Artificial Intelligence thread

[madhusudan.tim]

Eventually, the open source models will need to have easily accessible hardware stack (Superpods or superckusters) for inference for companies or nation outside of the West. We know by now that developing a really good model should bot be that difficult for China but it is extremely constrained by compute capacity for scaling up. If it wants the open source to suceed, it will need to have the capacity to launch and distribute Atlas 960 Superpods in thousands. No one else has the ability to deploy alternative to NVIDIA or huawei GPUs. New application scenarios beyond texts will emerge and the token demand will likely increase in high two digits in the near future. Can china deploy the compute in similarly increasing capacity? If not the open source will fail and closed eco system will win, just like how apple and android have duopoly today. Linux suceeded because of lower barrier of entry whereas you need supercluster class infrastructure to have meaningful capacity.

 

[meedicx]

For Alibaba for example, it has its chips/memory produced at Samsung. So there is actually no dependence on SMIC advanced node or CXMT HBM capacity. It's mostly a resource issue.

Is this just speculation or do you have sources?

There are public announcements of ADAS chip like the recent BYD one using Samsung, but high end AI chips like M890 will 100% fall under US export controls due to TOPS and HBM being above the threshold. It would be hard to imagine Samsung violating US export controls.

 

[tamsen_ikard]

In my opinion , there’s no doubt the US will ban Chinese models from US soil and might even extend that ban to other allied countries . The US seees Chinas rise as a threat the more China advances so no bets are off the table. So you shouldn’t be surprise when this happens , since AI is so important for the future Industrial Revolution taking place and will shape the next great power politics .

I dont think they can stop Chinese models anymore. China played a strategic masterstroke by making their models open source. We have to thank deepseek for inspiring this change.

Without open source, the debate in the west will be between western AI and CCP controlled AI, kinda like the Huawei debate about spying and data theft and so on.

But by open sourcing, the debate has become between closed AI vs Open AI and many western liberals are defending open source AI. So, they cannot defeat Chinese anymore with propaganda.


Moreover, as I said before, Open source is impossible to ban. Once software becomes open source, it becomes like math, book or ideas. Its just symbols.

You cannot ban an idea.

Yes, they will likely ban China hosted AI or AI hosted by Chinese companies, but China already kinda gave up their money making potential by open sourcing AI models.

They are giving up on IP and revenue but winning on user feedback and adoption.

 

[Michael90]

Is this just speculation or do you have sources?

There are public announcements of ADAS chip like the recent BYD one using Samsung, but high end AI chips like M890 will 100% fall under US export controls due to TOPS and HBM being above the threshold. It would be hard to imagine Samsung violating US export controls.

You mean Alibaba is using SMIC? I don’t think there has so far been any news showing that.

They are giving up on IP and revenue but winning on user feedback and adoption.

Agree, which is a good thing for the world . However, is it sustainable for Chinese companies investing heavily in this only to have to give out their IPs and not make much revenue ? Afterall, they are not SOE who don’t need to make profits and are more for public goods, they are private enterprises who need to satisfy their shareholders at some point . Is there away they can monetize their products /IP? Companies like Alibaba or tencent/Bytedance wont have much of this issue since it’s not their main business and they can use their AI models to increase their monétisation of their main business which they are already benefiting hugely from, but Chinese AI startups who are at the leading edge will have a hard time sustaining this for years to come no?

 

[tamsen_ikard]

You mean Alibaba is using SMIC? I don’t think there has so far been any news showing that.

Agree, which is a good thing for the world . However, is it sustainable for Chinese companies investing heavily in this only to have to give out their IPs and not make much revenue ? Afterall, they are not SOE who don’t need to make profits and are more for public goods, they are private enterprises who need to satisfy their shareholders at some point . Is there away they can monetize their products /IP? Companies like Alibaba or tencent/Bytedance wont have much of this issue since it’s not their main business and they can use their AI models to increase their monétisation of their main business which they are already benefiting hugely from, but Chinese AI startups who are at the leading edge will have a hard time sustaining this for years to come no?

All AI startups are doomed anyway. AI is a commodity that is given away for free. You cannot make money selling AI models.

The only way to make money is to become AI hosts, essentially data center companies.

I think most AI startups will either become data center companies themselves hosting AI and also developing AI, or they will be bought by the big tech data center companies like AliBaba.

Maybe they start gating their most advanced models and open sourcing their older models, kinda like patreon vloggers. Who put their latest stuff behind a paywall for a brief period and then release it for free when the next video comes along.

 

[tphuang]

Guys let's please stop with the useless one-liners. Not all one-liners will be deleted but if you have a one-liner and it has nothing useful in there that moves this topic along, I'm just going to delete them.

 

[Michael90]

All AI startups are doomed anyway. AI is a commodity that is given away for free. You cannot make money selling AI models.

The only way to make money is to become AI hosts, essentially data center companies.

I think most AI startups will either become data center companies themselves hosting AI and also developing AI, or they will be bought by the big tech data center companies like AliBaba.

Maybe they start gating their most advanced models and open sourcing their older models, kinda like patreon vloggers. Who put their latest stuff behind a paywall for a brief period and then release it for free when the next video comes along.

Yes , I was actually thinking that , the only future I see for this AI startups is to be bought off by the bigger more established tech companies with an already established business and source of revenue . Else they are doomed in the long run, since I see no other way out for them to have a sustainable income needed to keep running such private open source AI companies. So consolidation will be needed at some point I guess. Deepseek might be an exception since the founder already has his own established profitable business before even starting the company. So they should be okay.

 

[tphuang]

https://twitter.com/i/web/status/2079264404311335312

Kimi K3, GLM and Qwen news from today. Huge validation of K3 from Arena. That agentic ranking is basically the best measure of real world engineering task capabilities. It's basically on the same level as opus-4.8 and GPT-5.6 Sol for that. Remember. GLM-5.2 reaching Opus-4.6 level was the initial hint that told everyone it was really good.

GLM-5.5 basically verified and looks like the release is not too far away. Zai member is advising people to re-up their plans. Can't get better signal than that.

Qwen-3.8 preview continues check pointing. Pressure on them is high from Moonshot and Zai releases.

 

[tphuang]

Is this just speculation or do you have sources?

There are public announcements of ADAS chip like the recent BYD one using Samsung, but high end AI chips like M890 will 100% fall under US export controls due to TOPS and HBM being above the threshold. It would be hard to imagine Samsung violating US export controls.

it's pretty well know in China that both Alibaba and ByteDance picked Samsung. Keep in mind that you can get Samsung to produce small dies which you can then piece together in one large chip. There are ways around this. I generally don't ask too many questions for obvious reasons.

You mean Alibaba is using SMIC? I don’t think there has so far been any news showing that.

No I mean Alibaba uses Samsung. Seriously, did I write something that was hard to read? I thought I was pretty clear in there.

Eventually, the open source models will need to have easily accessible hardware stack (Superpods or superckusters) for inference for companies or nation outside of the West. We know by now that developing a really good model should bot be that difficult for China but it is extremely constrained by compute capacity for scaling up. If it wants the open source to suceed, it will need to have the capacity to launch and distribute Atlas 960 Superpods in thousands. No one else has the ability to deploy alternative to NVIDIA or huawei GPUs.

Alibaba, ByteDance, Google and others would disagree with you on this.

 

[bsdnf]

Agree, Given the size of the frontier models only companies from medium size can afford the hardware to run these models without heavy quantization, they could regulate US companies from serving these models. Personally I think will backfire badly.

First maybe companies will abide by the rules but normal users are likely to move to use these services outside the US, this is likely to damage the revenue of local inference companies.

Second and ironically this will do more damage to US open source AI efforts, because there will be less incentives to release open weight models and also create a regulatory environment were is dangerous to release open AI models. And even if for some miracle you manage to release a open weight model IT WILL NEVER BE even close to a frontier mode because is pretty clear that stooges Dario Amodei and Sam Altman will never let that happen.

Third and Finally will be likely be the final nail in the coffin of the AI bubble, you see Sam Altman and Dario Amodei they like the smell of their own biologically produced gases and the only reason people are not paying a fortune to access to a GPT 4 level LLM because the price pressure of Chinese AI labs. Little they know, I think, this pressure is contributing to a delay in the popping of the AI bubble that could produce capital destruction even bigger than the dotcom bubble, I think is going to pop anyways, but Chinese labs are keeping the flames from extinguish by keeping prices affordable and people interested in AI.

Southeast Asia will be the next data center hotspot (because the Middle East is a bit of a mess, and nobody will choose Europe), where a large number of Chinese companies are already building servers using Nvidia chips; the infrastructure is already there. If the executive order is indeed issued, some hosting providers will definitely choose the move there

The current executive order is insufficient to prevent such behavior.