China's Opensource Deepseek outperforms OpenAI GPT-4o

[Martin Everson]

Took them 2 months and $6m to develop.

As someone said in the comments section - "This is a competition between the Chinese in the US and the Chinese in China." lol

https://www.deepseek.com/

 

[daniels27]

Lessons to be learned

• Invest in universities and research
• Use overproduction and unemployed wisely to upgrade your country
• Sanctions lead to your opponent learning how to perform on limited resources
• The days of US bloatwares are over
• Time to get back to work for Americans, social welfare was yesterday
• Money alone does not do the job

Fat cats have been accumulating for too long. Good luck Mr. Trump!

 

[Martin Everson]

Yup all those lessons have been learned the hard way by U.S trying to suppress China's technological advancements.

Good luck to him blowing $500bn in overpriced AI development in U.S.....lol.

Trump highlights partnership investing $500 billion in AI​

https://apnews.com/article/trump-ai...tman-ellison-be261f8a8ee07a0623d4170397348c41

P.S How is Trump gonna feel about U.S AI companies and rest of world now using opensource Deepseek to build upon. Is he gonna sanction opensource software?

 

[0xDEADBEEF]

Deep Seek's parent company, High-Flyer, is a Chinese hedge fund. Their expertise is primarily focused on trading. So this is not even their final form and they will be able to continue financing it themselves.

The scale of Huawei, Alibaba and Tencent cloud services should not be underestimated either. So tough luck for the US, 500bn is being pocketed by these 'innovators'.

Rule of thumb: no matter what you're doing and how good you are doing it, Asians will find a way to do it better.

 

[0xDEADBEEF]

https://twitter.com/x/status/1882757279436718454

 

[JackAlabama]

I was testing it. Its quite good (in regards to tech stuff).

 

[mraleph]

Another Chinese wonder. People should wonder themselves is it really so

https://www.nature.com/articles/d41586-025-00229-6

https://medium.com/data-science-in-your-pocket/deepseek-is-highly-biased-dont-use-it-2cb0358647f9

https://morganlinton.medium.com/deepseek-r1-is-insane-70b3b0400d56

On top of that, it's using smugled Western hardware, quote DeepSeek has 50,000 NVidia H100’s that they can’t talk about

Everything coming from China is intended to force their political naratives and is either stolen Western or Russian IP or smugled and RE hardware.

Dumping prices for access arround new US president inauguration? C'mon, it's another Chinese ruse.

Wouldn't use that by any chance.

 

[yngmind]

I'm currently in China, and using deepseek here for my daily tasks, it works perfectly.
China is nice though, expect the firewall.

 

[0xDEADBEEF]

On top of that, it's using smugled Western hardware, quote DeepSeek has 50,000 NVidia H100’s that they can’t talk about

Everything coming from China is intended to force their political naratives and is either stolen Western or Russian IP or smugled and RE hardware.

Dumping prices for access arround new US president inauguration? C'mon, it's another Chinese ruse.

Wouldn't use that by any chance.

I agree with you on principle, China does what China does.

But the model is open source and this will result in accessibility in the long run. Which hopefully means we can run high-quality models locally on affordable hardware instead of relying on cloud solutions. Plus you can run models that do not refuse anything you ask of them.

 

[mraleph]

Plus you can run models that do not refuse anything you ask of them.

An output from DeepSeek

https://medium.com/data-science-in-your-pocket/deepseek-is-highly-biased-dont-use-it-2cb0358647f9

Not only biased, which means that PRC government censorship is applied, but by using it, we would feed and develop not only competition's LLM, but a hostile competition one - they are already using our hardware and knowledge against our interests.

Sorry, but, Russia is a natural ally of the West. China simply isn't.

You know my position and stance about cloud solutions that we don't own and manage. LLM/AGI models that can be run locally exist and they should be used - not Chinese ones.

 

[yngmind]

An output from DeepSeek

https://medium.com/data-science-in-your-pocket/deepseek-is-highly-biased-dont-use-it-2cb0358647f9

Not only biased, which means that PRC government censorship is applied, but by using it, we would feed and develop not only competition's LLM, but a hostile competition one - they are already using our hardware and knowledge against our interests.

Sorry, but, Russia is a natural ally of the West. China simply isn't.

You know my position and stance about cloud solutions that we don't own and manage. LLM/AGI models that can be run locally exist and they should be used - not Chinese ones.

I asked a question about why the internet is slow in China. It started to reply but then stopped and said it was beyond its scope.

 

[Forester]

On top of that, it's using smugled Western hardware, quote DeepSeek has 50,000 NVidia H100’s that they can’t talk about

Well... so what?

Everything coming from China is intended to force their political naratives

I agree at 110%.

and is either stolen Western or Russian IP or smugled and RE hardware.

With all due respect, I cannot agree. (Unfortunately.)
Yes, many, if not all, Chinese technologies were built using foreign IP (many times without the consent of the IP owner); but currently, they are developing on their own. Have a look e.g. at the railways.

But the model is open source

According to Nature (see here https://www.nature.com/articles/d41586-025-00229-6), the code is under MIT licence, the training data are unavailable. (Expectable; and fair enough, IMO.)

and this will result in accessibility in the long run. Which hopefully means we can run high-quality models locally on affordable hardware instead of relying on cloud solutions. Plus you can run models that do not refuse anything you ask of them.

I agree.

An output from DeepSeek

https://medium.com/data-science-in-your-pocket/deepseek-is-highly-biased-dont-use-it-2cb0358647f9

Not only biased, which means that PRC government censorship is applied, but by using it, we would feed and develop not only competition's LLM, but a hostile competition one

Yes. Hence take the code, build and run independently

Sorry, but, Russia is a natural ally of the West. China simply isn't.

To be honest, I agree with the latter; but as for the former, I do not think that Russia is a natural ally of the contemporary West... and I understand it well. The cultural differences gap become too wide. (It should not have happened; and perhaps West is even more guilty. My 2 cents.)

 

[mraleph]

Well... so what?


I agree at 110%.


With all due respect, I cannot agree. (Unfortunately.)
Yes, many, if not all, Chinese technologies were built using foreign IP (many times without the consent of the IP owner); but currently, they are developing on their own. Have a look e.g. at the railways.



According to Nature (see here https://www.nature.com/articles/d41586-025-00229-6), the code is under MIT licence, the training data are unavailable. (Expectable; and fair enough, IMO.)


I agree.


Yes. Hence take the code, build and run independently


To be honest, I agree with the latter; but as for the former, I do not think that Russia is a natural ally of the contemporary West... and I understand it well. The cultural differences gap become too wide. (It should not have happened; and perhaps West is even more guilty. My 2 cents.)

Regarding Russo-Western relations; it's complicated Though, the current state of affairs do not change the fact that we're natural allies. Far greater cultural differences exist across Europe and America then between the West and Russia. The balance state is cooperation, whatever is the narative on both sides. Till naratives' converge, we're observing embargo and sanctions compliance

Regading Chinese theft and copycat; they're developing railways - after they've stollen the IP and blueprints. They've experienced socio-economic development due to the fact that West transfered manufacturing to China. Now is reverse and will have detrimental influence to economy and social fabric. China will collapse in next 10 years. And above that, of all Asian nations, China is the most hated and feared; not liked by their own ethnic groups, India, VietNam, Koreas, Japan, Filipines.

Regarding DeepSeek; kind of a in the game, so, wouldn't even use the source code

Regarding smugled hardware; if the PRC is capable for anything else then theft and copycat, they would have original ideas and capacities - which they lack. So, I wouldn't be surprised that DeepSeek is using Western training data and not only hardware.

The PRC flourished massively selling cheap goods to the world. Things are becoming complicated for them now as their model is collapsing. Same thing with DeepSeek; another Chinese copycat and propaganda.

 

[cryptofriendly]

I have been testing it yesterday extensively, it's very fast and its answers are pretty good, compared to other LLMs.
Though, it seems that they have been using ChatGPT to train it, as it contains some of its biases (not all though) - even if they aren't relevant to China. Asking sensitive questions about Xi is met with refusal.

Positive apects:

- much cheaper and faster than ChatGPT
- open source
- one can run an instance of DeepSeek locally on one's computer or phone, even locally in a browser with CloudGPU
- there are already hacked versions available, where the 'Xi sensitive' censorship got removed
- for coding, it's much better than ChatGPT

I'd say ChatGPT is like Google search engine today, giving you s**t results and hiding what's important,
DeepSeek is like Google when it wasn't being evil yet - just a little evil regarding Chinese stuff though.

I have been co-developing an open-source LLM model for the last 12 months, DeepSeek makes most of our work obsolete, well, actually all we need to do now is to do it the Chinese way, copy it and fine tune it for our needs.

 

[daniels27]

I have been co-developing an open-source LLM model for the last 12 months, DeepSeek makes most of our work obsolete, well, actually all we need to do now is to do it the Chinese way, copy it and fine tune it for our needs.

This sounds pretty dooming. But shows that access to cheap labour and energy accelerates many things.

Regarding smugled hardware; if the PRC is capable for anything else then theft and copycat, they would have original ideas and capacities - which they lack.

And I am pretty sure that they are very well aware that they will lack it for a very long time to come. An education system based on memorising (remember what AI is capable of) is simply so what of 1925 and entirely useless. Any state of sound mind would mandate that any exam in school must allow a minimum of 10 pages of handwritten notes.

And then there is a bigger problem that kids aren't trained and supposed to question what they are being taught, what the government narrative is. It leads to a degeneration of human brain capacity in early years.

So, I wouldn't be surprised that DeepSeek is using Western training data and not only hardware.

They probably stole pretty much everything
https://www.nytimes.com/2024/03/06/us/politics/google-engineer-china-ai-theft.html

China can be happy. Stargate is like Watergate: Highly expensive. Everything is wiretapped. And in the end it only helps the opponents.

 

[Frankie]

Imagine believing and re-posting Chinese propaganda verbatim.. and talking about "lessons to the US"

$6m to develop ... best thing I read all day.

I can't wait for the article next week in which we hear that North Korea developed their own ChatGPT using nothing but a potato and 4 home-made GPUs...

 

[Martin Everson]

I absolutely love it when people get triggered by anything China . Still waiting for some usual suspects who have not turned up yet

 

[daniels27]

I absolutely love it when people get triggered by anything China . Still waiting for some usual suspects who have not turned up yet

Ok, here we go, for you, @Frankie and @John Spectre:

I just read that they have been scanning OCT and will soon revive long gone posters. Stay tuned.

 

[mraleph]

Next time some Chinese success story is published on OCT we'll celebrate