Support NeoGAF

[GHG]

"speculative crash"

The tool is still being assembled, and everyone wants to own a part of it.

I'd go a step further and say they are still figuring out the tool that is being assembled.

This will be the first of many instances of the apple cart being upset over the next several years.

 

[viveks86]

It would actually be more concerning if today's move in Nvidia's stock (and other impacted semiconductor stocks) wasn't actually because of deepseek.

I'm not seeing any other reason for it and would put the delay down to the claims needing to be verified because of their origin.

Based on today's activities, the benefactors of this news all mostly recovered or even ended green.

My tinfoil hat is vibrating. What if there was a concerted effort to short-sell and then repurchase by the same people? A ton of people could be getting rich over a blip this big.

 

[Mr1999]

Someone correct me if I'm wrong, but I've read here and there that the responses Deepseek generates aren't as accurate/high quality as at least the paid ChatGPT stuff?

I had a question that i would use to stump chatgpt, and even though the general idea of the answer was right, it wouldn't quite get the true meaning of it. I cant really say what the question was without doxxing but deepseek gave the correct answer and its something very niche and ethnic, I also learned something new. I'm not sure if upgrading to chatgpt would have helped but i got my answer, but I do remember trying an upgraded version but this was some time ago and the results were similar, but this could have changed since.

 

[Bernoulli]

 

[Cre8]

 

[mckmas8808]

if the chinese AI claims are legit, it means less hardware if required to produce current top-tier AI... which means fewer nvidia GPU sales.

but even if the chinese AI claims are legit (dont know yet), future top-tier AI may require tons of hardware... or it'll run on a potato, who knows.

timing's odd though... US announces $500 trillion on AI investment, then china shows it can be done cheaply. china trying to save US some money? dont think so.

1. The US announcement was BOGUS!!! Even Elon called it BS. They only have $100 Billion secured. And we aren't sure the whole $100 Billion is even there.
2. China was working on Deepseek before that bogus announcement. It's been public knowledge for at least 6 weeks prior.
3. If we could buy China's EV cars......that'll be saving US citizens money right now. But at the moment it's illegal to import Chinese EVs in America.

 

[GHG]

My tinfoil hat is vibrating. What if there was a concerted effort to short-sell and then repurchase by the same people? A ton of people could be getting rich over a blip this big.

That would require a sustained effort in buying which would need to be greater than the volume of selling we've seen today. There are now a ton of people (and funds) "trapped" above where it's currently trading who might look at any rise in the stock as an opportunity to offload (and/or sell calls against their positions).

Almost a billion shares being traded today in Nvidia alone is not to be sniffed at. That kind of volume and sustained selling pressure indicates there were plenty of liquidations on the way down (which has a cascading effect).

Tomorrow it will be short sale restricted so I expect it to be up for sure, but how it finishes the week will tell us more in terms of whether this is the start of a new directional trend (down) or just a major bump in the road.

 

[bender]

Enlighten me.

 

[Buggy Loop]

Bunch of gibberish because GHG cannot ever in his life let go an argument like its 1999 mIRC

I don't have time GHG for your nonsense.

You win

Nvidia down the toilet.

AMD can cancel their infinity fabric AFL

AMD Infinity Fabric AFL Scale Up Competitor to NVIDIA NVLink Coming to Broadcom Switches in PCIe Gen7

AMD's AFL Infinity Fabric scale-up competitor to NVIDIA NVLink is coming to Broadcom switches in the PCIe Gen7 era

www.servethehome.com

USA gonna crash

China won, USA lose

Stargate project in shamble!

Am I missing any other crazy shit you predicts? Peoples are gonna find AGI with a raspberry pi maybe?

 

[mckmas8808]

I think that only really Silicon Valley nerds worry about their lives having no meaning if people don't think they are the most intelligent form of life in the universe.

No man, this isn't true at all. Are you familiar with the A.I.? And this A.I. race?

yup

Not saying it’s a good move forward but realistically, they ain’t slowing down. If it comes to slowing down in public eyes, that’s one thing but behind the curtain they’ll be 100% continuing.

Would you yourself also agree that AGI is something that will be achieved by or around 2030?

Yeah people are getting a bit carried away. This is just usual leapfrogging with technology. My company is still training these AI models to properly read a goddamn pdf file with flowcharts in it. Singularity can (and likely will) happen, but not for a couple of decades. By then I'll retire and be fed by my robot slaves... or be food for them. Whatever they prefer.

I think you are thinking about ASI (Artificial Super Intelligence). Yeah that's probably decades away. But AGI isn't. We are looking at potentially seeing AGI in 2030. But definitely while the PS6 is Sony's main console.

 

[JohnnyFootball]

Im automatically in favor of anything that can take nvidia down a notch in AI.

Once the AI bubble collapses that will come running back to gamers.

 

[viveks86]

I think you are thinking about ASI (Artificial Super Intelligence). Yeah that's probably decades away. But AGI isn't. We are looking at potentially seeing AGI in 2030. But definitely while the PS6 is Sony's main console.

AGI is not singularity though. I was referring to the singularity, when we (will possibly) lose control.

 

[SHA]

It's simple, Turn on frame gen.

 

[rofif]

I think is all fake. All of these nonsense businesses

 

[Laptop1991]

They charged too much for GPU's that resulted in lower sales and put AI above actual gaming performance, it was going to happen at some point, and most of the newer games don;t even look like they need 1 to 2 grand GPU's, just like Ubisoft, self inflicted.

 

[Buggy Loop]

Would you yourself also agree that AGI is something that will be achieved by or around 2030?

Not so sure AGI by 2030. LLMs are not enough. But I mean, this advances so fast, I'm predicting on a hunch really. It doesn't seem to be very far and seems attainable by 2030 but there's probably a few hurdles in the way. The whole training approach might have to change. On top of government roadblocks slowing things down.

 

[fran.fernandez]

If u have spare money left. Buy buy buy

 

[Donald mac Ronald]

What's so good about this AI other than it's Chinese.

 

[FearlessBelgian]

What's so good about this AI other than it's Chinese.

You have the capabilities of Chat GPT pro version for free. (cost 200$/month)

 

[SidViscous]

What's so good about this AI other than it's Chinese.

It's actually open source, and gets amazing results on modest hardware. Nvidia were selling 40/50k servers for running similar models, you can run this model on a Mac Pro at home for fraction of the cost.

The crazy thing is it's being spun as somehow this is a bad thing, a completely open source model that uses far less energy, but we should be worried because it's from china and it's making poor Nvidia shareholders lose money.

 

[kutthoat5150]

This just in, Trump to double China Tariffs even more in an attempt to make American AI great again.

 

[Venom Snake]

It's actually open source, and gets amazing results on modest hardware. Nvidia were selling 40/50k servers for running similar models, you can run this model on a Mac Pro at home for fraction of the cost.

The crazy thing is it's being spun as somehow this is a bad thing, a completely open source model that uses far less energy, but we should be worried because it's from china and it's making poor Nvidia shareholders lose money.

Exactly, hell even Nvidia recognized it.

https://www.cnbc.com/2025/01/27/nvi...eek-r1-model-an-excellent-ai-advancement.html

• Nvidia called DeepSeek’s R1 model “an excellent AI advancement,” despite the Chinese startup’s emergence causing the chip maker’s stock price to plunge 17% on Monday.
• The comments come after DeepSeek last week released R1, which is an open-source reasoning model that reportedly outperformed the best models from U.S. companies such as OpenAI’s.
• Nvidia’s statement indicates that it sees DeepSeek’s breakthrough as creating more work for the American chip maker’s graphics processing units, or GPUs.

The way the market reacted tho..

 

[SHA]

What we see from openAI and other tech AI companies is what they want the public to see. There's models internally that we're not prepared for so USA in panic for world domination of AI is a cute facade, they'll use this to put fear into government and twist arms to force massive money injection into it.

The Tech field is the best field to face AI, who says it's useless to AI don't have a clue.

 

[StereoVsn]

It's actually open source, and gets amazing results on modest hardware. Nvidia were selling 40/50k servers for running similar models, you can run this model on a Mac Pro at home for fraction of the cost.

The crazy thing is it's being spun as somehow this is a bad thing, a completely open source model that uses far less energy, but we should be worried because it's from china and it's making poor Nvidia shareholders lose money.

It’s a bad thing because it came from China, was as always originally based on US tech and there is a good chance that Xi Jinping will be invading or blockading Taiwan within 2-3 years.

Mostly Anything that helps CCP and Xi is a bad thing. In this particular case the technology had to have government approval in China to be released unless the team really wanted to disappear one day. So why would CCP release anything that would help the West?

And personally I think Open Source AI is a Pandora’s box that can’t be put back and now we have hostile actors that are actively using it. And not just CCP.

 

[Wolzard]

Much older and cheaper Nvidia GPUs, showing that you do not need Nvidia's latest and greatest.

H800 is not old, it is exactly the H100, the latest model of Nvidia, but capped in the bandwidth (around 300 Gbps Compared to H100's 600 Gbps) to be sold in the Chinese market.

Open source > proprietary

Just need that code to be forked so it can learn about Tiananmen Square

I find it a little wrong to call open source. We only have the result, which is the model. The training code and data has not been made available , so it is impossible to reproduce and to make sure the alleged cost is true.

The Open Source AI Definition – 1.0

version 1.0 Preamble Why we need Open Source Artificial Intelligence (AI) Open Source has demonstrated that massive benefits accrue to everyone after removing the barriers to learning, using, sharing and…

opensource.org

 

[Buggy Loop]

Here's the founder of StabilityAI, Emad Mostaque, take on all this. He's a tad knowledgeable.

He knows of course very much the scene of distilled models















So yea

The "panic" on media is a psyops at this point. Experts aren't finding some secret recipe out of a quantic universe.

Media ran went with it, what's a bit of panic with your coffee today? While also placing Wallstreet to buy in for cheaper when dumb money gets scared away and government will inject funds in. But don't report on it with true experts in the AI field nor look at it holistically, nor the use cases for it, just slap benchmarks and scream in terror. Comparing these models for what they are and their limitation in parameters.

Also experts are finding a lot of distillation - i.e. using GPT-4o and o1 to train. Kind of funny USA restricts GPU usage but let AI models unhindered to China. Totally defeats the purpose of the export restrictions. You're basically paying the billions in infrastructure and they build their model by distilling leading edge american models.

BIG OUF USA

The $6M does not include the costs associated with prior research and ablation experiments on architectures, algorithms and data. But if they distilled american models, wouldn't surprise me they are near that target. So basically a big black box, we know how much a finished model costs to train. $6M is not impressive, once a model is done, distilled models cost roughly that and even less as Emad calculates.

No magic sauce. No look at our billions spent and they just did it with $6M drama, they are using the billions spent from previous work to their advantage. I think that pipeline will be cut off.


Another one appears like a mushroom

 

[digital_ghost]

Nvidia CEO:

 

[Inviusx]

Maybe I'm just retarded with stocks but they were trading for lower than this in September 2024. So why buy now when you could have bought then?

 

[HJuggernaut]

If a company can lose that much value over a competitor making a similar product then that just means the company's evaluation was too high to begin with. remember all these companies worth is based on speculation.

 

[ResurrectedContrarian]

To me, the big news with the Deepseek models is just that reinforcement-based learning for LLMs is even more powerful than previously thought, and simpler to set up now (since the Deepseek re-implementation of PPO uses fewer moving parts and trained components). And pure RL can be used to train LLMs to explore reasoning on their own, even without any instruct supervised stage. Great stuff.

 

[AeonGaidenRaiden]

market cap is a fake number. I hate this style of financial "reporting"

Literally. I hate when someone uses the market to say the worth of a company yet a single fart in the wind could knock off 100s of millions of $$$ and then a company goes from being worth 300 million to 90 million in a single day, then 5 days later its worth 150 mil etc.

 

[ChorizoPicozo]

guys, asmongold just made a video... i think this is a big deal.

 

[peish]

To me, the big news with the Deepseek models is just that reinforcement-based learning for LLMs is even more powerful than previously thought, and simpler to set up now (since the Deepseek re-implementation of PPO uses fewer moving parts and trained components). And pure RL can be used to train LLMs to explore reasoning on their own, even without any instruct supervised stage. Great stuff.

Sounds like current AI is overrated bullshit grifting

A simple change in coding and you get “better” results

 

[ChorizoPicozo]

You can say all the gobbledygook you want, but if it's not capable of searching the internet, it is basically useless.

 

[Haint]

Why do 7B's image examples still not look that great? Is this a different kind of generative model? The beautiful girl looks pretty terrible compared to what other modern models churn out, which are effectively perfect and photo realistic. 7B's eyes are still fucked up and the chin/jaw/mouth proportions seem off.

1. The US announcement was BOGUS!!! Even Elon called it BS. They only have $100 Billion secured. And we aren't sure the whole $100 Billion is even there.
2. China was working on Deepseek before that bogus announcement. It's been public knowledge for at least 6 weeks prior.
3. If we could buy China's EV cars......that'll be saving US citizens money right now. But at the moment it's illegal to import Chinese EVs in America.

Chinamen mobiles built in Mexico would bankrupt the entire US car industry overnight, and I wish they would be. We should not be paying $45,000 for every man shit mobiles.

 

[IAmRei]

At least now NV might have competitor, i guess? I'm not to fond of both tech currently. Although I mostly use NV laptop and once I use ATI, waaay too long ago.

 

[Sgt.Asher]

At least now NV might have competitor, i guess? I'm not to fond of both tech currently. Although I mostly use NV laptop and once I use ATI, waaay too long ago.

Deepseek isn't a competitor. They only released an ai model that is more efficient. An ai model that was trained on nvidia chips.

Nvidia designs the chips that run ai. This will only help nvidia in the long run as smaller companies can potentially train ai now.

 

[IAmRei]

Deepseek isn't a competitor. They only released an ai model that is more efficient. An ai model that was trained on nvidia chips.

Nvidia designs the chips that run ai. This will only help nvidia in the long run as smaller companies can potentially train ai now.

ah i see now, thanks for clarifying

 

[Johnny Concrete]

So time to buy buy buy!

 

[M1chl]

 

[StreetsofBeige]

Thank god I sold my Nvidia stock last fall making good money.

By the way, what proof is there deepseek is this good at dirt cheap costs and hardware needed?

Is it just verbal claims? Or they got proof? Just asking as I haven’t kept up.

 

[DoubleClutch]

Deepseek isn't a competitor. They only released an ai model that is more efficient. An ai model that was trained on nvidia chips.

Nvidia designs the chips that run ai. This will only help nvidia in the long run as smaller companies can potentially train ai now.

Except AMD, Intel, and cheaper chips can handle the processing meaning expensive Nvidia chips aren’t needed.

It’s obviously not a win for Nvidia because the market has reacted very negatively to it.

 

[Codeblew]

AI is way overhyped even though it is useful for many things. What I hate about these models is that they they censor certain topics. For example, ask deepseek about Tiananamon Square.

 

[bigdad2007]

AI since its resurgence has obviously been the next pump and dump scheme after cyrpto, NFTs, Web 3.0 etc all become saturated terms.

It turns out you don’t need massive new hardware to run a glorified chat bot that is really just the world’s biggest plagiarism machine.

 

[viveks86]

By the way, what proof is there deepseek is this good at dirt cheap costs and hardware needed?

Is it just verbal claims? Or they got proof? Just asking as I haven’t kept up.

Don’t know if this counts as proof, but most of the industry already have their hands on it and are testing it, my company included (based in the US). Shared what I know so far here:

Seems legit guys. Engineers are testing it out in the US on our own servers as we speak. Here is directly from my VP of machine learning. This is literally what he gets paid for. I'm quoting verbatim. Not posting a screenshot as I'm not comfortable with it on the internet.

Me: Hey have you looked into Deep seek R1 yet?

Him: yes. The model came out last week.

Me: Are the performance numbers and costs what they claim to be?

Him: Yes. Although I am surprised why the market is reacting to it this way today. By stock market standard, it is a week (eternity) old news. Why did it catch your interest?

Me: Seeing it hit mainstream news with the Nvidia price drop. Didn't know anything about it till then
Him: They spent $5 million to crater Nvidia by $600 B

 

[IDKFA]

Yeah people are getting a bit carried away. This is just usual leapfrogging with technology. My company is still training these AI models to properly read a goddamn pdf file with flowcharts in it. Singularity can (and likely will) happen, but not for a couple of decades. By then I'll retire and be fed by my robot slaves... or be food for them. Whatever they prefer.

Many experts in the field predict the singularity is actually closer than we think.

Ray Kurzweil is one who predicts we'll hit this level in 2045. Some have made predictions earlier than that, but I think 2045 is the most realistic prediction. Either way, it's happening in our lifetimes.

It could greatly benefit hunanity and lead us to a utopian future, or a be our catastrophic downfall. It all depends on how we prepare for and manage it. To do this, there needs to be global cooperation on its development

I just don't buy that AI will be able to infinitely code itself with no human intervention. Generally, I believe that the more complex the system is, the more fragile. Trip up in the wrong place will lead to cascade failures.

I understand your skepticism. You're not alone in believing the singularity will never happen. Complex systems can be fragile and prone to failures, however advancements in AI engineering are specifically addressing these challenges through redundancy, error correction, and simulation testing.

There has always been doubt about AI progress. We used to think that machines couldn’t handle the complex tasks like chess, Go, or medical diagnoses, yet AI has repeatedly proven capable of mastering and surpassing human expertise in tasks.

I think the singularity is a legitimate concern and one that people should be taking seriously, but that's only going to happen with global cooperation.

 

[Midn1ght]

huge 5090 price cut incoming, right ?

 

[hyperbertha]

I am angry and I will defend this company, like good people should. Those Chinese shitheads will be getting a strongly worded letter from my typewriter. On it as we speak.

I have been trying to get in touch with Deepseek's manager.

 

[GeeseGoose]

Chinamen mobiles built in Mexico would bankrupt the entire US car industry overnight, and I wish they would be. We should not be paying $45,000 for every man shit mobiles.

I moved to Australia and because they have no car industry to "protect" they allow Chinese cars here and they really are as good as major brands, in most cases with better features, but at a fraction of the cost.

 

[hyperbertha]

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.

OpenAI are now iterating every ~3 months on the o-series, they can incorporate whatever RL lessons there are to be had from the DeepSeek paper on o4. o3 got 25% on the research-level math FrontierMath benchmark, do you realize how much compute is going to be required for ~90% on this test -- with, or without, this new efficiency gain?

What evidence is there that simply injecting more compute will magically let it get 90 percent on frontier math? These models are not really reasoning. There is a high likelihood that we have hit an architectural brick wall, and some new breakthrough is needed.