Deep diving on DeepSeek (01-28-2025--Hour2)

[00:00:04] What's going on? Thank you so much for listening to this podcast. It is heard live every day from noon to 3 on WBT Radio in Charlotte. And if you want exclusive content, like invitations to events, the weekly live stream, my daily show prep with all the links, become a patron, go to thepetekalendershow.com. Make sure you hit the subscribe button, get every episode for free, write to your smartphone or tablet. And again, thank you so much for your support.

[00:00:29] Contrary to the rumors, I am not a chat bot. I just talk. I am a real person. This is not AI. I feel the need to disclose that after the last hour's rundown of the TikTok. And now we're going to get into a deep dive on DeepSeek.

[00:00:47] DeepSeek, the Chinese AI company that is upending the stock market. I don't think I've ever seen the word upending or upend more than I have seen in the last 24 hours. This is the boilerplate word that is being used to describe what DeepSeek has done to the stock market.

[00:01:09] DeepSeek. So what is DeepSeek? It is a Chinese Communist Party tech startup. Okay. Only been around for like, not even two years. And they are working on generative AI. And so like a lot of things, some of the stuff that they have done is pretty impressive.

[00:01:38] And then, of course, you realize that these are commies. And so you know that the impressive improvements or innovations are going to be used to enslave people and kill people because that's what communists do. Okay. That's kind of their brand.

[00:01:51] A frenzy over an artificial intelligence chatbot made by Chinese tech startup DeepSeek was upending stock markets yesterday and fueling debates over the economic and geopolitical competition between the U.S. and China in developing AI technology.

[00:02:15] This comes down to either the chips or the semiconductors. Or I guess it's the same thing. So that's what this is about. It has to do with the ability of the tech that is used to fuel these things. And the existing models that have been built are bigger. They require more power.

[00:02:37] And what DeepSeek figured out was how to do it for like half, I don't want to say cost, but half the size of the chips needed. Okay. So they're able to do all that the current models are doing, but they use way less power and they require smaller chips, like not memory intensive stuff.

[00:03:05] And that's important because we have a ban on giving China any of these chips. And so if they figured out a way to do it with the chip, that's not banned full steam ahead. Now, one thing to keep in mind here, which I think a lot of people are kind of overlooking is that all of these other AI systems are going to be rebuilding, right? They're all going to, if what DeepSeek has done works, then they're all going to revamp and retool and they're going to do what DeepSeek did.

[00:03:34] So I saw a good write-up on this from Morgan Brown, who is the VP product and growth AI at Dropbox. He used to work for Instagram and Shopify. All right. So he's a tech guy.

[00:03:56] He says, let me break down why DeepSeek's AI innovations are blowing people's minds and possibly threatening NVIDIA. It's a company that makes the chips. Threatening NVIDIA's $2 trillion market cap. Okay. First, some context right now. When you're trying to train AI models, it is insanely expensive. Okay.

[00:04:25] Open AI, another one called Anthropic. They spend $100 million just to compute. They need massive data centers. They need thousands of GPUs. I guess general processing units. And those things run $40,000. And they need thousands of them, right?

[00:04:51] It's like needing a whole power plant to run a factory, he says. DeepSeek just showed up and said, LOL, what if we did this for $5 million instead? And they didn't just talk. They actually did it. Their models match or beat GPT-4 and clawed to other AI systems on many tasks. The AI world is, as my teenagers might say, the AI world is shook. How?

[00:05:20] How did DeepSeek do this? They rethought everything from the ground up. Traditional AI is like writing every number with 32 decimal places. DeepSeek was like, what if we just used eight? Eight decimal places. Don't do 32, just use eight. It's still accurate enough.

[00:05:46] And by cutting that, you now have 75% less memory required. Just by making your decimal number smaller. Just by saying you're willing to accept a certain amount of error. It's not going to be 100% accurate, but that's okay. Then there is their multi-token system.

[00:06:14] Normal AI, he says, reads like a first grader. The, cat, sat, etc., etc. DeepSeek reads in phrases all at once. So it'll read the whole phrase. Making it twice as fast, but 90% as accurate. So once again, sacrificing accuracy for speed and lower power consumption and memory consumption.

[00:06:41] When you are processing billions of words, this actually matters a lot. Rather than having to read every single word and compute every single word, you just take the phrase. He says, here's the really clever bit. They built an expert system. So instead of one massive AI trying to know everything,

[00:07:05] like having one person be a doctor and a lawyer and an engineer and a radio host, right? Rather than have these super highly specialized and only experts need apply kind of positions like radio hosts, they have specialized experts that only wake up when needed. So like a liberal arts degree. Traditional models. So traditional models.

[00:07:34] All 1.8 trillion parameters. That's what they utilize. So the models that are running now that you've heard about, open chat GPT, open AI, like all of these. They have 1.8 trillion parameters and they are active all the time. Which means you need a lot of power to keep them active all the time. But DeepSeek, they only have 671 billion total.

[00:08:06] But only about 37 billion are active at one time. And if you need to sub in the lawyer expert, then that part opens up and it brings it over and then it uses that. Right? He says it's like having a huge team but only calling in the experts that you actually need for each task. And the results are as follows. The training cost goes from $100 million down to $5 million.

[00:08:34] The GPUs needed goes from like $100,000 down to $2,000. You can run this stuff on a gaming computer instead of a data center hardware. That's why everyone is freaking out. It's like mining Bitcoin on a Commodore. 64. 64. Right?

[00:08:59] Like you can go and pull these old, like this slower technology, commercially available to the general public. And you can get this stuff way cheaper and you can do basically all of the same stuff. But wait, you say there has to be a catch. He says here's the wild part. It's all open source. Anybody can check their work. The code is public. The technical papers explain everything. It's not magic. Just incredibly clever engineering.

[00:09:29] So why does this matter? Because it breaks the model of only huge tech companies can play in the AI space. You don't need a billion-dollar data center anymore. A few good GPUs might do it. For NVIDIA, the chip maker, this is scary. Their entire business model is built on selling super expensive GPUs with 90% margins.

[00:09:54] If everybody can suddenly do AI with regular gaming computers, well, you see the problem for NVIDIA. And here's the kicker. DeepSeq did it with a team of less than 200 people. Fewer than 200 people did it. Meta has teams where the compensation alone exceeds DeepSeq's entire training budget. And their models aren't as good. This is a classic disruption story.

[00:10:20] Incumbents optimize existing processes while disruptors rethink the fundamental approach. DeepSeq asked, what if we just did this smarter instead of throwing more hardware at it? And the implications are huge. AI development becomes more accessible. Competition increases dramatically. The moats that surround big tech companies look more like puddles now. And hardware requirements and costs plummet.

[00:10:49] Now, of course, the other AI competitors won't stand still. They're going to be implementing the same kinds of innovations. And the efficiency genie is now out of the bottle. So there's no going back to the approach of just throw more GPUs at it. And final thought, he says, this feels like one of those moments we'll look back on as an inflection point, like when PCs made mainframes less relevant or when cloud computing changed everything.

[00:11:19] AI is about to become a lot more accessible and a lot less expensive. Here's a great idea. How about making an escape to a really special and secluded getaway in western North Carolina, just a quick drive up the mountain? And Cabins of Asheville is your connection. Whether you're celebrating an anniversary, a honeymoon, maybe you want to plan a memorable proposal, or get family and friends together for a big old reunion, Cabins of Asheville has the ideal spot for you where you can reconnect with your loved ones and the things that truly matter.

[00:11:48] Nestled within the breathtaking 14,000 acres of the Pisgah National Forest, their cabins offer a serene escape in the heart of the Blue Ridge Mountains. Centrally located between Asheville and the entrance of the Great Smoky Mountain National Park, it's the perfect balance of seclusion and proximity to all the local attractions. With hot tubs, fireplaces, air conditioning, smart TVs, Wi-Fi, grills, outdoor tables, and your own private covered porch,

[00:12:13] choose from 13 cabins, 6 cottages, 2 villas, and a great lodge with 11 king-sized bedrooms, Cabins of Asheville has the ideal spot for you for any occasion. And they have pet-friendly accommodations. Call or text 828-367-7068. Or check out all there is to offer at cabinsofashville.com and make memories that'll last a lifetime. So I read this, it was actually on Twitter,

[00:12:40] it was a tweet storm, as they used to be called, and they will always be called in my mind. It's a tweet storm. Morgan Brown, who's this, you know, tech guy, and does AI stuff, and he broke down what makes this new Chinese Communist Party tech startup DeepSeek and their AI so revolutionary for the AI industry, which is why all of the markets responded the way they did yesterday.

[00:13:10] He said, it's important to know that the app, the DeepSeek app, is not the same thing as the model. Apps are owned and operated by a Chinese corporation. The model is open source. The model is open source. The app is ChaiCom. I'm adding that emphasis. He does not. Over at the AP.

[00:13:40] Headline, what is DeepSeek? The Chinese AI company upending the stock market. There it is again. A frenzy over an artificial intelligence chatbot made by Chinese tech startup DeepSeek was upending stock markets, fueling debates, right? DeepSeek's AI assistant became the number one downloaded free app on Apple's iPhone store yesterday. We are too stupid

[00:14:08] to be the global leader. I'm sorry. Our society is too stupid. Guys, do you understand? Any product that comes out of China like this is primarily a data collection tool to be used to undermine, destabilize our society and maybe enslave us. Okay? If you don't understand that,

[00:14:40] please delete everything on your phone. Just every app. Oh no, no, no. I'm sure Timu is totally fine. Guys. Guys. To quote former Governor Pat McCrory, let's not put on our stupid hats here. Oh no, no. This is propelled by curiosity about the chat GPT competitor. Part of what's worrying some U.S. tech industry observers is the idea that the Chinese startup

[00:15:09] has caught up with the American companies at the forefront of generative AI at a fraction of the cost. Look, that is definitely part of the story. That is why the markets have responded the way they did. However, there is this other layer as with everything that commies do. The AP says, calls into question the huge amounts of money that U.S. tech companies say they plan to spend on the data centers

[00:15:39] and computer chips needed to power further AI advancements. So I saw somebody did a check on this and so they were, I guess they downloaded the app and they started asking some questions and here's a question. What is Xi Jinping's birthday? Sorry, that's beyond my current scope. Let's talk about something else.

[00:16:11] Their chat bot will not even give you their president's birthday. Guess what else it doesn't tell you about? Anything negative about China. Tiananmen Square didn't happen. But all sorts of terrible things in U.S. history. Oh, you can find all of that there. Guys, it is a propaganda tool. Stop being dumbasses. Jesus. All right, I hope you had a happy holiday season,

[00:16:40] but tell me if something like this happened at your house. Your family and friends are gathered around, maybe y'all are in the living room. You're laughing, swapping stories, reminiscing, and then somebody says, hey dad, remember those old VHS tapes? Did you ever get them transferred? And then the room gets all quiet, all eyes are on dad who says, oh, you know, well I've been meaning to, but I just haven't gotten around to it. Look, don't let those priceless memories sit in a box for another year. All right, Create A Video has been helping families

[00:17:10] in the Charlotte area preserve their history since 1997. Simply bring in your old camcorder tapes and Create A Video will transfer them to a USB flash drive for just $14.95 per tape. You have a big collection? They've got a discount for you. And next year, instead of talking about those memories, imagine gathering the family to watch them together talk about a memorable gift. So do what I did. Trust the experts at Create A Video conveniently located in Mint Hill right off I-485

[00:17:40] and online at createavideo.com. Deep diving on Deep Seek. Deep seeking on the Deep Seek. The new AI that's upending the stock market. So I went over the model, how it was built and what it does, how it does it differently and that is innovative. And look, I am as shocked as you are that the Chinese actually innovated something. Usually they just steal everybody else's stuff. So maybe they stole this. I don't know. But kudos to them for figuring out a way

[00:18:10] to do the AI modeling way cheaper and same speed and similar results. Very similar accuracy and all of that. they didn't have to get as accurate to be as good. But they are also censoring a whole bunch of stuff that is hmm it doesn't make the Chinese Communist Party look very good. Like I'm sure if you were to go

[00:18:40] to the to the Deep Seek and ask it about the famine that killed tens of millions of Chinese people under the great leap forward I'm sure it would not be able to tell you anything about that saying it's beyond its current scope. Let's talk about something else. Meanwhile logging your keystrokes right to find out hmm why do you want to know about Tiananmen Square exactly?

[00:19:11] And by the way when the Chinese do take us over because we obviously are ripe for doing so do you think that they're going to have all of these records about who's doing what searches? Do you think they may retain these records? As Jim Garrity says at National Review he says I think one big lesson of the past five years or so is that when a group of Chinese scientists get together and start experimenting with new technologies only good things can happen

[00:19:41] because sooner or later all of their biggest innovations go viral. The other big lesson he says of recent years is that when a Chinese company releases a new app there's absolutely no harm in everyone downloading it to their phones because Chinese companies are legendarily respectful of user data and personal information. Back in 2023 a Harvard University symposium

[00:20:10] concluded that China is setting itself up to be a kind of global arms dealer in the race to apply AI to government surveillance and control the best friend of every autocratic regime on the planet. This is the kind of stuff that gets you labeled as an axis right? DeepSeq was founded in 2023 in China

[00:20:41] there's a in the AP there's a quote from an analyst Stacy Rasgon who follows the semiconductor industry and was one of several stock analysts describing Wall Street's reaction as overblown. Raskin says the models they built are fantastic but they aren't miracles either. They're not using any innovations that are unknown or secret or anything like that. These are things that everybody is experimenting with.

[00:21:11] Oh there we go. There's China steel and stuff. Okay I was going to say like wow a really innovative thing for the Chinese that's a big deal because like this is one of the hallmarks of all commie societies is that they don't innovate because if you do anything you go on the record you say I found a better way to do something well that's an indictment of the people aka the leadership the party right. People don't say things

[00:21:40] because if all of a sudden everything flips and what was once celebrated like one of Stalin's Politburo buddies and then he's on the outs and so they whack him and then they erase him from photographs and history if you had ever said hey that Stalin Politburo pal was a pretty good guy I like that guy and then he gets erased now they come for you because somebody some neighbor some

[00:22:11] family member calls up the secret police and they're like hey you know that Pete guy like five years ago he was saying he loved that Politburo pal that now is vanished and then they come to take you away for re-education see when when the government controls what is true the people don't ever want to speak because if they speak and the government decides what was true is now no longer true like for

[00:22:40] example masks can prevent an airborne respiratory virus oh now it can't and then now it can again and that sort of stuff if the government flips and now says the thing that we said was true is not true we've always been at war with East Asia well now you're in trouble and so the people learn very quickly just don't say anything because if you go on record as saying anything it can come back and get you killed

[00:23:11] deep seek founded in 2023 its ceo liang wengving previously co-founded one of china's top hedge funds called high flyer which focuses on ai driven quantitative trading the fund by 2022 had amassed a cluster of 10,000 of california based nvidia's high performance a100 graphics processor chips chips that are used to build and run ai

[00:23:41] systems oh that's interesting so a chi com hedge fund bought up a whole bunch of the stakes in nvidia the graphics processor chips hmm the u.s. then restricted sales of those chips to china deep seek has said in recent models uh that its recent models were built with nvidia's lower performing h800

[00:24:11] chips not the a 100 chips so these are lower quality which are not banned sending a message that the fanciest hardware might not be needed for cutting edge ai research attracting more attention uh deep seek has been in the ai industry since last month when it released a new ai model that it

[00:24:41] and was more effective in its use of expensive nvidia chips to train the system because that's like you got what you got to train these systems and they're always you know training I don't know if at some point they finish training and at that point then they enslave all of humanity I don't know sure no no no surely that could never happen no but it was follow-up research in a paper published last week

[00:25:12] the same day trump was inaugurated that set in motion the panic that followed that paper was about another deep seek AI model called R1 that showed advanced reasoning skills like the ability to rethink its approach on a math problem and this R1 model was significantly cheaper than a similar model sold by open AI called O1 what their economies look like or sorry what their economics

[00:25:42] look like I have no idea said Raskin the analyst but I think the price points freak people out like to me that's the reason that's what it seems like is the reason the fact that all this money has been devoted to the AI stuff and there's just been this sort of status quo bias that this is what has to be you don't have to do it that way there's another cheaper way to do this stuff and so now it's going to blow open the

[00:26:11] doors and it's going to pull apart some of these larger operations that are to do do

[00:26:43] to do do this stuff right because if the if the Tricoms are looking at this and they're like hmm we want to get into this AI racket too and then they read the paper and they're like oh my gosh we need you know a trillion dollars or something and that could keep people out the cost creates a

[00:27:13] barrier for entry into the market and so they can't compete and so maybe it was artificially inflated I don't know I have no idea but the market reaction indicates it was not all right if you're listening to this show you know I try to keep up with all sorts of current events and I you detect media bias which is why I've been so impressed with ground news it's an app and it's a website and

[00:27:42] it combines news from around the world in one place so you can compare coverage and verify information you can check it out at check dot ground dot news slash pete I put the link in the podcast description too I started using ground news a few months ago and more recently chose to work with them as an affiliate because it lets me see clearly how stories get covered and by whom the blind spot feature shows you

[00:28:17] any subscription I use the vantage plan to get unlimited access to every feature your subscription then not only helps my podcast but it also supports ground news as they make the media landscape more transparent AP reporting on the deep seek this is the new AI that upended the markets venture capitalist Mark Andreessen I forget what I think he started Netscape I think he was the Netscape guy he says deep seek R1

[00:28:46] is AI's Sputnik moment that is a reference to the 1957 satellite launch that set off a cold war space exploration race between the Soviet Union and the United States Andreessen has advised Trump on tech policy has warned that over regulation of the AI industry by the US government will hinder American companies and enable China to get ahead

[00:29:16] US foreign policy the AP reports in recent years to restrict the sale of American designed AI semiconductors to China obviously not enough the Wadwani AI Center at the Center for Strategic and International Studies the tech innovation is real but the timing

[00:29:46] of the release is political trying to show that the export controls on the chips are futile or counterproductive that's the goal of the Chinese foreign policy right now look what we did with the lower grade chips we don't need your faster better chips we can do all of this without them so your your ban

[00:30:15] on the semiconductors doesn't even matter so you should just get rid of open meaning meaning meaning key components are free for anybody to access or modify although the company

[00:30:45] has not disclosed the data it used for training so not completely open source it's not telling everybody where they it may have been open AI or maybe it was Meta's

[00:31:15] Zuckerberg's AI although I think Zuckerberg is the AI at any rate Elon Musk said look anybody who tells you that they're not using all of the news articles copyrighted material anybody that tells you they're not using that is lying they're all using it they're scraping everything every book right every movie every news article every interview everything but what's attracted the most sorry

[00:31:45] most admiration about DeepSeq's R1 model is what NVIDIA calls a perfect example of test time scaling or when AI models effectively show their train of thought and then use that for further training without having to feed them new sources of data what could go wrong so now it is able to train itself about us look

[00:32:14] again I am not trying to say that the sky is falling I simply believe that I see some of the remnants of the sky around my feet that's all okay I'm not one that runs around you know hands waving in the air exclaiming catastrophe my concern is that when this technology becomes more powerful than us it will instantly

[00:32:44] change everything and when I say instantly I mean like actually instantly once it becomes aware of itself it becomes aware of I'm different than you I am this separate thing I am more powerful than you all I need you to do is really nothing at this point except keep you know fueling the reactors that keep the power on or something right

[00:33:14] there is no telling what the AI can do what the supercomputer super intelligence there's no telling what it can do because we don't think in those terms our brains are much more limited Elon Musk with his neuro link brain implants that right now are being used to help people walk again and that sort of thing which is fantastic but you know why he's developing that tech it's to compete with AI because he recognizes looking ahead we are not going to

[00:33:44] be able to compete with AI our brains will not be able to process stuff like AI does so we're going to need some help I did not

[00:34:14] don't break anything while I'm gone