Transcription
So yesterday, Sam Altman made this post called "Three Observations," and it's an insightful document detailing life in 2035. I think most people are missing the mark with this document because it discusses things most aren't even looking at, things pertaining to exactly how AI will be integrated into society. So, without further ado, let's dive into these specific details because I think most people are missing a lot of the details when it comes to the beauty of this blog post.
Basically, this article starts with Altman saying that systems pointing towards AGI are starting to come into fruition. He puts asterisks here because of the Microsoft deal, which is referenced at the end, but that's not the main point. They basically say that the systems pointing to AGI are starting to come into fruition, and it's important to understand the moment we're in. He talks about how AGI is weakly defined, but generally speaking, it's a system that can tackle increasingly complex problems at a human level in many fields. I believe we'll get a truly encompassing system by at least 2027.
He continues—and this is something you shouldn't miss—because I think this AI stuff is more important than ever. I recently had surgery and was in the hospital. Luckily, I'm back now, still recovering. But being out of the AI space for a short time allowed me to understand exactly where we are in terms of the collective consciousness for artificial intelligence. He says each new generation builds upon previous discoveries—like electricity, the transistor, and the computer. These are tremendously important; you're probably using electricity, your phone, laptop—all powered by electricity. Transistors make these things work, forming the backbone of society, things we take for granted. We're moving into something 10 times more impactful, and the average person almost doesn't care, wrapped up in their own bubbles. This will be bigger than electricity, the transistor, the computer, and the internet.
He also says this time is different; the economic growth looks astonishing. We can imagine a world where we cure all diseases. Sam Altman has said this before, to some criticism. I think that's because, number one, Altman is a pretty hyped-up person; he goes on the hype train quite a bit, notably because his company's always trying to raise money, which makes sense. But also, if we think about what AI can do at a base level—connecting things, looking at connections between things—if we look at how certain diseases were solved in the past, it's quite likely we will get AI systems that can do that in the future. With Altman's hype, and considering how we've solved major diseases in the past (sometimes by accident, like penicillin), if we have multiple AIs submitting theories and a way to validate those theories at scale, then we could be in a world where we cure all diseases. Especially when we can simulate the human body, run a billion different experiments, and have intelligence 10 times smarter than us. Healthcare will change. Looking back on 2025, it will be like looking back on 1995 in terms of technology. The difference is stark, and in another 10 years, people might be like, "Whoa, things really did change."
Another crazy statement—and why I advocate that most people should be learning AI tools—is that in a decade, perhaps everyone on Earth will be capable of accomplishing more than the most impactful person can today. That's a crazy statement, but I think it's true. With an AI agent that can do a billion different things for you, multiple AI agent swarms, or a humanoid robot, there will be no limit to what you can do. You'll be thinking about *what* you should do.
There are three main statements he makes about 2035. Understanding these explains why this blog post is so profound. He says the intelligence of an AI model roughly equals the log of the resources used to train and run it; these resources are chiefly training compute. They are building this out—the Project Stargate $500 billion data center. People ask why invest $500 billion? It makes sense when you realize that this increases the intelligence of the model, not just brute force, but overall. You can spend arbitrary amounts of money and get continuous and predictable gains; the scaling laws are accurate over many orders of magnitude. This is important because they probably did some kind of test showcasing that the more compute you spend, the more intelligence you get. This is something everyone knows, but previously with a $3 million training run, they wanted to showcase that providing more compute power gets a better answer; the intelligence of the model goes up. They're stating this is true over many orders of magnitude. Maybe OpenAI tested this further. It's not feasible now, but provided that they know these models get smarter with more compute, this will change everything. Having that kind of intelligence is important to national security. If these scaling laws are accurate over many orders of magnitude, we're in the early stages of this intelligence era.
They talk about how the cost to use a given level of AI falls about 10 times every 12 months; lower prices lead to more use. This is relevant now and will change things in the future. 10 times cheaper every 12 months is incredible; you get frontier-level intelligence for 10 times cheaper every 12 months. What you can do next year will increase by 10 times in terms of USD value. The token cost from GPT-4 in early 2023 to GPT-4.0 in mid-2024 dropped about 150 times. Moore's Law changed the world twice every 18 months; this law is unbelievably stronger. In two years, we might use models as smart as GPT-3 mini for the price of GPT-4 now.
Rapidly declining price in AI means societal changes. If the price of AI becomes monumentally low, it's practically free. It's not just that the price is low; the price of intelligence has dropped to practically zero. Systems, companies, or businesses that used to require expensive intelligence (paying people $100K-$200K a year) will change. What happens when you can pay 100 people $100 a month? This rapidly declining price will be insane in 2035. If intelligence is abundant, we will have to see what society values most if it is not intelligence. This is a third- or fourth-order consequence most aren't paying attention to. It would be wise to learn how to use these intelligences now.
The socioeconomic value of linearly increasing intelligence is super-exponential. There's no reason for exponentially increasing investment to stop in the near future. For those who think we're in an AI bubble, I would argue that we are, and it doesn't matter. This happens once; the event horizon is completely different. While there have been past bubbles, this technology continues to build upon itself further. We don't know how far it can go. This is super-exponential. Companies are investing billions; they can cut marketing costs. Entire new industries are built; more AI investment follows, compounding continuously. There probably won't be any investment slowdown in the near future. The DeepSeek fiasco was more bullish for AI than bearish. People won't stop investing because AI will be one of the most powerful things across every area of society.
If these three observations continue to hold true, the impacts on society will be significant. With all three things happening, society will have to change. You don't get epoch-defining technology that doesn't change society. There will be a tipping point; society will demand change. I first realized this when I started AI research; if this continues, society will change. I want to be adaptable.
Another thing he speaks about is virtual coworkers. Right now they're not great, but they're doing a lot of work. In a few years, it'll be insane to see what they can do. Imagine a software engineering agent capable of doing most things a software engineer at a top company could do for tasks up to a couple of days long. It won't have the biggest new ideas and will require supervision, but imagine a thousand or a million of them in every field—healthcare, gaming. A million agents working overtime will create serious economic productivity. We're duplicating human intelligence by a thousand or a million, depending on compute. This will have serious implications for how much work gets done and how quickly. Most economies work from 9:00 a.m. to 5:00 p.m., but what if it's 24/7? While we sleep, AIs work faster, more efficiently, with no breaks or deadlines. How fast will the economy move?
The world won't change all at once, but the future will be coming at us in a way that's almost impossible to ignore. The long-term changes to our society will be huge. Correctly deciding what to do and navigating an ever-changing world will have huge value. It's always important to figure out what you do, not just do a bunch of stuff. Be plugged in to your industry and how AI impacts it. Those in this space have a tremendous advantage. The level of change coming will change things so much that society won't look the same in 10 years.
He talks about how the impact of AGI will be uneven. Society is already on a hyperdrive to where it will be so uneven that I don't know how we stop it. Remember how I spoke about how if the price of these models drops 10x every couple of months? The price of many goods will fall dramatically as the cost of intelligence and energy strains things. The price of luxury goods and limited resources like land may rise even more dramatically. The impact of AGI will be uneven, which isn't good for society. It will impact fields like software engineering, healthcare, and science a lot. The impact will be felt more in those tech-enabled, STEM areas. Later, I'll talk about how it will impact society and individuals unevenly.
They don't deploy insane models at once; they don't want people to freak out. The jump from GPT-3.5 to GPT-4 was severe; there was public backlash. People said they were moving too fast. A letter from the Future of Life Institute asked them to slow down. It was signed by Elon Musk and other notable AI figures. They're slowly dropping out models so society has time to adapt. At some point, we might not get certain models because they're too dangerous or too much for the average person to handle. With models like Sora, the difference was so much that the average person rejected it. They're trying to integrate it into society; initially, Sora was only let out to creative teams.
Sam Altman talks about open-sourcing more. I personally don't think OpenAI will open-source anything. I don't see what benefit OpenAI gets from open-sourcing technology, considering the AI race is more competitive than ever. The space is really competitive, even now with DeepSeek. I'm not sure how you do that with an advantage.
While they never want to be reckless, there will always be major decisions and limitations related to AGI safety that will be unpopular. As we get closer to achieving AI, trending towards individual empowerment is important. The other likely path is AI being used by authoritarian governments to control their population through mass surveillance and loss of autonomy. Most people either think it's a conspiracy theory or that it won't happen. It doesn't look good for those who believe it won't happen. We recently got information that billionaire Larry Ellison says an AI-fueled surveillance system will ensure citizens will be on their best behavior. This wasn't taken out of context. It's concerning, considering certain governments are already authoritarian. With AI surveillance, it will be impossible to have any privacy. The police will be on their best behavior because we're constantly recording everything; citizens will be on their best behavior because we're constantly recording everything. Cars have cameras; AI monitors the video. If an altercation occurs, the chief of police is immediately notified. It's AI looking at the cameras, not people. Drones can go out to situations way faster than police cars. There's no reason for high-speed chases. A drone follows the car. New generation autonomous drones are coming.
When we look at what people are saying and how society is changing, I think a lot of the technology developed will change how things are done in society. Drones are already being used in debatable ways, but drones being used to ensure public safety is something I can see happening. With advanced AI, this will be even more prevalent.
Ensuring the benefits of AGI are broadly distributed is important. AGI will be able to do a billion different things for a billion different people, but how do we ensure the benefits are distributed fairly? Sam Altman talks about Universal Basic Compute. People debated this, thinking it would be awful because AI overlord companies would determine how much compute we get and how we live our lives, like a universal basic income. The balance of power between capital and labor could easily get messed up and may require early intervention. Giving some compute budget to enable everyone on Earth to use a lot of AI is an idea. Relentlessly driving the cost of intelligence as low as possible also has the desired effect. This isn't concrete; we're still exploring ideas. If compute is tremendously valuable, does it make sense to give everyone a fair share? I wonder if the future looks more like Universal Basic Compute than Universal Basic Income. Everyone gets a slice of GPT-7 compute; they can resell it or donate it. You own part of the productivity.
What do you guys think about this? Is Universal Basic Compute the end-all solution? I think there will be a variety of solutions, but the solution will become clearer as time moves on. Hopefully, you enjoyed this video, and I'll see you in the next one.