DeepSeek becomes the top-rated free application

**possessor** · 28 January 2025, 01:49

Please tell me this is better then Gemini.

**chuft** · 28 January 2025, 02:41

It's a very interesting story.

I would not believe anything coming out of China however relating to costs etc. China is a totalitarian country and it lies constantly. Every analyst of China will tell you this, whether it's about business or anything else.

One conclusion you can draw is the Chinese government has something even better for military purposes, or they would not have allowed DeepSeek into the open.

A lot of media outlets are reporting DeepSeek is "open source" but it apparently isn't. I am not into AI, I dislike the concept intensely, but from what I read, there is a big difference between a "free model and open weights" vs "open source." Only the latter (including the code and data used to train it) would let you recreate the model itself.

This model has political "programming" relating to the Chinese government.

I will be interested to see where this goes. Nvidia is still absurdly overvalued by any sane metric. Most people panic selling Nvidia stock today do not understand much if anything about AI. They are just trend-following investors.

**BRBFBI** · 28 January 2025, 06:26

Originally posted by chuft

I would not believe anything coming out of China however relating to costs etc. China is a totalitarian country and it lies constantly. Every analyst of China will tell you this, whether it's about business or anything else.

If the cost were wildly misrepresented we would know. DeepSeek published a research paper on their methods and why they needed far fewer chips. The shocking part of all of this isn't just that a small Chinese company created such a powerful AI, but the realization that any relatively small business with a few millions dollars can do the same.

From the NYT: “It has become very clear that other companies, not just someone like OpenAI, can build these kinds of systems,” said Tim Dettmers, a researcher at the Allen Institute for Artificial Intelligence in Seattle and a professor of computer science at Carnegie Mellon University who specializes in building efficient A.I. systems. “DeepSeek used methods that anyone can duplicate.”

Originally posted by chuft

A lot of media outlets are reporting DeepSeek is "open source" but it apparently isn't. I am not into AI, I dislike the concept intensely, but from what I read, there is a big difference between a "free model and open weights" vs "open source." Only the latter (including the code and data used to train it) would let you recreate the model itself.

My understanding is that the code is open source and therefore anyone with the resources could train an LLM with it. Are you saying it "doesn't count" if they don't share the the data they used to train it with?

Originally posted by chuft

This model has political "programming" relating to the Chinese government.

For sure. I've seen people say they asked it about Tiananmen Square, etc... and it doesn't give a straight answer. I'm sure any of us could test it out if we made an account. All AI have "fine tuning" that adjusts how they respond to certain topics, which is concerning when you think of the influence these models have. I know people who use Chat GPT for advice nearly every day. I know people who talk to it like it's their therapist. If a small number of companies have control over millions of peoples' conversation partner I think that's bad for diversity of thought.

**chuft** · 28 January 2025, 12:36

I have to go to work and can't really respond to all of this in any depth. But there is interesting info here.

https://stratechery.com/2025/deepseek-faq/

For example

"DeepSeek is clear that these costs are only for the final training run, and exclude all other expenses; from the V3 paper:

Lastly, we emphasize again the economical training costs of DeepSeek-V3, summarized in Table 1, achieved through our optimized co-design of algorithms, frameworks, and hardware. During the pre-training stage, training DeepSeek-V3 on each trillion tokens requires only 180K H800 GPU hours, i.e., 3.7 days on our cluster with 2048 H800 GPUs. Consequently, our pre- training stage is completed in less than two months and costs 2664K GPU hours. Combined with 119K GPU hours for the context length extension and 5K GPU hours for post-training, DeepSeek-V3 costs only 2.788M GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M. Note that the aforementioned costs include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data.

So no, you can’t replicate DeepSeek the company for $5.576 million."

and

"So V3 is a leading edge model?

It’s definitely competitive with OpenAI’s 4o and Anthropic’s Sonnet-3.5, and appears to be better than Llama’s biggest model. What does seem likely is that DeepSeek was able to distill those models to give V3 high quality tokens to train on.

What is distillation?

Distillation is a means of extracting understanding from another model; you can send inputs to the teacher model and record the outputs, and use that to train the student model. This is how you get models like GPT-4 Turbo from GPT-4. Distillation is easier for a company to do on its own models, because they have full access, but you can still do distillation in a somewhat more unwieldy way via API, or even, if you get creative, via chat clients.

Distillation obviously violates the terms of service of various models, but the only way to stop it is to actually cut off access, via IP banning, rate limiting, etc. It’s assumed to be widespread in terms of model training, and is why there are an ever-increasing number of models converging on GPT-4o quality. This doesn’t mean that we know for a fact that DeepSeek distilled 4o or Claude, but frankly, it would be odd if they didn’t.

Distillation seems terrible for leading edge models.

It is! On the positive side, OpenAI and Anthropic and Google are almost certainly using distillation to optimize the models they use for inference for their consumer-facing apps; on the negative side, they are effectively bearing the entire cost of training the leading edge, while everyone else is free-riding on their investment.

Indeed, this is probably the core economic factor undergirding the slow divorce of Microsoft and OpenAI. Microsoft is interested in providing inference to its customers, but much less enthused about funding $100 billion data centers to train leading edge models that are likely to be commoditized long before that $100 billion is depreciated."

Lots more for those interested. Overall it actually seems like a gift to the tech industry, not a problem.

**chuft** · 28 January 2025, 22:53

Originally posted by BRBFBI

I know people who use Chat GPT for advice nearly every day. I know people who talk to it like it's their therapist.

That is truly disturbing.

**LazyPooky** · 29 January 2025, 12:34

Last week there was an annual meeting in Davos: The World Economic Forum. There it was said that it's an AI battle between the USA and China, and the battle will be decided in the next two years. One of them will win. Once you are ahead, you can no longer be overtaken, because then you can also use AI to get further ahead. They thought the biggest chance was that America will win because of all the investments in AI. US: "we are going to invest a lot more and that will ensure a new golden age". That was before DeepSeek came along.

Four things are needed to be able to win. You have to make sure you have the DATA. You have the CHIPS, used to make all the calculations. You have the MODELS that can do the thinking. You need a lot of ENERGY to make it possible for those chips to run. Because they use a lot of energy.

Those are important things in the battle for global dominance. China probably wins on data because they are a bit smarter with data processing and have less to do with privacy. Computing power, chips, the quality of chips, the amount of chips, probably goes to the United States because of those Nvidia chips. The models probably goes the United States too. Energy is not entirely clear.

Now one week later. The Americans are still going to invest lot of extra money in computing power and in the models. Although DeepSeek shows that you may need much less advanced computing power, and if you simply build on the models that are already there in the public domain (open source) then you can develop a new model much faster than anyone had previously thought. That is also much cheaper. This actually means that the thresholds to create these AI services, which are also very important for these companies to be able to make so much profit, are much lower than we first thought. That is of course a disadvantage for all those companies that have been working on this up to now. See how they lost a lot of money already.

Then, perhaps no one will have that dominant position anymore. It's an advantage of the rest of the world. If you can suddenly make those models cheaper and easier and use less advanced chips, then other countries or continents might be able to make models themselves. Everyone can design models very quickly - Then the price will drop and we will make more use of them with opportunities for the rest of the world countries as well.

The political significance of this might be even greater than the economic one. The US immediately gave a very clear message here: "America, we are going to do everything we can to become the winner in AI". If you want to prevent China from winning this race, then the United States must ensure that China cannot gain access to these chips and models. At all costs. Even if that is economically disadvantageous. It is protectionism. Although, some of the people in the government in the United States will say: "Let's not do any protection now, because we have already proven that it doesn't work." DeepSeek (China) somehow did have access to the chips. That debate is completely open in that respect.

China will then take political and economic countermeasures. That could also be favorable for other countries and continents if the United States and China are fighting so hard politically/economically. I don't think it will come to that though but I do like where this is going.

**boredjedi** · 29 January 2025, 23:58

Oppssss

DeepSeek in Deep Trouble! (Internal database exposed)

**possessor** · 30 January 2025, 00:00

Someone ask the AI about what happened in June 1989

**chuft** · 30 January 2025, 01:05

As I mentioned earlier, DeepSeek was trained using tokens from OpenAI's AI. This is called "distillation." It cheated. It was not trained entirely on raw data. It essentially asked ChatGPT millions of questions and got the answers and used those questions and answers to train DeepSeek. Training it from scratch would be much more costly. What they did violated the Terms of Service for ChatGPT.

https://www.inc.com/ben-sherry/opena...-work/91140698

So I would not make the mistake of thinking the Chinese found some amazing efficiency. The efficiency in this case was cheating and stealing, and then lying about it, as is so often the case with what China does. Always be skeptical of what the Chinese say.

Distillation is normally used inside a company to train newer models using older ones, with full access because it is internal. The question facing US AI companies is how to make their AIs available for use without opening them to distillation from competitors.

The irony here of course is that all AI is based on stealing, namely other people's copyrighted works from the internet. So some people are snickering about it.

https://www.msn.com/en-us/money/othe...ai/ar-AA1y4W8Y

**LazyPooky** · 30 January 2025, 01:25

I read what you wrote about distillation, and I know what it is. I've just pointed out the days before and just after DeepSeek's introduction, before it really became clear how they got these models and data 'distillated'.

Nothing is safe on the internet and no one can be trusted. I learned that years ago.

**chuft** · 30 January 2025, 12:27

Another interesting article.

https://darioamodei.com/on-deepseek-and-export-controls

It does not take distillation into account but discusses a lot of other things relating to this insane race to the cliff.

**LazyPooky** · 30 January 2025, 13:17

Originally posted by chuft

Another interesting article.

https://darioamodei.com/on-deepseek-and-export-controls

It does not take distillation into account but discusses a lot of other things relating to this insane race to the cliff.

It is good to see it from another angle. Interesting to read about shifting the curve. The article focused a lot on the export controls, which is a political decision. You can not determine in advance whether it works, as I described in my previous post there are also opponents who will say that it does not work. But according to this article the chips, mainly H800, are not illegally obtained so the discussion about export controls on those chips is not useful, and only applies to the chips that are not allowed to be exported to China. If done right, China will be stuck with the older H800 chips. I'm still hoping other countries and continents will join the 'race', because that would be a good opportunity now.

The more I read about it the more I think that the rumor about distillation has been spread mainly by OpenAI, Microsoft and The White House. So OpenAI, a company that has been obtaining large amounts of data from all of humankind largely in an unauthorized manner and, in some cases, in violation of the terms of service of those from whom they have been taking from, is now complaining that they have been robbed of their data. The story is ironic, the company is a hypocrite. 😋

**chuft** · 30 January 2025, 20:28

Well yes, if you recall I posted

Originally posted by chuft

The irony here of course is that all AI is based on stealing, namely other people's copyrighted works from the internet. So some people are snickering about it.

Oh Dear, Did Someone Steal Something From OpenAI?

**LazyPooky** · 30 January 2025, 20:46

Originally posted by chuft

Well yes, if you recall I posted

yes. I know you posted that. That doesn't mean you've claimed that, so I can't say it.
I made the sentence to use irony vs hypocrisy again, in the end.

DeepSeek becomes the top-rated free application

DeepSeek becomes the top-rated free application

Note

Note

Note

Note

Note

Note

Note

Note

Note

Note

Note

Note

Note

Note

Related Topics