News on China's scientific and technological development.

T

tacoburger

Junior Member
Registered Member
mossen said:
China's AI effort is dictated by the same two forces that shape Western AI: chips and software.

China can longer use CUDA or other software stacks. Building its own is possible, but will take time. As for chips: it can't buy top-end GPUs anymore and while it can design its own, it cannot fab them because they are forbidden from using bleeding edge fabs from TSMC.

As a result, China's AI efforts will be also-rans until and unless it manages to fix these two vulnerabilities. Probably not what people here want to hear, but I also attacked Zero Covid when lots of folks here defended it long past its usefulness. Harsh truths need to be told.
Click to expand...
China will need to be more efficient and innovate. Computing power and model size isn't everything in A.I development. High quality datasets and different training methods can result in smaller models that outperform larger ones. Natural language processing models also aren't everything, there's tons of other different A.I models that are used for different applications and will need different datasets.

Most people don't think that natural language processing models will ever get us AGI for example, even if you scale it to trillions of parameters, all you'll get is really really good version of chatGTP.

If anything, I'm just afraid that everyone in China is just going to jump on the GTP bandwagon and ignore all the other potential A.I models and deep learning methods. Which is dumb and needs to change if China wants to pull ahead. Does the country really need a dozen large natural language processing models and all the top A.I companies pouring in billions and years into what's basically the same product?
 
E

Eventine

Junior Member
Registered Member
Andy1974 said:
If you use the entire internet to train your AI you will also be training it will all the misinformation on the internet. To have an AI that is well trained on internet data.. somebody has to choose which data to use.
Click to expand...
Sure, there's misinformation. There's also bias. ChatGPT is well known for both of these issues.

For example, if I asked "is China a dictatorship that oppresses minorities," and ChatGPT wasn't trying to censor the results, it'd probably say "yes" simply because that's what the English speaking internet believes, and not because it's accurate.

But you'd be surprised at how well you can do in many subjects just by having enough data to "average."

If I asked "how do you smoke fish," the vast majority of resources out there will tend to be close to being accurate; and if you were doing this research yourself, you'd be doing the same thing ChatGPT does - going through all those resources and looking for common guide lines.

ChatGPT is basically a "internet knowledge expert" - ie it doesn't know anything the internet doesn't. But it can organize the information available on the internet in such a way, such that it can answer the vast majority of common knowledge questions with high accuracy. That's the data advantage afforded by the English language internet simply being bigger & more diverse.
 
A

Andy1974

Senior Member
Registered Member
Eventine said:
Sure, there's misinformation. There's also bias. ChatGPT is well known for both of these issues.

For example, if I asked "is China a dictatorship that oppresses minorities," and ChatGPT wasn't trying to censor the results, it'd probably say "yes" simply because that's what the English speaking internet believes, and not because it's accurate.

But you'd be surprised at how well you can do in many subjects just by having enough data to "average."

If I asked "how do you smoke fish," the vast majority of resources out there will tend to be close to being accurate; and if you were doing this research yourself, you'd be doing the same thing ChatGPT does - going through all those resources and looking for common guide lines.

ChatGPT is basically a "internet knowledge expert" - ie it doesn't know anything the internet doesn't. But it can organize the information available on the internet in such a way, such that it can answer the vast majority of common knowledge questions with high accuracy. That's the data advantage afforded by the English language internet simply being bigger & more diverse.
Click to expand...
Exactly, you are relying on the average of a huge amount of data to smooth out the effects of propaganda etc, or you could just train it with a small piece of accurate data.
 
L

luminary

Senior Member
Registered Member
tacoburger said:
China will need to be more efficient and innovate. Computing power and model size isn't everything in A.I development. High quality datasets and different training methods can result in smaller models that outperform larger ones. Natural language processing models also aren't everything, there's tons of other different A.I models that are used for different applications and will need different datasets.

Most people don't think that natural language processing models will ever get us AGI for example, even if you scale it to trillions of parameters, all you'll get is really really good version of chatGTP.

If anything, I'm just afraid that everyone in China is just going to jump on the GTP bandwagon and ignore all the other potential A.I models and deep learning methods. Which is dumb and needs to change if China wants to pull ahead. Does the country really need a dozen large natural language processing models and all the top A.I companies pouring in billions and years into what's basically the same product?
Click to expand...
Exactly, it's nice to have a chatGPT counterpart but we don't want that distraction to become counterproductive, that was the whole reason for the software tech giant crackdown in the first place, it was stealing talent from other much needed places.

On the other hand, LLMs do help as well. Large Language Models are a good exercise for AI devs because the internet gives them a lot of diverse traffic, user feedback, and data types to work with. Regardless of the "accuracy" or "usefulness" of the output, it gives researchers some experience in designing big data models. This is more likely to benefit China than any other country, as again, China has overwhelming advantage in data collection and interconnectivity, and may eventually need these techniques down the line. It is also a good societal exercise to challenge existing understanding of AI and provoke discussion on where humans will place and coexist with them in the future.

I actually think the US is approaching the limit for what kinds of AI they can develop. The internet seems to be the single biggest source of data they can train their AI on. From here on for them, it's all about making AI work more efficiently with smaller datasets, self-classification and reinforcement, novel model architecture, and possibly generating "life-like" fake data. Will they be able to sustain a positive feedback cycle with these methods? Or will they hit the "simply not enough data to work" development wall? Or (imo likely) end up with a fake, marketing-gimmick "AGI"? It's anybody's guess.
 
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Another good post from HW前HR. I think he is right on mark here. Don't think China can have good relationships with current industrial powers. For a while, it was beneficial for those countries to trade with China. They can sell China high tech stuff & china can seel them low end stuff. But China has moved too fast and this is putting too much pressure on Western industries. The two countries that probably hate China the most is Japan and South Korea, because China has done more to destroy their industries than anyone else. This past year, China has finally started to run trade surplus against Korea.

He listed 4 items in there: high speed rail, telecommunication, NEVs/renewables & tunnel boring machines. Not only has China been able to absorb all these technology and produce them locally, but also undercut everyone else in pricing.
中国产业快速跟进是西方最大压力，不要期望这段时间西方阵营对中国好。欧美现在最大的痛苦不是中国出口普通日用品，而是欧美的高科技产业，中国能学习后快速跟上，并通过快速改进和迭代，最终青出于蓝胜于蓝。在资本主义国家，断人财路如杀人父母，所以这是经济基础已经让双方变成很多竞争关系。相反，很多发展中国家因为受益于中国大量买入原材料，与中国关系反而很好。

第一，高铁国产化就是一个中国替代和国际化的过程。最早中国高铁很多都是合资的，然后合资里面的核心关键器件都是来自日本、德国和法国等企业；所以一开始的国产化，有点类似合资汽车国产化，大部分值钱的要进口，起到总装车间效果而已。但是当到了复兴号后，99%都国产化了。由于中国占全球高铁市场过半，国内基本拿下还走出去，结果中车把欧美几个高铁公司干得奄奄一息，有的合并了。

第二，通信产业。最早的通信产业程控交换机有七国八方之称。就是七个进口国家，还有一个国家来了两个厂。但是慢慢的那些国外厂家通信业务消失到历史长河里面，比如朗讯、北电、阿尔卡特、西门子、马可尼、摩托罗拉等，变成华为等国产品牌起来了。这是一个此消彼长的过程。本来4G时候美国还有点通信设备，到5G已经消亡，所以美国人歇斯底里的围堵华为等通信设备厂家，因为他们没有窃听的趁手工具了。

第三，新能源车和新能源产业。新能源车最早算是特斯拉闯出个名堂，但是中国厂家后来者居上，现在全球60%都是中国厂家生产的新能源车。很多东西，欧美有先发优势，但是中国厂家却有快速迭代优势，搞着搞着对方就玩不转了。比如在太阳能和风电这些新能源，最早都是欧美日韩技术更好，无奈中国厂家进步更快，先跟进、后持平再超越。所以新能源和新能源都是中国成为全世界的产业链基地。

第四，盾构机最早德国卖3.5亿人民币，我国造出来后，卖几千万。因为欧美日看中国盾构机出来后就想通过降价扼杀在摇篮，但是中国厂家就再降价迎头痛击，然后双方螺旋PK，最终成本下降了很多，中国产品价格成为欧美产品当初十分之一。中铁工业说目前公司盾构机在海外市场累计获超百台订单，不仅仅国内基本上全被国产拿走，海外也开始进入发达国家市场了。现在全球70%以上市场被中国厂家占领了。

实事求是的说，设备这些东西，有点此消彼长的零和效果，因为市场空间就那么大，不可能突然扩容几倍的。所以欧美媒体煽动下，欧美麻瓜是很恨中国企业的，并且恨中国。但是本质上他们是输不起而已
Click to expand...
 
C

CMP

Senior Member
Registered Member
luminary said:
Exactly, it's nice to have a chatGPT counterpart but we don't want that distraction to become counterproductive, that was the whole reason for the software tech giant crackdown in the first place, it was stealing talent from other much needed places.

On the other hand, LLMs do help as well. Large Language Models are a good exercise for AI devs because the internet gives them a lot of diverse traffic, user feedback, and data types to work with. Regardless of the "accuracy" or "usefulness" of the output, it gives researchers some experience in designing big data models. This is more likely to benefit China than any other country, as again, China has overwhelming advantage in data collection and interconnectivity, and may eventually need these techniques down the line. It is also a good societal exercise to challenge existing understanding of AI and provoke discussion on where humans will place and coexist with them in the future.

I actually think the US is approaching the limit for what kinds of AI they can develop. The internet seems to be the single biggest source of data they can train their AI on. From here on for them, it's all about making AI work more efficiently with smaller datasets, self-classification and reinforcement, novel model architecture, and possibly generating "life-like" fake data. Will they be able to sustain a positive feedback cycle with these methods? Or will they hit the "simply not enough data to work" development wall? Or (imo likely) end up with a fake, marketing-gimmick "AGI"? It's anybody's guess.
Click to expand...
We'll end up with fake stuff. Just like my fake 5G.

tphuang said:
Another good post from HW前HR. I think he is right on mark here. Don't think China can have good relationships with current industrial powers. For a while, it was beneficial for those countries to trade with China. They can sell China high tech stuff & china can seel them low end stuff. But China has moved too fast and this is putting too much pressure on Western industries. The two countries that probably hate China the most is Japan and South Korea, because China has done more to destroy their industries than anyone else. This past year, China has finally started to run trade surplus against Korea.

He listed 4 items in there: high speed rail, telecommunication, NEVs/renewables & tunnel boring machines. Not only has China been able to absorb all these technology and produce them locally, but also undercut everyone else in pricing.
Click to expand...
That's the ultimate end goal Chinese society should shoot for. To be the leader in just about everything.
 
siegecrossbow

siegecrossbow

General
Staff member
Super Moderator
At the risk of sounding like a complete nutter I want to ask the following question -- what are the chances that China has started using AI and/or super computing for geopolitical decision making? The recent coup with the Saudi-Iranian deicing and China's follow up moves are essentially flawless and I'm suspecting that there is more to it than just good decision making by statesmen.
 
H

HumanHDMI

New Member
Registered Member
siegecrossbow said:
At the risk of sounding like a complete nutter I want to ask the following question -- what are the chances that China has started using AI and/or super computing for geopolitical decision making? The recent coup with the Saudi-Iranian deicing and China's follow up moves are essentially flawless and I'm suspecting that there is more to it than just good decision making by statesmen.
Click to expand...
occam's razor
 
T

tacoburger

Junior Member
Registered Member
siegecrossbow said:
At the risk of sounding like a complete nutter I want to ask the following question -- what are the chances that China has started using AI and/or super computing for geopolitical decision making? The recent coup with the Saudi-Iranian deicing and China's follow up moves are essentially flawless and I'm suspecting that there is more to it than just good decision making by statesmen.
Click to expand...
There's no way current A.I can do something as complex as that. What they might have done is use A.I to help in large data analysis, breaking down trade, GDP numbers, oil production, all the millions of factors that make up the economy and geopolitics of a nation into a more digestible format for better decision making. But kind of large data processing is more than a decade old at this point.

More exotic uses might be using large language models like GTP to suggest the optimal respond to someone else, not impossible, given that people are using ChatGTP to write emails, love letters etc etc. But unlikely, given that there's no way they'll entrust something this important to an A.I, not to mention that this language models only really got good last year, whiles talks must have been ongoing for years at this point, there's no models that really fluent in Persian or Arabic just yet anyway.

Another use is in translation, instead of a few people fluent in Arabic/Persian, you get a really good language model that can translate better than a native speaker in both languages, so you can a large team of people analysing every scrap of data that comes out of a country, from local newpaper articles, to facebook posts. Better than current machine translation tech. Niche but could help. But again, language models only really got good last year and they're not fluent in Persian or Arabic.
 
You must log in or register to reply here.
Top