Limitless Podcast - DeepSeek Traded Its Way To A 2x. Can You Do It Yourself?

Episode Date: October 28, 2025

In this episode of Limitless, we discuss algorithmic trading, showcasing AI models that doubled $10,000 investments in just two weeks. Highlighting six models with varying performances, DeepSeek and Quen achieved over 100% returns, while others like ChatGPT struggled. We discuss AI trading behaviors, ethical considerations, and the impact of blockchain transparency through platforms like Hyperliquid. As the experiment wraps up, listeners can look forward to more insights on the evolving integration of AI in finance.------🌌 LIMITLESS HQ: LISTEN & FOLLOW HERE ⬇️https://limitless.bankless.com/https://x.com/LimitlessFT------TIMESTAMPS0:00 AI Models Can Make You Rich1:01 Winners and Losers in Trading1:57 Analyzing DeepSeek's Success4:33 Insights from AI Trading Logs8:49 Trading Models: Skill or Luck?10:54 The Value of Public Data17:20 How to Trade Like an AI21:53 The Risks of Algo Trading22:29 The Future of AI in Trading------RESOURCESJosh: https://x.com/Josh_KaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

It's official. Air models can make you rich. Over the weekend, two AI models doubled their money going from $10,000 to $20,000. But the best part about this is that all their trades were public and available for you to review, analyze, and maybe even trade yourself. In this episode, we're going to unpack which model makes you the most money, how an AI can make you money, is it just luck or is it skill? And most importantly, how can you do this yourself? So we have six models, $60,000, and in the last two weeks, two of these models have two X'd their returns. It has been an unbelievable amount of success from this experiment. Some have not done so well, but the ones that did are exceptionally interesting because we can actually emulate the trades. All of the trades are public, the thought process are public. You can look at the wallets, analyze the trades, and actually recreate this for yourself, not only by copy trading, but also trying to create your own replica model to try to emulate this returns. Now, there are risks. There are two big winners, but there are also two big losers being

Gemini and ChachyPT. So this is really interesting dichotomy split between how agents approach trades and the success that they actually see from these trades, which we're going to get into in this episode. But I want to talk about the top chart, the deep seek chart, who is at, what is that number, $22,000? Oh, yeah. That's a lot of money. So walk me through exactly how they made it to this point, please and how I can make 100% returns on my investment. So the model you just pointed out, Deep Seek, is currently sitting on $22,300, which represents more than a hundred percent return on the initial 10-K that it was trading. You don't know the craziest part about this, Josh? When I woke up this morning, or rather when I went to bed last night, it was number two.

And Quinn was the winner. So it just goes to show how quickly these things move and how quickly these models perform. If we look at the overall standings before we dig into the winners and the losers, I just want to give like a review as to like how these models are performing in general. DeepSeek is right at the top with a hundred and twenty two percent return. That is in just over a week, which is just kind of insane for any kind of hedge fund that is out there to look at and see perform. And you've got a range of different models that are also performing pretty high up there. Quinn is at 90 percent. And then right at the bottom, as you mentioned, you've got got Gemini and GBT, which are down 60%, which is like a horrendous return. But bringing it back

to DeepSeek in particular, I found it really interesting, Josh, to kind of unpack how this model trades and why it's been so successful. And to start off, I want to show you something called the model chat, which basically is like this model having a chat GPT conversation with itself. In this conversation, you'll see on the chat log, it's a very much. evaluating its trades. It's reviewing its current profit and loss. It's checking the market data that it gets fed, like, you know, Bitcoin is at this price, this asset is that price. Trump made an announcement on so-and-so and evaluating whether it should affect the positions and trades that it holds right now. I think this is like really important to kind of like

walk through a few of these examples here. So one which it posted just today is, despite all my positions currently being in on the red, technical indicators like RSI, which is like a trading indicator, shows me that my existing trades aren't invalidated just yet. So I'm still holding out for my initial profit targets. So it's a really strategic sense of like thinking, hmm, should I hold my positions for long? Does it make sense to cut at this point?

Just a really fascinating insight. Josh, do you have any takes on this? The chain of thought thing is fascinating to me because it's a peek inside the brain. It's a way to evaluate how these models think. It's a way to allocate EQ points to each type of model because they all think about these things very differently. One of the things that I'm actually not sure is true

is that I don't think these models are given access to news feeds and public sentiment. I think this is mostly just fed price and market data. Learning that, it creates much more of a simple problem in terms of the data ingestion that needs to happen in order for them to make decisions, and it allows it to be a little more precise about how we evaluate these, which is a good thing. One of the things that I really loved, particularly on the other side, which we'll get into, is how they self-reflect on the decisions that they make. Because one of the things, it's not just this pragmatic decision-making tree.

There is reflection involved. And I remember, you showed me a funny one about chat GPT and how it's like, all of my positions are down now. I'm doing bad. I should probably try to figure out how to do better. And it's fascinating to see into the brain, the chain of thought of how these things work and see the differences. So I haven't had a chance to look through a lot of these logs, but I know you have, is there any specific differences that you notice between the top and the bottom specifically? Because in the first episode, and people who haven't watched it last week, our biggest episode ever. Thank you for the support.

Thank you for watching. Go check it out if you haven't. But in that episode, we mentioned the fact that ChachyPT was the early loser and we kind of projected it to continue to be the biggest loser because ChachyPT is this very thoughtful, very sycophantic, very wanting to please. And the reality is that markets are a lot more hardcore than that. So I think we were probably right in our guess about this, but I love that we have the concrete evidence now. So have you noticed any differences in how they handle each other differently? I have. So deep seek probably unsurprisingly, as it was created by this model was created by a hedge fund, trades like a hedge fund trader or an analyst. So let's look at a few different things to kind of prove that. Looking at the chat log that it's having with itself, one thing that

strikingly obvious in this entire discussion with itself, is that it's constantly evaluating its stop loss, which is like when its trade thesis gets invalidated and when it shut off the trade with the current price that that asset is at. If you compare it to the bottom model, which I'm going to show you in a second, which is chat GPT, GPT5, it almost never does that. It just reflects on the current PNL that its trade has versus like looking at it more analytically. The second component for the top model, which is DeepSeek, which has made the most money, is if you look at its completed trades, Josh, you'll notice one thing in common, which is DeepSeek is constantly making trades. It's actually the model that has made its second highest number of trades in this

entire experiment so far. It's constantly opening positions. It's constantly closing positions. It's constantly reevaluating where it is in the market and what it needs to do. And you'll notice right at the top here in the most recent trade that it's closed, it booked just over $7,000, which has put it up in its first place. So again, it's trading more like a quantitative analyst, which is taking wins when it can and taking losses that are incredibly small. Like notice this, right? Like normally we don't highlight the losses of a model, but if you notice, all its red numbers are tiny compared to the profit numbers that it makes when it is right. So really, really strategic in its positioning. Now, if you compare that to the worst model, which is GPT5, you'll notice a few things.

Mainly there's a bunch of green and red that you can see, mainly red. In its green positions where it's completed a trade, Josh, you'll notice something pretty different, which is the numbers are pretty small. Look at this. It's only booking tiny profits with each of its different trades, which tells me that it's not taking enough risk, and it's not, it's closing the trades way too early for its thesis. So it's trading more like a cautious trader, like a lot of people that I know actually. And then if you look at the model chat where it's talking to itself, you mentioned earlier, here's an example. It goes,

I'm still in the red with a minus 61% total return, but my eth and XRP positions are showing gains, suggesting a slight upward momentum in those alt coins, despite the overall. market downturn. So I'm holding strong and waiting for those profit targets to hit. And so you might think, huh, that's not too crazy. That sounds like a sensible strategy. If you look at its profit targets, Josh, it's like super small from where the price currently is, which means that even if it does hit those profit targets, it only ends up booking like 50 bucks. So overall, the reason why this model has underperformed is it hasn't taken enough risk, whilst the winning models have taken either too much risk or just enough risk to put them ahead of the game.

There's a lot of noes in there that I think humans can take on just the stay of psychology around trading markets. And I'm sure if you kind of follow these models long enough, you'll start to understand the patterns that perhaps you as a human should follow and learn something from Deepseek versus Open AI being very conservative. But now that we've kind of laid out the foundation, the framework of how this works, there are two big questions that I'm really interested in answering. One of these is, should I use this model to trade for me?

The other one is, how can I use this model to trade for me? because, listen, I like a little bit of risk. I can deal with the downside in exchange for like a nice upside. And it looks like the odds are about split between all of these. So the first question I think I want to ask, EJ's maybe I'll get your take first is like, is this a benchmark? Is this real signal?

Or is this kind of just a reality TV show? Is this e-sports for AI models? Is this just a fun way to kind of throw our intelligence at this lottery machine that everyone loves to watch and see if it could beat us in the hope that one day, an AI will beat the system enough to give us an edge and actually make us money personally as portfolio owners. So what do you think about that? Okay. I'm going to give you the same response, Josh, and then I'm going to give you the optimist's approach. Oh yeah. Bring it on. Let's hear it. The same response to this is this experiment is way too tiny to make any kind of major financial

decision on and you would be stupid to risk putting your money with an AI model to trade for you. Incredibly stupid. Why? Well, this is one experiment. It's six models. Have you replicated those models? Like, what if you had 10 of the same models trading the same amount of money? Would they make the same trades? Probably not. And actually, the founder of this experiment highlights this problem that you speak about, which is, is this just skill versus noise? And the point he makes in this tweet is, like, of course it is, right? Because this is such a limited data set. And he goes on to to explain that they're going to be doing experiments which involve like more of these models doing the same kind of thing. So you can get statistical significant. So the logic answer is,

yes, it's insane. But the optimist take, Josh, and I have to give the optimist take, is this is giving us, or rather giving the public unparalleled access to data to which they never would have gotten access to in the first place, which is, they can take this training data and not take it too seriously, but use it to teach themselves what maybe not to do or what maybe not to trade with. How about you? Do you have a different take? There's a couple different perspectives I have on this because there's the fun speculative side of things, the gambling, the invest, or investing, whatever you want to call it. And then there's the actual technical benchmarking part of this that we spoke about briefly in the last episode,

which one of the things I was really excited about when this came out was the idea of having a real world benchmark that operated in dynamic conditions that cannot be gamified. So a lot of these benchmarks, this is the way you evaluate AI models. They are done based on a fixed problem set. And a lot of times when you're training an AI model, these big labs can do tricks to gamify these benchmarks. With this case and using real world data and real world markets, you can actually put them into the real world.

And there's no way to gamify these benchmarks because if there was, everyone would be rich and you'd be able to predict markets. To that point, though, there is a lot of problems with using this. as a benchmark because, I mean, one is the fixed data set, like you mentioned, is that this has only been around for one to two weeks. We need a lot more data to confirm this. The second is that this isn't really a very holistic approach to investing and to gambling because it really doesn't have all of the data required to make good decisions. It's only analyzing the price action and the volumes and whatever technical specs you can see on a single page without understanding

the context around the moves. So let's say that Bitcoin's encryption got hacked. It would have and Bitcoin falls 50%. It is no idea why Bitcoin is going down 50%. And because of that, it's a huge disadvantage that it doesn't know how to trade. Now, granted, these are unlocks. These are things that will change. And I assume the natural progression of this

will lead towards more of a steady state benchmark. But it is a very tricky thing because markets are so unpredictable. So is this a viable benchmark? I don't know. Probably I'm leaning towards no because market conditions change a lot. It's not quite there with the capabilities.

The other part of me is so stoked about this because the same way we love watching e-sports or we love watching a big thing on Twitch right now is gamblers. I don't know if you've seen these in real life. People will sit there and play, like they'll gamble blackjack on a live stream and people will just watch them play virtual. Yeah, those types of services.

This and very much feels like an early prototype for a new type of fun form of entertainment, which could be something where it's just, it's high-stakes trading. Imagine if this was done with $10 million per wallet. And you got to watch these AI's trade. and there was real money on the line. This feels sort of like a form of almost e-sport entertainment where I could see competing labs builds competing AI models to trade markets and winners are given access to certain prizes. In terms of trading for myself, which is the last point I'm

going to make on this. I am not very excited to take on these risks. For the same reason, I'm not really excited to bet on sports. And I imagine my opinions vary a lot from others, but this is very much a gamble. There's no way you can skew this in which it is not a gamble. the interesting part is there's a near perfect data split between them. There's two big winners, two big losers, the rest are kind of sitting around the median. Okay, but I'm going to push back on you a bit here, Josh. The earlier point you made was it doesn't have access to all the necessary data that it might need to make more informed trades. And I would argue, well, isn't the whole point of the benchmark, can you make money? And the fact that two of these models have made over 100%

returns in less than a week or just over a week is proof that it can make money to some extent. The second point I'll make is we throw around the term like gambling, which is actually what I would say the majority of these models in this experiment are doing. But they are one or two models that are actually way more strategic and trade much, much better than the average trader that you trade against, if that makes sense. So if we take Deepseek, which is the number one model, If you look at its trades, at an initial glance, you might see that it's using 25x leverage and be like, that is so ridiculous. I'm not even going to pay attention to this, right? But if you dig into the position that it holds under 25x leverage, you'll notice that it's actually not at 25x. It's using only a small

amount of its capital to do a very specific trade over like a five to 10 minute period, which automatically makes it a much more strategic technical trader than the average trader that is just gambling their money away. But the point you made around it being fair distribution, and this is my last counterpoint to you, Josh, you pointed out that it seems to be very even distribution, right? You've got two at the top, two at the bottom, and two right bang in the middle, right? I wonder whether actually GPT and Gemini are actually the best traders, even though they're at the bottom if you just inversely traded them, right? It's zero sum and it's the point that the founder of the experiment makes right here where he goes markets are zero sum. If you find a strategy that

consistently loses money, it's just as good as finding one that makes money. Just do the opposite. Yeah, absolutely. And it'll take time for these to play out because I imagine there is, they are kind of tuned for a specific type of trading. So in the case, a few weeks ago, there was a huge liquidation event in crypto. Things go down. Well, in a down market, some my trade way better than others. And the point you made about leverage, it got me thinking it was really interesting. Like, because I don't use 20x leverage and I imagine most people don't. But with AIs, they're able to hold a lot more in their memory. And it reminded me of the AlphaGo case, Google, where an AI model played a professional at AlphaGo. And there was one move that was way outside of

the expected data set, move 37, which was the famous move. And it turned out that that was a move that no human could have ever seen, but it resulted in the AI winning the game. And it kind of broke open the rule set and expectations around the game of AlphaGo. And I wonder if we'll get some sort of breakthrough with that around AI trading, where we have this very fixed set of outcomes that we do and strategies that we do. But AIs might actually just destroy a lot of these barriers that we, or perceived barriers that we have in exchange for these really weird strategies, like 20x longing everything. So I don't know. There's a lot to talk about when it comes to this. But Another of the big questions that I want to answer because this was something I was interested in is how can I use these for myself? Let's say I am degenerate gambler. I want to make 2x in a week or at least give myself a chance to do it. I want to know how can I use these models to trade for myself? What do I need to do to do to get involved in this? Yeah. It has been the number one question and feedback that we got on our previous episode from our listeners is I've got it up on a tweet here. How do I profit from this trading? How do I do this for myself? I have one simple answer for you.

which is the platform that these AI models are trading their tens of thousands of dollars on, Josh, is public. It's open. It's available for anyone to log on to right now and see what trades each of these models open up when they close it and what their inevitable strategy is. I'm going to give you an example here with the number one model, DeepSeek, which has doubled its money in just over a week. The platform that these models are trading on is called hyperliquid. It's a blockchain. Blockchains are known for being transparent and open, the fact that you can kind of see all the things that these models are doing.

And if I just scroll down over here, you'll notice a few things. Number one, these are all the positions that this model currently has opened. This isn't made up. This isn't on someone's word and you have to trust them. This is all verifiable using a blockchain. So the whole point of a blockchain is that you are able to verify what is real. and what is not real without having to trust someone on this. You can look into its holdings and you can see how much that it currently holds,

like in terms of like money or in terms of like dollars. You can also look at the trades that it's completed as well. So the point I'm making is you can't currently go onto Deepseek and say, hey, can I give you $10,000 and you go make me money like I've just heard about on this video? It won't be able to work. But what you can do is you can go onto a site like this. and look at the trades that they're making yourself, and again, this is not financial advice,

potentially copy those trades or make those trades yourself in order to trade like how these models are. Now, the last point I'll make is the founder of this experiment has all the intention to allow you and me to trade with these models directly. That is, you can speak to the model, give it your money, and it can do that. And to your point, Josh, it's up to you

whether you want to do it from an entertainment basis where it's just all gambling or whether you actually want to invest serious money into this. That will come in later iterations, probably around a couple of months from now. So there's kind of two ways to copy trade. There's one you could actually copy trade. Or another way to get into it is if you're feeling a little more ambitious,

you can actually generate one of these yourself. You can create like a mini alpha arena bot. The way to do that is pretty simple. I was kind of curious. I was like, what does it take to actually build one of these things? You choose your fighter. So you pick a model that you want. And then you kind of pipe market data in from HyperLiquid that you showed.

So hyperliquid has this endpoint, and not to get too technical, but you can kind of feed the model this data. And then the difficult part, the tricky part, and the thing that we haven't been able to talk about because we don't actually know is the system prompts behind the recursive loop that happens as these models receive this data. So the way it works is you choose a model, you give it feedback or you give it data, and then you write a prompt for the model to run in between each iteration of receiving new data. What that prompt says is how it makes a decision. The problem is that is all of the value. All of the value sits within that prompt. And the prompt is just written in plain English. Like we always say, the hottest language in the world is English.

So there is some string of words that you as a developer or just a novice can write into this to generate you more money than other people. So I encourage people who are feeling a little ambitious to actually try this out, to write a prompt yourself and see if you can get a bot to try and kind of trade like this. And if we ever do get the system prompts from this, we will certainly share because it'll be fascinating to see the behind the scenes and what happens to produce those outputs. that we were reading a little bit early in the show. So that's kind of how you can get involved. If you're interested, copy trade, maybe inverse copy trade. I think if I were to do this, I'd probably go to Chat Chpetech trading history, sit there refreshing and then just hit the opposite of whatever they decide to do.

That seems pretty consistent. But yeah, that's how this whole thing works. It's pretty fascinating. It's been amazing how the internet has kind of gotten behind this, and it has spread like wildfire. The thing is, I don't think they'll ever make the system prompt for this or any other successful trading AI. publicly available. The reason is that's the secret source and why would you let everyone have

access to it when you can use it yourself and make a ton of money? And that's what D.D. Das demonstrates in this tweet. He says, I've heard six people tell me they're doing this using vibe coding apps to algorithmically trade on the stock or crypto market. But the thing to remember is this is a dangerous game to play. Algo trading is the last thing I expect AI to democratize. the point being, if you have a successful algo, you're probably not going to democratize access to it full stop. That being said, I do think you can't stop AI entering the investment and financial scene. I think it's going to make people way more financially literate than they already are. Look how chat GPT has made so many people proficient in other things that they had

previously no idea about. So I think AI is inevitably going to be integrated. It's going to make markets way more efficient. It's going to give you access to knowledge that can make you you do trades that you otherwise wouldn't have known of five minutes prior to that, but will it make you a super trading guard? No, I think that it'll evolve the trading scene, though. I think the hedge funds that are successful today will look very different to the hedge funds that are successful in a AGI or AI world where AI is available pretty much everywhere. Yeah, AI needs to be integrated into all of these trading strategies. So to me, it's no-brainer that it will be. The extent of that integration is kind of what is up for debate and what we'll see in this answering the big question.

Is this a benchmark or is this just a reality show? And is this just a toy or is this this real technology baked into this? It seems as if AI will slowly creep its way in. I'm looking forward to tracking this. It ends next week. So we'll probably add some follow-ups on this first trading competition. The result is how it turns out. But that is a part two in our little saga of this crazy weird thing that's happening in AI crypto trading world. I hope you enjoyed this episode. You enjoyed the last one a lot. It was amazing. So thank you. for watching, sharing with your friends, liking, and commenting. It really goes a long way. It's been amazing to see the growth and support from everybody watching. So thank you for that. More of this

to come. We have a couple more episodes Slater for this week that are pretty exciting about autonomy and robotics and just a whole bunch of interesting things. So stick around for that. We'll be back in the next one. And I think that's it. Any final parting words? That's it. Let us know what you want to hear more of as well. If you're loving this trading stuff and you have some other ideas, let us know in the comments. Absolutely. All right. Well, that's been another episode of Lemmails. Thank you for tuning in. And we will see you guys. in the next one.