Artificial Intelligence thread

[Eventine]

Please, Log in or Register to view URLs content!

Hi. Can some experts help to explain the significance of
1) the slow speed for Deepseek R1 relative to other models?
2) the higher price of 500% more for Deepseek R1 when compared to Gemini 2.0?

How will these factors matter for end-users and determine the eventual winning model?
Thank you very much.

On top of what was already said, o3-mini and Gemini 2.0 Flash are both assumed to be significantly smaller models than R1; they are, however, highly optimized for specific use cases, especially coding and logic. R1 by contrast is more of a generalist model like O1 or Claude Sonnet. This is why R1 is much better at creative writing, for example, but it comes at the cost of model size and inference speed.

Google, Open AI, and Anthropic are the top Western AI companies, and despite all the jokes we're making, the reality is that they will remain tough competition for China in global AI dominance due to their vast financial resources and access to top end GPUs. A battle was won, but the war is just getting started.

 

[tphuang]

FT article about China having so much AI chips that H100 pricing dropped significantly last few months.

again, 250 public data centers being built. No chip access issues or funding issues. Biggest constraint is know how for people actually building them. Imagine that not all the SOEs building data centers has a clue what they are doing.

https://twitter.com/i/web/status/1893381944831869087

Steve with his chart of Nvidia revenue by region

https://twitter.com/i/web/status/1893393511694663798

More on re export to China via intermediate countries

https://twitter.com/i/web/status/1892693229822251432

 

[tphuang]

Videos generated by Bytedance new OmniHuman-1 model

https://twitter.com/i/web/status/1893352022503538953

New Kimi report on Muon scalability. Btw, I have no idea what Muon is

https://twitter.com/i/web/status/1893376737553326399

 

[SDtom]

Deepseek founder has been very transparent about his goal, to develop AGI. If he secures majority funding/mandate from the government to do that, he becomes the richest/most assets holding man in the world.

What assets? All of the source codes is open and shared with everyone. There are no patents, trademarks, copyrights, or trade secrets that Deepseek can use to earn monetary stream to do R&D to keep itself at forefront of AI development. So the Chinese government have to keep giving money and subsidize Deepseek to keep it competitive and then share all of its knowledge freely to everyone? Foreign AI companies can monetize Deepseek, comb through the source code and improve it all the while not sharing or contribute anything to Deepseek. I am all for open source and sharing Deeepseek for the betterment of the humanity, however I hope Deepseek can come up with a way to sustain itself without future money/subsidies from the government.

 

[Biscuits]

What assets? All of the source codes is open and shared with everyone. There are no patents, trademarks, copyrights, or trade secrets that Deepseek can use to earn monetary stream to do R&D to keep itself at forefront of AI development. So the Chinese government have to keep giving money and subsidize Deepseek to keep it competitive and then share all of its knowledge freely to everyone? Foreign AI companies can monetize Deepseek, comb through the source code and improve it all the while not sharing or contribute anything to Deepseek. I am all for open source and sharing Deeepseek for the betterment of the humanity, however I hope Deepseek can come up with a way to sustain itself without future money/subsidies from the government.

Not company assets lmao

What part of getting a blank cheque from the largest economy in the world don't you get? That makes you richest/most asset owning person in the whole world. And Liang Wenfeng could get that, if he proves he can make a team that makes AGI.

If deepseek projects are on logistics systems, in ports, in the military or even in strategic infrastructure, just how wealthy and powerful do you think that makes him and the other upper levels?

I don't think Liang Wenfeng wants to make himself a trillionaire, there's no point for people like him in having more money than he would ever spend anyways. What he wants is access to the state of the art facilities, ability to get any skilled workers he can, and access to data/testing infrastructure. If he's compelling enough using deepseek, he gets all that.

Come on, you can think further beyond some dumb patent royalties.

 

[AndrewS]

What assets? All of the source codes is open and shared with everyone. There are no patents, trademarks, copyrights, or trade secrets that Deepseek can use to earn monetary stream to do R&D to keep itself at forefront of AI development. So the Chinese government have to keep giving money and subsidize Deepseek to keep it competitive and then share all of its knowledge freely to everyone? Foreign AI companies can monetize Deepseek, comb through the source code and improve it all the while not sharing or contribute anything to Deepseek. I am all for open source and sharing Deeepseek for the betterment of the humanity, however I hope Deepseek can come up with a way to sustain itself without future money/subsidies from the government.

You do realise that it is Chinese companies that will most enthusiastically use Deepseek Open Source models?

Also remember that there are about 8 million Chinese software developers compared to about 4 million in the US.

 

[SDtom]

Not company assets lmao

Then what assets are you talking about?

What part of getting a blank cheque from the largest economy in the world don't you get? That makes you richest/most asset owning person in the whole world. And Liang Wenfeng could get that, if he proves he can make a team that makes AGI.

Have any Chinese companies ( like Baidu, Tencent, Alibaba..etc) starting giving money to Deepseek since they integrate it into their system? If not, then this blank cheque is from the Chinese government, right? But why would the Chinese government do that? Anything and everything that is Deepseek now and in the future will be shared freely to all American AI companies.

If deepseek projects are on logistics systems, in ports, in the military or even in strategic infrastructure, just how wealthy and powerful do you think that makes him and the other upper levels?

Eh, Deepseek is getting integrated into those sectors , so is he much more wealthy and powerful now? Is Deepseek getting any money from those integrations?

I don't think Liang Wenfeng wants to make himself a trillionaire, there's no point for people like him in having more money than he would ever spend anyways. What he wants is access to the state of the art facilities, ability to get any skilled workers he can, and access to data/testing infrastructure. If he's compelling enough using deepseek, he gets all that.

It is not about more money for himself, it is about sustainability. Money to further future AI development, money to hire and keep knowledge AI workers, money to buy hardware to run AI (so I don't keep getting service unavailable/down like right now), money to help developing countries to run and use Deepseek for good PR... etc.

Come on, you can think further beyond some dumb patent royalties.

again, it is not about patents, trademarks, copyrights, or trade secrets, it is about sustainability for Deepseek.

 

[SDtom]

Also remember that there are about 8 million Chinese software developers compared to about 4 million in the US.

I also remember that Deepseek didn't have as many software developers as Tencent, Baidu, or Alibaba etc and yet it is Deepseek that become most popular and most used in China right now.

 

[AndrewS]

I also remember that Deepseek didn't have as many software developers as Tencent, Baidu, or Alibaba etc and yet it is Deepseek that become most popular and most used in China right now.

My point is that by all those Chinese developers now have access to a State of the Art Model.

---

And I think it would be actually be worth the Chinese government subsidising an open-source Chinese model, given that most of the benefits will accrue to Chinese developers and companies. Hence the comment on 8 million Chinese software developers compared to only 4 million in the USA.

 

[SDtom]

My point is that by all those Chinese developers now have access to a State of the Art Model.

---

Well, so does everyone else, including those working in OpenAI and Elon' Grok.

And I think it would be actually be worth the Chinese government subsidising an open-source Chinese model, given that most of the benefits will accrue to Chinese developers and companies. Hence the comment on 8 million Chinese software developers compared to only 4 million in the USA.

It would work well as long as Deepseek become/stay as the best of best AI around. However, OpenAI and Grok can always look through Deepseek's code and copy/improve whatever feature they like into their own and without sharing anything back to Deepseek. How fast do you think Deepseek can come up with new feature/methods/ideas compares to how fast OpenAI/Grok can copy and improve them?