Pod Insights The Real Estate Espresso Podcast Who Burst The Artificial Intelligence Bubble? About this episode When a bubble gets burst, the pin usually gets the blame. But it's not the pin's fault. How do bubbles form? What is the definition of a bubble? Bubbles form when three things come together: A massive amount of capital injected into a segment A story which supports the unlimited upside potential which is fuelling the frenzy The income from the market is not enough to support the investment. On today's show we dissect the optimizations that Deepseek made to undercut the entire AI industry. These shortcuts are obvious in hindsight. ------------- **Real Estate Espresso Podcast:** Spotify: [The Real Estate Espresso Podcast](https://open.spotify.com/show/3GvtwRmTq4r3es8cbw8jW0?si=c75ea506a6694ef1) iTunes: [The Real Estate Espresso Podcast](https://podcasts.apple.com/ca/podcast/the-real-estate-espresso-podcast/id1340482613) Website: [www.victorjm.com](http://www.victorjm.com) LinkedIn: [Victor Menasce](http://www.linkedin.com/in/vmenasce) YouTube: [The Real Estate Espresso Podcast](http://www.youtube.com/@victorjmenasce6734) Facebook: [www.facebook.com/realestateespresso](http://www.facebook.com/realestateespresso) Email: [[email protected] ](mailto:[email protected] ) **Y Street Capital:** Website: [www.ystreetcapital.com](http://www.ystreetcapital.com) Facebook: [www.facebook.com/YStreetCapital](https://www.facebook.com/YStreetCapital) Instagram: [@ystreetcapital](http://www.instagram.com/ystreetcapital)
Transcript 0:01 Speaker 00 Welcome to the Real Estate Esprit Podcast. Good morning. Shout out what's new in the world of real estate investing. I'm your host Victor Minas. On today's show, we're talking about the seeming revolution that is taking place in the world of artificial intelligence. Now, this is a real estate podcast, and it might seem out of my lane to be talking about it. But since AI represents such a tectonic shift in the world, it's going to affect every single industry. I think it's important to talk about it. Moreover, since I come from a technology background, having previously designed microprocessors, I have a pretty decent understanding of how chips are designed and manufactured. I'm probably fairly qualified to talk about it even on a real estate podcast. This major lowering of cost for AI is something
0:43 Speaker 00 I've personally been predicting for some time. We're talking, of course, about Deep Seek having cut out the entire AI industry at the knees. Although I haven't been saying it publicly, let me draw a simple analogy. Imagine for a moment that you could walk up to the Lawrence Livermore Lab's supercomputer and ask it to perform some simple arithmetic. The supercomputer would do the job, and it would do it well. But then you could also get a desktop calculator from the dollar store, and it would do just as good a job on a limited set of arithmetic calculations. The supercomputer would be massive overkill for a simple calculation. We can clearly see that now, but there was a time back in the 1960s when you would seek out the world's most powerful computers to perform what today
1:25 Speaker 00 would be considered a routine calculation. Most of the time, you're not going to need to input the number pi to 100 decimal places. Just most calculations don't need that. Maybe my calculator analogy is a little too simplistic. Imagine for a moment you had a panel of the world's master scientists all working together on a problem. You would be harnessing the collective wisdom of that group. But you don't always need a panel of scientists to perform the most basic functions. They might have an apprentice that could be extremely productive and still do amazing work. And the reason the apprentice can be productive is because of the training they've received. It's narrower, it's more targeted, and they don't know everything that the panel of scientists does, but they know enough
2:08 Speaker 00 in some narrower restricted domains. The apprentice has distilled the accumulated knowledge of the masters. And this is precisely what Deep Seek has done. The language model created by Deep Seek is the apprentice. It is the calculator to the supercomputer. It cannot do everything, but it can do a lot of the fraction of the cost. A lot of attention seems to be focused on the fact that DeepSeek is based in China. There are assertions that the language model might have exploited the work of OpenAI. All of DeepSeek's work has been done using an open source model. Experts from all over the world have been dissecting the work from DeepSeek and validating their seemingly impressive results. All of the accusations are missing the point. DeepSeek has proven that something can be done, it actually doesn't matter who did it.
2:53 Speaker 00 See, the AI industry is in a bubble. And bubbles form when three things come together. Number one, you've got a massive amount of capital injected into a segment. Number two, a story which supports the unlimited upside potential which is fueling the frenzy. And then number three, the income of the market is not enough to support that balloon of investment. We've seen these before throughout history. There were tulips, and then the steam railway, and radio, then television, then the internet, and then the US housing market. These have occurred with alarming regularity throughout history. But when you're in a bubble, it's hard to see it. It feels really good to be swimming in the punchbowl. No, you're not drunk. The punchbowl is so warm and it tastes so good. How could anything like this could be so dangerous?
3:37 Speaker 00 A simple case in point is the market capitalization of Nvidia, which peaked at three and a half trillion dollars. How can one company with great market position be more valuable than the entire stock market for France and Germany combined? I mean, think about it. Nvidia would need to generate 140 billion in profit, not revenue, but profit, per year for the next 25 years in order for their stock to be worth anywhere near that amount. DeepSeek used a number of rational techniques to reduce the computational task. First, instead of using the large language model to parse every individual word, they use a chain of reasoning which focuses more on phrases. For example, if you take the word within, you will often find the word within followed by the word the or the word a.
4:22 Speaker 00 There's very few other choices in the English language that make any sense in any sentence structure. So there are many shortcuts like this in the English language that makes sense for you to have a much more efficient way of parsing the use of language. There are several optimizations that DeepSeek has made that Frankly, make perfect sense. These distillations of the general knowledge are confined to more specific domains. If you have a question about nuclear science, you would go to a nuclear scientist. You would not assemble a panel of doctors and dentists and hairstylists and nuclear scientists to answer a question of nuclear science. You'd go direct to the expert. And the folks at DeepSeek have taken this approach and created a much more efficient mechanism by segmenting the language models into more specialized domains.
5:03 Speaker 00 And in doing so, they reduce the computational requirements, and they can achieve results that are nearly as good as the generalized model for a fraction of the cost. And in this case, we're measuring the cost in terms of computational resources, memory, data centers, and electricity. So the savings are massive. It also means you can get this done for much less investment. Whenever a balloon is burst, people tend to focus their energy on the pin. They're looking to lay the blame on the pin, but you know what? It's not the pin's fault. The pin was simply the catalyst that caused the bubble to burst. What we need to be looking at is upstream, and why did the bubble form to begin with? As you think about that, have an awesome rest of your day. Go make some great things happen. We'll talk to you again tomorrow.
Previous Celebrating A Major Success Next Consequence of the Fed Divergence Page updated November 15, 2026 · transcript and summary are machine-generated