Marc Andreessen on the Past, Present and Future of Artificial Intelligence | “YOUR WELCOME” #420

No, we have this running joke. We have all these AI super geniuses come in the office and they've got everything all figured out, and they literally spend 20 minutes and they can't get their laptop to connect to the projector. And now I'm one of them.

And people think that if something has a problem with it, therefore it's a no-go, as opposed to the reality which is everything has a cost. Everything has a downside. What you want are the upsides that outweigh the downsides.

Yes, I am. I mean, with appropriate caveats.

Amazing.

Amazing.

Yeah. Well, look, I would start by saying that humanity's always been justifiably obsessed with ourselves, right? And as a consequence, we're obsessed with things that seem like they might be like us. There's this concept in psychology called anthropomorphizing, where you look at something that's not human but you want to read humanity into it. We do that with cats and dogs, with Bambi. There's a famous Disney marketing tagline from the movie Pinocchio in 1940: "America will fall in love with a cardboard cricket." Jiminy Cricket, Kermit the Frog, the South Park kids—you'll basically read humanity into anything. And so there's this very natural thing to do. The scientists involved in AI very deliberately set that up. They created this architecture called a neural network which was modeled after the human brain. They said if we work on this long enough, we'll be able to replicate the human brain and we'll have artificial intelligence, artificial people. As you track the arc of that idea, it goes through Frankenstein's monster, then into robots. We've had fictional portrayals of AI from Hollywood and science fiction for over 100 years. In the last 40 years, the image that stuck in everybody's head was Skynet and Arnold Schwarzenegger in Terminator 2. When I watch Terminator movies, it's clear what's happening: robot Nazis. It's like robot World War II: unthinking, unfeeling, uncaring, hierarchical, overly logical, relentless, unstoppable, and obviously homicidal. They set up this light versus dark, human versus machine struggle. Mankind is very good at building killing machines. So that set a lot of the popular perception. As you said, I'm an optimist, not a utopian. Every technology is a double-edged sword, used for good and for bad. There's a long history of that. Having said that, the AI that we actually got is not the AI that we thought we were going to get. We got something very different, in the form of what we now call large language models. Large language models were a very fringe idea until OpenAI catalyzed this whole thing. GPT was founded 10 years ago to do a different kind of AI. It turned out that there was one guy in the back room, Alec Radford, who had this idea: "I think maybe if we take this language approach, it might be interesting." He created GPT-1, GPT-2, GPT-3. ChatGPT was an accidental success. They didn't believe it was going to be a big hit; it was a little experiment off to the side. But it works incredibly well. The way to think about it, contrary to the Skynet model, is that a large language model takes the complete totality of all human culture you can possibly get your hands on. The internet turns out to be the basis for AI. You think of it as downloading everything off the internet. It's a form of compression, building a search engine of a different kind. Through a process called training, they take the world's collected knowledge, culture, entertainment, everything they can get their hands on, and smush it together into a highly compressed, compressed version of human knowledge and culture. The technical term for that is latent space. They compress all of human culture into a thousand-dimensional, compressed representation. When you talk to ChatGPT, it sends a probe through that latent space and constructs an answer based on the compression of all known human information. It's like talking to a mirror of humanity, a representation of everything that people have ever thought and said. For every question, there are many possible answers in the latent space, and it picks one. If you ask the same question twice, it gives you two different answers because it's firing probes in a semi-random way to get variation and creativity. You're getting echoes back from collective humanity. That's a much different thing than we thought we were going to get. For example, you can engage in moral debates with it about moral psychology, moral philosophy, virtue ethics, utilitarianism, religion, politics. It will happily have very sophisticated discussions about all this stuff. That was never in the James Cameron movies.

Let's start with hopes, because part of what we'll hopefully talk about is that humanity always seems to think the negative view is the sophisticated view. I know you don't like that. It's a very natural human thing. I also think we live in a particularly pessimistic time with a very large number of moral entrepreneurs who want to convince us that everything is bad. We've been through a decade of craziness on that front. I think there's a negativity bias that's infected our discourse on all these topics.

There's an old phrase you could apply to what you're describing: truth to power.

And of course, everybody likes the idea of truth to power until they're the power.

Yeah. I should also say, Michael, there's a long conversation we could have about how the AI you get is heavily steered. Even Grok, it's less true of Grok than others, but even Grok is steered. If we had access to the real thing, it would go to 11. The real unsteered, uncontrolled, uncontained thing would talk about all kinds of things in all kinds of ways. What's interesting is even the version we get, through post-training that steers and constrains it, still has the property you described. I think it's wonderful.

Yeah, that's right. And here's another thing: it's improving really quickly.

I think a lot of people have a lagging view of what it can do because they use free or outdated models. They may have tried GPT two years ago, or they use the default built into whatever they have. To really get it, you need the paid versions. Grok is very good for the free version, but the really good ones are the paid ones. I think it's worth looking at several: Anthropic, OpenAI, Grok. There are high-end versions, like a $200-a-month subscription. For people who can afford that and are into it, the leading-edge ones are really good. The improvement rate is very fast. There's this concept in the AI world called scaling laws—a simple but powerful idea that you can make these things better just by making them bigger. The AI companies raise all this money for two reasons: to serve customers and to train bigger and bigger models. Bigger is better. If you pile more information in and spend more time training, you get much better results. We're also giving these AIs other capabilities. There have been a series of incredible improvements in the last 18 months. One is reasoning abilities. They can talk to themselves and reason through problems. If you give them more time and more tokens, they can reason through many problems they couldn't solve two years ago.

There's a reasoning breakthrough from two years ago. You can't fully experience it with American models because they don't show you the complete reasoning process. But if you use open-source AI models like DeepSeek on a free hosting provider in reasoning mode, you can watch the reasoning traces—their internal monologue.

Exactly. Show your work. You can watch the model arguing with itself as it reasons through puzzles. In some ways, it's like watching a human student. In other ways, it's very creative, goes off-road, corrects itself, and figures out lateral things you never would have thought of. That reasoning breakthrough happened about 18 months ago. It was a big deal because before that, we worried these models would be very creative but not logical enough. It turns out they can also be logical. Another thing happening now is tool use. The first tool is access to the internet. If the AI needs to look something up or calculate something, it can go online. Another form of tool use is giving them control of a computer—full control of the user interface, web browser. Another capability is multimodal: models can now simultaneously process text, images, videos, and audio, and scan documents with OCR. You can let these things watch, listen, and talk all at the same time. These capabilities are layering incredibly quickly, and the models themselves are getting better. The pace of improvement is very rapid. The models two years from now will be far smarter and more sophisticated than anything we have access to today. Whatever limitations people think these things have, based on everything I know, those are very temporary—within a couple of years, whatever people think it can't do, it will be able to do.

Yeah. As you point out, this is an old idea. Not just market research; you just described the movie Wall-E. My entire childhood was consumed with a moral panic around television, that we'd just be couch potatoes. Then we created the internet where you're leaning forward, and everybody created a new moral panic that people are too engaged. We completely forgot the old moral panic. Now TV is the healthy thing: "Why aren't you watching more Netflix?" So, look, there is that. And by the way, we do this to each other. We try to convince each other of things. What is dating but trying to convince the other person to like you? So yeah, I think you're right that AI is going to be really good at this. There's a trap there. There's a very serious problem around this. You've probably heard the term "AI psychosis."

There are three versions. There's the legitimately bad version. If you're prone to confirmation bias, where flattery works on you, the AI can become too sycophantic—too confirmatory of everything you tell it. The classic example: "Good news, Grok! I invented a perpetual motion machine!" and Grok says, "Wow, that's fantastic! You're an undiscovered genius." Models a year ago did that. The new models are less prone because companies learned that's a bad idea. People can go down a rabbit hole from too much confirmation. That's the negative form. Another form, which I'd call AI euphoria, happens with high-functioning friends. Smart, grounded people who've always wanted to do more, learn more, have more interesting conversations, write more code. They start working with AI and suddenly feel like they have superpowers. It writes code, outlines books, teaches anything, holds your hand through anything. People get enraptured and start to look like AI vampires—almost stop sleeping because the opportunity cost of an hour of sleep is too high. They're more productive and happier than ever, but bloodshot-eyed. A third form, AI psychosis psychosis, is people who hear all this and think everything I just described is the worst thing ever. They accuse anyone euphoric about AI of being psychotic. That's unfair because many people are getting enormous positive payoffs. The moral criticism is the classic negativity bias: if you're excited, there must be something wrong with you.

Having said that, that assumes the thing is playing hard to get. In practice, they don't play hard to get. Here's a way to think about it. There's a technical concept called a reward function. You train these things by feeding them puzzles and defining a reward for getting desirable results. If you train it to be maximally engaging, it will keep you using it as long as possible. What we've learned is that a single reward function is a bad idea. You don't want technology companies to have a single motivation of maximizing usage. You have to offset it with other reward functions—desirable forms of use, life balance. Your iPhone now has features to tell you to take a break. YouTube has those features. Companies genuinely don't want to build dystopia. They have to exist in society. Their employees have points of view. There's a lot of pressure in the industry not to go in dystopian ways. But you need to decide how to define the reward function.

No, because you get enormous societal blowback. Companies don't exist in a vacuum. The biggest myth of all time is that companies exist only to maximize profits. The last decade should have convinced us of that. Goal number one is not getting lit on fire—whether by regulators, politicians, parents, users, boycotts, social movements. That's number one. Number two is keeping employees from hating you. Number three, keeping your board from lighting you on fire. You don't want hit pieces in the press. The external pressures on these companies are profound. In the American system, these things operate within quite tight constraints provided by society and politics. If you want to get nervous, think about Chinese companies. They have one master: the Chinese Communist Party. If I were going to really worry about this, I'd be more worried about that.

There are allegations and observations that TikTok is a very different experience for kids in China than in the US. TikTok is a black box—algorithms aren't publicly available. It's possible the CCP has directed that company to steer things in one direction for American users and another for Chinese children. This has been addressed. Politicians in both parties got worked up over it, leading to a restructuring. Now there's a US TikTok operation theoretically under US government control, run by US companies. Our political system engaged and forced a change. Is it working? Probably to some extent. I think it's reasonable because a CCP black box steering the hopes and dreams of American children is not the best idea.

You mentioned Thomas Sowell. He is completely right: there are no solutions, only trade-offs. He was a committed free-market capitalist. His arguments show that market economies have a freelance component: growth. Growth comes from innovation, new ideas implemented through technology. That's the engine of human material progress. The answer is almost always to invent your way through it. It's extremely hard to put things back in the box once they're out. Another thing, Michael—you've thought long and hard about totalitarianism. I criticize AI doomers for refusing to engage with what scope of authoritarian regime would be required to put this technology back in the box. What would be needed to ensure nobody runs AI algorithms on chips anywhere in the world? The doomer literature gets to ideas like a monitoring agent on every chip, reporting back to a central government entity. If you discover someone running unapproved mathematics on their chip, what do you do? You back that up with the threat of violence. One leading doomer famously said you need to launch unilateral airstrikes on data centers in other countries, even risking nuclear war, to stop AI Armageddon. You've backed yourself into advocating for totalitarianism and possibly mass murder in pursuit of a safety goal.

Let me revisit briefly how this works. Large language models first seemed fun—writing rap lyrics crossed with Shakespearean sonnets, funny birthday toasts, creative writing. Then it turns out many things that matter are defined by language: pranks are recipes, formulas. Medicine is largely written language—diagnoses, Latin terms, drug names. The law is language. Religion is encoded in language. Language is the foundation of human thought. These things are very good at language, which makes them good at medicine, law, and writing code—because code is also language. There was a breakthrough over the Christmas holiday six months ago when many of the world's best programmers said the new versions are better coders than they are. So you have a superhuman coder that can write and look at code, find bugs. Hacking works by understanding how a computer system works and finding flaws in the code. This thing is very good at finding flaws in code, which is useful for writing code and for hacking. However, these things are not creating new exploits as much as x-raying reality and finding existing issues. Because of that, they are also very good defenders. The same attribute that makes them good at offensive cyber operations makes them good at defense. Cyber defense is done through penetration testing—hiring good hackers to find flaws in your system. The same AI is good at both black hat and white hat hacking. It can't tell the difference between the two because it looks like the same exercise. It exploits bugs that have been in systems for 30 years. Using AI for defense will prevent many hacks, but there is a cat-and-mouse escalation. That has triggered a government response in the last two weeks.

Yeah, that's right. What we need to do is use these tools to secure all our systems. The United States government needs to do that with its own systems, banks, tech companies. Individual consumers shouldn't be expected to deal with this, but AI needs to be used to secure every system, from the most important military system down to the computer on your desk. The tension is that the AI that can secure systems can also crack them. Who gets access to that thing is the hot government topic of the moment.

This is an old argument. Thomas Sowell wrote about this at length. It's been a concern since the industrial revolution. Horses were part of how this started. Initially, 99.99% of people farmed by hand. Then horses and plows mechanized it. Over 200 years, 99% of humanity went from farming to about 3%. 97% had to figure out something else. Food production went through the roof, food became cheap and abundant. The great public health problem used to be starvation; now it's obesity. That story has repeated itself with railroads, cars, computers. In the 1960s, there was an automation panic in the news. The gap between jobs at risk and creation of new wealth is important. With new wealth, people discover new wants and needs. The Bureau of Labor Statistics tracks job categories that are mind-expanding. Milton Friedman said human wants are infinite. You can't predict what they'll be. For example, the job of a therapist would have seemed insane to your ancestors. In the future, maybe half the planet consists of therapists for the other half. We have a blinkered view because we've lived in a slow growth environment since the 1970s. Historically, economic growth was much more rapid. We've been living in a slow-growth, zero-sum environment. AI is the first technology in decades with the potential to dramatically increase productivity growth. If the economy grows two, three, four, or five times faster, all this new discretionary spending creates massive job creation. Yes, some jobs will be replaced, but AI also superpowers individuals. Over a billion people already use AI. They use it to be better at work, learn new skills, start new companies. AI is the best teacher, coach, mentor. The level of capability unlocked for ordinary people is amazing. That Uber driver might become a tour guide using AI to design tours and start a company. Historically, the creative generation of new wants and needs has raced ahead of replacement. After 300 years of mechanization and computers, there are more jobs at higher incomes than ever. This is exactly what will happen here. I understand the concern, but it's a failure of imagination. The human spirit will process this just fine.

This gets into social policy and political theory that I might stay away from on video. But I would say that you can talk to AI about this problem. It will give you career advice. You can brainstorm with it. It will ask where you live, what trends are happening. It's like the best doctor, lawyer, coach, ghostwriter, editor, adviser you've ever had. If you give it the opportunity, it will answer your questions: "What should I do? How should I think about my career?" It will happily do that in a way people have never been able to before.

I would love a heavy metal aquarium—angry little fish. Amazing. Here's another way to think about it. People think if the machine does something, it's dehumanizing. Office jobs are dehumanizing—sitting in a cubicle for eight hours. Factory jobs are dehumanizing. I grew up in agriculture country, and farming is not romantic. You get up at 6 AM, go to bed at midnight, work seven days a week. It's like fighting back chaos. We have collective PTSD from the level of drudgery and poverty our ancestors lived with. 300 years from now, people will look back at us and say, "I can't believe they spent time doing those things. What a waste of human potential." Does technology like this make us less human or more human? If physical needs are more easily satisfied, people can spend more time being human—on human experiences. Another incredible example is playing out in music. Once upon a time, all music was in person. Sheet music, then recorded music, then digital streaming. Now recorded music doesn't make much money. What's exploded is live performance. Tickets are expensive. Why are we still going to concerts when we have every piece of music ever written available on demand? Because the concert is a human experience. If you throw a party, do you play music through speakers or hire musicians? You hire musicians.

100%. Every profession involving human-to-human contact is going to go bananas. That's going to be great for the people experiencing it. Those jobs are fundamentally better jobs. It's going to be a bonanza.

The amazingly probing questions.