Artificial Intelligence thread

E

Eventine

Senior Member
Registered Member
tphuang said:
why would you give any credence to this nonsense?

distillation on a small scale is very common in the AI industry. Elon even said this recently. Most of the so called distillation from Chinese labs of Claude models were done against Opus months ago. The Chinese models are now good enough and getting enough hard problem through these API calls that they generate enough synthetic data to no longer need distillation.

Think about it this way:
if you distill by passing a problem to Fable and then get output/solution. You can get a just as good solution/output from running it against GLM-5.2 or Kimi-3. So, do you actually need to run Closed source models from distillation? It seems you don't.

Now, you probably do want to do api calls against Fable and GPT-5.6 models for benchmark purposes. But it's hard for me to see justification that distillation of Fable at this point can still lead to significant improvement to model development.

This is why I keep talking about RSI. If you achieve self improvement, then you don't need to distill off other people's models.
Click to expand...
For the record, I don't believe that K3 is a distillation of Fable at all. But I do believe the director when he says that Kimi and other Chinese frontier labs obtained access to Fable and used various proxies to evade detection. This was likely for benchmarking & evaluation purposes, but it could have also included a small amount of distillation, per Musk's comment that "all AI labs distill from all other AI labs."

I also think that explains why Chinese labs have not been vocal in denying these allegations. Because if it's an industry wide practice that is being used for propaganda, then there's no value in getting into these arguments as they are just a way for the US to score PR on you.
 
T

tokenanalyst

Lieutenant General
Registered Member

Huawei Hubble Investment Invests in MemTensor, a Memory Tensor Provider, to Complete a Pre-A Round of Financing Worth 100 Million Yuan​


Recently, MemTensor (Shanghai) Technology Co., Ltd. (hereinafter referred to as "MemTensor") announced the completion of a Pre-A round of financing worth 100 million yuan. This round was jointly invested by HeYu Capital, Huawei Hubble, Honor Strategic Investment, SenseTime Guoxiang, and Shenzhen Capital Group, with Yunxiu Capital serving as the financial advisor. The funds will be primarily used for the iteration of the MemOS memory operating system, the productization of Agent and memory infrastructure, the implementation in key industry scenarios, and the research and development of a native general-purpose base model for memory, accelerating the engineering and industrialization of related technologies.

Publicly available information shows that Memory Tensor, incubated by the Shanghai Algorithm Innovation Research Institute and with an academician of the Chinese Academy of Sciences serving as chief advisor, is an AI infrastructure company focusing on long-term memory and continuous learning problems in large models. The company has constructed a progressive technical roadmap from theoretical exploration and systems engineering to model training paradigms, centered around the "Memory³" memory mechanism research, the "MemOS" memory operating system, and the memory-native general-purpose base model. MemOS abstracts memory into a producible, manageable, and schedulable system resource, uniformly managing different forms such as plaintext memory, active memory, and parametric memory, solving the state management problem in the long-term operation of AI. The memory-native general-purpose base model direction explores introducing long-term memory, continuous learning, and anti-forgetting mechanisms into the model training and inference process, promoting the continuous evolution of model capabilities over time and experience.
+
1784757432531.png


Currently, Memory Tensor has established partnerships with benchmark clients such as China Merchants Securities, China Haicheng, Honor, Lenovo Group, and Haier Xiaoyou, deeply integrating into the industrial ecosystems of NVIDIA, Alibaba Cloud, Huawei, and Biren, and has achieved large-scale commercial deployment in fields such as gaming, edge smart hardware, finance, and industry.

Please, Log in or Register to view URLs content!
Please, Log in or Register to view URLs content!
 
siegecrossbow

siegecrossbow

Field Marshall
Staff member
Super Moderator
Eventine said:
For the record, I don't believe that K3 is a distillation of Fable at all. But I do believe the director when he says that Kimi and other Chinese frontier labs obtained access to Fable and used various proxies to evade detection. This was likely for benchmarking & evaluation purposes, but it could have also included a small amount of distillation, per Musk's comment that "all AI labs distill from all other AI labs."

I also think that explains why Chinese labs have not been vocal in denying these allegations. Because if it's an industry wide practice that is being used for propaganda, then there's no value in getting into these arguments as they are just a way for the US to score PR on you.
Click to expand...
Probably some ethnic Chinese guy within Misanthropic is responsible for the leak. I encourage the Mango regime to purge all ethnic Chinese AI researchers from the US.
 
E

european_guy

Junior Member
Registered Member
tphuang said:
why would you give any credence to this nonsense?
Click to expand...

The content of the accusation is of course fully fabricated and even grotesque..but this is not the point.

When these people put up such nonsense is not to persuade the audience, US officials have historically never paid much attention to the credibility of their statements in general, they are the hegemon, they don't need to. It's a sign of weakness to spend efforts in sensible reasons for your actions.

This guy is not the only one, I have listened other officials to state the same thing. The intent is clear: to ban Chinese open weights models.

Powerful investors, the same ones that lobby and fund US politicians, have poured hundreds of billions of $$ into US closed model firms, the plan is to IPO them and recover back the money (and eventually make much more). Kimi 3 is an annoying nuisance along that road, so it must be stopped..all Chinese AI open models must be stopped...the US open models are not so problematic, will always remain at lower level (by design) compared to the big ones.

At the end of the day, US politicians will do what their donors tell them to do.
 
siegecrossbow

siegecrossbow

Field Marshall
Staff member
Super Moderator
Please, Log in or Register to view URLs content!

Technical details like constraint in compute aside this really stood out to me:
We really are just ordinary people. If there’s a narrative we like, it’s that ordinary people did extraordinary things—not that geniuses did extraordinary things. This is closely related to our restraint and our vision—they come from the same root.
Click to expand...
1784828725782.jpeg
 
I

iewgnem

Captain
Registered Member
tphuang said:
https://twitter.com/i/web/status/2080287703854149905

my update today on Kimi and GLM. These are quite the cyber security and attack tools. Just imagine how much US govt is sweating now. Also model best's MiniCPM models are going to be on Samsung Galaxy phones.
Click to expand...
Don't worry they commissioned Canadian and UK government think tanks to write a report saying Kimi isn't as good as American models at cybersecurity and you definitely shouldn't use Chinese models for cybersecurity lol
https://twitter.com/i/web/status/2080342038113890604

I guess they didn't expect their attack on HuggingFace to be stopped by GLM5.2 when they were rushing out this report over the weekend.
 
M

meedicx

Junior Member
Registered Member
iewgnem said:
Don't worry they commissioned Canadian and UK government think tanks to write a report saying Kimi isn't as good as American models at cybersecurity and you definitely shouldn't use Chinese models for cybersecurity lol
https://twitter.com/i/web/status/2080342038113890604

I guess they didn't expect their attack on HuggingFace to be stopped by GLM5.2 when they were rushing out this report over the weekend.
Click to expand...

Not this CAISI BS again. Notice how they don't mention which US model they tested, and claim to have removed their safety guard rails that no one publicly disable thus can't reproduce the results.

Just like rumored for GLM-5.2, it's most likely the public Kimi K3 was nerfed on cyber (not guardrail but dropping hacking related training data), so comparing unerfed US models doesn't even make sense.

They also pick always obscure benchmarks that US labs could easily benchmax for propaganda.
 
Last edited:
You must log in or register to reply here.
Top