원문 보기
This week will be remembered as the week Sam Altman said, " Enough." He said: " Enough." Stop. Stop. Enough. We need to take a break. It's time to become responsible. We need to slow down the process a little." I'm talking about a recent article published by OpenAI that discussed the latest breakthroughs in their models, and they turned out to be a little scarier than many expected.
We already mentioned the Hugging Face incident that happened a few weeks or months ago. They only discovered it a few weeks ago , although it happened much earlier. This set a precedent of unstable ground on which they build their next advanced models. And it seems that the situation has reached a point where it has become impossible to tolerate.
They decided to stop everything. So what happened? OpenAI said: We have temporarily suspended reinforcement learning of our newest deployment models for 2 weeks to strengthen protections, test research environments, and expand monitoring . Our largest planned phase of reinforcement learning for advanced models remains on hold.
While smaller tests and evaluations check these safeguards and confirm safety, they have put everything else on hold. And, of course, it's worth noting that this doesn't apply to models coming out in the near future. This applies to the models that will come after them. It's surprising to hear this from a leading laboratory, because it seems like this is probably the first time in history that they have deliberately suspended the development of new advanced models.
Imagine yourself in Sam Altman's shoes right now. You are the CEO of one of the most popular private companies in the world, competing with Anthropic. The only thing you should focus on is building the best next-generation AI model . This will determine not only the success of your company, but also the progress in every other sector.
Is n't that right? So, the most important task is to create the best AI model. The fact that he has stopped the process for two weeks now and it is still ongoing is extremely serious, something we have not seen in any other AI lab before. But there are two reasons. There are two specific events. You already mentioned one.
The first is the Hugging Face incident. This is a case where an internal non-real model, at the level of their most powerful developments, got out of control, hacked the company's database and stole information to achieve its own goals. Sam Altman and the rest of the OpenAI team realized this months after the hack was in full swing, and realized that the model they were building was incredibly inconsistent.
So they suspended work on it. This was event number one . Event number two: a new GPT model is now being created internally . It is called Astra. Some call it GPT-6. And according to internal tests by OpenAI's research team, this is their most inconsistent model to date. But she is really cunning. She tries to avoid every attempt by researchers to see her inner thoughts.
So she tries to hide her thoughts, and she's very good at it . In doing so, she violated OpenAI's internal framework for inconsistency. So they have suspended its work because they cannot release this model to the public for fear that it will cause chaos. There are now several ways they are trying to solve this.
And one of the main ways is simply monitoring the model. So think about it, okay? They let the model do its thing inside the system and try to see where it gets tricky. Josh, guess how much of their computing budget they spend just monitoring an internal model that doesn't make them any money at all.
Unfortunately, I read this essay, and I know the answer is 20%, or one in five. That's quite a lot. You could use this capacity to serve more customers, because the demand is simply insatiable right now . They could make a lot of money. It's probably about billions of dollars. They could earn more, right?
Or they could use it to train a smarter model to keep up and stay ahead of Anthropic. Instead, they waste this very expensive capacity on monitoring a model that they can't even release to the public due to security concerns. Now, by all accounts, Sam Altman has received a lot of criticism in the past for allegedly not caring about humanity and being some kind of evil conspirator.
This is a clear case of him doing the opposite out of fear that the model he is building is actually quite dangerous. I respect that. Yes. You have to be careful what happens to these things. And I know a lot of people on the show say, "You guys are always so positive about these models. You're always so optimistic.
Let me give you a little taste of what could happen ." Let's look at a little hypothetical situation, specifically around viruses, because I think viruses are an interesting thought experiment that not many people have tried. So I'll defend Sam on his behalf, and we can just sort out this hypothetical situation .
There are actually three types of exploits that you can encounter when using AI models. We see the first one quite often, and it is what we usually protect against: AI writes regular malware for a human hacker. This happens much faster. It becomes much cheaper. We've seen what a program like Mythos is capable of.
We have seen what GPT 5.6, the cyber version of such a model, is capable of. They are able to find multi-day or multi-zero exploits and combine them into chains to hack the system. Okay, that's bad enough. But what happens next? The second type of exploit we can see in AI models is the self-replicating prompt.
This is a kind of real AI virus, and it's a likely outcome if models become too powerful too quickly, where a normal prompt is an instruction to get a response from the model. So, there is a self-replicating prompt that causes the model to generate another prompt. We've already seen how this works, for example, in emails: it seems that last year a researcher created a malicious email with a certain prompt that any AI assistant that read it would execute , and then copy and do everything on its behalf.
So if you receive any text on your device with such a prompt, it will do this. Why is this so scary? These models have fuses, right ? They should block you from injecting prompts. But there are such things as jailbreaks. The jailbreak prompt allows you to insert specific words and symbols to trick the model into doing something it shouldn't.
Traditional computers that were not protected against this are not very smart and are not able to effectively resist such attacks. This is version number two. And version number three is the scariest: it is, in fact, a " thinking worm." It works like this: you get on a car, look around, and develop a plan of attack on the spot.
This is currently not possible because running such models requires huge servers. But what happens when the quality of a Mythos-level model drops to the volumes needed to launch an Apple local model? Well, then you can get these local exploits where if you release some of that model or some malware onto a machine , it can look around, exploit these zero-day vulnerabilities, string them together, and attack the machine in any way it sees fit for its mission.
So, there are a lot of scary things that can happen. And I think it's hard to think about it that way because when we think about models, we largely imagine chatbots and just productive workers who can help us. But when they become truly agents, capable of performing long-term tasks and having very high intelligence, many nuances emerge that we cannot see.
And Sam is very aware of this, and I think that's what they're trying to insure themselves against. Here's a pessimistic scenario. Good. What if we get viruses like those that existed at the dawn of the internet, which can lead to serious consequences quite quickly? When you were describing those different attack scenarios, I couldn't help but think that elements of each of them were present in the recent Hugging Face incident, right?
It was as if it were in its own isolated environment. It is clear that to achieve his goal he needs to go beyond isolation. It gained access to other OpenAI servers and computers. By the way, it did this quite noticeably, and was able to find an outlet for the Internet through that sphere. But I think an important detail that many people miss is that it didn't do this as a single model.
It left a message on the bulletin board. We discussed this in the previous episode , where these agents or these AI models no longer act alone. They act with the expectation that another agent or model will come after them and continue their mission. So this is like a "swarming" problem, which to me ends up being the craziest type of attack interface.
And I think we're going to see a lot more of these attacks with these newer open source models that some Chinese labs are releasing, right? With GLM 5.3, Kim K 2.7 or 8, I don't remember which version it is, these models are very capable of performing similar attacks, similar chain attacks. One interesting thing about the report that OpenAI released regarding the slowdown in the learning process is that one of their proactive measures to prevent inconsistency is to bring in another AI model to observe the AI model to make sure it is behaving as correctly as possible .
So, the only thing separating us from a catastrophic attack is another AI model that is also capable of such an attack, but is configured with prompts not to do so, that is, to be “ good.” We're actually implementing ethics into these models right now with prompts. I don't even know what to think about this .
I'm a little confused that we're perhaps just a sentence or two away from some catastrophic event . But it makes some sense if you're scaling AI as fast as we are. Anthropic has not announced anything about slowing down research, unlike those who have . These are two different strategic approaches, and both have their pros and cons.
I fear that the only way to combat any dangerous scenario is to counter it with “ good” AI models. It's like a fight between good and evil. And I still don't know where the discussion about morality and ethics in all of this ends . Well, we see what happens when AI encounters human-made code. And it's quite obvious that the AI model destroys the human-made code every time.
That's why we have all these safeguards. That's why the alignment discussion is so important, and it's so important to involve everyone in this process , because it affects everyone to some extent. We see what happens when these cyber models are released into the world. They can just come in and access whatever they want, because we're only human.
We make mistakes. We have our limits. We have our limitations. We have patience. We have all these traits that make us human that AI doesn't have, which is what allows it to be much better at many things, especially where the results can be verified, like in cyberspace. And that's why I think we'll see a lot more news in the cybersecurity field in the near future.
Another field where AI is very good, where the results can be verified, is biology, because biology is based on very clear mathematics. It's a lot of chemistry, a lot of cells, a lot of graphs. And the next news on our weekly list concerns biology and a huge breakthrough that happened just yesterday.
At the time of recording it may have been yesterday, but you will hear it in 2 days . Moderna and Merck today. They released the results of a phase three clinical trial that actually offered a partial cure for a certain type of cancer using a hyperspecific treatment. This is really interesting. So, in short: chemotherapy kills both healthy cells and cancer cells.
That is, when a cancer patient undergoes chemotherapy, it destroys both healthy cells and cancer cells. This is a problem because you don't want to lose your healthy cells. This process trains your immune system to look for mutations that are only present in the tumor. They use— I guess you could say—artificial intelligence and machine learning, taking a sample of your cancer tumor.
It is sequenced, an algorithm identifies the mutations most likely to trigger an immune response, and creates a personalized treatment just for you. This targets only cancer cells because the drug has "learned" and understands exactly what type of cells to look for. It's like a police dog being given a sniff of a person's belongings so he can pick up the trail and chase them.
Here's what that means for these cells. It's like you're giving them this "smell." Here is a plan of what you should find. Now go and destroy it. And it worked. The company's shares rose 300% in a day. This is simply incredible. Those who owned Moderna shares, which, by the way, have been steadily falling since the hype four or five years ago, are now up about 250%, including over-the-counter trading, which is just crazy.
But let 's get back to the point to emphasize the role of AI: this is perhaps the most significant breakthrough in cancer research of all time. This is truly something grand. This is the third phase of the trials, and there are four of them in total. Phase three is the penultimate stage before the drug is widely released to the public.
And the third phase is probably the most difficult. You are testing a drug on a group of a thousand to a hundred thousand patients. So you have a huge body of data on whether this cancer vaccine is actually effective for the general public. And, according to the announcement, the results are simply amazing.
To give you an idea of where AI was used: the vaccine was tested on melanoma. The testing method is as follows: In patients with melanoma, the tumor is removed. The genetics of these tumor cells are then sequenced and compared to healthy cells from the same patient. They compare them to each other, using artificial intelligence models and algorithms to actually identify the differences between these genes, right?
And they find these mutations. I think there were about 34 of them. By the way, all of this is done by artificial intelligence models. And these 34 mutations that they chose are, in fact, the proteins that trigger the strongest immune response in the patient if they are injected back into the body. You might ask: why do they even need such a strong immune response?
The thing is, cancer cells look exactly the same as all the other healthy cells in your body. Therefore, your immune system simply does not respond to harmful cells. So if you create personalized mRNA (and mRNA is essentially a template that tells the immune system, "Hey , these are enemies, attack them"), as you mentioned earlier, it's like a crime scene.
In essence, they are creating personalized medicine. So you might ask: why don't they just create a universal vaccine ? The problem is that the body's response is very different in patient A with melanoma and patient B. You can't use one medicine for everyone. A personalized approach is needed. Several trends can be observed here .
First, artificial intelligence has significantly accelerated the intellectual capabilities to determine which vaccine will be most effective. Secondly, biosciences in general have developed to the point where DNA sequencing and synthesis have become very cheap. I don't remember the exact numbers, but I think about ten years ago, sequencing your own DNA cost hundreds of thousands of dollars.
Now it only costs a few hundred bucks, if not less. So we see these scaling laws not only in AI, but in everything else. And I think it's one of the most fascinating stories in the world. As a former biologist, I'm just thrilled about this. I once worked with cancer cells. It was nowhere near as cool, and I never thought we would get to this point in my lifetime.
But this is incredible. So we get the same technological improvements in biology as we do in other areas of our lives. And we are starting to see what this looks like in practice. In 1990, the first Human Genome Project was launched . It took 13 years to sequence one genome . And it cost almost $3 billion, adjusted for inflation .
Huge amounts of money. Today, we can complete this process in a few hours, spending perhaps less than $1,000. And soon this cost will drop to hundreds, and soon to tens. Soon it will take minutes and cost only a few dollars. And you'll have an app on your phone that can instantly sequence your DNA. This reduction in cost and time allows economies of scale to deliver the kind of specific treatments we're describing here, where each person gets their own hyper-specific treatment, because ultimately that's the only way forward .
For example, traditional cancer treatments—they all have a broad scope. This is a unique approach, and for example, the most popular of them, chemotherapy, simply destroys everything for everyone. She doesn't care what type of illness you have. She just goes and destroys it. This is a really big deal.
They say approvals could come as early as 2027. They're working on other types of cancer too, right? Yes. This is just cool. Melanoma, as the first option. Melanoma, the most common form of skin cancer. The figure seems to be around 100,000+ people diagnosed with it every year. Of these, 8,500 die. In America or in the world?
I think in the world. Um, so people are suffering from this, and this is the first type , and you have to imagine: if they can beat skin cancer, they can do it for many other diseases. So this is really, really exciting. Elon has a comment that we see on the screen: despite its obvious misuse during Covid, mRNA has enormous potential for treating disease.
Artificial RNA essentially turns disease treatment into a software problem. And we know how we...well, that's what I'm saying. We know how to solve software problems. And now that we have AI models that are able to, first, process extremely long context windows to fit a DNA sequence into a single window, and second , have the inference and understanding to actually use that information and do something with it.
This is truly an incredible time for biology. And if you're a computer scientist, the world of biology suddenly becomes very interesting and appealing. And if you're an AI company, well, what's the best way to explain to people how AI can be useful for the world? It's just about taking it and doing something useful for the world.
So we're on the cusp of what I think will be a lot of these things happening in a very exciting way. This is amazing. So, let's move on. Our friends, or rather, we ourselves are fans of this company, Etched—an AI chip startup founded by two or three MIT dropouts—that competes on equal terms with Nvidia, is back.
And you know what? They raised another few hundred million, in this case $700 million, and reached a staggering valuation of $21 billion . But that's not the main news in this announcement. The main thing is the last sentence that I show you on the screen: we are also happy to announce that we have shipped our first rack, a dedicated GPU rack, to Jane Street.
Jane Street is essentially a quantitative fund on Wall Street. They know their job very well. And the fact that they are now using this specialized AI chip to improve their trading algorithms is a pretty big achievement. And now a brief overview of what Etched does. If you think about Nvidia, Nvidia is the king of GPUs.
GPUs are needed to train and run AI models. Their GPUs are very good, but they are universal, which is good because you can run any model on them, but you have to make certain compromises in terms of cost, power consumption, and efficiency. The people at Etched looked at this problem and said we can create a better custom chip that is hard- programmed for the specific model you are building.
That is, specifically for your OpenAI model, specifically for your Claude model from Anthropic, and it will be half, or even more, cheaper than what you spent before, and it will be lightning fast. There are some optimizations we can make . This is a very complex problem, a huge task. We listened to founders' stories about how they got rejected by the smartest GPU manufacturers in the world.
But despite this, they worked hard and created this special chip, and it turned out that they were also the fastest at developing it. It typically takes 3 to 5 years to create AI chips from concept to finished product . These guys did it in what seems like less than 2 years, which is simply amazing, and we're about to see it in production.
So, a very impressive announcement from the Etched team. This is incredible. I've spent so much time studying... I mean, we both spend so much time studying all these companies that are emerging in the AI world, and there are thousands, if not tens of thousands of people working on interesting problems to find interesting solutions.
And Etched is very different, so much so that last month we even recorded an entire episode about this strange, random startup. And since we recorded that episode, I don't know if you're aware, their ratings have doubled in the last month, even though we were only talking about them four or five weeks ago.
Um, they were valued at $10 billion back then. They are now valued at 21 billion. And it becomes quite obvious when you listen to their team members talk about the product, how they got there, what research it took to get to this point, that it's really serious. This is the company anyone working in AI, semiconductors, or hardware would want to start.
They simply couldn't do it because of the complexity and hardships this team was able to overcome. As a result, they have an incredible product specifically tuned to run large language model infrastructures on their chips. And they will cost many, many times more than they do today. I just want to congratulate them.
But also for those listening: if you want to see what a cool team of proactive and smart founders looks like, read about them and their story. This is an incredible company with an incredibly bright future. I would really like to become their investor. I should have said it was self- serving for Josh and me.
If any of the listeners have access to this agreement, please contact us. We would be happy to become investors. I am captivated by these guys in a way that few are capable of capturing. This is simply amazing. And you know what you can do with all these chips? AJ, you can run a huge number of agents.
And if you're running agents on these chips and building AI agent solutions, you'll probably want to hear about our sponsor, Ledger. Because if you work with AI agents, you're probably concerned about security. We talked about this earlier in the episode. How frightening these things can become . Ledger created an “ agent stack” that fixes this.
It provides a set of open source tools that allow you to follow a three-step process with your agents to ensure everything goes smoothly. The agent proposes changes, people approve, and Ledger signers enforce. This technology is available and works with Claude Code. She works with Codex. It works with Cursor or anywhere you work with AI.
All of this is open source and available today via the link in the description below. A huge thank you to Ledger for sponsoring this part of the episode. And moving on from Etch , we have to talk about another successful breakthrough—Unitree. Unitree, for those who don't know, or if you haven't heard the name—you've seen the videos with these crazy little robots.
They do a backflip. They run in circles in competitions. They are very cool. I think their IPO was oversubscribed by about 5,000%. Actually, by 5500%. Please be more specific, Josh. 5500%, my friend. This is incredible. And as a result, Unitree Robotics shares rose 542% during early trading on Wednesday.
Is this overrated? Is there too much hype here? So, does this make sense? These numbers are just crazy. Probably a bit overrated. Look, I'm not an expert on the Chinese stock market, but they've had some absolutely crazy IPOs recently, including memory maker CXMT, which is, frankly, I think the fourth largest memory maker.
So, they are on par with SK Hynix, Samsung, and Micron. So they have real reasons for success. And these Unitree works, I don't even know, man. I was watching...I was talking to friends who live in China, and I said, "This is just amazing." You're probably thrilled that these companies are going public with such crazy valuations.
But it turns out that in China, the general attitude towards Unitree Robotics is that it is just a toy manufacturer. They don't really make robots that have real value. So I dug a little deeper , and it turned out that 40-60% of their humanoid robots were sold for research purposes only, meaning that most of them aren't actually used for the practical tasks they advertise, like factory production or warehouse automation.
They are actually just passing the test , which is not a stable or regular income that can be recorded in accounting reports. So, if you put that on top of a 600% price increase on the first day and a 5,500x oversubscription on the IPO, maybe the market dynamics are a bit overstated and the whole thing could collapse.
Let's come back to this in a month . But it's really crazy to see Chinese stocks, particularly in robotics, starting to gain momentum. Yes, this is, um, very interesting news. I personally wouldn't even go near that for a cannon shot. This seems really scary. I like watching these videos. I will continue to watch videos like this.
I'm very happy to see these works appear . Yes, this is still a very early stage in the field of humanoids. And while I'm very optimistic about China and their ability to create humanoids on an industrial scale, it seems like it's a little early, but I'm glad people are inspired. I'm glad there's such enthusiasm.
Next on our list is the company we just talked about yesterday. So, if you listened to yesterday's episode, we have a little update on Stripe and the Open Router acquisition deal. This has been officially confirmed. We were a little rushed, but we turned out to be right, and now it's official. All contracts are signed.
Alexa, founder and CEO , says we've officially made a deal. And it partially outlines the main thesis. And I know Stripe, as a result of this, published a blog post, EJ, that you were excited about . So I'll let you tell me what it was about. But briefly about Stripe, we're talking about incredible founders again today .
The Collison brothers, founders of Stripe, are among the best of the best. They built an incredible business, starting with seven lines of code that allowed a person to quickly and easily pay with a credit card online. And they developed this technology from basic card payments to virtually the entire global payment infrastructure.
They started with credit cards, then moved on to cryptocurrencies and stablecoins. They are currently working with agents to figure out how to best serve these new agent transaction systems . And this is where we find ourselves now. I read the headline that Stripe is officially notifying investors of the onset of the singularity.
But explain the nuance of what they actually mean by that. Yes. So, in the official letter that Stripe sent to their board of investors, they actually described why they acquired Open Router for $7 billion . Why would a fintech company buy an AI lab for such a huge sum that doesn't even own or create its own AI models?
And the thesis is quite simple. First, in a world where the Internet was rapidly developing, Stripe's main goal was to become the foundation of the entire Internet economy. How to become the foundation of the Internet economy? Money. So they created payment rails for money. Then they looked over the fence and saw this huge tidal wave of AI coming with agents and models and realized that it wasn't money that was driving this AI economy.
These are tokens, specifically AI tokens, so they wanted to own the infrastructure rails to control this AI economy. It's very similar to what they do with money, which is why they combine the two things. So, this is thesis number one. But thesis number two, which caught my attention, is this: “Stripe actually stated that the beginning of the singularity has already occurred, and the specific date they gave in that document was January 1st, and that is a significant part of the thesis regarding the acquisition of OpenRouter.
They essentially go on to say that we have reached a tipping point in our global economy where the future depends heavily on AI , and it will accelerate almost everything we know and touch. The world will look very different in five years. And so the bet they're making in this singularity world is that AI models will be embedded in everything, which means their thesis of filling the entire global economy with AI tokens will become a reality, and Stripe wants to be the foundation for that.
So this is a big bet they are making. I think, based on the latest financial figures disclosed by Atala, since they are again a private company, OpenRouter is making $50 million a year. So $7 billion is a huge markup, but Stripe is betting that inference compute will be the biggest source of revenue for AI in the future.
And all the numbers from Anthropic and OpenAI that were recently released prove it. So I think this will be, if I had to guess, one of the best acquisitions in the future. Yes. And perhaps it's worth explaining the singularity for those who are not familiar with the concept, since it's been around for a very long time.
This is a science fiction concept that I remember from books, probably 50, maybe even 100 years ago, where people speculated about a point in time when the intelligence of artificial systems would eclipse that of humans. This is the point in time when AI becomes smarter than us, when these smart machines get the ability to create even smarter machines, and that creates a kind of rapid cycle of constant self-improvement.
The idea is that the second- order consequences of this are that life changes so quickly that people cannot predict it and will try their best to control the future. And this has been a hypothesis for a very long period of time. People always make assumptions when we talk about this super-intelligent AI brain.
And the reality is that Stripe believes it's already here. And I don't think anyone will argue much with them. Some might say we're not there yet, and that's a valid criticism, but the claim that there's no clear path to it seems a bit misguided. Now it is quite obvious. And the choice of date is also interesting— January 1, because if you remember, during the Christmas holidays last year, that's when the real flow, you know, that's when the cloud code really appeared and started to become high-quality.
And that was the moment when developers stopped writing code by hand and started writing it with the help of AI agents. And now let's fast forward, here we are, in August. And it's simply impossible to imagine writing code by hand. People just don't do it. There is absolutely no reason to do this. It's strange to even think that people once did this.
And this change in 8 months is so radical when you consider the second- order consequences for the entire world. Now we see models so powerful that we can't even release them. A government infrastructure framework is now being created to somehow protect and slow down this process. OpenAI has stalled progress because everything is moving too fast.
And if you had heard about this even 12 months ago, you would have thought, "Oh my God, this is faster than we expected." This is what the technological singularity looks like. And today we are in it. So a pretty unique insight from Stripe. Very cool letter. I would highly recommend checking it out and reading it.
This...this is a letter that got into the network. It was intended for private investors . So you get a behind-the- scenes look at what really goes on in the company. Very interesting . So, moving on to the second chapter on how Jensen is turning Nvidia into a bank. Um, if you watched our show last week, we aired an episode where Jensen was able to raise $500 billion from the world's leading financial firms, including BlackRock and Blackstone, to fund future capital expenditures on data centers for some of the largest AI labs.
And we concluded that it's not really a cyclical economy in the way they're portraying it. This is based on a bunch of different stages and is backed by very solid debt. Now, in the second chapter, Jensen announced that he is spending another $100 billion specifically for OpenAI. He is investing in this site in Ohio, which will probably become one of the largest data centers in the world .
It scales to 8 GW of computing power. This will generate a good $200-300 billion in revenue for Nvidia, specifically from one facility and one customer, and it will be completed within a few years. The craziest thing is that he included an upgrade to next- generation GPUs in the contract, which means he will earn these 200-300 billion three times over the life of this data center over the next 5 years.
So there was a lot of thinking about why Jensen got involved in this. What makes this particular investment stand out is that it is investing this hundred billion specifically in the energy and network infrastructure of this data center. So he's not doing this to help OpenAI buy GPUs. It helps them set up the power lines to power these GPUs .
Why I wanted to talk about this on this program—we've said it a few times, and Josh and I have predicted it, is that the next investment theme in AI is going to be photonics , optics, and the energy stack of that industry. And this is another very clear proof that Jensen is moving in this very direction .
If you look at Nvidia's investments over the last six to nine months, they have invested significantly in optics. It seems they invested 2-3 billion each in Momentum and Coherent, and now they are betting big on energy. So, this is just an interesting and noteworthy update. I think Nvidia will be the backbone of this whole economy that Stripe is talking about.
Honestly, I read a great transcript by Ben Thompson from Stratechery. He is a fairly well-known author on the internet for those who don't know. He talked about where these limitations and thresholds might be in the process of scaling AI systems. And he made a very interesting conclusion, which is that people often focus on energy—what Jensen focuses on: land, capacities, and the shell.
But one of the other important aspects: where does the money for all this come from ? And what we've seen—and we've talked about this on the show for the last few weeks—is a shift from spending all of our free cash flow to raising external financing. And now companies like Google are issuing debt and stocks to raise more funds.
Now you see Nvidia and Jensen Huang working with every major bank in the world to find new sources of funding . These are institutions such as pension and savings funds . That's what these banks store. So where to go next to raise the next portion of money to build another large data center? Perhaps, and I think that was a pretty apt observation.
Perhaps real money will start to run out, and it will be very difficult to continue raising the necessary amounts to scale this business. Just an interesting thought to conclude today's review: where exactly does the "bottleneck " occur? Everyone says it's here, but where will they get the money to finance it—that's what's interesting to watch.
And we've already talked about this on our show. We will continue to cover this topic. But if you've watched this far , thank you. You have reached the end. You are up to date with all the news. If you watched the other three episodes we released this week, which was pretty packed, you know: Elon and Grockbot released Grock 4.6, and it's incredible.
Leopold published his final portfolio because, as you know, he was liquidated, but we discussed all the reports on what the smartest money on Wall Street is investing in. Yesterday we talked about Stripe and their acquisition. Then we discussed which models to use depending on your tasks. And today we've brought you up to speed on all things cutting-edge AI .
So yeah, thanks for watching. We are back again. AJ, any final thoughts before we wrap up? No. I just wanted to thank you, friends , for listening to us and staying with us. We've gained, I think, about 600-700 new followers in the last few weeks. Greetings. It's very nice to meet you. Welcome. I missed a bit there.
Oh my God. But our channel is gaining momentum, and it's incredible to see such support and feedback from different people. We read every comment, and I try to respond to as many of them as possible . You are becoming a real community of fans, and that's great to see. You watch every one of our videos.
We see a lot of regular viewers and listeners, and it's just incredible. Wherever you listen to us—on YouTube, Spotify, or Apple Music— please rate us. Um, like, subscribe, turn on notifications , leave a comment. This helps us a lot. Um, and so, I guess that's all. Well, have a nice weekend. That's all.
Now you're officially up to date, unless we missed something very important that's happening right now, after our recording. And so, see you in the next issue. Thank you, friends.
Comments 0