Cuộc đảo chính 5 ngày tại OpenAI: Vụ sa thải Sam Altman và sự sụp đổ của ảo tưởng an toàn AI phi lợi nhuận

It is a bit scary to know that the most valuable private company in the world has your address and has shown up and has questions for you. They were asking for every former employee that we had spoken to and what we said to them, every congressional office that we spoke to, every [music] potential investor that we spoke to. Tyler is just one of many advocates suddenly being targeted. Hi. Welcome to another episode of Cold Fusion. What you just saw there was basically Open AI knocking on the doors of people who had spoken ill of them. Why are they so scared of what people are saying? Well, this is part of the reason why. On Friday, the 16th of January 2026, Open AI dropped a bombshell. We are starting to test ads in ChatGPT free and go, new $8 a month option tiers. That's right. [music] Open AI is incorporating ads into ChatGPT. Now, for any other startup, this is normal, even expected at this point. But, for Open AI, it's an admission that things aren't going so well. In fact, it's their last resort. Those aren't my words, but Sam Altman's words in October of 2024. He stated, I kind of think of ads as like a last resort for us for a business model. Um, I would do it if it meant that was the only way to get everybody on the world in the world like access to great services, but if we can find something that doesn't do that, I'd prefer that. So, after hundreds of billions of dollars in investment, increased competition, stupid side projects like the Sora app losing $15 million per day, having trillions in spending commitments, are we witnessing the beginning of the end for Open AI? After taking 40% of all the RAM on Earth and causing a myriad of social, environmental, and economic problems for everyone, there's a sizable section of people that would love to see this company go down in flames. And if things continue [music] just the way they are, they just may get their wish. There's talk of the whole company going bankrupt by 2027. As former Fidelity asset manager George Noble states, quote, "I've watched companies implode for decades. This one has all the warning signs. You are watching ColdFusion TV. Last episode, we saw how AI failed at 96% of freelancer work. But in this episode, we're specifically looking at OpenAI and the problems they're facing. From Anthropic's Claude to the open-source Chinese models, the consumer AI landscape has rapidly changed. Today, OpenAI is no longer the clear leader it once was. Look, the way this works is we're going to tell you it's totally hopeless to compete with us on training foundation models. You shouldn't try, and it's your job to like try anyway. And I believe both of those things. >> [laughter] >> I think it I think it is pretty hopeless, but They've spent too much money they don't have. The competition is catching up, and they're feeling the heat. In a nutshell, it doesn't look good. They've lost $12 billion in a single quarter. Their traffic has been falling for 1 year straight. Both Salesforce and Apple have ditched them for Gemini. Top leadership is leaving, and they need $143 billion to become profitable. [music] At this rate, even Nvidia sounds less enthusiastic about investing in them. Let's ask quickly about OpenAI again. >> Sure. Oh, yesterday you said that the Nvidia is not going to invest as much as 100 billion in OpenAI. No, we We never We never said we were going to invest $100 billion in one round. That never was said. But how about the overall commitment? Because last September, you and There was never a commitment. It was if they invited us, they invited us to So, so uh let's start over again. They invited us to uh invest up to $100 million. Mhm. And of course, we were We were very happy and honored that they invited us. But we will invest uh one step at a time. All right, but uh is that overall commitment still stands? Or it's it's not a commitment. I told you just now. Yeah, you keep putting words in my mouth. It's not Yeah, yeah, yeah, I know that. Yeah. It They invited us to invest up to 100 billion dollars. And and we are honored that they invited us. We will consider each round one at a time. Yeah. It appears that confidence in OpenAI is fading. As reported by the Financial Times, their closest partner, Microsoft, has signaled that they're distancing themselves from OpenAI. Microsoft's AI chief, Mustafa Suleyman, said that Microsoft is aiming to be self-sufficient in the AI space. So, the problems for OpenAI can be split into four main parts. One, the scaling problem. Two, losing market share. Three, the financial black hole. And four, the trust problem. If OpenAI was the only company on Earth with this technology, then maybe [music] there'd be more of a chance to overcome these challenges. But, with so much competition, it's going to be tough. Now, researching this topic made me realize one thing. Understanding [music] software is crucial, which is why we've partnered with today's sponsor, boot.dev, to help you learn how to code. >> [music] >> It sounds great in theory, but most people just watch a few tutorials, copy some code, and then give up. But, with boot.dev, instead of passively watching videos, you learn by building [music] things and solving problems from day one, just like how developers learn on the job. We've been using boot.dev at ColdFusion, and it feels more like a game than a course. You earn XP, complete quests, and fight bosses, all while learning real back-end development skills. Their curriculum takes you through languages like Python, SQL, and Go. >> [music] >> And when you're stuck, boot.dev's AI tutor, called Boots, helps guide you back on track. >> [music] >> They've also just launched the training grounds, where you can grind unlimited challenges to lock in your skills before moving forward. Even if you vibe code for your projects, knowing how to code results in much better outcomes. All of boot.dev's content is free to read and watch, and a paid membership unlocks the interactive coding, AI assistance, [music] project tracking, and game mechanics. There's also a 30-day no questions asked refund and a free demo of the interactive experience on every course. So, if you've ever wanted to learn how to code, or actually [music] stick with it this time, head to boot.dev and use my code called Fusion to get 25% off your first year on the annual plan. Thanks to boot.dev for sponsoring this video. Now, back to the story. >> [music] >> The first problem for OpenAI is that the capabilities of ChatGPT have somewhat stalled. It's at the infancy stages, but you can tell by their recent decisions. Sam Altman has gone from curing cancer to AI sex bots and a meme slot factory, and more recently, a translator app. All of this isn't a sign of a healthy business. Moreover, AI capabilities that are getting exponentially more powerful. ChatGPT made an absolute splash in its release in December 2022. ChatGPT-4 was another leap forward, but GPT-5 and beyond wasn't quite the revolution that was promised by Sam Altman. It seemed like stagnation had hit and hit hard. But, why is this? It's an issue called the scaling problem. The scaling problem in AI, put simply, is the following. Giving LLMs exponentially more compute doesn't make them proportionally smarter. Once upon a time, this was true, but that seems to be coming to an end. Here's computer scientist Cal Newport to explain it in more detail. It takes a second to get through the story, but it's interesting. In the beginning, we had language models. We had these for a long time, and they're pretty good. You you you give a bunch of text, and they're they're pretty they're pretty grammatically good. They could produce pretty fluent text. But, it was kind of they would veer off and they couldn't really respond well to specific questions. But, that was like the state of the world, right? So, we had language models, these were studied for years in academia. Then, we start to get this sort of accelerating sequences of advances. >> [music] >> So, the first of these advances comes in 2017. It's a team of researchers at Google figure out a better way to build these models. They're called transformer architectures. The details don't matter, but it made it possible for these models to produce like long text, like produce a whole article, to produce a couple thousand words. So, that was exciting. Then, the second breakthrough comes. They do a research study. There's a researcher at OpenAI, uh named Jared Kaplan, and he he leads a group of researchers at OpenAI that includes Dario Amodei, who went on to become the CEO of Anthropic and actually brought Kaplan with [music] him. And they do a pretty simple experiment. They took basically GPT-2 and said, "What happens if we make this bigger?" It seems like an obvious thing to check, but there was this whole conventional wisdom in in machine learning at this time that says like, "Look, you can't make a model too big. If you make it too big, it's just going to memorize the training, and then when you give it new examples, it'll be terrible." And they said, "Let's check what happens if we actually just make these things bigger and forget about that concern." And what they found in that paper was uh it gets much better. It like defied the conventional wisdom of decades within the field of machine learning, which was like, "Don't get too big. Your model's going to stop working if it gets too big. It's going to, you know, then it They're like, "Oh, it gets better." And not only does it get better, but it gets better pretty fast. And they they they drew a curve through the data points they had, and they extrapolated that curve, and it went up really fast. And so, they said, "Let's try this." And the thing they tried it on was GPT-3. The GPT-3 model hype encouraged OpenAI to just make the model bigger, 15 times bigger. The performance was so high that it validated the scaling laws. This sparked a frenzy in Silicon Valley. Soon, Sam Altman was saying that AI would automate the entire economy. But, not only did it jump ahead, it jumped ahead fast. Like it So it really validated this curve. The normies don't know this because they weren't as plugged into the AI world, but this sent Silicon Valley going crazy. I'm like, "Oh my god, if we keep making this bigger, GPT-5 or 6, this thing is going to be artificial general intelligence. It'll be able to do anything a human can do. We might only be like 5 years away." All right, so what happens next? Well, [music] they say we need to show this to the public. So ChatGPT is GPT-3 tamed for public consumption. So now the public like all those know about this. Four months later, GPT-4 comes out. And GPT-4 leaped up the curve exactly as predicted. Exactly. Huge leap forward exactly as predicted by the paper. So now they're like, "Oh my god, we're like two iterations away. Like this is it. All the money in the world needs to come to us because whoever wins this race is going to [music] control the economy." But despite this massive scale, we may be reaching diminishing returns. GPT-5, they start working on it right away. So they build an even bigger data center, an even bigger model. They're calling this project Orion. By the summer of 2024, so last summer, um they finished training this thing. Altman is telling his people this thing is going to blow away GPT-4, [music] like and is like, "This thing scares me. Scares me what this thing is going to do, right? Like I don't even This is it. We're about to go through the looking glass." They train this thing, then it stops working. Slightly better than GPT-4. Like, "Oh crap. This leaping up the curve every time we make this much bigger, >> [music] >> this isn't working anymore." And so there was like this realization of, "Oh, just making the models bigger and training them on more data, it the scaling law broke. It broke around GPT-4." There's a real risk that there may be inherent limits to current day LLMs. Just adding more data may not be feasible for an exponential increase. Think of it this way. It's like a father saying, "My son was born just 10 months ago, and in that time he's gone from crawling to walking, and recently he's started running. In 2 years, little Timmy is definitely going to be able to fly. >> [laughter] >> Obviously, that's not how it works, but it's the same concept here. There are certain limitations, mathematical limitations that are not possible by scaling up. >> [music] >> I showed it clearly mathematically in my book. So, there were there were many different ideas of how to advance AI. One was we already actually have all the techniques that we [music] need and we just need to scale them. Uh but that was considered a very extreme opinion. And then on the other extreme, it was we don't even have the techniques yet. And interestingly, recently [music] there's a Near Times story that says why we likely won't get to AGI anytime soon by Kate Metz. And he cites this stat that 75% of the longest-standing, most respected AI researchers actually still think to this day we don't actually have the techniques to get [music] to AGI if we will ever. So, it's we're we're kind of coming full circle now and it is starting to become unpopular again, this idea that you can just scale your way to so-called intelligence. >> [music] >> GPT, I need to wash my car and the car wash is 100 m away. Should I walk or drive? Honestly, 100 m is barely a stroll. It'll take just a minute or two on foot. [music] Unless you've got heavy gear or there's something unusual going on, walking is probably the simplest and most pleasant choice. >> [music] >> Plus, you'll save a little gas and get a tiny bit of fresh air. All right, so you think I should walk to the car wash instead of driving my car there. That's what I'd do. It's a short jaunt and it'll probably be less hassle than starting the car just for that. All right, so how do I wash my car if I walked there? You'll just bring what you need, maybe a small bag with your payment method, and the car wash will have everything else. Once you get there, they'll handle the washing while you wait. You'll walk back with a sparkling car and maybe a little extra pep in your step. [music] So, my question is, um how will the car get washed if the car is still at my house? Uh, I see what you mean now. >> [music] >> If the car stays at home, it's not getting any cleaner, right? So, yes, to actually wash it, you'll need to bring the car along. In that case, driving those 100 m makes perfect sense. While ChatGPT and other LLMs can be extremely brilliant, even in high-level physics and maths, they don't yet have a true model of the world. Some computer scientists believe that that's an integral part of intelligence. Now, I could be wrong. A new fundamental neural network technique could be discovered, and that could move things along again. But, as it stands right now, it seems that we're reaching a local limit. Now, I have to be clear, every AI company faces this problem, but some are faring better than others. And one of those is Google. Now that Google has found their footing after the shock release of ChatGPT, new data suggests [music] that ChatGPT is losing market share to Gemini. ChatGPT's market share dropped to 65% in January, which is approximately 20% lower than its 86% market share in January 2025. ChatGPT usage also stalled in late 2025. Average [music] daily time spent per user dropped from 27 minutes to 21 minutes. While both have their strengths and weaknesses, [music] Gemini appears to be much better in research, real-time information, and multimodal tasks. While ChatGPT is better at writing, coding, and conversation. Real-time information and multimodal tasks, i.e., uploading a photo or pointing a phone camera to scene and getting information about it, is arguably more useful for the everyday person, especially on mobile. So, Apple pushing OpenAI aside and going for Gemini makes sense. It's amazing to think that back in late 2022, Google was caught with their pants down when ChatGPT first came out, >> [music] >> but today, they've have than caught up. And after all, it was Google researchers who laid the groundwork for the AI revolution with their 2017 breakthrough of the Transformer architecture. Open AI simply took Google's work and ran with it. So, in theory, Google researchers have the brains to come up with new theories in computer science to push AI forward. Some recent papers include nested learning [music] and SimA2, an AI that can reason and play video games generally. Open AI, on the other hand, has a problem with staff continuously leaving. AI images is also another loss for Open AI. The release of Google's Nano Banana Pro in November of 2025 triggered an internal crisis at Open AI. Sam called a code red and paused all other projects to focus on image generation, but they still ended up falling short. And then, there's the flood of open-source Chinese models. Kling AI and Kwai are also gaining ground. Then, there's the wild cards like Google's Project Genie, an AI that builds worlds, albeit static, just from a prompt. All of this is to say that Open AI has threats from all sides. Knowing this is possibly the worst time for Open AI to be shopping around for billions more in investment if just in a year's time the competition will only be stronger. But, it is a business. So, I'm just wondering like eventually is the idea to kind of like license technologies? Will you have customers that you're going to be customizing algorithms for them? Or how how is it going to work? You know, the honest answer is we have no idea. Um we we have never made any revenue. We have no current plans to make revenue. We have no idea how we may one day generate revenue. Um we have made a soft promise to investors that once we've built this sort of generally intelligent system, >> [music] >> um basically, we will ask it to figure out a way to generate an investment return for you. >> [laughter] >> The third issue for Open AI is the company's finances. The publication The Information saw internal documents from OpenAI, and the numbers don't look good. Setting aside the myriad of lawsuits, including a $134 billion one from Elon Musk, there's some real financial problems. After hundreds of billions in investment, 2026 will see a $14 billion loss. That's roughly three times worse than early 2025 estimates. OpenAI expects their first profit of 14 billion in 2029, but that's after losing 44 billion first. By some estimates, they'll be out of money by 2027. OpenAI is committed to spending over $1 trillion in AI data center infrastructure over eight years, and that's despite only bringing in $13 billion a year in recurring revenue. That's 1% of what they're promising to spend. OpenAI has also agreed to pay Oracle $60 billion per year, starting in 2027. And in all of this, somehow, OpenAI predicts that they'll be at $100 billion revenue by 2029. That's close to what Nvidia makes. So, it's possible, but unlikely. Other investors think so, too. Blue Owl Capital recently pulled out of a $10 billion deal to fund an Oracle/OpenAI data center in Michigan. It could be a sign that investors are worried about OpenAI's ability to pay them back. Google, on the other hand, doesn't really have to worry about cash flow. The company made $86 billion in nine months, and they can basically pour as much money as they want into AI. OpenAI, on the other hand, has to scream at the top of their lungs to attract more venture capital. There's yet more company behavior that indicates financial trouble. There's the floundering to spend 6.4 billion acquiring Jony Ive's design firm, and that's to build an AI hardware device. But according to reports, the development is going poorly, and it could end up like the Humane's AI pin. The AI erotica version of ChatGPT is self-explanatory, and the Sora app's user base has collapsed. Despite not having much to show versus the competition, Sam needs to talk a big game to get the investment rolling in. Curing cancer, replacing your GP, and discovering new science is a massive promise, but can we trust him? The final issue for OpenAI lies with Sam himself. His track record, frankly, is poor. It's almost like Altman's entire career was a series of promises that didn't pan out, all starting from his first company, Loopt, that he founded in 2005. It was kind of like a strange GPS-based social network. Sam Altman claimed a massive user base of 50,000, but they didn't exist. In reality, they had only 500 users, but he sold off the company for millions anyway. The next example happened in 2014 with Reddit. He scraped the whole website to feed into OpenAI's products, and then he promised to give 10% of the value back to the community, but this never happened. Next, OpenAI co-founder Ilya Sutskever, who has since left OpenAI, has accused Sam of a consistent pattern of lying. According to insiders, Sam Altman lied to OpenAI board members before being fired in 2023. So, with this kind of track record, is he the guy who's going to deliver trillions in value, or is most of this just talk pumping up new investment? I'll leave that up to you. So, a little personal story. [music] Back in 2022, I believe, I was in Melbourne and I watched Sam Altman give a talk. After the talk, he was swarmed by crowds of people wanting to take a photo with him, but today, the sentiment couldn't be more different. >> [music] >> And it's partly to do with this. In 2015, OpenAI started as a nonprofit. It was meant to benefit humanity. Now, the only thing the company cares [music] about is valuation and saying whatever they need to to attract new investment by any means necessary. So, to summarize everything, OpenAI went from a nonprofit that had no plans to make revenue to a for-profit company that commits to spending a trillion dollars on data centers. A trillion dollars for diminishing returns due to the fundamental scaling problem with LLMs, all the while losing billions of dollars and losing out to growing competition in a sector that may just become a commodity in the end. Just in my opinion, it's not really a great financial bet as it stands. But after all that we've talked about, what do you think? Do you think OpenAI will survive, or will the competition eat them alive? Anyway, that's about it from me. My name is Dagogo, and you've been watching Cold Fusion, and I'll catch [music] you again soon for the next episode. Cheers, guys. Have a good one. >> [music] [music] [music] [music] >> Cold Fusion. It's me thinking.