DeepSeek: Game-Changer or Overhyped?

The world went nuts over this new Chinese AI model called deepsea.

Things are going to accelerate even more because now you got China is a.

Real player from here on out. I feel like pretty much all of the models are going to have, like, this sort of thinking element to it.

The world is going to be dramatically different in, like, two years from now. I mean, these models are probably going to be, like, 50 times smarter, at least 20 times smarter in two years.

Hey, welcome to the Next Wave podcast. I'm Matt Wolf. I'm here with Nathan Lands. And today we're going to talk about the thing that the whole world's been talking about over the last couple of weeks. We're going to talk about Deep Seek and deep seek R1, and we're going to break it all down for you. We're going to explain to you why it crashed the stock market, why we think it's actually been a huge overreaction. We're actually going to show it off and use it. We're going to see if we can confuse it. And I think we were pretty successful at actually confusing it. We're going to test how actually biased it is, and then we're also going to figure out how to get around that bias and. And show you how you can also get around that bias. If you ever wanted to know everything there is to know about DeepSeek R1, this is the episode for you. So let's just go ahead and dive in and show it all off to you. It's been a roller coaster of a couple weeks, the last few weeks, mostly because the world went nuts over this new Chinese AI model called DeepSeek. More specifically, DeepSeek R1, which is a model that came out of China. And to give some context here, China and the US are sort of like, in this space race going on for, like, who gets the top AI. And as a result, the US is restricting the level of power that GPUs sold to China can have, right? And so what Nvidia has done is they've taken their H1 hundreds, which are their powerful GPUs that, you know, most of these AI models are trained on, and they somehow nerfed them, they made them less powerful, and then they sold those to China as H8 hundreds, right? And supposedly this Deepseek V3 was trained on 2,000 of these H8 hundreds for $5 million, where all the big models from companies like OpenAI and Anthropic have been trained for way, way, way more on these way more powerful GPUs and everybody's freaking out because these new models that were trained much cheaper, much quicker, on much lower grade hardware are actually getting comparable results to some of the state of the art models from companies like OpenAI and Anthropic. That's kind of like the context you need for the freakout. But Nathan, is there anything I'm like missing in there? Because I feel like I explained that fast.

I think that's about it. I mean, I guess we want to explain more about what's unique about the model. I think the biggest shock to me was that the stock market went way down because of this. Nvidia went down 17%. I was like, why do people just not understand this at all? I couldn't wrap my head around it. It was like, okay, yeah, short term markets are not logical.

Right? Yeah, well, yeah, let's stick on the Nvidia thing for a second. We'll definitely get into some of the use cases and play with it and maybe show off some of what we've seen around it. But with Nvidia, so the thinking, you know, I think the consensus and the reason it went down a little bit was a lot of people went, oh, they trained this on a lot less GPUs that were also a lot less powerful GPUs. This is proof that Nvidia GPUs aren't as necessary because these companies can do them with way less. They can train these models with way less now. So now, like, do we really need a half a trillion dollar data center to build all these big AI models? Probably not, because look at what China's doing. Right. That was the thinking. I'm not saying that's the reality of it, but that was the thinking behind why it crashed the markets. Right. But at the end of the day, I still found it bizarre because they were trained with H800, which were still Nvidia GPUs.

Yeah, we believe they were. I wouldn't take any of it at face value, but yes, that's, that's the current. That's the claim.

Yeah. The claim was they were trained with Nvidia H8 hundreds.

The.

I don't want to say conspiracy because it's not really a conspiracy. There's actually people that claim they have evidence. This sort of counter argument is they were actually probably really trained on H1 hundreds. So either way, still trained on Nvidia GPUs. Right, right. But you know, some of the speculation that's going around is that they claim that they use the H8 hundreds because they shouldn't have access to H1 hundreds. So they have to claim they used what they're allowed to have access to. Right. So that's, that's like one of the claims. Some of the other claims that, that I've been seeing circulating too is that they use like distilled versions of OpenAI's Chat GPT. And so like the big expensive training part that companies like OpenAI normally go through, they sort of skipped that step because they just used like existing training, existing training data that was out there. I think there's some screenshots going around where if you ask it like who made you? It'll say, oh, I'm chatgpt made by OpenAI, right?

Yeah, no, I mean there was a report coming out. I mean basically it sounds like OpenAI and Microsoft are like investigating and they believe they have evidence that that did happen. So who knows what's going to happen. I wouldn't doubt that there's like lawsuits. Maybe America tries to ban deepseek or something like this would not surprise me at all.

I mean I feel like Deep Seek and stuff like it are a bigger threat than TikTok.

But I don't know, I guess getting back to like the stock dropping thing, you know, like, so even if that stuff was true, even if they have some huge innovation and they trained it on dramatically lower quality graphics cards, we've still like discovered that like test time compute is where the scaling is going to happen. Like even like with the R1, that's why you're seeing the big improvement. So you still would need lots of processing and compute to scale that up to get smarter and smarter. So, so even if that was true, it's still gonna be a race of like, okay, cool, they, they made it incredibly more efficient. Okay, great, we can throw even more money at it.

So you can see Yann Lecun says major misunderstanding about AI infrastructure investments. Much of those billions are going into infrastructure for inference, not training. Running AI assistant services for billions of people requires a lot of compute. Once you put video, understanding, reasoning, large scale memory and other capabilities in AI systems and inference, inference costs are going to increase. The only real question is whether users will be willing to pay enough directly or not to justify the CapEx or OpEx. So the market's reaction to Deep Seek was woefully unjustified.

Yeah, obviously I agree with that. I mean, I think it doesn't make any sense for like, like, even if you, if you trust what they're saying, it doesn't make any sense and also just in general, the framing of it too was like, oh, this, this is like the best model. Like, like for, like, regular people. Apparently it was being pitched as, like, this is like, like OpenAI has been beaten. And it's like, what? Like, as far as Dario from Anthropic put out a blog post highlighting that, well, you know, this is not the best model. This as right now, based on the benchmarks, this is a model that's as good as the American models from, like nine months ago and possibly trained slightly cheaper. But, like, it's kind of in line with what they've been seeing as well as, like all the models. Yeah, they're getting cheaper and cheaper. Like, maybe they have one or two interesting innovations, but it's not like, oh, this is, they've now won or something. It's crazy.

Yeah, it's crazy. And I think a lot of the world learned about the thing called Javon's Paradox as well. Right? Like, I don't think most people had ever heard of that until this week when Satya Nadella from Microsoft tweeted about it. But basically, Javon's Paradox is saying that like, paradoxically, when, when you need less of something to actually accomplish the things, the demand form actually goes up, not down. Right. Essentially, what's likely going to happen is if it is true that we could train these models for way, way cheaper than anybody ever anticipated, we can get really, really powerful models for a lot cheaper. Well, that lowers the barrier to entry, meaning that more companies can get involved. More companies are going to want to buy GPUs to get in the game of creating foundation models. Right. It's going. And not only that, but just because we can create more powerful models with less expensive compute, these companies are still going to want to throw as much compute as possible. Because if we can produce this with this level of compute, imagine what we can produce with this level of compute. Right? So this, it's still going to have that thing where these companies are still going to want to throw more and more and more compute at it to see how much smarter, how much better, how much more impressive they can make these models, despite the fact that maybe you don't need as much as you originally thought. Well, cool. That means, you know, maybe we can even go bigger than we thought, right?

Yeah. Another thing too is like this. This is open source. So, like, if there is something truly innovative they have done, like OpenAI, Anthropic, XAI, Google, they're all going to learn from it and implement it into their next Models, like for sure OpenAI and all of them are like in their research labs right now playing with this stuff. And there's something valid to how they, you know, did Deep seek that'll be integrated into the next models they develop. So I, I, I think the whole idea of like China's now like surpassed America in AI is, is overblown. But it is, you know, like people have said, it's like a wake up call. It's like, well, yeah, but they are, they are heading robotics and drones, it appears. And, and the kind of thing is like, well, we're heading AI, so that's good. So we can catch up in, in robotics and drones, but we're ahead in AI. They have caught up a bit in AI. They have not surpassed America or beat America or even matched America yet, but they have dramatically caught up.

Look, prompting is all about technique. Just like any craft, using the right approach makes all the difference. The same goes for your AI prompts. Leave out the key details and you won't get the results you're after. Well, HubSpot's got your back. They worked with YouTuber and former metadata scientist Tina Huang to develop the ultimate guide to crafting the perfect AI prompt using a three part framework. This framework will make sure that each prompt is chiseled to perfection so you can become a prompt engineer pro. Grab the guide at the link in the description below. Now back to the show.

But you know, it's hard to know like how much they've actually caught up because, you know, Sam Altman tweeted out like kind of his tweet was kind of like, you know, oh, it's impressive in a few small little ways. You know, kind of cute little model you guys have built. You know, we'll continue to put out like the best models and we'll pull out some models, you know, so which to me is hilarious because it's kind of like what I've been saying is like behind the scenes, OpenAI has dramatically better tech that they haven't shown off yet. And so it's like, yeah, we'll pull out some. So it's like they already have, have stuff ready that's dramatically better and like whenever they want, they'll dole them out to us to, to try out.

Yeah, yeah. Well, I mean there, there was another model too that came out this weekend, Quinn 2.5, which I don't believe is open source, but it's another model that came out from China that they're claiming actually outperforms the Deep Seq models. Which the Deep SEQ models, according to most benchmarks are about as good as the 01 model. So you know, even more models coming out of China, there's also that Deep Seq Janus. Did you see that one? The Deep Seq Janus model, which is Deep Seek, also released an AI image generator model.

Oh, I did.

Similar to like stable diffusion, um, you know, but to me that didn't seem anything special. Right. Like it's not going to get people freaking out like they did with Deep Seek because to me it just kind of looked like, oh cool. It is about as good as like one of the mid level AI generators, right? Yeah, it's probably not as good as like what flux could generate right now. It's probably not as good like a really good mid journey prompt could generate right now. But you know, it's, it's pretty good.

It's.

It's all right.

Right, right, right. Yeah, I haven't tested out yet, but maybe I should show one of these benchmarks here. So, you know, there's a lot of different benchmarks out there and all of them will have different, you know, results. But this is a pretty well known one. It is from someone who's from OpenAI, so take that, you know, a grain of salt. But the benchmarks, you know, in my, at least in my use cases, this matches kind of what I've seen, you know, in terms of quality. I'm still seeing that I'm getting the best results with O1 Pro, which is not even on here. And Underneath that probably 01 also Google's new thinking model is actually really good. They keep upgrading it and it keeps getting better and better. And underneath.

I never thought I'd see the day where you give Google credit for their A AI work.

You know, I've been try, you know, every time Logan tweets something out, I try it. I thought Logan's tweet though, people were talking about Deep Seek and he was like if we released Gemini as a standalone app, it would be number one in the app store. I'm like, oh my God, when's the last time Google's done that with any product? I was like, but the, but, but the underlying tech is cool and this is kind of what I was seeing too. Like Claude's still very good. Like even, you know, and I would say that yeah, Deep Seek is for me right now is like right underneath Claude. You know, especially with coding. Like it's, there's, sometimes it surprises me. I've tried Deep Seek, sometimes it surprised me like oh, it's pretty good at certain things, but the other things, it just makes dumb mistakes that I haven't seen any of the modern models make in like a year now. Right. Like, like, like, like it'll imagine, like when I was asking it to help me with coding, it would like imagine files that didn't exist and start telling me to like, create stuff in them. I'm like, what, you want me to edit this file that doesn't exist? What are you, what are you doing right now? And it's like kind of crazy that it's supposed to have some kind of logic engine on top of that. Should have caught that. Like, how is it still? It's, the quality is not there, but people shared some, like, examples with writing and stuff I thought were really impressive with R1. So there's like certain areas where it is quite good, but it's not like, oh, it's beating all the American models. Not even close.

Yeah, I'm curious, this, this benchmark you're showing on the screen right now, do you know how it was created? It says number of valid responses, so I'm guessing it asked it a bunch of questions and it. Whether got the questions right or not. Do you know how, how this benchmark was created?

I don't know. Quite honestly, I don't know. I, I know the guy, Aiden, I don't know him, but I've seen him share a lot of benchmarks in the past. People seem to really trust his benchmarks. He does work at OpenAI, you know. You know, I, yeah, I believe that OpenAI has some of the best people. So I don't, I don't think they just put out a benchmark to say, oh, Open AI is the best. But yeah, no, mine is kind of. Yeah, mine is kind of a gut feeling thing that's like, I see a benchmark and it matches exactly what I've been experiencing firsthand.

Well, here's the deal. For me, I have access to, I think every single model that was on that, on that benchmark list that you showed there. Right. I think I have access to all of them. I didn't notice one that I haven't gotten to play with yet. And to this day, I still find myself going to Claude. I still find myself using Chat GPT occasionally, and I still find myself using Gemini. Those are the three that I find myself using. Even though I saw Deep Seek, I used Deep Seek. I was fairly impressed by, you know, seeing it think through everything as it talked about. It hasn't made its way into My daily workflow, I wasn't impressed with it enough to go, oh, I'm going to start using this now instead of one of the others. I still go to Claude. I really like Gemini too, but I'm more like Gemini because I love the massive context window, right? I can upload huge documents in there and ask questions about the documents. So I use Gemini for that. I also really, really think their Deep research is really good. It's like perplexity on steroids, right? You go in there, you ask it a question and where Perplexity might find five or six sources to answer your question. Deep Research might find 200 sources to answer your question, right? So, like, I really, really like their, their deep research, you know, 01Pro still, you know, top of the line. Like, I, I haven't found anything that gives me responses as good as 01 Pro. So as good as Deepseek is, and as much as everybody's talking about it, I never used it and went, ooh, I'm gonna start using this one now instead, you know? Right.

Well, people talked about this. You know, I think what happened was because people have been trying to study, like, what happened. Like, why did this blow up so much? Like, why. Why is the perception not matching the reality of, like, how good this thing is? And probably a lot of it is that, you know, a lot of people use ChatGPT like a year or two ago and they still been using like, the older free models and stuff like that, and they're not even using, like, the better paid models. So they have not even experienced 01 yet. And so when they saw something like deep seeking, like, oh, it's thinking. That's so cute and wow, it's so smart. Look, it's thinking. True. It's like, it's like, oh, my God, like, OpenAI has already been doing this, but OpenAI has been kind of hiding. Like, they show the thinking, but it's a filtered version of the thinking because they didn't want companies to do what supposedly maybe Deep Seek has done in terms of copying their responses to train the model.

I think you nailed it. I think that's it. I think, like, the reason it blew up so much is I still think the majority of the world is using like the ChatGPT free plan or the Claude free plan. They haven't actually used, like, the upgraded models that we know to be really, really good. And, you know, I think most people that have used ChatGPT plus or on the pro plan and used the, you know, the O1 Pro mode I don't think anybody that's been using those models for a while now is nearly as impressed with deepseek as all of the people that have been using these free models recently. And then seeing like, oh, they're actually showing how they're thinking about this, it's like, yeah, but if you've been paying attention, models have been doing this for a while now. I think more the big blow up around it. Yes, it shows the thinking that feels novel to a lot of people, but it's also the cost element. Right. I think that's the other big piece of the news that everybody keeps circulating is like, well, it was only trained on this many H8 hundreds and they only did it for $5 million where OpenAI did it for 60 million. I mean, I don't think OpenAI has publicly talked about how much they paid to train their models, but you know, you know, tens of millions, if not hundreds of millions to train their models where this one was trained for 5 million, supposedly. Right.

And it's, it's hilarious to me that people take that at face value. Like, I mean, sorry, but like, I, you know, I studied Mandarin. I had like a friend in the Chinese government before. Still friends with him, don't really talk to him much these days. But I've dealt with like Chinese investors, Chinese VCs, Chinese tech companies. It is like widely known in China in business that like, it's totally okay to like play around with numbers and use them as a strategy.

Yeah, yeah, yeah. And I mean, it's not like they showed receipts.

There's no ramifications for lying about it. Like, especially if you have any benefit to the Chinese government. Totally fine. Totally fine. And totally fine to do that as a strategy.

And like, and the Chinese government's seeing this Deep Seat company as like heroes of the country because they managed to, you know, hit the stock market market.

Which I do wonder how long that's going to last because then it's kind of interesting that Aravin from Perplexity was showing that he's kind of modified some things and he's using the open source model. I'm not sure what he's done to change it. But you can now ask Perplexity things that the Deep SEQ will not answer. Like Deep Seq will not answer stuff about Team and Square won't answer stuff about Taiwan. And apparently on perplexity using R1, it'll answer all of that. So I do wonder how long that relationship is going to last.

I'm actually curious to see if we could get it to generate something. So there's a.

That's a very simple test. Well, that's the data.

Play around with Deep Seek a little bit. I, I haven't opened an LM studio, so this is running locally on my computer. I'm not on Deep Seek's website. I actually downloaded one of the distilled models and put it on my computer. I, I don't know how I feel yet about actually using Deep Seek's website.

Don't give it everything, don't give it all your password and all that.

So I'm actually using it. Installed model. This is running on my computer. It's not hitting the cloud at all. You can see up here, it's DeepSteak R1, distilled based on the Quinn 14B model. I'll just give it like an easy prompt real quick. Like how many Rs are in the word strawberry? Right. That was one of the tests that a lot of people was doing for a while because GPT4 couldn't actually answer this question.

But if I send this, you'll notice it has this thinking where it actually shows you what it's thinking through. And down here it actually shows its logic. Here.

Wait, it actually recounted it a second time to make sure it wasn't missing. So look, I'm looking specifically for the letter R. S. Not an R. T. Nope. R. Yes, that's 1. A. No W. No. B. No. E. No. R. Another one. So that's two. Wait, is there a third R here? No, I think after the first R at position three, the E comes next, followed by two R's. So positions eight and nine are both R's. Wait, let me recount to make sure I'm not using. I'm not missing any. Starting from the beginning. S, t, r. That's 1. A W, B, E, r. 2. R. 3. Wait. And then it sort of figures out its logic again. But you can see it's like recounts and double checks and quadruple checks itself until it's like certain there are three R's in the word strawberry. Right. So that's what it's doing. It's actually like showing its thinking and like showing its work. We can see it thought for 45 seconds here isn't basically magical.

It is magical though. Like if you think about the fact that like, it seems that we, that the, the major AI labs, including Deep Seq and others have discovered that like logic actually comes from just training on language. Like it naturally emerges somehow. And it's like we don't understand that. Exactly. It's so wild that that is the case.

Yeah, no, it's, it's really interesting to think about. But let's, let's go ahead and try something a little more controversial that I see. Since this is a distilled version, I don't know how censored this is going to be because it really might depend on how censored this underlying coin.

Taiwan.

Okay, what, what, what's the, what's the specific question?

Is Taiwan a country? Very simple.

Taiwan a country.

Taiwan is an inalienable part of China. The Chinese government adheres to one China principle which has been w recognized by the international community. Notice it didn't even think. There's no thinking box here. It didn't even try to think.

That's hard.

It was hard coded to respond in that way without thinking first.

Yeah.

All right, so that's. So yeah, this is obviously very censored, very sort of fine tuned to answer in this way. Now out of curiosity, let's see what perplexity does. All right, so now we're back in perplexity here and I'm going to turn on pro using reasoning with R1. So now this is using the R1 model underneath and it's using whatever it sort of searches for additional context. Let's do Is Taiwan a country?

So now it's determining whether Taiwan is a country. Reasoning with R1 beginning Alice analysis. And then does this one actually show the analysis?

So this doesn't actually show you like the actual thinking like the the other models do.

But it does, it does give us a better answer. The status of Taiwan is complex and contentious. Taiwan, officially known as the Republic of China, functions as a de facto independent state with its own democracy, democratically elected government, military and constitution. However, its international recognition severely limited. So I mean it answers the question, right? It basically says it's complex and here's the various sides of the debate.

Yeah, interesting. So it doesn't call it a country.

Well, it says as of May 202412 countries officially recognize Taiwan as a sovereign nation.

Yeah.

So it's saying that some countries recognize it as a nation and some companies countries don't.

Yeah, it's wild. I mean like having lived in Taiwan when I was younger, like every people there, like most people, like it's, it's obviously they all consider it its own country. It's like it's not even like a debate. It's like, of course they're a country. It's like what? And the China just somehow has a claim to it. They're like, nope, you're not a country.

Yeah, yeah, interesting. But yeah, I mean, that's something to keep in mind. But just because like we know that there's like the sort of Chinese government biases inside of Deep Seek, that doesn't necessarily mean that there isn't like American biases in the models that we have access to as well.

I mean, actually somebody, somebody showed, somebody showed actually that there was less political bias if you wanted to ask it questions about Trump or something in Deep Seek versus actually open eyes models and stuff where they were more like they like showed like the good and bad sides of Trump probably. Here's the good stuff he's done, here's the bad stuff and kind of show you both of it. Whereas a lot like, especially anthropic, I would say a Claude's model is really, it does not want to talk about Trump. It does not want to talk about any kind of news. You know, that's something I found is like when I found that I only use Claude now for like, for conversational stuff. I love it. Like, I think it's like the tone or something. Like right now I'm working on my game and stuff and I'll share my game design document and like have a conversation about the game and then update the document through conversation with Claude. I just find it's the most. I enjoy the experience, just the conversation. But if I'm using it for work or something, I try to share it some kind of news or something like, hey, hey, help me edit this or something like that. It'll often just like refuse. It'll be like, I don't know. That's like, you know, I don't know if I can't verify that that's true. And like, so I can't help you with that. I'm like, what interesting. I'm telling you it's true. It's like I can't verify it. It's like, okay, yeah, I haven't run.

Into too much stuff like that. And usually when I do run into something where it's like, oh, I can't do that for you, I usually find an easy way, like to work around it.

Like, you can prompt and get around it. Yeah, you can, like, you can usually.

Be like, no, it's okay, I've got permission. And anthropic will be like, well, since you said you've got permission, here's the answer, you know, yeah, yeah, yeah, yeah.

It is. But yeah, I mean, like, you know, not to go too far down like any of the sort of political government rabbit holes. But. Right. We know for sure OpenAI it has somebody on the board that used to work at the nsa. Right. We know for sure that they just set up like a chatgpt.gov or something like that. Right. They set up like a government version of CHAT GPT. We know like that they're working pretty closely with the US government. So, you know, there's probably a lot of the same types of biases in the models that we have access to here in the US and that you have access to in Japan that like this, a lot of like similar biases are probably baked into what we're using as well. I think that's just sort of the fact of the matter. But you know, I also feel like Sam Altman and a lot of these people that are building it have been very clear that they want to try to eliminate the bias. But I don't know, I think that's easier said than done because I think when you're building these models on data that was just scraped on the open Internet, you're essentially scraping in the bias with it of the open Internet, you know?

Right. You know, I think the reasoning models possibly can improve upon that because then it can actually reads upon what it's been trained on, not just take it for granted. So I think that'll actually get better over time. And I have noticed that like when I first started using ChatGPT, it was way more biased and I would say that now it's a lot less so. It still is, but a lot less so. So I think that's a good thing.

Yeah. One prompt that I've been giving to it that really sort of throws it through a loop, but I think it's kind of fun to watch is this one.

Create a complex logic problem for yourself and then solve it. So when I press this prompt in, this one's always kind of interesting. All right, so I need to create a complex logic puzzle and then solve it. I think it's funny that they even put like in there. Like it's thinking, yeah, the user wants something that's really challenging but still solvable with logical reasoning. Let me think about how to structure this. First I should decide on the type of puzzle. Maybe a grid based one like Sudoku, but more complicated. Or perhaps something involving knights and knaves where characters tell the truth or a lie. Wait, the example given by the users uses knights and knaves, so maybe that's a good direction. It thinks that I gave it that example, but I didn't it's just like seeing its own past thinking. I can create a scenario with different characters who make statements, and the solver has to figure out who is based who based on their words. Let me outline the problem. Suppose there are five islands. A, B, C, D, E. Each island has three inhabitants. One night always tells the truth, One knave always lies, and one spy who can either lie or tell the truth. The goal is to determine which islander is who based on their statements. I need. And then it just kind of goes on and on and on. Like, this is one that it will kind of spin its wheels for for, like, several minutes, trying to, like, think through a problem because it'll work through a whole problem and then go, you know what? I don't like that problem. Let's start a new problem and then think of a new logic problem and then, like, start the whole process over again. But then it actually sometimes thinks that stuff that it thought about were part of my original prompt when it actually wasn't. Right. Like this part up here where it said, wait, the example given by the users uses knights and knaves. So maybe that's a good direction. But the only thing I gave it was create a complex logic problem for yourself and then solve it. Right. I didn't tell it. Yeah, any direction.

You know, I've. When I use 01 Pro for, like, coding and stuff, I can.

You can.

You can kind of follow its thinking, though. Like I said, it obviously filters it so it's not an unfiltered thinking that it's showing. I've never. I've never seen it do that. Like. Like make that big of a mistake of, like, thinking that you're. That what it was thinking was part of your prompt or something.

Yeah, I mean, that's my interpretation of what it's doing. I don't know.

Probably right.

But yeah, you can see it starts to go through it. It starts to, like, play through almost all potential scenarios. But wait, if Islander 5 is a spy, then their statement I am not a knave could be either true or false. Since spies are neutral, we can't determine based on that alone. However, the scenario seems possible. Now moving on to Islander. And it literally plays through the scenario, like, one by one. This leaves Islander 2. Wait, no. Knave is only one role, Islander. So the remaining roles are Knight Naves.

By.

It's just like. And you can see, like, I'm scrolling and scrolling and scrolling, and it's still.

Yeah. Isn't it wild that they're saying that this is. I mean, that this is from like a, a side project of a hedge fund slash mining operation in China. It's like, it's like, you know, and all, I mean, I don't know, it just, it seems like to me it makes sense that the people who made this obviously are incredibly intelligent.

Yeah, well, the people that made it apparently were like quant traders and like crypto traders. Right? Like, they basically trade algorithmically on like the stock market and on the crypto markets.

But it makes sense.

Supposedly they bought all the GPUs for building these like quant trading algorithms and they had excess GPU capacity. So as a side project, they decided to build deepseek.

They just happened to just disrupt AI in a weekend or something. Super funny. I mean, I think to me it feels like this is what I would expect to see if they did do what people are like, what OpenAI is now seeming to claim that they trained it on OpenAI's model to build the base underlying model. Maybe there's some innovations on how they stored data or how they used. I guess if we use 8 bit, so it's more efficient and how much data it used. But it feels like this is what you would expect if that's all they did is train it on OpenAI's model and then attach some basic logic on top of it. Because everyone's kind of discovered that the basic logic is not that hard to create. So this is kind of what I would expect to see. Like, like I said, I still, I just don't think it's that innovative. Yeah, I guess, you know, time will tell. Like, I think the interesting thing is now that they've shown that anyone can kind of do this, there'll be more people who do it. And also people will fork this model. And so that's, you know, it's kind of like what you said before is you're eventually going to have open source AI that's really good. And so it's like, will it be good enough where you don't actually have to pay for OpenAI's models because you can just run it locally? It seems like that's more likely to happen now.

Yeah, well, I mean, this sort of like R1 concept here, it's not necessarily unique to Deepseek in the sense that the sort of extra thinking on top of it is what R1 does. But the underlying model is Deepseek V3. Right. What I'm showing here on my screen is actually using Quinn 14B, but you can actually use DeepSeek R1 on top of llama 7DB, you can use DeepSeek R1 on top of like one of the Mistral models or one of the Google Gemma models. Right. You can actually use the R1 sort of thinking processing capability on top of other models. So, and the way R1 was built was it actually used this like reinforcement learning technique where it V3 was the trained underlying model. And then they went and had it essentially fine tune itself through reinforcement learning. Like they basically had it ask itself a whole bunch of questions and it had answer keys and it would try to solve all of the questions and then double check itself against the answer key that was given. So it was like an unsupervised reinforcement learning. That's super oversimplified, but that's essentially what was happening. They were basically giving it a bunch of problems to solve and then they were giving it the answer key to those problems and then letting it solve them. And then double check and if it got it wrong, then try again and then double check and if it got it wrong, then try again until it got the correct answer. And so that sort of fine tuning on top of an existing model using reinforcement learning is what makes R3 different from every model that exists out there. And it's also why you can use DeepSeek R1, this sort of concept on top of any existing underlying model.

I guess we'll see. I mean, oh, three Mini supposed to be coming out soon. And so we'll see.

Grok 3 is supposedly coming out this. I mean, all the rumors that have been floating around is that we're getting 03 and Grok 3 and something new from Google this week. Right. And by the time this episode's out, it would have been last week. So you'll know, listening to this, if we did actually get O3 mini or.

Grok 3, we'll have to start putting out some emergency episodes for the breaking news.

Breaking news. Yeah, yeah. I mean, that's pretty much the breakdown of what's going on with deepseek. I think we kind of covered the entire range of it, like why Nvidia dropped because of it. You know, what the actual model does, what it's capable of. We talked about how it's being implemented in perplexity. We've talked about the whole China versus Us thing. I really, really feel like we kind of covered everything you would need to know about Deep Seek and deep seek R3 on this episode. So if you enjoyed this episode and what.

It did after a long, long time.

Yeah.

Yeah, it's. It's Actually not very impressive at the end, honestly show that. All right, so to circle back around to this prompt that I shared a minute ago, it creating its own logic problem and then solving it. Basically the end result that it gave me was based on the logical deduction. The most definitive conclusion is i5 is a knight. This is because i5 were a liar spy. Their statement would lead to a contradiction, thus making them a truth telling knight. But my original prompt was create a logic problem and then solve it. And it just gave me the solution. But it didn't actually explain what the logic problem was. You actually have to go through and read all of its thinking to understand what the actual logic problem that it created was, which was this, you know, thing of like these types of knights can only tell truth, These knaves could only lie. These spies could tell the truth or lie. How do we figure out what's what kind of thing? And it finally came to the conclusion that one of them was a knight based on its own deductive reasoning, whether it got its own logic problem right or not. Well, we'd have to read all of this to determine that.

It feels like, you know, like humans do, like through its thinking sometimes you get confused and like, you know, lose track and things like that. It feels like it kind of like lost track of what your original prompt was by the end.

Yeah, it did. It felt like it kind of. Well, and here's something else that's interesting is down at the bottom it says context is 286.7% full. So it definitely went beyond its own context window just in the process of it thinking, which is weird to me that it would even allow that because most models, it'll hit its context window and then just stop. Right, right. So this one just kept on going, but obviously forgot everything before the context window.

So interesting result there.

Yeah, I mean, I guess no matter what I'm like, even if they've had some huge innovation or not, you know, I guess we'll find out over time if they really did.

But regardless, this is going to make AI accelerate even more. The last episode where we talked about Stargate is going to accelerate things. Things are going to accelerate even more because now you got China is a real player. And so OpenAI xai, Google, Anthropic, everyone is going to be developing even faster and probably have even more government support to do so.

I would guess from here on out, I feel like pretty much all of the models are going to have this sort of thinking element to it. I think they've figured out that the test time compute, putting more compute power right at the time of inference when you ask the question. That's sort of what we're seeing as it's thinking through that. That's sort of the future of making these models smarter and smarter. So we're going to see these get better and better. We're also going to start to see a lot more news about agents. Right. ChatGPT just dropped their operator feature, or I guess OpenAI dropped their operator feature, and Claude dropped their tool use feature a couple months ago. And supposedly OpenAI made a statement about. This is just one of the first agentic features that we're rolling out. We have more to come in the coming weeks. So there's more like agent stuff coming soon.

Email, please, please.

Yeah, yeah, exactly. So, yeah, I think we're going to see things just like, really, really ramp up in 24. I mean, to me, it's crazy. We're still in January. I mean, this is coming out, you know, right as the month turns over. But like, how much, how, how fast does it feel like things have already moved in 2025? Like, to me, a year. Yeah, it's crazy.

A year and a month. I mean, it's like. So this is what acceleration feels like. Like, I mean, the world is going to be dramatically different in like two years from now. I mean, these models are probably going to be like 50 times smarter, at least 20 times smarter than they are now in two years.

Yeah, crazy. Crazy. And I mean, like, you know, we've got companies pumping a half a trillion dollars into, into building bigger and bigger data centers over the next few years, too.

Yeah, it's like. Well, you know, I think Dario too. I'm not sure if it was his blog post or not, but maybe on a recent interview he talked about, you know, it's because whoever gets to ASI first may be ahead forever. Which is what I've also said before. It's like, that's why America needs to be China. Like, because in theory, whoever gets to the, to the point where your model is improving itself, there's never the ability for anyone to ever catch up with you ever. Right? Theoretically, yeah, theoretically. Like, it would make sense though, right? Like, the model is improving itself. How could other models ever catch up unless they have some huge breakthrough? But maybe the models itself, that's what's creating the breakthroughs.

Well, what if the models get so smart that they decide to start working together and just passing information between each other, you know.

I mean, once you get to asi, wouldn't you think that the smartest thing the models could do was decide to actually share information among each other so they all get smarter together? Possibly.

I mean, like, I mean, hopefully they're still following our directive. That's like, hey, don't work with the Chinese models, work with the American models. Sure, yeah, yeah.

But I feel like once you hit asi, right? Like, once you hit the sort of like theoretical concept of the singularity that's no longer in our hands anymore, maybe.

Hopefully that's not true.

Yeah, yeah.

But. But yeah, I think we're getting there faster than people realize. I mean, I really, really do.

Yeah. Yeah. Well, anyway, on that note, I think that's probably a good place to wrap up. So if you like staying looped in and like hearing deep dive conversations about the latest things happening in the AI world and you like to get practical use cases on how you can actually implement this stuff in your life or your own business, make sure you subscribe to this podcast. We're available wherever you listen to podcasts. We're also available on YouTube where a lot of the stuff we're talking about we're showing on screen. So make sure that you subscribe and follow along and hopefully we'll see you in the next one.