the release of deep seek AI from a Chinese company should be a wakeup call Chinese AI just dunked on Silicon Valley so hard the market cap of Nvidia lost the equivalent of the entire GDP of Sweden it’s fine it’s Sweden yes as we’ve all heard China spent peanuts to build what cost American companies hundreds of millions of dollars the markets went into panic mode and US tech stocks got wrecked but is it so over for American dominance of AI now is open aai cooked what about Nvidia how did China do it and why are my altcoins down to well we know you have a lot of questions so my name is guy and stick around as we unpack them and figure out what China’s deep seek means for AI geopolitics and crypto but first a refresher on how we arrived at the present moment in AI if you already know your stuff well we’ve timestamped this video just for you so you can skip ahead now for as long as we’ve had computers people have been dreaming of AI neural networks were conceived way back in the 1940s as a way to mimic the human brain’s pattern recognition they paved the way for deep learning whereby computers learn to extract answers from data much like our own brains do of course back then we didn’t have Hardware powerful enough to do any such thing and the rest of the 20th century was a story of computers catching up to our big Ideas they became smaller cheaper and vastly more powerful but funnily enough in the end it was the Advent of 3D video games that got us over the line you see in the 1990s demand for 3D video game Graphics brought about the invention of the graphics Processing Unit or GPU specialized Hardware designed to handle parallel computations efficiently little did anyone know at the time that gpus would become the back phone of modern AI in the early 2000s it became apparent that parallel processing had use cases far beyond rendering 3D banana peels in Mario Kart as it happens AI algorithms particularly those used in deep learning rely heavily on the same parallelizable operations that gpus were built for by the late 2010s breakthroughs in deep learning architecture and natural language processing made possible the first large language model or llms in particular Google’s invention of Transformer architecture and its pioneering Bert llm in 2017 paved the way for the generative AI that we’ve all become so familiar with soon ever more performant llms started proliferating demonstrating remarkable abilities in coherent and contextually relevant text generation but it wasn’t until open AI had the bright idea of turning a GPT model into a public facing product that the world took notice the launch of chat GPT in November 2022 changed everything AI quickly took on a new strategic importance in geopolitics and open AI became the poster child for American leadership in the field less than three years on open ai’s latest funding round has the startup valued at around $340 billion investment continues to four in 10 figures at a time and open AI has monetized access to its Flagship models with a $200 per month subscription service the US government seeing America’s lead in AI as a matter of vital importance has sought to strangle China’s ability to compete in AI by cutting off its supply of high performance Nvidia gpus the picks and shovels of the AI Gold Rush president Trump picked up where President Biden left off making a point of hailing a new joint venture between open AI Oracle and Japan’s soft bank shortly after his inauguration The Venture entails up to $500 billion do of spending on infrastructure namely data centers and electricity generation needed to maintain America’s Edge in the race to scale Ai and herein lies a certain vulnerability in the headlines open AI is always seen adjacent to one eyering sum of money or another whether it’s a new raise valuation or or capital expenditure one thing you never see though is profit because open AI is still burning through cash and treading water now this alone is not a gotcha moment because it’s typical for startups to not be profitable in their formative years but open ai’s unique ability to raise and burn seemingly unlimited sums of money now looks like a double-edge sword Capital has replaced Innovation as open ai’s biggest Moe and within the finds of the game that open AI pioneered that is exactly what is needed the game is simple you just need more of everything more compute power in the form of more gpus more energy to power and data centers to house them and more data to train your models on there are no shortcuts only an unquenchable thirst for dollars data and electricity to satisfy it’s blunt and extremely resource intensive and Silicon Valley had all of us persuaded that it was the only way that is until deep sea came along and upstaged the lot of them apparently on a shoestring budget and using pretty flimsy gpus turns out that hyperscaling through Brute Force isn’t the only way to build AI so then let’s talk about deep seek it began Life as a pet project of Lang wenfang co-founder of Chinese Quant hedge fund high flyer after spending 7 years using AI powered trading algorithms to slay the markets and build a $14 billion portfolio he decided to go all in on AGI research according to a 2023 post from highflyer official WeChat account quote highflyer will concentrate its resources and strength wholly devote itself to serve AI technology that benefits all of humanity create a new independent research group and explore the essence of AGI the company said now this research group became deepsee although it’s not clear what high flyer meant by independent as deep seek is 100% funded by the hedge fund now this is a totally different funding model to the startups of Silicon Valley where external venture capital is the lifeblood of every startup it put deep seek at Liberty to pursue long-term research goals without the pressure of having to deliver quick commercial returns to Greece VC exits and this ethos is spelled out in the tagline of deep seeks website which reads quote unravel the mystery of AGI with curiosity answer the essential question with long-termism and Lang does appear to be playing the long game as he had personally accumulated around 10,000 Nvidia gpus by 2021 a collection he described as driven by curiosity rather than business logic well it proved to be a serendipitous stockpile though as the US banned exports of powerful gpus to China in 2022 by 2024 Lang’s gpus were out ofd and not fit to compete with the likes of open AI meta and Google who gatekeep the frontier of AI with massive data centers full of the latest Nvidia gpus but this disadvantage was ultimately how Lang and the Gang shocked the world with the 20th of January unveiling of the flagship deep seek R1 model like open ai’s Flagship model 01 deep seek R1 is a reasoning model built on top of an underlying large language model dubbed V3 now V3 was trained for 55 days using approximately 2,000 of nvidia’s less powerful h800 gpus at a cost of around $5.6 Million by comparison open ai’s 01 was built at top the gp4 large language model according to gp4 it cost somewhere between 100 million and several hundred million to train now you’d think that a model that cost at least $100 million to train on the world’s most powerful arsenal of gpus would be incomparably superior to one trained on toy gpus for $5.6 million but this wasn’t the case by almost every metric R1 is as performant as or Superior to 01 this is not a very good look for open AI especially since its models are closed Source whereas deep seeks is open source to put it bluntly deep seek just dumped a miraculous Leap Forward in AI on the internet for anyone to use and modify however they please completely for free whereas open AI is closely guarding inferior architecture behind a $200 per month payall however there has been some skepticism about deep seeks resource limitations or at least their framing for example the $5.6 million headline figure is impressive by any measure but it has been presented misleadingly in the news media the attention given to this number could leave you with the impression that deep seek accomplished for basically peenuts what Silicon Valley had spent billions to do for example a passage in the guardian read quote the upset has come from Deep seeks claim that it trained an AI model on a par with open ai’s chap GPT with less than $6 million for comparison Microsoft Open ai’s primary partner plans to invest around $80 billion in AI infrastructure this year now this looks like a willful misunderstanding intended to dunk on Silicon Valley that $5.6 million was the cost of a single step in the process of building deep seeks llm and there’s no reason to compare it with infrastructure spending they’re apples and oranges but even the timing of deep SE R1 rollout had the same effect R1 was unveiled the day before President Trump trotted out Sam ultman and SoftBank CEO Maya shoi son at the White House to hail their planned $500 billion investment in AI infrastructure in the United States the announcement was widely interpreted as a move to counter China’s advances in AI but it was quickly overshadowed by Deep seek Trump’s touting of the $500 billion invest made for an irresistible Jer position with deep seeks $5.6 million spending on training as if to say what the US is burning hundreds of billions of dollars and draining the national power grid to build something China made for peanuts and dumped online for free well that’s embarrassing it’s worth repeating that such comparisons are unjustified and betray either a poor understanding or bad faith you don’t need to invoke hundreds of billions of dollars in infrastructure invest ments to get the point across you can just say that GPT 4 cost at least $100 million to train whereas V3 cost $5.6 million this fact has the same effect without overstating it another critique meanwhile relates to the compute resources used by Deep seek scale AI CEO Alexander Wang claimed on CNBC that deep seek owns 50,000 state-of-the-art Nvidia Hopper gpus which it can’t disclose because to do so would be to acknowledge a violation of Washington’s export ban disclosure could cause trouble for anyone the Americans can get their hands on who was allegedly involved in getting those Hoppers into China and this would not be in deep seeks interest Elon Musk responded to wangk claim on x quote obviously although neither Wang nor musk have provided any evidence or sources but it’s kind of a moot point anyway that’s because the release of R1 was a accompanied by a technical paper spelling out exactly how deep seek accomplished its results with resources far inferior to those controlled by silicon Valley’s Tech Giants there doesn’t seem to be any doubt about the content of this paper which was by all accounts very well received by the global AI Community for its actionable insights last July Lang told Chinese media that there were no plans for a deep seek IPO because quote the problem we face has never been money but the Embargo on high-end chips which of course begs the question of what deep seek would be able to accomplish with resources on par with its Silicon Valley counterparts the flip side of this question is where all of this leaves said counterparts just two years ago Sam ultman told an audience of VCS that quote there is no hope to compete with us on training Foundation models but as the saying goes necessity is the mother of invention while open AI was busy collecting VC checks deep seek was for forced by the US export ban to come up with new ideas in the short term deep seek wiped $1 trillion in market cap off of the US Stock Market it’s a good thing that open AI is not a publicly traded company because that could have been really ugly the stock worst affected was invidious whose market cap took a $593 billion haircut roughly equivalent to Sweden’s annual GDP as investors considered the implications of deep seeks accomplishments on nvidia’s Hardware Monopoly as the markets digested the news tech stocks recovered though to a significant extent the day after eating the biggest daily loss in history nvidia’s market cap regained $260 billion the second biggest daily gain on record meanwhile attention was drawn to outperformance by meta the company responsible for the previously leading open- Source AI model llama in the medium to long term the Deep seek moment is likely to be remembered as a big win for open- Source AI rather than as a victory or loss for any specific company the exception however is open AI who as the leading developer of closed Source AI now looks a little bit silly Sam ultman has since publicly stated that open ai’s decision not to open source its AI has placed the company on the wrong side of History this is perhaps Deep’s greatest accomplishment assuming Sam’s words become actions that is all the grumbling about how much how much they spent on training or what gpus they had or which other companies models they benefited from will be forgotten soon enough what will be remembered though is deep seeks technical breakthroughs which were made publicly available for free for the benefit of all people everywhere including residents of Silicon Valley the so-called hyperscalers will now be able to benefit from the efficiency breakthroughs introduced by Deep seek this is likely to alleviate some of the tension built up on Supply chains in the race to secure the most gpus and electricity and build the biggest data centers it’s now clear that the era of intensive pre-training of foundation models like GPT 4 where vast resources are being expended on diminishing returns will come to an end sooner rather than later the pre-trained models will be integrated into more applications and products and run locally say on a smartphone rather than on a remote server they will continue to grow learn and enhance but by interacting with their users and their digital and physical environments rather than by further training in the cloud in a process known as inference now ai for self-driving is a good illustration of what inference means the AI runs in vehicle learning about its physical environment and self-improving without input from the company who pre-train the model as pre-training becomes less Central to AI advancement the question question for Silicon Valley is what is it going to do with the enormous infrastructure that is being built for this purpose data centers represent vast amounts of valuable compute power so it shouldn’t be too difficult to find a monetizable and worthwhile use for these resources Nvidia is likely to continue dominating and any reduction in demand for its most powerful Hardware used for pre-training may be replaced by new demand for its more affordable offerings which will be more appropriate for inference Sam ultman and Microsoft CEO satian Adella have praised deep seeks breakthrough as impressive and a wakeup call a sentiment even echoed by President Trump should be a wakeup call for our industries that we need to be laser focused on previous assumptions about China being years behind the US in AI development and chronically inclined to copy and steal American intellectual property rather than conduct original research and innovate have been exposed as naked propaganda and hubris while the dust is still settling from the deep seat drama it’s not yet clear if or how America’s AI strategy will change in response to the Newfound competition at the very least we can conclude the US export ban which sought to handicap China has not succeeded and underestimating China is not a good idea now all of this is very interesting but the real question is of course what does any of this stuff mean for crypto the crash in US tech stocks was so violent that it brought crypto down with it despite there being no apparent relationship between deep seek and crypto as with every macroeconomic shock to financial markets crypto was hit because of its correlation to trafi whenever there’s a general shock to financial markets BTC dumps and when BTC dumps the rest of the crypto Market inevitably tanks even harder that said apart from the knee-jerk Market reaction deep seek should be good news for AI crypto projects which can now try to incorporate it into their TCH tax and you don’t have to look too far for instances of this already happening for example deep in Project nilon subsequently integrated the Deep seek R11 14b model into its privacy enhancing computational Network coin Bureau is invested in nilon by the way then there’s Venice AI a privacy enhancing and censorship resistant chat GPT alternative which is now offering its users censorship free access to deep seeks most powerful biggest model R1 671b as deep seek is now open source it’s also inevitable that AI agents in crypto are going to start leveraging R1 to further enhance their posting onchain analysis and well whatever else the market deems necessary now the arrival of deep seek was likened by some to a Sputnik moment which refers to a shocking Revelation that a foreign power can compete with the United States on a matter of great strategic importance well you have to think pretty little of China though to be shocked that it could contribute so significantly to the advancement of AI we think the real takeaway here is the big boost for open source over closed Source AI in general deep seek has also accelerated the time Horizon for pre-training to give way to inference as the main Vector of AI advancement US tech Giants will have to think carefully about the meaning and use of the vast infrastructure Investments that they are already committed to their share prices are going to survive but open AI may suffer a bit of an identity crisis at the time of making this video crypto is getting hammered for unrelated reasons but if you’re bullish on AI cryptos you may want to keep your eyes peeled for any projects integrating deep seek because if nothing else it makes for good Marketing in this attention-driven sector of ours and if you have a hotter take about what deep seek means for AI geopolitics and the markets then please do let rip in the comments below if you got something out of this video go ahead and drop us a like and while you’re at it make sure you’re subscribed and have your bell notification switched on so you’ll catch our next upload before anyone else and you can check out our latest video over here if you’re looking for what to watch next as always thank you for watching and I’ll see you next time this is guy over and out [Music]
Add A Comment