Support NeoGAF

[A.Romero]

lol . This is like the SuperZip of AIs. An AI program that’s probably actually using other AI sources and just scraping. Ofcourse it uses less resources.

“Unlike many Chinese AI firms that rely heavily on access to advanced hardware, DeepSeek has focused on maximizing software-driven resource optimization,” explains Marina Zhang, an associate professor at the University of Technology Sydney, who studies Chinese innovations. “DeepSeek has embraced open source methods, pooling collective expertise and fostering collaborative innovation. This approach not only mitigates resource constraints but also accelerates the development of cutting-edge technologies, setting DeepSeek apart from more insular competitors.”

How Chinese AI Startup DeepSeek Made a Model that Rivals OpenAI

When Chinese quant hedge fund founder Liang Wenfeng went into AI research, he took 10,000 Nvidia chips and assembled a team of young, ambitious talent. Two years later, DeepSeek exploded on the scene.

www.wired.com

Can't read the article because of the paywall but just going by the quote, it doesn't sound like it's just scraping. For example, scraping wouldn't help with reasoning nor solving mathematical problems.

Just a few posts above yours Bernoulli posted a benchmark posting how it compares to other models with apparently impressive results (check this resource if you want to learn more about AI benchmarks: https://hai.stanford.edu/what-makes-good-ai-benchmark)

If you want to know more about deepseek, you can check the paper they released together with the model: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf

 

[StereoVsn]

Forget the coders. That NVidia CES presentation was nightmare fuel. AI cameras watching warehouse workers. AI robot chef arms.

If robots can efficiently do skilled labor jobs like a chef, we are so fucked as a species. It’s gonna eliminate a ton of middle class jobs and make billionaires more rich.

And you know how everyone now goes “Go to trade school “. Guess who these robots will stop replacing…

Well, and not to mention millions of white color folks.

But this particular sell off is kind of stupid hysteria.

The Chinese model was based on ChatGPT/Meta models. This isn’t a small startup but a bunch of quants using a lot of … Nvidia cards.

Yea, model was cheaper to train, but it’s not nearly as good as latest from OpenAI/Gemini/Anthropic.

I am seriously wondering if the hysteria is for some people to make bunch of $$$ on shorts and then buy up tech to make more on the way back up.

As well as to tank financial capabilities of competitors. Musk got to love this.

 

[Kilau]

 

[TVexperto]

I dont get it. We can use both OpenAI as well as deepseek here in Germany?!

 

[Fabieter]

Nvidia is done for if its true what the Chinese claim. They are gonna drop another 50% if it's true.

 

[StereoVsn]

I dont get it. We can use both OpenAI as well as deepseek here in Germany?!

I think the point is that EU in general is way behind in AI research. There is Mistral (I think that’s the name) but it’s nowhere near the front runners.

Separately as far as gaming goes, by far more interesting development is new Tencent model that turns images and text prompts into full on 3D models that you can use for various purposes.

Read up on Hunyuan3D, it’s pretty damn wild. I feel sorry for 3D artists .

Edit: It’s also free and on Hugging Face and GitHub.

 

[ChuckeRearmed]

I like that the bottle cap is literally his brain

There is Mistral (I think that’s the name) but it’s nowhere near the front runners.

Anecdotally I think it would have achieved much more if it was based in USA.

 

[Bernoulli]

They just released open source image generation

 

[StereoVsn]

I like that the bottle cap is literally his brain


Anecdotally I think it would have achieved much more if it was based in USA.

Maybe? It’s not really clear what they bring to the table really. I think the whole point was that this was going to be European OpenAI and got bugged down in usual EU bureaucracy.

I actually would have preferred that everywhere vs current no holds barred crazy Wild West that may kill us all yet. But it is what it is and EU needs to get off its ass.

 

[Fess]

In other words, time to buy some stocks again soon

 

[StereoVsn]

They just released open source image generation

GitHub - Tencent/Hunyuan3D-2: High-Resolution 3D Assets Generation with Large Scale Hunyuan3D Diffusion Models.

High-Resolution 3D Assets Generation with Large Scale Hunyuan3D Diffusion Models. - Tencent/Hunyuan3D-2

github.com

Here is Tencent’s 3D models, also free.

And Google made their 2.0 model with up to 1mil context free (not open source) as well just recently.

We are really in crazy times. I recently had to write and test some scripts for a deployment and it would have taken me a few days previously. I literally got this done in a couple of hours.

I have no idea what young folks will do for work in 5-10 years.

 

[GHG]

No one is "waiting on" better hardware, the megacaps are simply spending exponentially more each turn (yes, obviously this will hit a wall sooner or later). And, yes, there are also the 3-4x generational leaps in hw every couple of years.

It's cool that we got a one-time improvement like this, but the vast majority of compute in the future is going towards test-time inference compute. We don't even know the inference cost/performance curve of DeepSeek-R1... more data is needed.

That's exactly the point. They've constantly been chucking money at it with little to no financial pay-off. Ultimately that's the only thing that matters.

Look, you're hoping for cheap GPUs and that's fine

I'm actually pretty happy with what I have right now but thanks for your concern.

But you're way off with "the time of fully equipped AI server racks is gone", like completely delusional.

They aren't going to be selling them at the volume they have been up until now. That's for sure.

Has nothing to do with Nvidia or AMD. Tech firms will pick the best inference available and scale it as high as they can with whatever power they have available to make it run, nuclear power and all.

Can the shovel seller switch hands in the future? Sure. But as of now, even AMD for free is still more expensive than Nvidia for these firms when they consider the whole picture of total cost of ownership.

Subject to change. It doesn't even need to be AMD either by the way, but their tech will likely become far more viable as a result of this.

You always need an interlink for any massive GPU scaling. What are you even on about. You're talking about things you don't understand. DeepSeek doesn't change GPU scalability. Its incredible the misinformation.

Oh look

https://www.amd.com/en/developer/resources/technical-articles/amd-instinct-gpus-power-deepseek-v3-revolutionizing-ai-development-with-sglang.html

AMD Instinct™ GPUs accelerators are transforming the landscape of multimodal AI models, such as DeepSeek-V3, which require immense computational resources and memory bandwidth to process text and visual data. AMD Instinct™ accelerators deliver outstanding performance in these areas.

Requirements for bandwidth and computational ressources not going away.

Oh amazing, using AMD's marketing are we? How does their bandwidth compare to the bandwidth on Nvidia's solutions? It is there that you will discover my point instead of attempting to twist my words as if I said bandwidth is no longer a consideration.

Tech firms forever have managed to do without Cuda, how else is AMD selling anything?
That's not what they buy Nvidia for, not the big tech firms at least. Even OpenAI which uses Nvidia GPUs has its inhouse solution, Triton.
When you're that big you have the software inhouse. Cuda again, you're thinking advantages like someone making hentai porn in his basement.

Jesus christ. Nvidia currently have a 85%+ market share in the AI space, that doesn't equate to AMD selling something. As for Triton, it's ~80% less efficient than Cuda, why do you think that is? What does it ultimately convert to when running it on a Nvidia GPU?

Deepseek are using their own low level solution, it does not have any reliance on Cuda at any level - hence the efficiency gains.

But yes, anyone who sees the benefit of what's been introduced here and understands the threat this poses to Nvidia's business specifically is a "midwit" who likes to make hentai porn in their basement. Perfect summary.

 

[Guardian Monkey]

Can't read the article because of the paywall but just going by the quote, it doesn't sound like it's just scraping. For example, scraping wouldn't help with reasoning nor solving mathematical problems.

Just a few posts above yours Bernoulli posted a benchmark posting how it compares to other models with apparently impressive results (check this resource if you want to learn more about AI benchmarks: https://hai.stanford.edu/what-makes-good-ai-benchmark)

If you want to know more about deepseek, you can check the paper they released together with the model: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf

Fair enough, It’s already known it’s trained by using other models and known it was created for speed optimization. You can ask it that and it will tell you that itself. It won’t tell you what other ai models were used but will refer you to a faq. But what is its scalability with the more complex routines. Are we saying that the optimization was so bad in the other models that a bunch of recent graduates that were 1 to 2 years out of school were able to produce something better. I think what we are seeing is that a.i. models will all be specialized in routine training and usage. Also what validation have we on how much was really used to train and operate.

 

[Sentenza]

It was bound to happen.

It's bound to bounce back.
Especially when people will realize that even Chinese software will need the hardware to run on and Nvidia in the one positioned to offer the best one.

 

[ChuckeRearmed]

I actually would have preferred that everywhere vs current no holds barred crazy Wild West that may kill us all yet.

So what do you prefer?

I personally despise EU bureaucracy and prefer unrestrxistricted unlimited pursuit towards the sky.

I have no idea what young folks will do for work in 5-10 years.

Ask for advice from seniors because they won't have any idea how anything works

 

[ReBurn]

Sounds like a good time to invest in some of these dropping tech stocks. When everything shakes out they'll climb again.

 

[StereoVsn]

So what do you prefer?

I personally despise EU bureaucracy and prefer unrestrxistricted unlimited pursuit towards the sky.

Personally I would prefer a world that takes a much more measured approach to potentially dangerous new technology.

But we live in a world where other powers don’t care and are going full tilt.

EU has issues with bureaucracy, no question, but that also brings more level headed measurement of new tech. Somewhere between US recklessness and EU bureaucracy walls would be good.

 

[MarV0]

I have no idea what young folks will do for work in 5-10 years.

Get high and play video games on universal income.

 

[Bernoulli]

 

[StereoVsn]

Get high and play video games on universal income.

Maybe in EU and considering budget issues, need to spend on infrastructure, defense, and falling exports; this is a big if.

US and universal income is lol-worthy. But yeah, we are heading into full on dystopia, with huge corporations controlling everything (more than now) and world in the shitter with climate change.

 

[Guardian Monkey]

there we go.

 

[Buggy Loop]

They aren't going to be selling them at the volume they have been up until now. That's for sure.

Jesus Christ. Talk about missing the timeline of AI.

So cute

Yup. LLM is the end game. That's it. AI stops there. Was a good run. OpenAI and Stargate was just aiming for youtube AI videos.

Subject to change. It doesn't even need to be AMD either by the way, but their tech will likely become far more viable as a result of this.

Best - inference - no - matter - the - vendor

DeepSeek-V3 is dependent as much as any previous model before it on inference speed.

Oh amazing, using AMD's marketing are we? How does their bandwidth compare to the bandwidth on Nvidia's solutions? It is there that you will discover my point instead of attempting to twist my words as if I said bandwidth is no longer a consideration.

AMD always claims higher bandwidth than Nvidia. The opposite of your DeepSeek claim you made that "no longer need to make sure you have access to the massive bandwidth Nvidia". I discover that you do not know what you're talking about.

Jesus christ. Nvidia currently have a 85%+ market share in the AI space, that doesn't equate to AMD selling something.

your maths are something else

As for Triton, it's ~80% less efficient than Cuda, why do you think that is? What does it ultimately convert to when running it on a Nvidia GPU?

Less efficient for what GHG? Which tasks are you trying to program? Or benchmark?

Unsloth AI last year raised 2.2x inference in LLM with 70% less RAM usage by using converting sequential code paths into parallel ones using Triton kernels.

Is your whole premise of Nvidia getting 85% of market share because of Cuda? Look at the big tech firms. None are using cuda. You're missing the whole reason why Nvidia is selling like it does. Total cost of ownership. Not Cuda. Even with AMD getting better theoretical number on papers, their racks and implementation at these big AI farm levels is total shit.

Deepseek are using their own low level solution, it does not have any reliance on Cuda at any level - hence the efficiency gains.

That's nothing new and has nothing to do with inference speed or computational parallelism. They went to the metal with near assembly language on GPU. What does that have to do with GPU vendor? There's a mountain pile of alternatives to Cuda.

Peoples pick Cuda / Pytorch not for the best performance, they pick it for ease of implementation and easy to code. Do you have any idea how it is to implement DeepSeek? No you don't. It doesn't remove the needs of broad and easy coding languages. This is equivalent to saying that we should never have made an API because assembly. Sure, nothing beats that, but good luck coding.

But yes, anyone who sees the benefit of what's been introduced here and understands the threat this poses to Nvidia's business specifically is a "midwit" who likes to make hentai porn in their basement. Perfect summary.

Yup

Also a business dimwit that has never heard of the Jevons paradox. Its quite cute.

 

[StereoVsn]

Peoples pick Cuda / Pytorch not for the best performance, they pick it for ease of implementation and easy to code. Do you have any idea how it is to implement DeepSeek? No you don't. It doesn't remove the needs of broad and easy coding languages. This is equivalent to saying that we should never have made an API because assembly. Sure, nothing beats that, but good luck coding.

Yeah, using CUDA means easy access to support, developers who are familiar with CUDA, a ton of resources and more. It’s not the most efficient but it’s so much easier to hit the ground running with it.

 

[AmuroChan]

Corrections are healthy for the market. Also, I always welcome days like this where I can buy the dip.

"be fearful when others are greedy and be greedy when others are fearful.” - Warren Buffett

 

[mckmas8808]

It's quite convenient that Trump gets into office with this tech bro support, he announces an initiative for AI development, which is fine, and then IMMEDIATELY after, this Chinese model arises out of nowhere that is somehow vastly superior to what the tech bros have been doing.

This is space race, "missile gap", Cold War type shit all over again. People need to smarten up!

That annoucement was just that. An annoucement. Project Stargate offically started in March of 2024. It was already known. There's no conspiracy here. The Chinese right now are just more efficient at A.I. coding than us in America.

 

[Buggy Loop]

Also, as an electrical engineer working for 20 years in power distribution, the number of Chinese papers with wild claims being quite bullshit is not lost on me (among many other countries in that region of the world). The benchmark is one thing, but training with H800 and $6M budget is what they want you to believe.

In the meantime :

https://wccftech.com/chinese-ai-lab-deepseek-has-50000-nvidia-h100-ai-gpus-says-ai-ceo/

“The Bureau of Industry and Security (BIS) of the U.S. Department of Commerce banned the export of Nvidia’s A100 and H100 chips to China in October 2022. This prohibition was extended to the A800 and H800 chips in the following year. However, the results of these restrictions have been unsatisfactory. Some of the chips have been privately imported into China through shell companies.“

https://thediplomat.com/2024/06/4-ways-china-gets-around-us-ai-chip-restrictions/

Yea yea, totally not H100

They respected the commerce laws

By using H800

China would never break commerce laws

 

[readonly]

None of this makes sense to me. Okay it's 20x more efficient. Then they will write something that needs 20x more power. This almost feels like the bill gates we will never need more than 640K memory. Once you give someone more power they will find a way to utilize it.

 

[mckmas8808]

The arms race looks a lot different now though.

Due to the open source nature of deepseek it means companies can adapt and intergrate it at will. You no longer need to use Nvidia hardware, you no longer need to go to one of the big players for an AI solution, and you certainly don't need huge nuclear power plants running in the background feeding it energy in order for it to be viable.

Yes you can sit there with millions of Nvidia chips and say you have the most capacity - but who exactly are you selling to and at what price?

Some aren't ready for what you're saying right now. This will become a China vs. America thing and sides will be taken. But you are right. The race is going to look a lot different going forward.

 

[Rat Rage]

Everything related to A.I. keeps on being terrible. A.I. will fuck you all. Even the billionaires, because once they start replacing many jobs with A.I., society will have less and less purchasing power, because many people will lose their income. Capitalism needs customers, though. Once people can't afford shit anymore, it doesn't matter if there are companies who can produce huge amounts of goods with minimal human labour, if there is no one who can afford to buy them.

This change won't come overnight. It will happen gradually, as gradually as the civil unrests will happen.

Seriously, fuck A.I.. Ya'll better start anti-A.I. movements, before it's too late.

 

[mckmas8808]

None of this makes sense to me. Okay it's 20x more efficient. Then they will write something that needs 20x more power. This almost feels like the bill gates we will never need more than 640K memory. Once you give someone more power they will find a way to utilize it.

That's the correct way to understand it. YES! You got it! So why do you sound confused?

 

[mckmas8808]

Everything related to A.I. keeps on being terrible. A.I. will fuck you all. Even the billionaires, because once they start replacing many jobs with A.I., society will have less and less purchasing power, because many people will lose their income. Capitalism needs customers, though. Once people can't afford shit anymore, it doesn't matter if there are companies who can produce huge amounts of goods with minimal human labour, if there is no one who can afford to buy them.

This change won't come overnight. It will happen gradually, as gradually as the civil unrests will happen.

Seriously, fuck A.I.. Ya'll better start anti-A.I. movements, before it's too late.

U.B.I. Learn it and remember it well. Universal.....Basic......Income!

The elites will have no answer, but to implement UBI laws all across the world. Even in America. Unless......they want chaos and outrage.

 

[Guardian Monkey]

The arms race looks a lot different now though.

Due to the open source nature of deepseek it means companies can adapt and intergrate it at will. You no longer need to use Nvidia hardware, you no longer need to go to one of the big players for an AI solution, and you certainly don't need huge nuclear power plants running in the background feeding it energy in order for it to be viable.

Yes you can sit there with millions of Nvidia chips and say you have the most capacity - but who exactly are you selling to and at what price?

They were limited to nvidia because of price versus performance. AMD didn’t just magically become viable in that sense, even if you believe the new ai model is more efficient nothing changes. There will still be a massive deprivation of computational output that is needed and Nvidia is still the leader there.

 

[MarV0]

Corrections are healthy for the market. Also, I always welcome days like this where I can buy the dip.

"be fearful when others are greedy and be greedy when others are fearful.” - Warren Buffett

LMAO this is fearful? When dumb comments such as "bought the dip" are nowhere to be seen and everyone is shitting their pants because they literally lost their life savings you'll know what fear is.

Nvidia's valuation before the AI bubble was $2. Let that sink in. You think it can't happen? Then take a look at Cisco before and after the dot.com bubble burst.

 

[readonly]

That's the correct way to understand it. YES! You got it! So why do you sound confused?

The confusion is why people think this will ultimately effect Nvidia and its ability to sell as many video cards as they can produce. If they were actually worth 3.4trillion before then they are still worth it today. This news shouldn't change that.

 

[dorkimoe]

There is nothing more american than this

The company said it had spent just $5.6 million on computing power for its base model, compared with the hundreds of millions or billions of dollars US companies spend on their AI technologies.

 

[Buggy Loop]

None of this makes sense to me. Okay it's 20x more efficient. Then they will write something that needs 20x more power.

Yup

For 160 years this has been observed in many technological advancements, power being one of them.

Jevons paradox - Wikipedia

en.wikipedia.org

The roadmap of AI is so far into the future and needs so much computational work that DeepSeek is like a water drop in the ocean of the whole picture.

Also its "free" and good compared to "free" models, but the API logs and trains on data. Kind of wild. Even the non-paid versions of GPT and Claude let you turn that off. This will break TOS of many companies and countries.

 

[Hawking Radiation]

Christ on a bike, China stays winning. They're already turning the auto industry on its head.

 

[mckmas8808]

Yup

For 160 years this has been observed in many technological advancements, power being one of them.

Jevons paradox - Wikipedia

en.wikipedia.org

The roadmap of AI is so far into the future and needs so much computational work that DeepSeek is like a water drop in the ocean of the whole picture.

Also its "free" and good compared to "free" models, but the API logs and trains on data. Kind of wild. Even the non-paid versions of GPT and Claude let you turn that off. This will break TOS of many companies and countries.

What do you mean this will break TOS of many companies?

 

[ChuckeRearmed]

I have no idea what young folks will do for work in 5-10 years.

Ask for advice from seniors because they won't have any idea how anything works.

but to implement UBI laws all across the world

Never going to work. There was UBI in the past - when everybody was a peasant and grow their own food and could be self sustained. And even that relied a lot on high mortality rate and weather. The moment there was less rain or sun, you got deaths and famine.

Just like with people advocating for communism, social safety nets do not work the moment you lose access to cheap resources or you have the overwhelming amount of non-contributing folks. We see it from the Europe and Canada. It is cool to pay everybody nice money, but the moment the company goes down those same people become a burden on a budget. And trying to keep the company afloat just drags down the economy.

EU has issues with bureaucracy, no question, but that also brings more level headed measurement of new tech

Nah, EU's motto - "regulate before it achieves anything".

Personally I would prefer a world that takes a much more measured approach to potentially dangerous new technology.

This approach does not work in a world where not everybody is playing by the rules.

 

[mckmas8808]

Never going to work. There was UBI in the past - when everybody was a peasant and grow their own food and could be self sustained. And even that relied a lot on high mortality rate and weather. The moment there was less rain or sun, you got deaths and famine.

Just like with people advocating for communism, social safety nets do not work the moment you lose access to cheap resources or you have the overwhelming amount of non-contributing folks. We see it from the Europe and Canada. It is cool to pay everybody nice money, but the moment the company goes down those same people become a burden on a budget. And trying to keep the company afloat just drags down the economy.

I agree UBI can't work for the reasons you presented. But it's the only thing we will have LOL! What other solution will there be? If many won't implement HUGE regulations to protect human societies?

 

[nikos]

Buy the dip?

 

[cinnamonandgravy]

2025 getting spicy

will be interesting to see people attempt to verify the claims and what comes of it
wouldnt be surprising if a bunch of extra nvidia hardware and govt funding is found

hope it's 100% legit though--fun to see progress

 

[AmuroChan]

LMAO this is fearful? When dumb comments such as "bought the dip" are nowhere to be seen and everyone is shitting their pants because they literally lost their life savings you'll know what fear is.

Nvidia's valuation before the AI bubble was $2. Let that sink in. You think it can't happen? Then take a look at Cisco before and after the dot.com bubble burst.

If a person lost their life savings over one stock, they did something seriously wrong. $NVDA has been a great play for me in the past 18 months (6 figures in realized gains and more than that in unrealized gains). I'm happy to buy the dip here as I needed to replenish the shares I've sold off anyway.

 

[ChuckeRearmed]

I agree UBI can't work for the reasons you presented. But it's the only thing we will have LOL! What other solution will there be? If many won't implement HUGE regulations to protect human societies?

UBI will be available in very small areas of the world at best. Maybe in some small small countries with low populations like within city-border and so on. But overall you won't see it on a country-wise basis. You just cannot maintain any social safety net due to non-existence of anything that it is unlimited (whether it is a success or a resource).

There are two trends developing right now in the world - some areas are becoming heavily urbanized to filled with so many people that it is unrecognizable with the rest of the country. And you have growing rural areas that develop independently. With new tech like drone delivery, more effective farming etc. It is going to be very interesting society in the future.

 

[mckmas8808]

UBI will be available in very small areas of the world at best. Maybe in some small small countries with low populations like within city-border and so on. But overall you won't see it on a country-wise basis. You just cannot maintain any social safety net due to non-existence of anything that it is unlimited (whether it is a success or a resource).

There are two trends developing right now in the world - some areas are becoming heavily urbanized to filled with so many people that it is unrecognizable with the rest of the country. And you have growing rural areas that develop independently. With new tech like drone delivery, more effective farming etc. It is going to be very interesting society in the future.

So what do you believe the long term solution (Year 2040+) will be when AGI is reached in A.I.? And then combining that with robotics that will be physically working in your average store in the 2030s?

 

[Buggy Loop]

What do you mean this will break TOS of many companies?

Most companies would run a model but not share the data they send for further training.

 

[cartman414]

And you know how everyone now goes “Go to trade school “. Guess who these robots will stop replacing…

Lol at that whole argument. Like everyone can physically do manual labor or go through an apprenticeship.

 

[Topher]

There is nothing more american than this

The company said it had spent just $5.6 million on computing power for its base model, compared with the hundreds of millions or billions of dollars US companies spend on their AI technologies.

Nothing more American than massive corporations spending lots of money? Don't know much about America, do you?

 

[dorkimoe]

Nothing more American than massive corporations spending lots of money? Don't know much about America, do you?

How you managed to miss the point is pretty astonishing.

 

[Topher]

How you managed to miss the point is pretty astonishing.

Enlighten me.