Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Deepseek: The quiet giant leading China’s AI race

chinatalk.media

Post reply on HN
Print view View original

11–20 of 476 posts

Re: Deepseek: The quiet giant leading China’s AI race

#11

Post by rahimnathwani »

NitpickLawyer wrote: post #4
mentalgear wrote: post #3

Impressive to think about how DeepSeek achieved: ~ Parity with o1 and Claude with > 10x less resources. Better algorithms and approaches are what's needed for the next step of ML.

While impressive, the deepseek models aren't really "on par" with either oAI or Anthropic offerings, right now. The models seem to be a bit overfitted in the post-training step. They are very "stubborn" models, and usually handle tasks well if they can handle them, but steering them is quite difficult. As a result, they score very well on various benchmarks, but often times perform slightly worse in real-life scenari…

  They are very "stubborn" models
Have you found this to be the case even when using the recommended temperature settings (ranging from 0 for math, to 1.5 for creative tasks)?

Re: Deepseek: The quiet giant leading China’s AI race

#12

Post by wodderam »

lomkju wrote: post #6

I feel the GPU restrictions created an environment for Chinese Devs to be more innovative and do more with less. Kudos to the deepseek team!

Kai-Fu Lee describes the culture so well in AI Superpowers. The roots are well before GPU restriction. Absolute cut throat competition.

Imagine Sam Altman throwing a chair out a window in a meeting lol.

The message of AI Superpowers is that China will lag the US at first but once things stabilize this will happen because China has a lot more engineers and a lot more data.

Anyone who hasn't read AI Superpowers should really make it a point to read it in 2025. It is an incredible book.

Re: Deepseek: The quiet giant leading China’s AI race

#13

Post by nsoonhui »

I find that the gushing around deepseek is fascinating to watch.

To me there are a few structural and fundamental reasons why deepseek can never outperform other models by a wide margin. On par maybe--as we reach the diminishing returns with our investment in the models, but not win by a wide margin.

1. The US trade war with china which will place deepseek compute availability at disadvantages, eventually, if we ever get to that.

2. China censorship which limits the deepseek data ingestion and output, to some degree.

3. Most importantly, deepseek is open source, which means that the other models are free to copy whatever secret source it has, eg: Whatever architecture that purportedly use less compute can easily be copied.

I've been using Gemini, chatgpt, deepseek and Claudie on regular basis. Deepseek is neither better or worse than others. But this says more about my own limited usage of LLM rather than the usefulness of the models.

I want to know exactly what makes everyone thinks that deepseek totally owns the LLM space? Do I miss anything?

PS: I am a Malaysian Chinese, so I am certainly not "a westerner who is jealous and fearful of the rise of China"

Re: Deepseek: The quiet giant leading China’s AI race

#15

Post by logicchains »

nsoonhui wrote: post #13

I find that the gushing around deepseek is fascinating to watch. To me there are a few structural and fundamental reasons why deepseek can never outperform other models by a wide margin. On par maybe--as we reach the diminishing returns with our investment in the models, but not win by a wide margin. 1. The US trade war with china which will place deepseek compute availability at disadvantages, eventually, if we ever…

>I want to know exactly what makes everyone thinks that deepseek totally owns the LLM space?

It achieved competitive performance to the competition at literally 10x less cost of production (training). That's an incredible achievement in any industry, especially given they have such a small team relative to competitors. Their API is 20-50x cheaper than the competitors, and not because they're burning cash by charging less than costs, but rather because their architecture is just that much more efficient.

They already achieved the above in spite of sanctions limiting their availability to top-tier GPUs, and the gap between Chinese domestic GPUs and NVidia is getting smaller and smaller, so in future the GPU disadvantage will be less and less.

Re: Deepseek: The quiet giant leading China’s AI race

#16

Post by NitpickLawyer »

rahimnathwani wrote: post #11
NitpickLawyer wrote: post #4

Earlier quoted context omitted.

While impressive, the deepseek models aren't really "on par" with either oAI or Anthropic offerings, right now. The models seem to be a bit overfitted in the post-training step. They are very "stubborn" models, and usually handle tasks well if they can handle them, but steering them is quite difficult. As a result, they score very well on various benchmarks, but often times perform slightly worse in real-life scenari…

They are very "stubborn" models Have you found this to be the case even when using the recommended temperature settings (ranging from 0 for math, to 1.5 for creative tasks)?

I use 0.05 for math, just did a 5k problem set, trying to fine-tune a smaller model with the outputs. It has some very interesting training, borrowed from r1 per the tech report, where it does the o1/qwq "thinking steps", but a bit shorter. It solves ~80% of the problems in 4k context, while qwq would go on for 8k-16k. It's very good at what it does.

But as soon as I need it to do something other than solve a problem - say rewrite the problem in simpler terms, or given a problem + solution provide hints, or rewrite the solution with these , etc. it kinda stops working. Often times it still goes ahead and solves the problem. That's why I'm saying it's stubborn. If a task looks like a task that it can handle very well, it's really hard to make it perform that other, similar but not quite the same task.

In a similar vein - https://github.com/cpldcpu/MisguidedAttention/tree/main/eval...

Re: Deepseek: The quiet giant leading China’s AI race

#17

Post by nsoonhui »

logicchains wrote: post #15
nsoonhui wrote: post #13

I find that the gushing around deepseek is fascinating to watch. To me there are a few structural and fundamental reasons why deepseek can never outperform other models by a wide margin. On par maybe--as we reach the diminishing returns with our investment in the models, but not win by a wide margin. 1. The US trade war with china which will place deepseek compute availability at disadvantages, eventually, if we ever…

>I want to know exactly what makes everyone thinks that deepseek totally owns the LLM space? It achieved competitive performance to the competition at literally 10x less cost of production (training). That's an incredible achievement in any industry, especially given they have such a small team relative to competitors. Their API is 20-50x cheaper than the competitors, and not because they're burning cash by charging…

But like I said, deepseek is open source so why can't the competitors copy whatever source that makes the cost of production 10x cheaper ?

Re: Deepseek: The quiet giant leading China’s AI race

#18

Post by viraptor »

nsoonhui wrote: post #13

I find that the gushing around deepseek is fascinating to watch. To me there are a few structural and fundamental reasons why deepseek can never outperform other models by a wide margin. On par maybe--as we reach the diminishing returns with our investment in the models, but not win by a wide margin. 1. The US trade war with china which will place deepseek compute availability at disadvantages, eventually, if we ever…

> The US trade war with china which will place deepseek compute availability at disadvantages

Will it? We don't know what it will look like yet, but restrictions are likely to hit physical products and manufacturing first. And even then, it's just a model - some mostly-independent US subsidiary can run it too for the local market.

> China censorship which limits the deepseek data ingestion

Deepseek has been improving through training, architecture, and features. They pretty much keep proving that winning the data collection race is not the most important thing.

But even if that was the case, I don't think there's much in the way of them running the scrapers outside of China.

> Most importantly, deepseek is open source,

OpenAI relies on burning cash and creating huge, expensive models. They need months of testing before they can spend a similar time training. Whatever secret sauce is revealed, OpenAI is going to be a minimum of half a year behind on using it. (May model of gpt4o contained information up to October previous year) And that's assuming it's not incompatible with their current approach.

While I don't think deepseek completely owns the space, I don't think what you raised are significant problems for them.

Re: Deepseek: The quiet giant leading China’s AI race

#19

Post by tossandthrow »

nsoonhui wrote: post #13

I find that the gushing around deepseek is fascinating to watch. To me there are a few structural and fundamental reasons why deepseek can never outperform other models by a wide margin. On par maybe--as we reach the diminishing returns with our investment in the models, but not win by a wide margin. 1. The US trade war with china which will place deepseek compute availability at disadvantages, eventually, if we ever…

> ... why deepseek can never outperform ...

This read more like a "western supremacists" post.

1. Only until China produces more compute than the west.

2. You don't have to ask ChatGPT / Claude many questions before realizing the grave censorship these are under - DeepSeek has access the roughly the same corpus of data as their western counter parts.

3. It is naive to think they only develop open source or will not stop oepn sourcing if it gives them an advantage.

Re: Deepseek: The quiet giant leading China’s AI race

#20

Post by arisAlexis »

logicchains wrote: post #15
nsoonhui wrote: post #13

I find that the gushing around deepseek is fascinating to watch. To me there are a few structural and fundamental reasons why deepseek can never outperform other models by a wide margin. On par maybe--as we reach the diminishing returns with our investment in the models, but not win by a wide margin. 1. The US trade war with china which will place deepseek compute availability at disadvantages, eventually, if we ever…

>I want to know exactly what makes everyone thinks that deepseek totally owns the LLM space? It achieved competitive performance to the competition at literally 10x less cost of production (training). That's an incredible achievement in any industry, especially given they have such a small team relative to competitors. Their API is 20-50x cheaper than the competitors, and not because they're burning cash by charging…

Of course if you arrive last and copy all the existing architecture you can train it cheaper
Post reply on HN