DeepSeek, China's AI model: News & Discussion

[UKBengali]

DeepSeek R1 has essentially destroyed the American monopoly over AI before it even began. The reason being cheap asf inference and training costs, and it being open source. Now most people have been clinging to the Tianeman Square and Xi Jin Ping questions but miss out on a major point. This censorship applies only to the model interface hosted within China so of course whatever the CCP approves of, the model will be allowed to talk about. Similarly, OpenAI ChatGPT does not want to talk about the Rothschild family at all. Now the point where DeepSeek takes the cake is it being open source. A person with the know how could adjust the weights and remove any and all censorship from the model. It is available to everyone to use for free and for any application. You can host it offline on your own hardware so that no data goes to China and neither does any data get censored. Americans are trying to downplay this HUGE blow to their billions of dollars invested into AI because a minor company has come and destroyed their flagship reasoning model for a fraction of the cost.

And for people still skeptical about it, Hugging Face is currently cloning (re training from scratch with help from DeepSeek itself) the model in its entirety so that it is free of any censorship, and hosted in America. Perplexity already has done that but with minor censorship. OpenAI essentially has to produce a miracle such that a better model than R1 is made and is made open source just like it, else why would any person in their right mind pay for O1 or O3 when they can just host it on personal hardware for a fraction of of the cost.

Also, this model introduced a new type of compute, and hence is groundbreaking technology. It is using Reinforcement
Learning unlike O1 that used supervised learning. Reinforcement will always be superior to supervised learning so until OpenAI works on a RL based model, R1 is king.

I have Ryzen 5700X, RTX 3070, 32GB DDR4 RAM and 1TB Gen 3 NVME SSD.

Any ideas how a local version will run on my machine?

 

[iblini]

All the Indian posters here think DeepSeek answers should be pro-India , pro-West and anti-China, lol.

we have this two or three BJP boys and two MAGA boys on this forum, they are not here to discussed anything.

 

[iblini]

I have Ryzen 5700X, RTX 3070, 32GB DDR4 RAM and 1TB Gen 3 NVME SSD.

Any ideas how a local version will run on my machine?

This post went viral, you can take a look.

To view this content we will need your consent to set third party cookies.
For more detailed information, see our cookies page.

Accept third party cookies

 

[j_hungary]

DeepSeek R1 has essentially destroyed the American monopoly over AI before it even began. The reason being cheap asf inference and training costs, and it being open source. Now most people have been clinging to the Tianeman Square and Xi Jin Ping questions but miss out on a major point. This censorship applies only to the model interface hosted within China so of course whatever the CCP approves of, the model will be allowed to talk about. Similarly, OpenAI ChatGPT does not want to talk about the Rothschild family at all. Now the point where DeepSeek takes the cake is it being open source. A person with the know how could adjust the weights and remove any and all censorship from the model. It is available to everyone to use for free and for any application. You can host it offline on your own hardware so that no data goes to China and neither does any data get censored. Americans are trying to downplay this HUGE blow to their billions of dollars invested into AI because a minor company has come and destroyed their flagship reasoning model for a fraction of the cost.

And for people still skeptical about it, Hugging Face is currently cloning (re training from scratch with help from DeepSeek itself) the model in its entirety so that it is free of any censorship, and hosted in America. Perplexity already has done that but with minor censorship. OpenAI essentially has to produce a miracle such that a better model than R1 is made and is made open source just like it, else why would any person in their right mind pay for O1 or O3 when they can just host it on personal hardware for a fraction of of the cost.

Also, this model introduced a new type of compute, and hence is groundbreaking technology. It is using Reinforcement
Learning unlike O1 that used supervised learning. Reinforcement will always be superior to supervised learning so until OpenAI works on a RL based model, R1 is king.

It really depends........

Because as a general rule, the more processing you have it exponentially expanded the processing power. Which mean unless DeepSeek can really squeeze out that much processing power with less with hyper efficiency, at some point, their processing power is going to reach their limit.

Again, look at the output token, you are talking about 8K vs 100K, which mean if you are doing some deep diffusion algorithm, you can't use DeepSeek, because it won't give you enough output for Machine Learning to "remember" the shape, which mean DeepSeek won't be able to correctly predict pattern which is going to open up a whole lot of inaccuracy for any Machine Learning Model.

But then not everyone uses AI need MLM, so for the majority of people who just going to maybe open a phone and then take a picture and your phone tell you what it is or send you a link to buy it, DeepSeek would be enough, may even be a godsend for phone company like Samsung (Whether or not they will use them is another issue) because it's license free. But for bigger AI operation like Neural Network, DeepSeek can't replace Open AI, but what it did is DeepSeek WILL take away a bunch of Open AI client because, as I said, most of them aren't going to need MLM for their application.

What I think OpenAI will do to counter this is probably push out a lit version and make it open source, to directly counter DeepSeek.

 

[Musings]

DeepSeek generates good responses for general queries and sometime they are even more detailed than ChatGpt.
However, it is controlled by CCP and it goes blank when asked about anything not approved by Papa Xi. Example - Type Tianmen and it would have a brain fade.
That’s a bummer considering that an AI assistant should be completely free to generate responses and not biased or controlled like DeepSeek is.
A good example of CCP restricting a wonderful creation.

Bro
Ask yourself is ChatGpt going to tell you about all the fukc ups by the US? Are they going to give an objective response on Gaza for example?
Which nation has no skeletons in their cupboard?
Look at the difference in finance - one cost 6 million and the other cost billions. Enough said

 

[j_hungary]

Bro
Ask yourself is ChatGpt going to tell you about all the fukc ups by the US? Are they going to give an objective response on Gaza for example?
Which nation has no skeletons in their cupboard?
Look at the difference in finance - one cost 6 million and the other cost billions. Enough said

I think the response is quite neutral....

 

[r3alist]

hahaha. QA jobs would be the first to go as AI is the most efficient at creating test code vs development code.

The consultancy types will need to position themselves somehow, ai based qa I think will be an attempt, watch below

the first guys use case was hilarious ......use ai in the hiring process

The next lady effectively admitted India does not have the higher institutions to produce its own competitor ai model, tbh I am sure it does

To view this content we will need your consent to set third party cookies.
For more detailed information, see our cookies page.

Accept third party cookies

@BananaRepublic

 

[mulj]

Cheap chinese AI ! View attachment 97648
@Hamartia Antidote @F-22Raptor

idiotic objection, like chatGTP and open ai do not have biased preprograming for western political ideology.

On the other hand the biggest consequence of deepseek success would be that the world realize on massive scale that China do things simply better then USA and it would be unreversable effect.

 

[r3alist]

See above

Frankensteining software will only get you so far. Eventually you hit a wall without the compute

An the US has a clear moat in compute.

Ok ok compute compute compute

I partially agree with you, but what use for more compute, does anyone know?

You have to concede that deepseek is putting a question mark over the breadth and depth of the USA ai supply chain

Now maybe in years to come when ai has been absorbed and envisioned by actual corporate top managers on where they want to go will compute be a question. It could be just you need something good enough for your needs


And Tata consultants will ofcourse help maga happen, with their deep knowledge

DeepSeek generates good responses for general queries and sometime they are even more detailed than ChatGpt.
However, it is controlled by CCP and it goes blank when asked about anything not approved by Papa Xi. Example - Type Tianmen and it would have a brain fade.
That’s a bummer considering that an AI assistant should be completely free to generate responses and not biased or controlled like DeepSeek is.
A good example of CCP restricting a wonderful creation.

Imagine an Indian ai model, like their own deepseek, should we ever get there, some of the questions I think it won't handle well


Why is modi called the butcher of Gujrat?

Explain in detail the greatness of the mughals?

Is india a superpower?

 

[mulj]

View attachment 97711

I think the response is quite neutral....

perfectly in line with covering all usa **** ups, his point still stands.

 

[大汉奸柳传志]

Apparently someone took the challenge to find the cheapest way of running DS locally

an antique X79 MB 2 CPUs with 768GB Mem, bought from flea market, cost $150

Runs pretty slow at 1 token/s. Not an efficient way compare to a modern gaming rig or Mac mini clusters, but a lot cheaper, and demonstrates the endless possibilities.

 

[大汉奸柳传志]

I have Ryzen 5700X, RTX 3070, 32GB DDR4 RAM and 1TB Gen 3 NVME SSD.

Any ideas how a local version will run on my machine?

at least 16GB of VRAM is needed, from what I heard

 

[Developereo]

Is india a superpower?

This is a topic about China and AI.

Why do you keep bringing India in every post to turn this into a troll fest?

 

[RAJPUT]

I have Ryzen 5700X, RTX 3070, 32GB DDR4 RAM and 1TB Gen 3 NVME SSD.

Any ideas how a local version will run on my machine?

It has multiple versions. Basic difference is in number of parameters on which it is trained. The full fledged model has upwards of 680 billion parameters but the lower end ones are for lower powered devices. It ranges from 1.5 billion to 3.2,7,32,70, and finally 680 iirc. Your hardware could handle the 32 and 70 billion ones but the 70 one would generate response slower as lesser tok/s. 32 billion one should run nicely. Just download Ollama and install it or try the 1.5 billion one as a test run. If it runs fine then you could download the better version.

 

[RAJPUT]

View attachment 97711

I think the response is quite neutral....

Dont be naive. Israel is committing genocide, straight up. They have stopped food, water and medical supplies, shot children in the heads with sniper rounds, drove tanks over dead bodies, flattened housing complexes with people still in them, starved people to death, occupied territory of the Palestinians, Jordanians, and Syrians. All these are textbook definition of genocide and all these events are reported by international media yet you are happy with a ‘neutral’ response. Shame on you. If it were not censored, it should have mentioned the overwhelmingly one sided death toll of the supposed ‘war’, dead children, lies told, and exposed, by the Israeli government, using fake news of dead beheaded babies as an excuse to start this genocide and more. All these are cold hard accepted truths worldwide.