsure what degree of hardware it would take to effectively run something like this locally,
Click to expand...
Basically take the number of parameters and then if it is BF16, 2 bytes per and if it is FP8/Q8 1 byte per. That's the minimum just to load all the parameters in RAM. It'll actually take more for the conversation cache and so on but it gives you a quick idea.

RanceJustice said:
The one issue people always cite with Chinese models is the concern over censorship, and I don';t know much about Z but I remember DeepSeek R1 's officially hosted instances did shy away from certain steretoypically "sensitive to the Chinese gov't" issues.
Click to expand...
I don't know about GLM but when I asked the local version of Qwen (which is Alibaba's model) about Tiananmen square massacre it answers. You might consider the answer a little "censored" in that it doesn't dump out a lot of text and it says that the Chinese government's position was that it was necessary to maintain order, but it doesn't deny the event or deflect. Funny enough the uncensored (abliterated) version actually spouts propaganda whereas the regular releases just talks about it in a fairly neutral (if abbreviated) historical context.

While it is a real issue you can face with LLMs, it really isn't that much different than anything else where you need to check their output and be careful what you use them for. Never mind deliberate censorship, they are prone to straight out make up shit sometimes. They are always the kind of thing to be checked on and used under expert supervision, if the job matters. Training them not to make up shit is really hard, since they are basically a probabilistic "make shit up" engine.

It is also the kind of thing that can be modified in open weights models. There are ways to basically train smaller subsets and load those with the model to change behavior. So it would be perfectly feasible to get one of these, hammer on it to find censorship you aren't ok with, run training to produce a LORA on more modest hardware, and use that.

RanceJustice said:
Ultimately, if China (or anyone else for that matter - I hope those in Japan, Korea, and Europe will also be able to contribute) release top-graded performance models under FOSS licenses, for all manner of usage (agentic/chat/LLM, image/video creation, audio/voice creation etc), that are not encumbered by sanitized / censored restrictions, we'll all at least have a chance of the benefits of AI to be more widely distributed and accessible, rather than it being in the hands of a few who dole it out drip by drip to the many at increasing cost - especially once its use and interest becomes apparant.
Click to expand...
At the moment only the US and China really seem to be playing the LLM game. We'll see if that changes. China seems to largely be going the open source route (though some are hybrid, small versions of Qwen can be downloaded, the big ones are only on Alibaba), the US largely going proprietary.
 
RanceJustice said:
The one thing I have to give Chinese models credit for is releasing them open source and ideally open weights/training data. Like DeepSeek's models before, Z's GLM-5.2 appears (from the article) to be FOSS as well and it takes a little poking around to find the github/huggingface where the Z related open source projects are available. It seems there's both a BF16 and FP8 precision model, and are at the same size of "744B-A40B" . Off the top of my head I'm unsure what degree of hardware it would take to effectively run something like this locally, but for the hell of it asking DuckAI (which pulled from GPT5.4 for this) seemed to suggest that it requires an absolute fuckton of hardware (8x H200 class 80GB GPUs) , and that even a heavily quantized 2-bit variant would need like 250GB unified RAM (which I assume it counts as VRAM + system RAM), so yeah..not likely to be plopping that on my new home server. Still, I'm glad that something like this is out there for those that can run it locally. Of course, thanks to quantizing and distillation as well as increasingly performant models there ARE variants of sufficient capability that can be locally hosted on common hardware. Still, AI certainly makes one feel like having even a 5090 + 64GB RAM system feel low end, but I digress...

The one issue people always cite with Chinese models is the concern over censorship, and I don';t know much about Z but I remember DeepSeek R1 's officially hosted instances did shy away from certain steretoypically "sensitive to the Chinese gov't" issues. Then again, US based models are often "censored" or limited and while concern for getting the current administration's disfavor is a recent issue, the longer standing motivation is more about marketing, avoidance of liability/bad PR, and avoiding easy justifications for regulation (as with the recent suicides and subsequent lawsuits). However, when it comes to the Chinese models (at least DeepSeek when I checked) self hosted versions did not have the censorship hard-coded into them as it was instead something added to the governance of the model's output after the fact - a good thing. You could ask DeepSeek's various local models about Tiananmen Square or Taiwan's independence without it refusing to discuss or spouting the ideological approved view unlike iif you asked the officially hosted variety. If Z is at least like this, then somuch the better.

The question over AI development "arms race" between the US and China is not going to be helped if the vast majority of high performance models don't just require expensive hardware, but are basically framed as a proprietary, software-as-a-service subscriptions to access. With a few exceptions (I rarely like to give Meta credit given so much they've done wrong, but some of their FOSS models and related software like the Llama series is greatly appreciated) , the open source offerings from Western companies are much lower performance compared to what is ofifically hosted and proprietary. There's also the issues about copyright and training models in the US, surrepticiously or otherwise ; China is unlikely to have to worry about this the same way, but at least is up front about it. I think we'd be in a much better situation in all regards - including those concerned about power and other resource use - if companies were all collaborating on improving FOSS models, but sadly we're seeing the handful of big tech all striving in hopes to be rulers of the new digital fiefdom.

Ultimately, if China (or anyone else for that matter - I hope those in Japan, Korea, and Europe will also be able to contribute) release top-graded performance models under FOSS licenses, for all manner of usage (agentic/chat/LLM, image/video creation, audio/voice creation etc), that are not encumbered by sanitized / censored restrictions, we'll all at least have a chance of the benefits of AI to be more widely distributed and accessible, rather than it being in the hands of a few who dole it out drip by drip to the many at increasing cost - especially once its use and interest becomes apparant.
Click to expand...
Yeah I'm basically running full Qwen right now. 122B on my DGX Spark, and 27B on my 5090. In theory I can also run a heavily quantized Deepseek V4 on my Spark as well (or get 2 Sparks and run it much better, but that's another 3.5k investment minimum). These Chinese models are basically what most people are using out in the wild these days. Otherwise from frontier models, all I really use is ChatGPT, because I can access it for free. I'm not paying a sub for this shit. Take your huge datacenters and go shove them up your ass. I hope the whole huge datacenter industry at large implodes and we're left with local inference instead, which will then be optimized further.

That said it's important to not mistake what China is doing right now for their long term goal. Anything that involves this much money and research can never really be taken at face value. I don't think this free and open source landscape from them is going to remain like this. I think the thing is that they have nothing to lose by releasing it and making it free. That's the idea. Then they can leverage community efforts for free, and it's not like their models are more advanced, so they only have to gain from this. They're also just straight up competing on cost because they can. Labor over there is cheaper, industry standards are looser, etc. They're not winning on cost because of "good" reasons. They're just winning because their culture is technically more dystopian than anything we can even fathom.

I think we can thank the rot of late stage capitalism for the position we're in now.

Sycraft said:
Basically take the number of parameters and then if it is BF16, 2 bytes per and if it is FP8/Q8 1 byte per. That's the minimum just to load all the parameters in RAM. It'll actually take more for the conversation cache and so on but it gives you a quick idea.
Click to expand...
Yeah but heavily quantized models don't actually lose that much fidelity, and you can save a lot more ram by further quantizing the kv cache and whatnot. Realistically to run even some higher end models locally... it's possible with about 128 to 256GB I guess. I have a DGX spark that I got for about 3k, and I'm somewhat idly contemplating getting another because it's realistically all I would really need for probably the rest of my local LLM days. But honestly even 1 of them can run Qwen 122B at pretty good speeds (~45-60 tok/s), and this model is still quite intelligent. The next step up would be running Deepseek V4 on 2 sparks, but that's ~3.5k more for that privilege. Hard sell.


Realistically what people have to realize is that for most things you do NOT need a frontier model, frankly. Even 27B, which is something that can run really fast on a 5090, is absolutely plenty to help people vibe code surprisingly large things.
 
Templar_X said:
Could I trouble you for what tok/s you're getting and which inference engine you're using on those?
Click to expand...

I mean, I'm still pretty much a layman myself, but sure. On the DGX Spark, I think this is the current easiest high performance "all in one package":

https://github.com/Entrpi/qwen3.5-122B-A10B-on-spark

It's probably one of the more painless installs I have had. It will take a while to finish, but overall sets everything up. Total throughput varies a lot based on content. I haven't actually tried it for coding tasks as much, but for my document processing it has been anywhere between 20 to 100 tok/s, and on average after a very long runtime it was at about 49-50 tok/s. It loads up fairly quickly, and uses the most modern VLLM version. Note a lot of stuff on DGX Spark is VLLM based, and sometimes it's a bitch to get running. You'll be hitting up random ass repos out in the middle of nowhere if you want the performance, and then it's a hellscape of broken dependencies if some dipshit decides to update some package that's a dependency and/or suddenly just take a version that you needed for your special snowflake install out.

https://github.com/albond/DGX_Spark_Qwen3.5-122B-A10B-AR-INT4

This one is much more stable (40-50 tok/s) but uses an older VLLM package that needs a bit of tinkering to get up and running. On average it's slower, and also much slower to load, so I don't know if I would bother. The creator doesn't seem to be actively maintaining it.

You can also try some others. Better get used to browsing the Spark forums: https://forums.developer.nvidia.com...ng/dgx-spark-gb10/dgx-spark-gb10-projects/723

This 122B model provides both LLM and image processing (image descriptions recognition, etc, including reading text off of an image). You can have it translate fairly long documents (I've input 30 images into it before to translate an entire comic chapter from Japanese, and it did it, though it won't actually edit the images), help you code, hook it up to an agentic flow, whatever you want. It's pretty smart, but it does loop sometimes (ie fall into toxic token repeat patterns, so literally an endless loop, usually while reasoning, sometimes during output), especially in very high context tasks depending on your temperature and other settings. Everyone is waiting with baited breath for a 3.6/3.7 variant of this. Widely considered one of the "sweet spots" for a single spark with 128GB ram. Other ones include 35 A3B, highly quantized Deepseek, Minimax (think single spark can do 2.5 or 2.7, idr), etc. The nice thing about this platform is that it's STILL being optimized. As good as it is now, you can bet your butt people will have even better recipes later, and better models will release for it. Full cuda support, so you have exactly what would run on a much larger and more expensive Nvidia stack (just obviously much slower).


For the 5090, I just use llama.cpp, pretty much because it's one of the first things I came across. I don't even know if it's optimal. I use my 5090 for mostly Stable Diffusion and Lora training lately since my DGX Spark (well, ASUS Ascent) does all of the LLM shit, so I pretty much just used the first config I tuned. It's not optimal for speed or anything, it just works:

This was the model:
https://huggingface.co/RDson/Qwen3.6-27B-MTP-Q4_K_M-GGUF

The runtime was llama.cpp.

These are my start params:

Bash: 
#!/bin/bash
# MAX-PERFORMANCE: Qwen3.6-27B-MTP on RTX 5090 (31.4GB VRAM budget)

MODEL="/models/Qwen3.6-27B-MTP-Q4_K_M.gguf"

docker run --rm --gpus all \
  --shm-size=28g \
  --ulimit memlock=-1:-1 \
  -v /opt/app/models:/models:ro \
  -v ~/llama/cache:/cache:rw \
  -p 8080:8080 \
  ghcr.io/ggml-org/llama.cpp:server-cuda13 \
  --model "$MODEL" \
  \
  --n-gpu-layers 999 \
  --no-mmap \
  \
  --ctx-size 196608 \
  \
  --batch-size 16384 \
  --ubatch-size 2048 \
  \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  \
  --flash-attn on \
  \
  --threads 16 \
  --threads-batch 16 \
  \
  --parallel 1 \
  \
  --spec-type draft-mtp \
  --spec-draft-n-max 1 \
  --spec-draft-p-min 0.60 \
  \
  --temp 0.6 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.05 \
  --repeat-penalty 1.05 \
  --presence-penalty 0.05 \
  \
  --host 0.0.0.0 \
  --port 8080 \
  --jinja \
  --chat-template-kwargs '{"preserve_thinking":true}' \
  --cache-reuse 8192 \
  --rope-freq-base 10000000 \
  \
  --log-file /cache/server.log \
  --log-prefix

This config is optimized for high context size (that 196k is real). The speed is about 60-80 tok/s in my experience, depending on data and prompt size. You have about 1 gig of leeway maximum on the GPU with this loaded so either run a headless Linux install (just SSH in) or a lightweight distro like Mint XFCE, or just use a CPU with an iGPU and use it to drive the desktop.

For llama.cpp, the built in user interface is just fine in my experience. For vLLM, you have no user interface lol. I suggest "Cherry Studio". But lock it down via firewall so it can't phone home (me just being paranoid).

If you want more intelligence and/or speed at the expense of context size, I'm sure there are options out there. The 5090 is far from incapable. The 27B model won't wow you right away but any mistakes it makes are very quick to correct since it outputs very fast. It's perfectly cromulent.
 
Last edited:
Thank you very much for all the info, it's appreciated.
Also a layman here, so every bit helps.

StoleMyOwnCar said:
For the 5090, I just use llama.cpp, pretty much because it's one of the first things I came across. I don't even know if it's optimal. I use my 5090 for mostly Stable Diffusion and Lora training lately since my DGX Spark (well, ASUS Ascent) does all of the LLM shit, so I pretty much just used the first config I tuned. It's not optimal for speed or anything, it just works:

This was the model:
https://huggingface.co/RDson/Qwen3.6-27B-MTP-Q4_K_M-GGUF

The runtime was llama.cpp.
Click to expand...

I can share my experience with a 5070 ti and the same Qwen 3.6 27B but with a more aggressive quantization that still retains good accuracy - Qwen3.6-27B-3.30bpw-exl3 with Q4 for the KV Cache.
I started with llama.cpp but then changed to ExLlamaV3 to be able to use better quantization and get way better throughput (better kernels). My tok/s almost doubled, sometimes hitting just shy of 70 t/s, I would expect you to get around double that on the 5090.
If you do decide to try it, I would suggest going with TabbyAPI which was incredibly easier than configuring ExllamaV3 by hand (damn dependency version missmatch hellscape)

Anyhows, thanks for sharing.
 
RanceJustice said:
The question over AI development "arms race" between the US and China is not going to be helped if the vast majority of high performance models don't just require expensive hardware, but are basically framed as a proprietary, software-as-a-service subscriptions to access. With a few exceptions (I rarely like to give Meta credit given so much they've done wrong, but some of their FOSS models and related software like the Llama series is greatly appreciated)
Click to expand...
Nvidia is actually all open source and data, meta (like many chinesse one, like Qwen, Deepseek, Kimi, GLM ) are just open weights I think, the MIT license used for those weights is the most open licenes if you use some too.

I am not sure if framed as proprietary would hurt the arms race, specially if you can make them large revenues that put back large ressource into making them, open need some people ready to burn giant money on it like Musk did back in the day on OpenAI or an Nvidia that want to sell the hardware to run them, which what would happen the second Chinese model made to run on chinese ASIC get over the proprietary American one, Nvidia would release much better version of their open one they make (now they do not want to compete too much with their clients)

One issue to make chinesse fully open source, they almost certainly train on gpt-claude-gemini, etc... so they can"t
 
"American Closed-Source AI company is doing everything that they accused of Chinese Open Source AI is doing."
Anthropic has embedded hidden spyware-like code in Claude Code that covertly targets Chinese users. It then sends information regarding every user by injecting it into their prompt message. Claude Code is sending info like timezone, proxy and possible AI Lab connections into the system prompt in ways Chinese users can't notice. A coding agent with repo and command permissions should not silently hide routing metadata inside prompts. This is a serious breach of user trust.
Click to expand...


1782879672243.png
1782879710131.png


Anthropic:
1782879509419.png


1782880008958.png
 
Been playing with Ornith1 in BF16 and Q8 for agentic work with Hermes on my Ryzen 395/128GB today. The 8Q seems to be a nice sweet spot. Seems to be very fast and unlike any LLM I have used before in how it verifies its work and moves forward on its own given the right prompts.
 
Delicieuxz said:
"American Closed-Source AI company is doing everything that they accused of Chinese Open Source AI is doing."



View attachment 812417 View attachment 812419

Anthropic:
View attachment 812416

View attachment 812421
Click to expand...
No, it's not. That's not what distillation is. I see no problem trying to catch the perps.
China is doing what they do best - stealing IP.
Dumping their knockoffs on the world to steal marketshare and crush the US economy.
The world needs to find a way to protect against this. I do think Anthropic and Claude making their best models closed source and limited release will help.
China claims to be running on lighter hardware ...sure they probably are for inference or fine tuning after distilling the US models. China was caught smuggling Blackwells through Supermicro (Taiwanese company) - I'm sure they've been buying them on the secondary market for some time.

That said, I will run the best open weight models locally. If they are knockoffs made by China, so be it. Qwen 3.6 27B is decent with Hermes. For reasoning, it's still a far cry from Claude. I also have a Claude subscription for general stuff. I don't trust Anthropic or OpenAI to keep my important data private.

But do yourself and the world a favor - don't help China train their models or steal your IP by using their APIs (Deepseek, Z.ai, MiniMax, etc.)
 
Last edited:
sleepeeg3 said:
But do yourself and the world a favor - don't help China train their models or steal your IP by using their APIs (Deepseek, Z.ai, MiniMax, etc.)
Click to expand...
I mean... the US models stole IP to train themselves. They are arguing in court they they should be allowed to and indeed MUST steal copyrighted works because otherwise they won't have enough training data and the US will lose the AI war! Now I'm not trying to defend China or anything, but if protection of intellectual property is your thing, the US AI companies are not the heroes and I don't feel like you can really point a finger and say "Sure they stole a bunch of IP to train their models, and that's fine, but China is then stealing from them and THAT'S wrong!"
 
Sycraft said:
I mean... the US models stole IP to train themselves. They are arguing in court they they should be allowed to and indeed MUST steal copyrighted works because otherwise they won't have enough training data and the US will lose the AI war! Now I'm not trying to defend China or anything, but if protection of intellectual property is your thing, the US AI companies are not the heroes and I don't feel like you can really point a finger and say "Sure they stole a bunch of IP to train their models, and that's fine, but China is then stealing from them and THAT'S wrong!"
Click to expand...
As it's been made abundantly clear millions of times, none of the data they harvest steal for training is private or protected.
 
Sycraft said:
I mean... the US models stole IP to train themselves. They are arguing in court they they should be allowed to and indeed MUST steal copyrighted works because otherwise they won't have enough training data and the US will lose the AI war! Now I'm not trying to defend China or anything, but if protection of intellectual property is your thing, the US AI companies are not the heroes and I don't feel like you can really point a finger and say "Sure they stole a bunch of IP to train their models, and that's fine, but China is then stealing from them and THAT'S wrong!"
Click to expand...

This is why I don't have a single care to give about AI models being stolen. They stole the data to get where they are and now complain about people stealing from them? The irony is delicious. If anything, their actions are worse because they stole from the people in their own country. They fucked over their own. China, at least, is screwing over its enemies so you can at least give them props for doing that even if we are (collectively) the enemy.
 
And this is not been about stolable data that much now (and the part they accuse other of stealing from them as very little to nothing to do with it), they generate their own valuable data now via their own monthly users accounts usage and synthetic data.

And they do ofte pay for copyrighted books and others source of data, anthropic paid authors $3000 per books and now it is often opt-in at publication time that authors decide, OpenAI paid more than 250 millions for wall street journal, post and other publication, Reddit has deal, politico has deel, closed scientific paper repositories.

The fair use case of crawling the open internet to learn things to make something new to the world from it, repackage it in a novel way, than trying to make the most exactly the same product that exactly exist by copying it.

Would anthropic complained about anti crawling-scrapping system trying to stop them to learn things it would be quite hypocrite to then implement their own (which if they want to have the right to operate in the US to start with need to be strong, it is not like they have a choice at all to let the chinese model train reasoning using claude, model would get closed by the whitehouse), that not really the case, they respect robots.txt crawling instruction.
 
You must log in or register to reply here.
Back
Top