DeepSeek R1 Reactions Explained - Who Is Right??

0:00 deep seek is completely appending the AI industry in real time right before our eyes the US tech Market took a massive hit erasing trillions of dollars in value in just one day this incredible model created by a small Chinese company as a side project on a Sho string budget has completely annihilated all of our assumptions on what it takes to build state-of-the-art artificial intelligence and the reactions from the AI industry range from this is incredible because it's open source to America just fell behind China to the CCP is lying about how many gpus they have and trained with let me break it all down for you so the first thing I want to show you is a post

0:45 by Andre karpathy from December just about a month ago if you're not familiar with Andre karpathy he is one of the leading Minds in AI he is a co-founder of open AI he worked at Tesla on their full self-driving and just generally understands AI better than most people so this is his tweet and remember this is about deep seek V3 which at the time came out but didn't really Garner that much attention this was a more traditional model not a thinking model but Andre karpathy actually predicted how important this model was and nobody really gave it the gravity that this prediction really warranted deep seek a Chinese AI company making it look easy today with an open weights release of a frontier grade llm trained on a joke of a budget 2048 GPU use in 2 months $6

1:31 million so he saw it at the time but even I didn't really realize what this meant and many others didn't either for reference this level of capability is supposed to require clusters of closer to 16,000 gpus the ones being brought up today are more around 100,000 gpus for example llama 3 405b used 30 million GPU hours while deeps V3 looks to be stronger at only 2.8 million GPU hours 11 times less compute so even a month ago deep seek was making more with less if the model also passes Vibe checks basically the arena rankings human evaluations it will be a highly impressive display of research and Engineering under resource constraints now here's the key and I'm going to

2:18 touch on this point multiple times through this video does this mean you don't need large GPU clusters for Frontier llms no but you have to ensure that you're not wasteful with what you have and this looks like a nice demonstration that there's still a lot to get through with both data and algorithms very nice okay so remember that was a month ago and now just a few minutes ago as of recording this video he actually made a follow-up post to that post let me show you here's the key I will say that deep learning has a legendary ravenous appetite for compute like no other algorithm that has ever been developed in AI deep learning this is the technique that R1 used to get so so good and it did not have human feedback that's the key so they removed

3:05 the biggest bottleneck in really any system but specifically deep learning which is the human feedback element now with that removed compute can go infinite you can use unlimited compute to do deep learning it's essentially self-play kind of how alphago was able to become the best go player in the world without actually having any human examples of what worked you have a really strong well-defined reward mechanism that teaches the model a good response and a bad response you may not always be utilizing it fully but I would never bet against compute as the upper Bound for achievable intelligence in the long run so the fact that the US tech stocks have taken such a big hit today really shows that Financial analysts and investors maybe aren't truly

3:50 understanding what's happening here and so he goes on to explain in detail why reinforcement learning deep learning is so powerful and how it really creates Crees unlimited data for these models to train on as long as you have that reward function and it's well defined then these models will create a ton of data and all of that data requires compute and then to train the model back again requires compute and so let's talk about the stock market now because as you see the market has erased trillions of dollars of value in a single day because of deep seek really because of this one open- source model created by this tiny Chinese company but again I personally do not agree with the assessment that this devalues gpus that this devalues The Cutting Edge US tech sector it just

4:39 doesn't and it has to do with javon's Paradox and I'm going to come back to that in a moment but for now let me show you a report from JP Morgan on deep seek 95% of the questions we received were about the Chinese AI company deep seek particularly regarding how they managed to develop state-of-the-art models on a modest budget raising questions about the billions being poured into Cutting Edge AI capex as reflected in Nvidia -3% today it's down like 15% in actual trading hours and VRT down 5% we believe these concerns are overblown and would like to share why deep seek is a 2-year-old AI research and development Venture startup based in China fully funded and operated by high flyer a

5:25 leading Chinese Quant fund they go on to briefly describe a little bit of history of the model releases by Deep seek but I want to get to the important part what has drawn the most attention however are deep See's research papers and the efficiency of its models it is unclear to what extent deep seek is utilizing High Flyers approximately 50,000 Hopper gpus so they said they're using a lot less but according to Goldman they have 50,000 a cluster comparable in size to the one open AI is rumored to be using for GPT 5 training what is evident though is that they are dramatically reducing cost which is great for everybody because they open- sourced it now here's the key point this does not signify the end of scaling the end of needing more computing power or that the

6:10 most Capital intensive players won't win remember Mark Zuckerberg announced a sharp increase in AI capex yesterday instead it suggests that competitors across the Pacific have been forced to pursue efficiency due to export restrictions now that I believe completely constraint is the mother of innovation and deep seek had to do more with less unless you believe that they're actually using many more gpus than they're leading on but I personally don't because they released such a detailed research paper really showing exactly how they created the model and there are many companies right now reproducing the results and personally I'm on the vibe that Gary tan president of Y combinator is on have to admit I just bought the Nvidia dip not

6:55 investment advice but I personally agree with this now here's the last point I want to make on the AI industry from an economic perspective this is Jonathan Ross the CEO and founder of grock grq now quick disclosure I am a very small investor in them but I really do believe what he's saying let me read this so the original post is the market is deeply seeking an answer to deep seek they are not finding those answers in the semiconductor stocks all right so Jonathan Ross replied is that the sound of the market sloshing from pre-training to test time compute yes yes that is that sound now if pre-training a state-of-the-art model now cost just $5 million the real Leverage is post training during inference we can scale

7:41 that up seemingly indefinitely there may be a limit but we certainly haven't seen that limit yet let me just break down what that means during inference which just means when you prompt the model and it's thinking about the result and it's preparing the response to that we now have thinking models which are like 01 03 three from open Ai and now R1 from Deep seek and so these thinking models during that time where it's coming up with the response to your prompt it's using a ton of tokens and it's reflecting on its response it's trying a bunch of different responses it is truly thinking about the response and so all of that thinking uses a lot of tokens all of those tokens require a lot of compute so just because everything got cheaper doesn't mean we're going to use less of it we're going to use more many

8:28 use cases that were not Roi positive previously now became Roi positive javon's Paradox so what is jeevan's paradox well intuition might tell you that when something becomes cheaper and more efficient the market for it is going to shrink because everything is less expensive so of course people don't spend as much but it's actually the opposite let's think about energy as a unit of energy becomes less expensive the amount of use case cases it can apply to profitably increases so the total market and the total spend greatly increases even though the per unit price decreases now there are a number of examples throughout history of jeevan's

9:13 paradox the steam engine being one energy and gas prices being another and now we're going through another with artificial intelligence now although deep seek made the cost of pre-training much less expensive and it's even much more efficient during inference time that doesn't mean the total market for artificial intelligence compute is going to decrease in fact quite the opposite it is going to allow us to accelerate so much more quickly towards AGI and Asi artificial super intelligence and thus the spend is going to increase greatly so anybody who is shorting companies like Nvidia and other major tech companies you probably have a pretty painful future ahead of you not

10:00 investment advice so that's jeevan's Paradox so let's think about this in a positive way now there have been a number of leaders in the AI space who are really thinking this is a very positive thing and again that is something I agree with now here is Dr Jim fan leading AI researcher at Nvidia and obvious we are so back moment in the AI Circle somehow turned into it's so over in the mainstream unbelievable short-sightedness the power of 01 in the palm of every every coder's hand to study explore and iterate upon ideas compound the rate of compounding accelerates with open source the pie just got much bigger faster javon's Paradox we as one Humanity are Marching towards Universal AGI sooner now obviously there is competition between China and America so that might be the

10:46 one line that I don't necessarily agree with yes sooner you read that right Zero Sum game is for losers now I hope he's right but I don't think China is thinking that way they are thinking it is a zero some game whoever reaches AGI first reaches ASI First artificial super intelligence because of the intelligence explosion which we've talked about on this channel from the paper situational awareness and Gro who I just spoke about just released their own version of Deep seek hosted by them a US company the 70 billion distilled version of llama which is insanely fast I tested it out I was getting 250 tokens per second so imagine all of that thinking time can be done in an instant and you can start giving

11:32 these models harder and harder problems that might take minutes hours days to actually go through the thinking process now it might just take seconds perplexity one of my favorite AI tools I use it all day every day just included deep seek R1 in their service here's arvind the CEO of perplexity the world's most powerful reasoning model deeps R1 with reasoning traces is now on perplexity for supporting your daily deep web search so already we are benefiting from the opsource deep seek R1 and here's another positive take on it Bill Gurley famed venture capitalist in Silicon Valley has this to say it's a better world if the most disruptive llm model is foreign and open source versus domestic and proprietary I wonder who

12:20 he's talking about better for safety security Free Speech Innovation startups cost performance protection from sovereignty and Global Prosperity so a lot of people are saying China just jumped ahead they just ate our lunch but they open- sourced it everybody wins that is why I love open source now speaking of geopolitics let me show you a few takes on that here's David Sachs the AI and cryptoart for the United States here's what he has to say deeps R1 shows that the AI race will be very competitive and that President Trump was right to resend the Biden executive order he's referring to the executive order that Biden put in place right at the end of his presidency putting in place regulation on how we develop artificial intelligence so which

13:06 hamstrung American AI companies without asking whether China would do the same obviously not I'm confident in the us but we can't be complacent if that is the main takeaway that we get from this that we need to double down on how quickly we accelerate AI research and development I'm all for it Aaron Levy CEO of box has this to say the reason you don't want overly aggressive AI regulation right now now is the deep- seek problem one day there will be a breakthrough somewhere else and having your arms tied behind your back trying to accelerate past it will destroy your chances of doing so that is kind of what's happening in the EU right now they are basically watching from the sidelines when they could be participating in this incredible world changing technology and now here's kind of a crazy take and it completely got

13:52 destroyed on Twitter so this is Steven heidle who works on the API atop AI Americans sure love giving their data away to the CCP in exchange for free stuff immediately Community note deep seek can be run locally without an internet connection unlike open ai's model so he really tried to jab at Deep seek but immediately was corrected it's open source you don't need to use deep seeks hosted version and just to let you know I'm not going to be using deep seeks hosted version I'm not going to be downloading it on my phone it is just too much of a privacy and security risk giving the CCP all of my data I feel the same way about Tik Tok but grock is hosting it and I'm sure many other inference companies are going to be hosting it very soon so I've talked

14:38 about javon's Paradox quite a bit in my previous video and this video already now let me give you a perfect example and a recent example this is emad who is the founder of stability AI they created incredible open-source text to image generation models so javon's Paradox is real after we release stable diffusion as an open source smaller f faster better version of the closed image models out there it completely used up all GPU capacity on clouds that is what happens when Tech gets more efficient it gets used more you couldn't get Nvidia chips for inference market for thinking AI way larger completely agree and this was a quote of Satia Nadella CEO of Microsoft saying essentially the same thing javon's Paradox strikes again all right so now let me talk a little bit

15:24 about how the actual AI industry is reacting to it or some predictions of what will happen so suale a well-known founder in Silicon Valley says whether or not profitable all AI inference prices shall go down in a week just watch all margins eviscerated you can't let them have any distribution Advantage due to price Scorch Earth policy leads to the app layer winning ultimately and that's the key the app layer the cost of intelligence is truly going to zero that is not just coming from the open- source Community Sam Alman said the same thing the unit price of intelligence is going to go down significantly Dave Freeman says seems like bad news for NVIDIA no that's the wrong way to think about it suale replies going to be great for them

16:10 yes we're going to be using so much more compute even if the cost goes down and especially because the cost goes down but not everybody believes deep seek achieved what they did now I don't believe this but I want to show you a few takes including one from Elon Musk where they're basically basically saying no deep seek is lying but the first conspiracy I want to show you actually might make sense I don't know if it's true or not but it kind of would make sense now high flyer the Chinese company that owns deep seek is a hedge fund meaning they do Investments and so Chris Camilo says likelihood deep seeks $7 billion hedge fund parent is shorting Nvidia and considered this into their decision to open source their model while emphasizing its $5 million

16:56 training cost essentially they open source this incredibly efficient model and they thought once the market sees this Nvidia stock price is going to go down so I'm going to short it and if they did that maybe one of the best trades of all time who knows but again that is making the wrong assumption that this is bad for NVIDIA because it isn't it's good now Nikita be on the conspiracy side and this isn't really a conspiracy I kind of believe it although not to this degree reminder US tech companies want your data so they can get rich nation states CCP want your data to kill you now a close friend of the channel chubby posted this on X and quick reminder he writes for the Ford future newsletter he does incredibly deep dives into different AI topics so if you want to stay up toate on the latest AI topics

17:43 definitely go to forward future AI to find Chubby's writings it's really incredible so he posted this billionaire and scale AI CEO Alexander Wang deep seek has about 50,000 Nvidia h100s that they can't talk about because of the US export controls that are in place I talked about that in the last video but here's the thing let's scroll down a little bit Elon Musk replied saying obviously so he believes that they are not telling the truth and they actually used many more gpus than they said now Elon Musk is known for taking a pretty negative view on anything that he deems competition but it might be true I don't know here's the thing we're going to know soon enough it's all open source the recipe is open and is being

18:29 replicated right now Mark benof CEO of Salesforce posted about deep seek Roy Abrams said Mark you take the claims that this was done on a shoestring at face value and then Elon Musk once again l m Ao no so Elon Musk does not believe it for a second all right last I want to show you the memes cuz the meme game on this topic have been incredible so here's growing Daniel Finance guys are like I effing knew you nerds were full of it about needing that much money and that is so funny here's another one remember just last week open AI released Operator by the way it's incredible and the agentic future is very promising but this is funny so Sam Alman fun watching

19:15 people react to operator reminds me of the chat GPT launch zaum says WTF is operator deep seek D's Nuts bro got him here's another great one from cheddar flow they fell for de deep seek good buy 20 million more of calls and that's Nancy Pelosi who has been known to absolutely outperform the market being a politici and as I said the markets have been deeply in the red today so here is autism Capital they call it deep seek because everyone will be deeply seeking McDonald's employment and every time crypto goes down and that's what you're seeing here that's the meme got to get a job at McDonald's again and the last meme I want to share because I actually agree with this this isn't necessarily as funny as it is true Adam D'Angelo board member of open AI

20:03 shared this so on the left AGI will drive GPU demand that is the kind of simpleton take on it then in the middle we have deep seek efficiency will reduce GPU demand a bunch of analysts real angry about that then the Jedi on the right the Smart Ones cheaper AGI will drive even more GPU demand completely agree if you enjoyed this video please consider giving a like And subscribe and I'll see you in the next one