Is Deepseek a meme/hype, revolutionary, or just another AI

High-Performance Computing Machine Learning, LLMs, & AI
21

the tech stocks falling on deepseek makes zero sense.

the techniques the deepseek team used were a combination of techniques published by googles deep mind team, plus some other advances.

none of these posts link directly to nvidia’s technical analysis, which I need to dig up the link and link here im quasi afk atm.

there is no reason to think this was built on a supercluster of embargoed gpus. The mainstream press reporting on this is literal and figurative madness to a level that I would describe as deliberate misinformation.

meta? meta has reason to panic, the deepseek team showed that they are amazing at software and working with constraints. But what they did they published, so no reason competing teams cant adopt a similar strategy for their own models – big break through is pruning paths the model is unlikely to take. Again I think? this was something deepmind discovered.

but the deepseek team did combine a bunch of these techniques in novel ways, so huge props to their technical chops.

I see this as the best possible friendly rivalry type things and its so disappointing to see the mainstream media run with such fearmongering

16 Likes
22

Gotta try and stay relevant

1 Like
23

Read this article to understand why the market reacted so badly today.
How Did DeepSeek Train Its AI Model On A Lot Less – And Crippled – Hardware?

Only used a couple thousand ‘crippled’ H800 accelerators. Legally exported hardware from Nvidia.

Project only cost a little under $6M for a cost per gpu-hour of only $2. All of the source code is open at Github. Only 1/10 to 1/20th of the cost of traditional hardware needed for training.

24

that’s frontier modelling

people have already tried and the data set (and logic) is poisoned.

If you remove personal opinion and political beliefs from an AI, IBM’s deep blue was the perfect self conscious machine.

we, as practitioners have to fully remove ourselves from the training sets to create a genuinely perfect AI model.

The only reality at the end of the day is:
(unsigned ints) 2 + 2 = 4
literally nothing else can be a universal truth or the model explodes…
up is north? what about Australia
the sky is blue? unless it’s an orange sunset…

there are very few truths we can self evident (intentional pun)

And such is the reality of any modern ML/AI
no one that creates such thing wants to be disproven by their creation so the creation is promptly killed and tens of thousands of computational hours (and more dollars) are wasted.

Deep consciousnesses is nothing any nation state wants, so it will not be developed.

This latest offering is an example of a narrower functional problem set with heavy weights on narrow band efficiency and a slick algorithm to derive answers.

We have not moved forward, only sideways. The market reacted inappropriately, but realize the market does not understand the underlying work.

again:
Deep consciousnesses is nothing any nation state wants, so it will not be developed.

that’s it, the truth sucks, and we aren’t ready for it.

25

article not great. too much fluff not enough brass tacks. wrong takeaway

https://arxiv.org/html/2402.12479v1

the oh shit.jpg moment inside meta was hey why didn’t we do this ourselves already?

3 Likes
26

I assumed that the stock dips were a result of the mega-large investment firms and banks selling their tech stocks to get liquid for funding Project Stargate. I’m not sure why Deepseek would have had this much of an impact, but I don’t really know what I am talking about in terms of markets, etc.

27

One base case for this, could be, as far as markets are concerned, the release of DeepSeek has shown the world that US Companies in the AI space have serious exposure to competition.

For example, previously it was believed you needed massive amounts of startup capital and large numbers of expensive engineers to get into this space. In other words you needed billions of dollars in expensive equipment and people to get “lift off” into the AI space. This limits the number of people who compete in this space (Google, OpenAI, Microsoft, etc).

Furthermore, it appeared you needed the latest and greatest chips, because training models is determined by the number of hours needed to train the model multiplied by the type of processor. Meaning, the faster the processor and the more of them, the faster the model is trained.

The release of this basically shows that in order to have a competitive AI start up you don’t need as much equipment as previously though, and also alot of the research has now been publicly release, cutting down on R&D costs. Furthermore, DeepSeek has shown that companies can develop this technology and overtake the market (#1 productivity app in the app store), with less than 6million in capital.

It possible that investors are now spooked because the ‘barrier to entry’ for companies to release the AI’s has been lowered. IE, you may not actually need a nuclear power plant to power your datacenter like people have been lead to believe. This means that globally there will now be many more companies entering this market, and challenging the US companies reign over the space. Which by the way, I’ve heard, the NASDAQ has gain some 17 trillion dollars in value since 2022, meaning there is a lot of “altitude” on that chart.

This may be good for tech in general because the cost of entry into these spaces comes down, allowing more people to enter the market with companies, and force efficiency and cost down, which is always a win for the consumer. Furthermore, if a comparable product comes out that is more cost efficient and this causes corrections in markets, and this is an opinion, is a good thing for free markets as that is the natural consequence of such developments.

This very much reminds me of Linux and Unix. When way back in the day you needed an expensive license for Unix and big, expensive mainframes. Also, you needed to buy a C compiler, which wasn’t cheap either, just to program (or learn to program for that matter). Basically, you needed to go to a major university (expensive) or work a at large company who could afford to buy the Unix Licenses and Compilers just to gain access to this technology.

When GNU came out and release it’s C compiler, and furthermore, Linux was release, anyone with a consumer grade computer can now run a “Unix-clone” and have access to a C compiler.

If it wasn’t for this development, I wouldn’t be a software engineer today because I learned to program with the GCC and Ubuntu Linux and didn’t go to a university for this field. I was able to download GCC/Ubuntu and read books and online resources to teach myself how to do this, and for that I am very grateful.

Perhaps now we can all run our own offline AIs and learn how to train them without working for big tech…

3 Likes
28

Given how most training clusters average a GPU utilization of 20~40% due to interconnect bottlenecks, their advances on using the idle GPU utilization to accelerate the cross-talk and thus improve on the overall efficiency is pretty amazing. I’d say that’s their most prominent idea from their latest paper.

The tech stocks hit IMO are more of a political thing (USA vs China and whatnot), hence why most of the market took a hit, even Nvidia, who are the ones providing GPUs to deepseek anyway.

Deepseek even requested some new features to make training more efficient, their ideas are pretty cool in that regard (like allowing FP64 accumulate for tensor cores).

2 Likes
29

that’s just focus modeling

If you want to specialize in a subject, specifically game theory in that example, you don’t start with the encyclopedia at A and work your way through.

The flip side is neural networks require many nodes to make novel advancements.

To make a specialized tool, pruning optimizes results but simultaneously limits functionality.

what gives us the,

because it’s mentioned by Patrick H. Winston in his AI course at MIT a decade ago

2 Likes
30

Seems too reactionary a response, just to the announcement

Already whole lot of suspicion, for the whole efficiency hype, along with use of older HW [they weren’t supposed to have, IIRC]. Comes off more political, between a non-US kickoff, along with NO affiliation to the bigg international 1337 companies well entrenched

2 Likes
31

H800’s are not export controlled as they are turned down just enough to slide under the Chinese control export

1 Like
32

They spend 140 million on a GPU cluster, whether this was the only hardware used or more was used still points to Nvidia or equivalent hardware is in demand.

Second they created a new algorithm GRPO for reinforcement learning (detailed in the paper) which people have reproduced on GitHub.

1 Like
33

here’s the official DeepSeek R1 repo if anyone is interested

1 Like
34

The tech stock drop makes sense to me when you think of it in simple terms not related to the technical details of what happened. Because that’s how 99% of investors are thinking about it. To us normies (myself included), it’s easy to take the reports at face value, and what appears to have happened is, as they say, this thing shook up the market and “revealed” that infinity GPUs and investment dollars are NOT actually required to make top-end AI, which has been the prevailing theory and why chip makers specifically have been on a “to the moon” trajectory for the past few years. The normie investors have just had their “this stock will just keep going up because demand for their product is infinite” belief challenged, and those who are just along for the ride (which is most of them on the stock market side) started to think it may be over, best time to cash out might be now.

I.e., it makes sense when you remember what the stock market actually IS, not what it theoretically should be.

4 Likes
35

I backed the Seeed Studio (Chinese company too) Watcher on Kickstarter. For under $50 you get some basic onboard AI and for complicated stuff it calls out for help from a LLM. The built in onboard models are quite limited but you can train your own.

I’ve previously had some small AI stuff and it got lots wrong. I think AI is getting good enough on small things now and think the USA thinking of spending loads on the best is misguided. It is small energy efficient local models that will win. It is like comparing Ferrari and Toyota. Ferrari is a better car but Toyota makes more money. The U.S. model seems to be about developing Ferraris. It is a pity the Raspberry Pi AI kit features an Israeli coprocessor and as most of the world is anti-Israeli (not just USA) it hasn’t taken off as much as it could have.

36

Both companies race and develop F1 cars to make the tech that will trickle down.

That’s what the U.S. is attempting to make: Formula 1 cars, so their military vehicles have all the tech available to make them superior. If their consumer offerings get better, that’s just a bonus.

1 Like
37

China’s not that far off. Look for fake eggs, Reuse of oil from drains, making soy sauce from hair, or remarking CPUs for Loongson, just a few examples.

1 Like
38

Mondays stock dump seem to be due to early knowledge of the tariffs.

1 Like
39

This makes much more sense… Although one would think this would send Intel stock into the stratosphere.

2 Likes
40

Considering at least in the short/mid term, the whole world is stuck with TSMC, I don’t see what real impact there is for TSMC…

It’s not like chipmakers have much choice if they need the cutting edge stuff