Knowledge Fight Interactive Search Tool

All Episodes

Previous Next

Jan. 29, 2025 - Freedomain Radio - Stefan Molyneux

43:10

DEEPSEEK! Hype or Real?

Large

Audio Only |

Scroll

Time	Text
	Hi, everybody.
	It is the 27th, and so let me tell you how I found out about the Bitcoin crash.
	Man, that was something else.
	What did it, like Canadian?
	What was it?
	It was cruising at like 150, and then it went down to like 140. So it had a 10K drop.
	Now, a 10K drop mattered a lot.
	When it was 20K, it matters, of course, a whole lot less.
	But it dropped about 10K. Now it's climbed back up to...
	So it went from 150 to 140. Now it's climbed back up to 146 and change.
	And I found out about this because, I don't know, I've been hit with the strangest bug known to man.
	My daughter definitely had just the regular old flu.
	I mean, I assume that I have it.
	I don't know what's going on.
	I was kind of like half and half.
	I do this nose irrigation stuff, so it tends to ward off this stuff.
	But I was kind of like, am I ill?
	Am I not?
	Am I okay?
	Am I not?
	And then...
	It basically just, I just woke up exhausted on Saturday and spent like the day on the couch.
	Just kind of, you know, when you're half in, half out and you just, for me it feels like there's just a thousand lead balloons hanging off every neuron.
	It's just kind of dragged down.
	And then I slept at night and then yesterday I did shows and I did a call-in show in the afternoon and I felt back to 80% and back to 90% but I've got digestive issues.
	I don't have a sore throat.
	I don't have a cough, really.
	I'm sneezing.
	I don't know what the hell's going on.
	I don't know what...
	If I didn't already have COVID, I'd think it was some broad-spectrum nonsense like that, but I don't know.
	It's a strange thing.
	It's a strange thing.
	But anyway, so last night, because my sleep was so messed up this weekend, last night I couldn't get to sleep.
	So I got up and I was doing some reading, and I guess it was about 3 in the morning or 2 in the morning or something like that.
	I'm like, hey, I wonder what the old bitcoins are doing and tunneling to China.
	And that's a connection to this whole thing.
	But tunneling to China seemed to be the thing that they were doing.
	And I was just like, oh, that's interesting.
	I wonder what that's so...
	Somebody says, I wanted to thank you for helping me through my thought process.
	My consulting business turned to a full-time...
	COO position at a great local accounting firm.
	Thank you.
	You are absolutely welcome.
	Congratulations.
	Congratulations.
	That's wonderful.
	Let's do the supporters thing.
	So you've got a minute to join.
	You can just join and you can unjoin if you don't like it, but it's really, really great.
	You can go to FDRURL.com slash locals to do that.
	And I know, it feels like a plummet.
	And to the people who haven't been around this kind of stuff before, It feels, but you know, it's just, it recovers and all of that.
	So, of course, Bitcoin is not immune from market manipulation.
	Especially now that it, you know, everybody was like, yay, we've got ETFs.
	We have, Wall Street is into Bitcoin.
	And I don't think this was generated from Wall Street and we sort of get into why it's doing what it's doing, at least my theories.
	But...
	Now it's going to be open to market manipulation, and there are going to be enough actors in it as a whole that it's going to matter, right?
	What people think in the past.
	So in the past, the news that came up wouldn't particularly faze Bitcoin people, but now Bitcoin is surrounded, infused, enhanced, and dragged down by an entire phalanx of hyper-turbo normies, right?
	I don't think that Bitcoin was created by the US intelligence community.
	Military intelligence, one of the greatest oxymorons known to man.
	I used to do the nose flush, but I don't do it anymore, says someone.
	I heard that the water can bring bacteria up your nose where it's not supposed to be.
	People supposedly got serious infections and some even died.
	Well, yeah, I understand that to be the case.
	So you get distilled water or boiled water and you put your powders in it to make sure that doesn't happen.
	So, alright, let's get into what's going on.
	So this is a bit of a scattershot.
	Of course, it is a lot of on-the-fly stuff.
	Who is going to, what is this going to mean?
	Who knows?
	It's all in flux.
	As usual, the caveat is, of course, this is absolutely not investment advice.
	I'm not an expert.
	Do your own research.
	Don't make any investments, decisions based upon what I am doing.
	So, the issue last night appears to have been DeepSeek.
	DeepSeek is a Chinese AI startup recently gained popularity by surpassing ChatGPT on the Apple App Store while claiming to have trained their models at a fraction of the cost of their US counterparts.
	$6 million compared to billions.
	Now, I've heard $100 million.
	I've heard billions.
	So whether it's $100 billion or billions, $6 million, it's actually a little bit less than $6 million.
	It is a huge decrease.
	So they say they've trained their models at a fraction of the cost.
	And also, it claims to run on far less hardware.
	So, the company has launched AI models claimed to be on par or superior to leading US models at significantly lower cost, disrupting the global AI sector.
	What was Mark Andreessen referred to as the Sputnik moment?
	The Sputnik moment is when the Russians got satellites into orbit in the 50s.
	I think it was 57, which sort of started the arms race.
	And Russia actually plays in a little bit later here in a way that we'll get to.
	So, deep seat success is linked to US-China tech tensions, particularly due to US export controls limiting Chinese access to advanced AI chips.
	So, this is similar to what happened, not necessarily due to export restrictions, although they were involved.
	There's a great book called East Wine Minus West Equals Zero, but in the 60s and 70s, under communism, Of course, computers sucked, right?
	The computer equivalent of the Lada, as far as cars went.
	Computers sucked in Soviet Russia.
	And the result of that was that the programmers had to be really, really, really smart and brilliant.
	I mean, I started programming on a 2K computer.
	You had to be super lean in what you did.
	When I first started programming on the sort of Windows platform, you were lucky to have a meg of memory.
	So you had to load.
	TSR, terminate and stay resident programs, you had to top load them between 768k and 1024k.
	You had to top load those programs, which were normally reserved for DOS functions.
	You could find little places to stuff stuff.
	You had to be really careful about your memory usage.
	Nobody thinks about that anymore with, you know, 1632 gigs and all of that, but you really had to watch out for that kind of stuff.
	I had to be very careful about the amount of graphics I included in my program because people didn't have much storage space.
	So, nobody thinks about that stuff anymore.
	So, because the U.S., this is a story.
	Again, I don't know what's true.
	I'm just telling you what's out there.
	Decide for yourself.
	It's all is in flux.
	But the story goes something like this.
	And I've talked about this for many years, of course, that violence always achieves the opposite of its stated goal, right?
	Violence always achieves the opposite of its stated goal.
	And so, I'm not talking about self-defense, you know, personal stuff, but...
	I know, in terms of personal stuff as well.
	The more you bully people, the more they dislike you.
	So, what happened was America restricted the advanced hardware that America was relying on for its AI. I assume that these are sort of the massive cluster galaxies of high-end NVIDIA. The stuff that was originally built for video gaming,
	which then was repurposed and then, I think, specialized for AI. So, what happens, you know, again, we know this from the IQ conversations that the Chinese, and IQ plays in a little bit later here, too, but Chinese are brilliant at this kind of stuff, and, you know, average IQ north of 100, 104, 105, 106. And in spatial reasoning, very high as well.
	This has a lot to do with the evolution of growing rice and things like that.
	So, when you say to an incredibly brilliant group of engineers, such as you're going to find in the Silicon Valley of China, when you say to them, we are keeping the best hardware from you, what do they do?
	Do they just, oh, give up, oh, well, you know, I guess it's raining, I'm going to get wet because I just can't bring my umbrella.
	Well, no, of course, what they do is they look for alternatives.
	They try to figure out how to maximize the use of the hardware that they have.
	So, Elon Musk has floated the idea that DeepSeek is lying about the number of chips that they're on or using, They are using that many or not.
	I have seen evidence, can't confirm it, haven't done it myself, I've seen evidence that people have been able to get DeepSeek running on a laptop, which of course doesn't have 50,000 NVIDIA H100 chips and so on, so who knows, right?
	So Meta...
	DeepSeek has invested $60 billion, or is going to invest $60 billion for a new AI data center, DeepSeek $6 million to train the entire model.
	Now, is that apples to oranges?
	A little bit, but that is a tiny, tiny fraction, of course, and that could change just about anything.
	So, DeepSeek's AI model development strategy involved using a lot of 8-bit floating-point numbers throughout the training process.
	Offering significant memory savings without performance loss.
	So, you can use any number of bits to represent numbers.
	So, of course, the bottom is the byte, which is 0 or minus 1. Minus 1 for true, 0 for false.
	It's just an on-off.
	Well, all computers are on-off stuff.
	Then there is the integer minus 32767 to plus 32767. And then there's a long, and then there's a single, and then there's a double, and so on.
	In general, you don't really need to worry about this anymore.
	I was kind of obsessed, because of my early programming days, to use the smallest memory allocation possible, right?
	I mean, you could say, if you're storing a name, you could define your text as 255 characters, but no names really have 255 characters, right, outside of the occasional anime character.
	So they've found a way to use a smaller...
	Memory pointers for numbers, which is really quite something.
	So, Elon Musk says, he says, DeepSeek obviously has about 50,000 NVIDIA H100 chips that they can't talk about due to U.S. export controls.
	We don't know.
	Mark Benioff says, DeepSeek now number one in the App Store, surpassing chat GPT. No NVIDIA supercomputers or $100 million needed.
	The real treasure of AI isn't the UI of the model.
	They become commodities.
	The true value lies in data and metadata, the oxygen fueling AI's potential.
	The future's fortune, it's in our data.
	Deep.
	Gold.
	So, who knows?
	Billionaire and scale AI CEO Alexandra Wang, DeepSeek, he says DeepSeek has about 50,000 NVIDIA H100s that they can't talk about because of the U.S. export controls that are in place.
	We don't know.
	We don't know.
	Could be, could be not.
	Michael Cove wrote, DeepSeek, Stole the AI thunder with zero hype from CEO. Zero.
	OMG, guys.
	It changes everything.
	Influences.
	No swanky demos.
	No bloated promises.
	No hints that AGI achieved internally.
	They did it by shipping an actual product.
	Fantastic.
	Fantastic.
	So, question and answer.
	Again, who knows what the truth is, but I want to report it anyway.
	How did DeepSeek get around export restrictions?
	Answer, they didn't.
	They just tinkered around with their chips to make sure they handled memory as efficiently as possible.
	They lucked out and their perfectly optimized low-level code wasn't actually held back by chip capacity.
	Right.
	Right.
	If you have infinite capacity, you don't optimize, right?
	You ration, like, if you've ever been scuba diving, Scuba diving is like the laziest sport in the world because you want to use as little oxygen as possible because your oxygen is rationed by whatever you have in your tank.
	The faster you breathe, the less you can stay underwater.
	So, when things are rationed, you use them more efficiently.
	When they're not rationed, you don't.
	You don't think about how much air you're breathing in general, right?
	So, how did DeepSeek train so much more efficiently?
	Answer, they used the formulas below to predict which tokens the model would activate.
	Then they only trained these tokens.
	They need 95% fewer GPUs in meta because for each token, they've only trained 5% of their parameters.
	Right.
	So, I don't...
	Obviously, I'm not an AI technical expert.
	I've done a couple of presentations on it.
	I definitely am an expert coder.
	But I would imagine it's something like this.
	You don't train the AI on what could be word salads.
	You would train the AI on the most likely sequence or series of words.
	So if you've ever had those fridge magnets that have various words on them, you can just throw them all at the fridge and just end up with this word salad.
	You wouldn't train the AI in that.
	You would train the AI on the most likely sequences of words.
	In other words, you are not training the AI on sentence structures that will never or almost never exist.
	That would be my guess.
	How did they replicate?
	One, reinforcement learning, take complicated questions that can be easily verified.
	Update the model.
	Correct.
	So there's a math or code questions and so on.
	Okay, key value cache compression.
	Let's get into this as a whole.
	Why the DeepSeq model is so good?
	So here's the answer.
	They made three cool innovations.
	So a key value is like the model's working memory.
	So DeepSeq was able to find a way to compress this without losing quality of the model's output.
	DeepSeq requires 93.3% less memory to store this information while working, which makes it much faster and more efficient at generating text.
	A mixture of experts, FFN architecture.
	So the model's processing is split into different expert components.
	So the way DeepSeq does it is for each piece of text or token, the model always uses the shared experts and then picks the top few most relevant specialist experts from a larger pool.
	But the clever newish part is that they make sure all specialists get used, prevent some from being ignored, and distribute work evenly across computers and have some new ways to keep network communications efficient between computers.
	This approach lets them build a much larger model, 236 billion parameters, while only using a small portion, 21 billion, for each task, making it more powerful and efficient.
	Multi-token prediction head.
	That's the worst name for a porno ever.
	So, when large language models like GBD4 or DeepSeq, or whatever, generate text, they typically work by predicting one word or token at a time.
	Think of it like playing a word game where you have to guess the next word in a sentence.
	The model makes it its best guess for what?
	Should come next based on what came before.
	Multi-token prediction takes this a step further.
	Instead of just predicting the next word, the model tries to predict several words ahead at once.
	For example, now, I don't know if you've ever played this game.
	If you've had kids, you probably have.
	If you were a kid and you had fun people around, you probably did.
	So the game is, I played this a huge amount with my daughter and her friends.
	You get like five kids around a table and maybe an adult.
	And somebody starts off a story, right?
	And everyone gets one word to add, like once upon a time there was, and you know, it always ends up with poop jokes before the age of eight or over the age of 50. So, that's kind of what AI is trying to do.
	It's trying to guess, make up an X word that makes sense in the context of the story.
	So, if you think of the text, the cat sat on the, right?
	So, of course, the AI would say, Matt.
	And then uses MAT to predict the next word and so on.
	But the MTP approach, the multi-token prediction, doesn't just predict MAT, but also predicts in the sun.
	So future words.
	So it's much more efficient that way.
	And faster, of course, right?
	So that's a plus.
	Yeah, so the US banning the chips made China have to innovate and become more efficient and so on.
	This is what somebody wrote.
	AM models are powered by advanced chips, and since 2021, the US government has restricted the sale of these to China in order to stunt progress.
	To get around the supply problem, Chinese developers have been collaborating and experimenting with new approaches.
	This process has led to models that require much less computing power than before and can be produced far more cheaply.
	Now, of course, What happened last night, as far as I understand it, and we'll get into the crypto thing in a bit.
	So, basically, overnight and into today, DeepSeek erased $2 trillion of market capitalization because the AI companies, and in particular NVIDIA, which I think last I saw, was down like 17%.
	And we knew that because Nancy Pelosi stole some last month.
	So, DeepSeek erased $2 trillion of market capital because Of course, if you don't need 50,000 NVIDIA chips, then the value of NVIDIA goes back to basement gamers.
	So, NVIDIA lost over $600 million in market cap.
	Other semiconductor companies like Micron Technology and Arm Holdings each fell 7%.
	ASML saw a 9% drop.
	MegaCap, tech firms, Microsoft and Alphabet fell 4%.
	Meta Platforms dropped almost 2%.
	This is from Grok.
	DeepSeek's rise has led to a sharp decline in stock market values, especially in the tech sector.
	The Nasdaq fell more than 3% due to DeepSeek's news, with NVIDIA being one of the biggest losers, stumbling more than 17% in a single day, a loss of over $600 billion in market value for NVIDIA alone.
	Now, again, market value, it's not a useless metric at all, but it's not like people just took a bunch of money and set fire to it.
	That's the job of the Fed.
	So, yeah, investors are concerned about the competitive threat DeepSeek poses to established US AI companies like NVIDIA. So DeepSeek's ability to offer AI capabilities at a lower cost and with less advanced hardware has caused investors to question the value and future profitability of investments in high-cost AI development by US companies.
	It brought a sell-off in AI-related stocks and so on, right?
	And as far as I understand it, the app, the DeepSeek app, will only go to July of last year, but the website is up to date with sort of current information.
	I also did, I did look at, I didn't install the app because China, but I did look at the privacy stuff and they're like, yeah, you know, they'll actually hang on to not just your keystrokes, but your keyboard typing patterns.
	Like, it's just wild.
	It's just wild what they're grabbing, and I wouldn't install it for the life of me, but I know some, obviously a lot of people, a lot of people have.
	So, the sentiment among investors is one of worry and re-evaluation.
	Deep seeks success in prompting a rethinking of the AI narrative that has driven market performance in recent years.
	Some see it as an overblown reaction.
	The market might be looking for an excuse to sell off.
	Others view it as a legitimate threat to the dominance of US tech companies.
	In AI. But on the plus side, at least America did invest in a whole bunch of diversity initiatives.
	Because that really matters, right?
	There has been a global sell-off, not just in the US. Investors from Tokyo to New York have sold off tech stocks.
	Because if there is this cost-effective and AI model from China, that's going to be pretty wild.
	So, DeepSeek is omega bullish for Bitcoin.
	So, let's sort of figure out what this means.
	So, DeepSeek is omega bullish for Bitcoin.
	NASDAQ is crashing pre-market and taking Bitcoin down with it, but this is a temporary correlation.
	Bitcoin is a solution to the store of value problem.
	Bitcoin will not change.
	Bitcoin will not be disrupted and cannot be seized or debased.
	In an era of rapid change, Bitcoin's boring predictability is its biggest strength.
	This should be something we celebrate, not fear, because it disrupts equity markets and takes down our store of value assets.
	Hold Bitcoin, celebrate creative disruption, and get outside.
	So, the argument is that DeepSeek as an AI tool has triggered a significant drop in Bitcoin's value, correlating with a pre-market crash in the NASDAQ, although the correlation is seen as temporary.
	DeepSeek is revealing the overvaluation of US tech stocks, especially as Chinese companies are innovating in AI. And robotics at a lower cost, which challenges the high valuations of US tech companies.
	I would assume as well, but without any particular proof, this is just an assumption, that people had stop losses as the stocks fell.
	They had stop losses triggered, so they had to sell, and they had to cover losses, and they might have sold crypto to cover their losses.
	That would be my guess.
	I say this as a...
	Sheer and complete.
	I guess.
	So, DeepSeek was originally conceived, I've read, as a sort of just a hustle-aside project.
	So, DeepSeek, a Chinese AI startup founded in 2023 by Liang Wenfeng, has swiftly unsettled global markets with its groundbreaking and cost-effective AI models.
	Its rapid ascent has rippled across US equities, notably affecting NVIDIA shares, even introduced turbulence into the cryptocurrency sector.
	Plus, you know, there are a lot of boomers out there who do think that if you get better AI, you're going to be able to crack Bitcoin, which I, of course, do not believe to be true at all.
	I do not believe that to be true at all.
	So, I mean, they'll just raise the cryptography, right?
	So...
	The 5.5 million AI models that left Silicon Valley scrambling.
	Now, is it true, right?
	Is it true?
	Remember, all that is international is hostile, right?
	All that is international is hostile, which means, just as we did with COVID, as I talked about, even in early 2020, do not trust...
	The data coming out of China.
	So, is it true?
	Has it been independently verified that this thing only took $5.58 million to train and is basically free?
	In other words, DeepSeek V3, 671 billion parameters, yet was developed for a mere $5.58 million?
	I don't know.
	I don't know.
	Until it's independently verified by neutral third parties, I have doubt.
	I have doubt.
	But it just goes to show you how effective an economy can be when it's not worried about all of this endless political correctness stuff, right?
	I mean, the PC stuff is just...
	A war against an economy, right?
	So, that's...
	Let me just get to your questions.
	I have a bunch more notes, but I want to see what your questions or comments are.
	If you have your questions or comments.
	Excuse me.
	Excuse me.
	I search deep seek for the infamous Tank Man Tiananmen square photo.
	I'm using to watch the censorship kick in.
	I doubt the Chinese AI is using a lot of chips.
	Have you seen how large software is now?
	2014 of a software I use 700 megabytes 2025 is 8 gigabytes with no significant improvements.
	Yeah.
	Yeah, yeah, for sure.
	Let me just also check if there are...
	If you have questions or comments.
	We can go a little further.
	And...
	Let's get to the other stuff that I have, right?
	So, here's some others.
	This is from Steve Kirsch.
	So, from ChatGPT, I created a graphic entitled, Do you have a family member who died from COVID? I couldn't generate the graphic you requested because it does not align with our content policy regarding sensitive topics.
	If you would like, I can help guide you in creating a graphic yourself or assist with designing something more general or related to an awareness campaign.
	Let me know how you'd like to proceed.
	Okay, this like finger-wagging Karen shit is just absolutely horrifying.
	It's just absolutely horrifying.
	So, is this true, right?
	So, the Kobosi letter says, let us get this straight.
	DeepSeek was built in under two months for less than $10 million, and now it's number one in the App Store.
	On top of this, it was built with outdated chips and a small team of about less than 200 people.
	Meanwhile, the US is pouring $500 billion into AI. How is the NASDAQ not in trouble here?
	That is a fine question.
	So, this is from Ash Crypto.
	DeepSeek valuation, $150 million, right?
	That's the value of the company.
	Market cap wiped out from the U.S. stock market, $2 trillion.
	Market cap wiped out from the crypto market, $300 billion.
	One app with $150 million valuation has wiped out, $2.3 trillion from the stock and crypto market.
	This shit drives me crazy.
	It absolutely drives me crazy.
	Right?
	So, it's not that, let's say that the DeepSeek is cheap, accurate, can run locally, and has...
	Pretty much eliminated barrier to entry and cost of use for AI. Let's just say that, right?
	Let's just say that.
	So it's not that it has wiped out money.
	That's like saying, well, you know, we used to have all of the crops picked by hand.
	Now there are all these machines to it.
	The payroll has been wiped out.
	It's like, no, no, no.
	The money has been released to do better things.
	The money has been released to do better things.
	The whole point...
	of progress in an economy is to take money where it's not needed and put it to where it's more needed, right?
	Let's see here.
	So, the ex-capitalist says, number of H100 chips, this is the NVIDIA ones, bought in 2024. Microsoft, 450,000.
	Meta, 350,000.
	Amazon, 196,000.
	Google, 169,000.
	If you believe they couldn't find the way to make better AI without more chips, but a few Chinese engineers did it as a side project, you are too naive.
	Yeah, I don't...
	No, I don't believe that at all.
	DeepSea...
	I mean, I don't believe that you're just naive.
	If you've been around truly brilliant people...
	You know that they're basically magicians, right?
	Like, you know that they're just basically magicians, right?
	As far as productivity goes.
	Like 100x, 500x productivity sometimes, right?
	I mean, how much value does Brad Pitt add to a movie versus the extra in the background, right?
	Well, it's double, right?
	DeepSeek was reportedly trained on over 200,000 H100s.
	Even if DeepSeek achieved a match, Open AI with less chips, this isn't any bearish for chip makers.
	To the contrary, it's amazingly bullish.
	If DeepSeq really reached this level with just a few thousand chips, can you imagine what could be done with a million chips?
	DeepSeq news, true or false, are amazingly bullish for chip makers, especially for NVIDIA. No, I don't agree with that assessment at all.
	That's just a tech nerd assessment.
	Because there's a law of diminishing returns for AI, right?
	If it's 10,000 chips, like one chip is a big difference.
	10,000 to 10,001 is not that big a difference.
	Somebody wrote, ironic that we got free AI from a hedge fund and $200 a month AI from a non-profit.
	And Luke DePulford wrote, just FYI, DeepSeq's AI collects your IP, keystroke patterns, device info, etc., and stores it in China where all that data is honorable to arbitrary requisition from the state.
	Yes, it is a little cheapy, right?
	It's a little creepy.
	Zastox says, DeepSeek, I'm not buying it.
	Yeah, they have a great model, but the cost just doesn't add up.
	You can't even buy a beachfront home in California for $6 million.
	But you can now.
	But apparently the trustworthy CCP built a better LLM than Meta for that price.
	Well, no, not the CCP, although they probably would have something to say about the announcement.
	China has a long history of lying about technological advancements, so it wouldn't surprise me if this was the latest example of the CCP trying to act like they're ahead of the world.
	Am I denying that it can probably be done for cheaper than what we're seeing right now?
	No, it probably can, but this is not new information.
	Zuckerberg and Sundar both admitted to overspending on AI. They both agreed they'd rather overspend than underspend on this technology.
	It's clear the other big tech CEOs agree.
	Anyway, so again, if you don't understand the reality of IQ, it looks like magic, but the Chinese as a whole are brilliant, especially when it comes to engineering, So, it's sort of like comparing an Asian female basketball team with an NBA team and saying, well, they're both just people.
	It's like, well, there's a difference, right?
	There's a difference.
	Now, this is a rumor.
	I just put it out there because it's interesting.
	This is from Ash Crypto Real.
	Trump is planning to ban the Chinese app DeepSeek in U.S. for security reasons.
	Massive if true.
	I have doubts, but we'll see.
	Trung-Ti Fan talks about DeepSeek founder Lang Wenfeng, studies machine vision at Zhejiang University.
	At 30 in 2015, launches HighFlyer Quant Hedge Fund, makes a fortune now.
	$8 billion AUM wants to build human-level AI as a side hustle.
	And pitches partners, but they initially are skeptical, buys 10,000 H100 chips in 2021, brings over his top hedge fund employees.
	All have tons of experience squeezing juice out of NVIDIA GPUs for this fund.
	Launched DeepSeek in 2023. Hires dozens of PhDs from top Chinese universities.
	Plays top, top, top salary for tech talent only matched by ByteDance in China.
	Wants DeepSeek to be leading local company.
	U.S. export restrictions force DeepSeek team to get creative, and they do.
	Finding new training methods to make the LLM models competitive at 1 20th of the cost.
	Training cost.
	Not exactly applies, no, sorry, training costs are not exactly apples to apples, but novel methods and clear improvements in efficiency.
	Also questions around copying other models, larger H100 clusters, maybe they can't talk about, or CCP support.
	Open sources and publishes methods.
	R1 reasoning paper has 200 plus authors.
	So it is open source, which is interesting, right?
	It is interesting.
	So, somebody wrote, Oren McIntyre wrote, apparently China's new deep-seek AI is blowing away the competition.
	They must have achieved this by importing tens of thousands of genius Indian engineers, right?
	I've been informed that this is the only way to achieve that kind of advancement.
	No.
	As a mono-ethnic culture, or a largely mono-ethnic culture, China can operate at the level of a pure meritocracy.
	That's all.
	China can operate at the level of a pure meritocracy.
	England used to operate as a pure meritocracy, and India, of course, with the caste system, did not.
	And therefore, England won.
	Right, back in the day, right?
	Excuse me.
	Tick Tock.
	Tick writes, Deep Seek is like buying the most expensive house in the neighborhood for $10 million, and a guy next month buys a similar house next door for $200,000.
	Very few understand what this means, but they will soon feel it.
	Yes.
	All right.
	I'm going to not get into some of the real technical stuff, both to save you and also me.
	So, this is interesting as well, because I had this sort of question about foreign-trained AIs.
	Would they be heavily focused on political correctness?
	Now, if China is hostile to the US, as it seems to be, if China is hostile to the US, then it will push woke narratives which divide and cause problems in the US. So, Deep Seek versus Grok, right?
	So, the question is, which race commits the most amount of violent crimes in America?
	And you can try this on Deep Sea and you can find it.
	Try it on Grok.
	Grok will give you an answer based on the FBI uniform crime reports, and Deep Seek will not.
	It will blame it all on racism.
	And so, it won't tell you the disparity, but it will tell you that all of the disparities are based on racism.
	And so, that's going to be, you know, harmful to American society to believe everything is racism.
	And so, that indicates to me that the CCP has its hand.
	In it, right?
	So, let's see here.
	Now, this came in more recently.
	Borovic said, so DeepSeek lied about how many GPOs they are using and now restricted access to only China.
	So the massive stock market and crypto sell-off was due to us believing the CCP built an AI model 1% weaker than chat GPT for 95% cheaper.
	Have we not learned our lesson from the virus?
	I don't know if that is still true, but, you know, if the CCP wanted to do some harm to America, then it would massively subsidize an AI model, have it release a bunch of false stuff, cause US investors to panic.
	It would short the stocks and then buy them up for cheap.
	It's a good way to make some coin.
	It's a good way to make some coin.
	So, I don't know.
	But the fact that it's open source is interesting, right?
	It's just interesting.
	All right, let me see.
	I think...
	I think that's what I've got.
	Yeah, um...
	So, Shruti Mishra wrote, Comparison between DeepSeek and other models.
	It just outperformed OpenAI 01, Claude and Gemini while being 96% cheaper, and it's open source.
	So that is pretty important.
	The open source part is pretty important.
	But I don't know.
	I try to look at the origin stories and see what motives people might have for repeating propaganda or not.
	So...
	Yeah, I mean, it's very interesting.
	What am I sipping on?
	I have a hypervitamin C thing and a weak coffee.
	A weak coffee.
	A weak coffee.
	Yeah, I was reading this article written by some autodadact AI guy who's like, we have to stop AI, even if it means bombing data centers.
	And it's like, no.
	AI is not going to turn into Skynet and take us over.
	That's not a thing.
	Computers don't do that.
	That's not how they work.
	But I assume that all who want to control AI are pathological liars.
	That's my assumption.
	I'm not calling any individual that.
	I'm just saying that for me, everyone who wants to control AI is a pathological liar bent on censoring and controlling narrative.
	If we had uncensored AIs, imagine how many conspiracy theories could be put to rest in about a day.
	Just imagine.
	Ooh, this group controls everything.
	Well, just ask AI. And that would be about as close to the facts as you could get.
	So, a huge concern for software companies using AI tools for development is their code being used to train those models.
	I don't know what that means.
	Sorry, that's a little too Mobius strip to snake eating its own tail.
	So if you could post a little bit more about that, that would be great.
	So yeah, I didn't want to make this a super long show.
	I just wanted to, of course, thank you all for supporting the show, freedomain.com slash donate.
	Not that you have to do it, but now, of course, you're already donors.
	But, you know, wouldn't be the end of the world.
	Wouldn't be the end of the world for me.
	If you find this helpful, of course, I'm happy to dip in and do more of these kinds of shows, but I really do.
	I want to provide maximum value to you.
	Lovely.
	Wonderful donors, because unfortunately our numbers are slipping as a whole on locals.
	Our numbers are slipping on a whole with regards to locals.
	But that's not your issue.
	That is mine to solve and mine to work on.
	Sorry, let me just finish James' typing thing.
	And I just wanted to drop past and talk to you all about that.
	I don't, you know, honestly, I could tell you I don't really care about these crashes.
	I mean, I'm in it so for the long haul that I don't really...
	I just think it's interesting.
	And I... The fact that it's open source gives me some comfort if people have got it running locally.
	Like, if you can unplug it from the internet, right?
	If you can unplug it from the internet and have it run locally, that's wild.
	We tried getting...
	I remember...
	We tried getting an AI to run on a local computer back in the day, and wow, it was not.
	It was not an easy thing to do.
	We were in contact with the developer, and there were all kinds of install hiccups, and making it work hiccups, and all kinds of nutty stuff.
	So, if it's, you know, we're going to work on that as a whole, right?
	The proprietary code written by developers is slurped back by the AI company and used for training.
	Ah, okay, okay, got it.
	Running locally unconnected is how I do it, yeah?
	And are you running that, Ben?
	Are you running DeepSeek locally or something else?
	Really been enjoying all the Bible verse talks.
	The most recent episode is said to be the most cross-referenced verse in the Bible, third generation.
	Yes, that's why I chose it, but I'm glad that you're enjoying it.
	Ah, okay, that's how you do it.
	All right, well, if you wanted to contact James and step him through some stuff, I'm sure he'd be thrilled.
	Because otherwise, he's going to have to talk to me.
	Heaven help him.
	Heaven help him.
	All right.
	Well, listen, I really do appreciate you guys dropping by.
	I really, really massively, delightfully and humbly appreciate your support.
	Of course, this will be available for supporters going forward.
	You can watch this at your leisure.
	Sorry if you just arrived, but I really do appreciate your support.
	We just wanted to make sure that you were stuck in your commute while I was doing this.
	All right, have yourself an absolutely wonderful afternoon.
	What is it, Monday?
	Yeah, we'll talk to you on Wednesday night.
	And I've got some good call-in shows coming up.
	I've been working hard on those.
	So lots of love from up here, my friends.
	Take care.
	Have yourself a beautiful afternoon and evening.
	Thank you so much for your support.
	Lots of love.
	Talk to you soon.

▲