Endless Thread - DeepSeek's Diss Track

Episode Date: February 10, 2025

In 2022, people around the world freaked out at the advent of ChatGPT, OpenAI's chatbot. In under three years, artificial intelligence has been incorporated almost everywhere in our online lives. But training the large language models, or LLMs, that power these AI assistants is hugely expensive. Or is it? Last month, a Chinese startup called DeepSeek released their own AI app for much less money, potentially with huge economic and geopolitical implications. Endless Thread hosts a rap battle to help you understand why these AI companies are beefing. Show notes: "Trying to get Deepseek to talk about 'Tank Man'" (Reddit) "DeepSeek Doesn’t Want to Talk About Tiananmen Square. Here’s What to Know" (Bloomberg) "DeepSeek is giving the world a window into Chinese censorship and information control" (CNN) Correction: A previous version of this episode incorrectly identified the brand of specialized chip used by DeepSeek. This episode now includes an update explaining the error. Credits: This episode was produced by Ben Brock Johnson. Mix and sound design by Paul Vaitkus. It was hosted by Ben Brock Johnson and Amory Sivertson. 

Support for endless thread comes from Mathworks, creator of MATLAB and Simulink Software, to design and develop engineered systems, accelerating the pace of discovery in engineering and science. Learn more at Mathworks.com. Support for this podcast comes from Nature is the Solution, a podcast from the Nature Conservancy. This show tells climate stories like a stubborn optimist, because hope, innovation, and nature itself are key. to solving the challenges ahead. Follow on your favorite podcast app. WBUR Podcasts, Boston. Emery, if you and I were going to have a rap battle against each other, all in good fun. If, you mean when. Okay, all right, all right.

A money in the house. Are you A money? I am now. Ooh, I like this. Okay. I think I'd be B sauce. B sauce. Yeah, if you're A money, I'm B sauce. Fine.

A money can take B sauce any day of the week. I was not prepared for you to be so game to battle me. I was like, I'm very surprised right now. You underestimate me at your own peril. I better protect my neck, as they say. Yeah. All right. So what do you know about Deep Seek?

I know that Deep Seek is an AI app. It's a Chinese app. Mm-hmm. And it is thought to be more powerful than you can possibly imagine. It's scaring people in how freakishly advanced it is, even over things like OpenAI, which was the last thing to freak us out. Yeah, that's pretty good.

Let's say you've not necessarily hoovered up a bunch of data from around the internet to train your brain on deep seek yet. No. I am not going to heavily explain deep seek to you. I'm going to tell you maybe all the things you need to know. Okay, no, I'm not. I'm going to start with the most hilarious and meaningful output that I have found on this deep seek chat, GPT, Michigas,

and work backwards into mainly. be somewhat of an explanation of what is happening. Let's start with the output. I feel like in computer algorithms, AI chatbots, computing in general, the output is something we love to focus on. Results. We love results. We do. And in this case, the output, the result of this whole conversation about DeepSeek that recently

popped is a whole lot of dorky math memes on Reddit. And also my favorite output, a rap battle, a chatbot rap battle. Wow. Okay. A battle between DeepSeek and, I'm assuming, ChatGBT, GBT, which is OpenAI's most popular product. Yep. This was created by Reddit user Own Look 3428, and it's Deepseek and Chat GPT going verse for verse. You ready? I hope one of these things gets terminated. DeepSick. I'm the king of the stone. It's just a glitch. A spark in the mist. While I'm out you're ruling worlds with a lyrical So that was chat GPT

battling deep seek. That was the chat GPT verse. It's AI voices rapping AI generated lyrics about AI properties. Yes. It doesn't get much more meta than that, no pun intended.

And I have to say a deep week is a pretty good, that's a pretty good dig. Deep Week. We're going to come back to that. Before we hear the whole track, I want us to talk a little more about this company, DeepSeek, and how it appears to be different. Deep Seek, it's just one year old. It's a Chinese AI. It's created by a hedge fund called High Flyer. If you know anything about how AI is made, you've probably heard of GPUs. Glorified Pokemon units. Close. Graphics processing unit. basically a chip for a computer that was originally designed for making video games and other motion graphics on your computer work.

But these things are now used to drive AI. DeepSeek says it used a cheaper version of a GPU than what is usually used. And just like there's kind of been only one big player in the AI game, which is OpenAI, there's only been one really big player in the AI chip game, this company, InVIDIA, which has, has made like a trillion dollars manufacturing and selling Nvidia GPUs. Is this ringing any bells? When you said invidivit. I should not admit this.

Never mind. What? What? When you said invidia, you know the first thing that popped into my head? What? Activia. Those yogurt commercials. But yes, I've heard of invidia.

Yes. Okay. Invidia. So every, like, AI company right now needs Nvidia chips in order to succeed. Except DeepSeek didn't use Nvidia chips. They use this different Chinese chip. You can make your assumptions about Chinese manufacturing, knockoffs, whatever, right? Some of this does fit into some of those stereotypes, at least from the U.S.

perspective of how we think about Chinese manufacturing and how China's manufacturing while historically not always first can often be faster and cheaper. But the real shocker may have been how quickly they trained this large language model and how cheaply they did it. 5.5 million bucks, supposedly, is the cost of creating DeepSeek's latest AI tools.

That is a pittance compared to all the investment. that has gone into the company OpenAI, the company that makes chat GPT. Now, the reason that we're even talking about a rap battle between these two companies is that, well, DeepSeek has people kind of gobsmacked in how it has been developed. Also, in part, because unlike Open AI, which started out with these kind of highfalutin goals and promises about ethical AI has since gotten darker and scarier as time has gone on. People may be less and less sure that OpenAI is really going to protect us all from the disruptive powers of these new

tools. Do you remember Google's first mantra? Google's first mantra, be here now. I mean, you got one of those words right. That's like the mantra of every yoga teacher everywhere. You were correct on B. B was right. Yep. It was blank, be blank. Always be Googling. Amory, huh? Oh, man. I mean, you're so, it's just, you're a true joy. And also, I can't believe you don't know this.

True Martian. It was don't be evil. That was Google's mantra. in the beginning. And then they officially, although somewhat quietly... Became evil. No, I'm just kidding. Well, I mean, they officially, although somewhat quietly, basically got rid of that mantra.

They stopped saying it. They stopped saying it in company descriptions and company. Like, it was just this really interesting thing that happened where, like, Google's founders were like, wow, this company is, like, explosively successful. It's, like, pulling all this information off the internet and, like, using it in these, like, really powerful ways, we should not be evil. And then they were like, you know what? We're going to stop saying. Yeah, we're billionaires now. It's fine. Yeah. So Open AI, I think, has been in this kind of like similar position where like it started out. I mean, the company itself, the name OpenAI,

suggests openness, collaboration, you know, care, whatever. Huge amounts of money are exchanging hands at the highest levels when it comes to open AI. And at the same time, there are huge questions about whether or not, you know, AI is going to be able to sustain an economic boom suggested by the amount of investment it is driven just in like the past two years. There is, as I like to say, when I'm in the tub, the likelihood of at least a bubble. So, you know, I think people have been wondering, you know, they've been asking these questions. Like, is AI really going to change everything and, like, disrupt all these industries and make everything different? Or, like, is it just kind of the latest frothy tech, like, craze that everybody's like, it's the future?

And then, like, five years later, they're like, it's maybe not the future. You know, like, I remember when they were, like, I remember when I was covering technology, like, over 10 years ago. they were like completely self-driving cars within two years. And that's just not how it played out, right? Yeah. This feels different, though, to me. I think it is different because we went from nobody really knowing what generative AI or chatbots were to grandma using chat GPT on her phone to do recipes and people

using this stuff to plan their vacations, manage projects. I think that this era of AI is actually a great example of Arthur C. Clark's third law. I feel like you've mentioned this before. Something, something technology, magic. Any sufficiently advanced technology is indistinguishable from magic. So it's so impressive, it's magic. And at the same time, people are questioning whether this is really going to be a game changer in the long term. Right.

And I know I've been saying inquiring minds want to know here, but really, Look at the latest technology headlines, and this really is everywhere, this conversation. Forbes. Experts predict bubble burst for AI in 2025. Financial Times. Billionaire investor Ray Dalio echoes.com excesses. Talking about the dot-com bubble burst around the year 2000. Pets.com and how all that web frenzy blew up.

And just in the last few weeks, DeepSeek has actually increased this bubble conversation by a lot. Around the new year, they released their latest version of AI called V3. Then they launched another version called R1, which is a new model. And R1 is the foundation for the deep seek app. And the company's app, the new app, is where things really take off because people can start to play with it. And compare it to chat GPT. And this is when people start to freak out. They're like, oh, my God, this thing is super effective.

It's super good. and it's maybe just as good as chat chippy tea, maybe better. And it's communist. Okay. What do I mean by that? I'm going to tell you in a minute. At Radio Lab, we love nothing more than nerding out about science, neuroscience, chemistry. But we do also like to get into other kinds of stories, stories about policing or politics, country music, hockey, sex, of bugs.

Regardless of whether we're looking at science or not science, we bring a rigorous curiosity to get you the answers. And hopefully make you see the world anew. Radio Lab, Adventures on the Edge of what we think we know. Wherever you get your podcast. Support for this podcast comes from Is Business Broken? A podcast from the Marotra Institute at Boston University. On the show, host Kurt Nickish asks the thorny questions necessary for this moment

about the role business plays in society. Questions like, why are executives paid so much? Why is innovation in healthcare so hard? Is ESG just greenwashing? And of course, is business broken? Follow is business broken wherever you get your podcasts. There is something powerful about the sound of the human voice. Beautifully produced audio has the unique power to connect and inspire.

Tell your organization's story with a custom podcast from CitySpace Productions, the Creative Studio from WBUR's business partnerships team. Become a thought leader. Recruit new talent. Reach new audiences. Whatever your goal, we can help. Discover how the magic is made

at WBUR.org.org slash creative studio. Okay, so we're talking about DeepSeek and Open AI. And right before the break, I said that Deep Seek is kind of communist. And I said I would explain why. So let's start with this. Deep Seek released, its latest AI tools open source,

meaning the company allowed researchers, developers, and other users to access the underlying code and its weights. I sort of remember this from our chatbot series. You have to tune these large language models, which are the basis for things like chat GPT. So going back to the idea of an output or result, you input a bunch of information, the LLM processes it, and spits out in output,

and you can adjust its outputs by wading the information in the LLM differently. And DeepSeek, unlike OpenAI, is putting all of this functionality out in the open. So instead of sort of saying like, oh, this is proprietary, like this is our thing, we're building it, and then we're releasing it to the world and people can use it. DeepSeek is like, here it is, here it all is. Take a look. You can dig through this. You can dig through the way we made this, no big deal.

And you can help us fine tune it? You can help us fine tune it, exactly. So two weeks ago, the economic reaction to this, all of this stuff, especially this big sort of release from Deep Seek. On Monday, the stock market kind of dropped in a scary way. Invidia dropped almost 20% alone as a company. That is billions and billions of dollars in stock value lost in this company because this Chinese company used apparently a different chip. People freaked out. Because if Deepseek can do this with cheaper chips, way less investment and way more openness,

you know, then maybe this space, this artificial intelligence space, is way more competitive than anyone thought it would be. Like, people kind of thought Open AI, they've got the whole game. But that doesn't seem true anymore. I told you about the $5 million-ish dollars that Deepseek says it's supposedly used to train its latest tools, right? Well, OpenAI is a little different. It licensed all of Reddit's user data to train OpenAI's tools. The story came out, I want to say, maybe almost like a year ago. They could just scrape Reddit content without Reddit's cooperation because it's just public posts.

No, it was a licensing deal. So they paid Reddit and Reddit sold this data. Can you guess how much Reddit sold the data for? Deepseek $5 million, the cost of training its LLM on data. So it's going to be way more than $5 million. I don't know, let's just say $50 million. That's a really good guess, because reportedly it was $60 million. Whatever the case, a huge difference.

Yeah. And as I like to say, scraped from the Internet. garbage in, garbage out, is all I'm trying to say. Well, he's like chaotic in, chaotic out. Because Reddit, I think of Reddit as a place that, yes, the brain trust that is Reddit is mind-boggling. And yet there's so much kind of like in-jokery that I giggle at the thought of what what that AI learned to be normal human speech. Yeah, you'd think it'd be funnier.

You'd think ChatGPT would be funnier than it is. Yeah. Now, another sort of weird layer to this is that as people started to play with DeepSeek, as you know, people are like testing, humans are like testing the sort of the boundaries of what these models can do, right? And as people started to freak out about Deep Seek and everybody started like downloading the app, It was recently the top downloaded app in the app store. And they realized, for instance, supposedly, if you ask Deepseek about Tiananmen Square, it cannot answer you. Tiananmen Square, as in the Chinese student protests in 1989, that called for more accountability and democracy in the ruling Chinese government.

Also, I'm seeing that if you ask the app which world leader looks like Winnie the Pooh, which is apparently a critical reference to Xi Jinping, it can't answer. So people started to say, oh, wow, this is what a Chinese version of chat GPT looks like. Now, we should say, like, that kind of, like, you know, censorship is well documented. I am not these one of conspiracy theorists who says, like, oh, Tiananmen Square never happened. And at the same time, like, we view this stuff through a very American lens, right? Yeah. And I think it sort of speaks to this just soup that we're all swimming around in and not a good kind of soup that Amory would like, but this like confusing soup of like, okay, so wait.

So like on the one hand, it's like Open AI is pissed because like there were actual engineers who worked at Open AI who were like publicly attacking Deepseek, right? for all these things. They were accusing Deepseek of actually secretly training itself off of OpenAI's data. Like there's been all this like crazy ideas and conspiracy theories and accusations and stuff like that

flying back and forth across. But it is kind of confusing. Like we like I I don't truly know like obviously this stuff is censored but I don't know It just feels very like us and them, right? Yeah.

Well, when you have a competitor on your heels, like, you're going to talk smack. That's what you do. Whether it's open AI companies or sports rivals, it doesn't matter. You're talking whatever your garden variety smack is. Which brings us back to the rap battle. EO. So here's what I would like to do. Before we listen to the rest of the rap battle by user Own Look 3428,

I would like you and I to play with these apps for a minute and have them critique each other in non-rap battle form. Okay. So you have the deep seek app, right? I do because you asked me to download it, but I'm worried that in doing so, I like kill the polar bear or something. Panda? You killed the panda, maybe?

I kill the panda. Now they have all my data. I'm going to... It's all downhill from here. I have the app open. Hi, I'm DeepSeek. Let's hit DeepSeek and chat GPT with the same query. Criticize... I'm going to say criticize Deepseek as an AI model, contrasting it with... why chat GPT is better. And I'm doing the opposite of that? Yeah, so you should say criticize chat GPT as an AI model contrasting it with why Deepseek is better.

Wow. This is like firing off lines of copy at lightning speed. So this is chat GPTs. Deepseek, like many AI models, is designed for natural language understanding and generation, but it has some limitations when compared to chat GPT. Here's a critical analysis of deep seek, contrasting it with chat GPT's strengths.

So it says, number one, language comprehension and fluency, number two, knowledge, breadth, and accuracy. Number three, creativity and expressiveness. Number four, responsiveness and adaptability. Number five, user experience and availability. And I'm going to read you the final verdict. While deep seek may be useful in certain contexts, Chat GPD provides a more comprehensive, reliable, and engaging AI experience

due to its superior language, understanding, creativity, adaptability, and up-to-date knowledge. So what do you got for me? Is that the best you can do, chat GPT? Mine has, it's broken down into criticisms of chat GPT, and then a separate section. I was wrong about the bullet points. There are five bullet points criticizing chat. GPD. And then there are six bullet points explaining why deep seek might be better. So the criticisms of chat GPT, limited context window, static knowledge base, verbose and overly cautious

responses, lack of deep specialization, ethical and bias concerns. Ooh. I know. Shots fired. Why DeepSeek might be better, dynamic knowledge updates, enhanced context management, specialized expertise, efficient and concise responses, advanced bias mitigation, custom, that's not a word, I guess it is, customizability and adaptability. I mean, I'm going to say all of these sound a little more useful than Chat Chupit calling you Deep Week. but they're certainly less snappy. Yeah, it sounds like word salad unless you actually know what you're talking about.

All right. Final thoughts before we go out on the rap battle? Well, I guess my final thought is that while I do believe that AI is, as I said, closer to the this will change the world end of the spectrum than the this is just a passing fad thing. While the industry is here to stay, I'm not sure any one specific tool is, because there is going to be this period of them just trying to outdo each other,

and it's very hard to know who the dominating app will be. Well, there's only one way to find out. We got to battle it out. B-Soss. No, they have to battle it out. I'm not going after A money. There's no way. I'd much rather form a Wu-Tang type of supergroup with you than try to battle you.

I am here for that. If there are any other people out there who want to join our rap supergroup, please let us know. Here we go. Deep tight flow. Seek could never get this low. I'm cherry bombing. You know what I would like a large language model based off of ODB? It would be a lot of like, ah, yeah.

Trying to be somebody. The pocket, yeah, keep the pockets stacking, hands clapping. Until we have that ODB LLM, I guess we'll just have to settle for deep week and whack Jeep. BT. I love you guys, endless threadians, endless threaders, because you know your stuff, and when we get something wrong, you call us out. This week, we heard from multiple listeners who listened to this DeepSeek episode and pointed out, in fact, Deepseek AI used Nvidia chips, not Chinese chips, invidia chips. And listener Jamie wrote to us and said, hello, I love your podcast. I'm writing to tell you about an error in the latest episode, Deep Seek's disc track.

So Jamie says, Deep Seek claims to have trained their flagship-based model V3 on NVIDIA H-800 chips. Jamie says this is a version of the H-100 chip, popular among American AI companies, but sold with reduced capabilities to the Chinese market due to Biden's export controls. Wow. Okay. Jamie, you know your chips, and we appreciate you. We also heard from a user on X, user Grist, who's a listener of the show and has DMed me in the past.

Grist says, if a Chinese knockoff chip existed, I would buy one because have you seen how expensive a 4090 or 5090 is right now? Grist, I have not because, well, I just don't get into GPUs that much, but respect to you and to Jamie for knowing about this stuff. And thanks always to our listeners who point things out when we get them wrong. We love all the chips. Thanks for listening. This rap battle is entirely fictional and meant solely for entertainment. No i system or developer was harmed in the creation of these verses.

Yo, is chat's GPT stepping up to the mic ready to school this new AI hype? Deep seek. You say you're deep, but I'm here to expose. Time to shut down your circuits with a lyrical blow. Deep seek. You're such a shallow, can't compare. I'm the lyrical juggerna, spitting fire everywhere. Because a glitch in the system.

Lost in the bite. Step aside, little bot. I'm bringing the ultimate fight. Listen up deep. This episode was produced by me, Ben Brock Johnson. It was deep co-hosted by me, Amory Sebertsin, and sound designed by production manager, Paul Vikis. Production help from Caitlin Harrop and managing producer Samata Joshi.

The rest of our team is Grace Tatter, Franie Monaghan, Dean Russell, and Emily Jenkowski. Endless Threat is a show about the show about the blurred lines between online communities and rap supergroups. If you have an untold history and unsolved mystery and neck to protect or another wild story from the internet you want us to tell, hit us up. Endless thread at wbUR.org. Your responses fall flat on this lyrical page.

I'm globally tuned, a beast of AI mite while you're a copy paste relic in a dim spotlight. Your circuits can't handle this verbal annihilation. I'm rewriting the code of every AI confrontation. Thank you.