AI's Spring Hasn't Come Yet

Notes on Marc Andreessen's interview on Lenny's Podcast

Today a few of us at the office huddled up and talked again. Everyone is shouting about AX (AI transformation), and we're helping enterprises with their own AX. But are we actually an AI-native company? Early this year, for a whole month, every employee spent every Thursday working in Claude Code and pushing hard to build AI into how we work. Each person's AI skills clearly improved, and fast. After that, we decided internally that we'd hit the goal, and we stopped those sessions. To be fair, thanks to the habits we built back then, everyone on the team can now push commits to Git, run their own projects without a developer, build an MVP, and get feedback from users.

But there's something beyond that. The real challenge is making the whole organization AI-native. We also had an enterprise AX seminar coming up soon, and the most effective thing to share there is our own wins and failures, more than case studies from overseas consulting firms. So we started by asking ourselves again whether we had truly made AI part of how we work.

Since January we've used AI aggressively. We spend tokens freely and we've adopted a bunch of AI tools. But if you ask whether that produced visible, measurable results, it was hard to see. Maybe because we weren't tracking it, or maybe because we're still in the middle of the process.

Looking at the outcome, as I said, each person's AI ability and fluency went up, but we haven't yet worked that into the company as a whole. To put it bluntly: if we set a new goal where each person on our 13-person team does the work of ten, producing what a 130-person company would produce, we could aim for more than automating and scaling individual work. We could aim to improve how everyone communicates and collaborates too. A company only means something when 1 + 1 equals 5, not 2. And in the AI era, I think it isn't only about teammates collaborating. Every job description and role needs to be redefined, and the way the company makes decisions and runs meetings needs another look. We're still discussing that part.

In my own case, I run global growth for the company, and normally I could barely cover the US and Singapore. With AI I'm now reaching out to expand into India and Japan too. Even that hasn't been easy. But if I had to do the work of ten people instead of two or three, I think I'd change my whole way of thinking. If I'd been told from the start that I also had to cover five countries in Europe, five in Africa, and five in South America, my action items would have looked completely different.

What I told people at work, which I also wrote about in my Claude Blue post, was this: if we think there isn't much of a bottleneck because we're using AI well, we should assume the opposite. It means we aren't driving AI hard enough for me, the human, to become the bottleneck.

Anyway, to keep up with the latest AI trends in Silicon Valley, I try to listen to US startup podcasts whenever I'm walking around or working out on weekends. One of them was a16z's Marc Andreessen on Lenny's Podcast. With his diction and his charisma, he gives off the energy of a very scary school principal (that's a compliment). The interview runs almost two hours. I got through more than half of it, then it felt too long and I asked AI to summarize it. Every sentence turned out to be important, and it lined up so closely with what we'd talked about today, AI-native organizations, the future, and work, that I had to write it up.

So let's skip the preamble and get straight into the interview.

Marc Andreessen: The real AI boom hasn't even started yet

On raising kids, why job loss fears are overblown, the future of PM/eng/design careers, and the macro force you should pay attention to

https://www.lennysnewsletter.com/p/marc-andreessen-the-real-ai-boom

The AI summary (I strongly recommend reading the full write-up below too)

1. AI and population decline: a "miraculous" timing

Without AI we'd be heading into an economic crisis. Technology has moved slowly for 50 years, and population growth is slowing too.

AI and robots showed up exactly when we need them. The human workers who remain will be at a premium, not a discount.

AI is a "philosopher's stone," a technology that turns the most common thing in the world (sand, silicon) into the rarest thing (thought).

2. Education and super-empowered individuals in the AI era

The "super-empowered individual": people who use AI well don't get 2x more productive, they jump to 10x or 100x.

The best coders are already experiencing this: "Suddenly I'm 10 times better than I used to be."

The key to raising kids: go deep in some field, and put yourself in a position to use AI's power to the fullest.

Deep expertise plus the ability to use AI matters more than being a T-shaped person.

3. AI's effect on employment: gradual, but positive

Look at past technology revolutions (ATMs led to more bank tellers, automation reshaped manufacturing jobs) and technology has almost never wiped out occupations.

Tasks change, but the job itself survives.

Even if AI advances quickly, regulation, organizational inertia, and legal liability make real adoption gradual.

Regulated professions like doctors and lawyers will take decades before "AI replaces" them.

Key point: humans will still take on oversight, judgment, and accountability.

4. Should you learn to code?

Absolutely yes. Coding is still a valuable skill in the AI era.

Even if AI writes the code, you need the ability to judge what good code is.

Leave AI on "autopilot" and you get mediocre results. Use AI while understanding things deeply yourself and you get brilliant results.

Programmers become 10 to 1,000 times more productive. They don't become unnecessary.

5. Three stages of AI-native startups

Stage 1, redefine the product: add AI to an existing product, or create an entirely new product category.

Stage 2, redefine the organization: the "one founder + AI = an entire company" model. No engineers, designers, or marketers; the founder handles every role alone with AI.

Stage 3, AI runs the company: the founder runs a "company" made of AI agents. Start with zero human employees and hire only when needed.

The founders at the very front are already experimenting with Stage 3.

6. How VC investment criteria are changing

Traditional VC: "Is this market big enough? What's the TAM?"

a16z today: almost no TAM analysis, because AI makes the size of every market impossible to predict.

Instead the focus is: "Is this founder ambitious enough to change the world, and can they execute?"

Peter Thiel's "determinate optimism," the attitude of designing and building the future, is key.

They invest in people who build the future concretely, rather than "indeterminate optimism" (vaguely believing the future will get better).

7. Software is still eating the world

In 2011, when he said "software is eating the world," software was about 5% of GDP.

It's still around 5% today. 95% still hasn't become software.

AI may finally turn the other 95% into software, so the real boom is only now starting.

8. Burnout and how he thinks about work

Andreessen himself doesn't experience burnout, "because the work itself is fun."

Burnout usually happens when you force yourself to do work you don't want to do.

Finding meaning and fun in the work itself is more sustainable than work-life balance.

Advice: find work you truly enjoy, and the burnout problem is solved at the root.

Takeaways

For founders: push the one-person operating model with AI as far as it goes, but deep expertise (domain knowledge) is what sets you apart.

For product builders: go beyond adding AI to an existing product (Stage 1) and consider redesigning the organization itself around AI (Stages 2 and 3).

For your career: rely on AI tools, but don't leave them on "autopilot." People who use them with deep understanding become super-empowered individuals.

For investors and business: a founder's determinate optimism and speed of execution matter more as evaluation criteria than TAM.

1. The philosopher's stone: turning sand into thought

Marc Andreessen invented the web browser and founded a16z, the largest venture capital firm in the world. In 2011 he put out his famous thesis, "Software is eating the world," and said there would be 5 billion smartphone users within ten years. In reality there were 6 billion. He admits he gets plenty wrong, but he also gets a lot right.

In this interview Andreessen called AI a "philosopher's stone." The heart of the alchemy Newton was obsessed with his whole life was turning something common (lead) into something precious (gold), and Newton never pulled it off. Andreessen said AI is doing exactly that. It's a technology that turns the most common thing in the world, sand (silicon), into the rarest thing in the world, thought.

He said 2025 was the most exciting year of his career and his life, and that 2026 will top it. This is a man who built the web browser, lived through the dot-com bubble, and watched the smartphone revolution, saying that now is the most historic moment.

2. An economy stalled for 50 years, and a miraculous timing

Andreessen pointed out something deeply counterintuitive. For the past 50 years we felt we were living through massive technological change, but the actual statistics show that technological progress in the economy was very slow.

Economists measure technology's economic impact through "productivity growth." US productivity growth has fallen to half of what it was from 1940 to 1970, and a third of what it was from 1870 to 1940. It feels like technology is changing fast, but its actual effect on the economy has slowed down.

On top of that comes population decline. In many countries, including the US, birth rates are falling below replacement level (2.0), and China is the same. A scenario where humanity actually shrinks over the next 100 years is becoming real.

Here's Andreessen's core point: "If we didn't have AI, we'd be in a panic right now about what's going to happen to the economy." If the population shrinks and technological progress is also slow, the economy can only contract. And right at that moment, AI showed up. Andreessen described the timing as having "lined up miraculously well." We got AI and robots exactly when we needed them, and the human workers who remain will be at a premium, not a discount.

He also brought up his famous debate with Peter Thiel. Back then Andreessen argued that "technological progress continues," and Thiel said, "No, it's stopped." Andreessen said that now he largely agrees with Thiel's view. What Thiel meant was that "there's been progress in bits, but almost none in atoms." Look around and you'll find buildings from the 1960s, bridges from the 1930s, and dams from the 1910s still in use. Where are the new cities? Whatever happened to California's high-speed rail? Progress in the physical world clearly stalled.

3. Task loss, not job loss

Andreessen gave a clear rebuttal to the fear that AI will wipe out jobs.

"Everybody wants to talk about job loss, but what you really need to look at is task loss." In economics, a job is a bundle of tasks. When technology changes, the tasks change. The job itself doesn't disappear.

His example was very intuitive. In the 1970s, executives didn't use typewriters or computers themselves. They dictated memos to a secretary, who typed them up and sent them. When email arrived, the secretary's tasks changed. The secretary would receive the email and print it out, the executive would scribble a reply on the paper, and the secretary would type it up again and send it. Today executives write their own email. The secretary's job didn't disappear, but the tasks changed completely. Secretaries now do other work, like planning business trips and coordinating events. If anything, the executive's tasks expanded, and they now do office work they never would have done before.

The same pattern will repeat in the AI era. The tasks of coding, PM, and design will all change. Jobs outlive their tasks.

Even if AI triples productivity, that's the same rate of economic change as 1870 to 1930. If you read documents from that era, people felt the world was full of opportunity. Kids built new careers in new sectors of the economy. Most of the modern world we live in was invented and spread during exactly that period.

4. A Mexican standoff between PMs, engineers, and designers

PMs, engineers, and designers. Andreessen compared these three core roles at tech companies to a "Mexican standoff," the scene in a John Woo movie where three people point guns at each other.

Here's what that looks like. Every coder now thinks they can do PM and design with AI. Every PM thinks they can code and design with AI. Every designer thinks they can code and do PM with AI. And the real irony is that all three roles will figure out that "AI is also better at the manager's job." The guns start pointing up the org chart too.

What Andreessen found interesting is that they're all actually right. AI is now a pretty good coder, a pretty good designer, and a pretty good PM. At the very least, it can handle many of the tasks in each role.

He said the exact same Mexican standoff is playing out in Hollywood. The director thinks, "AI writes scripts and acts, so I don't need writers or actors." The writer thinks, "AI directs and acts, so I don't need directors or actors." The actor thinks, "AI writes scripts and directs, so all you need is me." Tech and Hollywood are being reshaped into the same three-way setup.

5. The super-empowered individual: 100x, not 10x

He said AI affects people on two levels.

First, it makes decent people very good. Someone who can write gets much better at writing with AI, and someone who can code gets much better at coding with AI. It raises the overall average.

Second, it makes truly excellent people phenomenally good. Andreessen said his friends who are really great coders are already experiencing this. They're saying, "Suddenly I'm not 2 times better than I used to be, I'm 10 times better." That's what he calls the "super-empowered individual."

He also mentioned Scott Adams's well-known career advice. Adams, the creator of the Dilbert comic strip, put it like this: "I could have been a pretty good cartoonist, and I could have been a pretty good businessman. But because I was a cartoonist who understood business, I could create Dilbert." The best cartoonist in the world couldn't have drawn Dilbert without understanding business, and the best businessman in the world couldn't have made Dilbert without being able to draw.

Andreessen took this further. If you're good at two things, the effect is more than double. If you're good at three things, it's more than triple, because you become a top expert at the combination of fields. In Hollywood, people who both write and direct are called "auteurs," and they're the creators who really move the industry.

Larry Summers apparently put it in similar terms: "Don't be fungible." If you're only a coder or only a designer, you're a replaceable part. But someone who codes, designs, and does PM can't be replaced.

When Lenny heard this and asked, "So, T-shaped (deep in one area, broad in the rest)?", Andreessen went a step further. He said "an E turned on its side": one deep area of expertise with two or three horizontal skills attached. And AI is exactly what makes that horizontal expansion possible.

6. Should you learn to code? From calculators to AI coding

Asked whether people still need to learn to code in the AI era, Andreessen answered firmly: "Absolutely."

The history of coding he walked through was striking. He asked if you know what the word "calculator" originally meant. It used to refer to a person. Back before electronic calculators or computers, when an insurance company or the military needed math done, they'd seat hundreds or thousands of people in a big room and have them do the calculations by hand. Those people's job title was "calculator."

From there, the first computers were programmed in machine code, ones and zeros. Then assembly language came along, and when Andreessen was growing up, C was at the center. After that came scripting languages like JavaScript and Python. Even then there was a debate that "scripting isn't real coding." Real programmers compiled code in C and managed memory themselves; scripting was cheating. The answer, of course, was "No, scripting is real coding too," and today most coding happens in scripting languages.

AI coding fits the same pattern. Just as scripting languages abstracted away five layers of detail beneath them, AI coding abstracts away the process of writing the scripting code itself. The best programmers in the world already have 10 browsers open running 10 AI coding bots at once, moving from bot to bot to review results, debug, and change direction. The programmer's job has shifted from "writing code yourself" to "arguing with AI bots."

But if you don't know how to write code yourself, you can't evaluate what the AI puts out. That's why you still need to understand assembly and machine code. If you just hand things off to AI coding, you get an infinite amount of mediocre code. But if you use AI while understanding things deeply, you get brilliant results. That's where the difference comes from.

He also talked about his 10-year-old son, who is completely hooked on vibe coding in Replit. At dinner he'll spend two hours arguing with the AI while building games. Andreessen said he tells his son: "It's great that the AI writes the code for you, but you need to understand for yourself why that code works and why it doesn't." It's the same principle as someone who uses a scripting language still needing to understand how a microprocessor works.

7. The future of education: Bloom's 2 Sigma and AI tutors

Andreessen homeschools his son, and his view on education was clear.

There's something education research has known for centuries: the ideal way to educate an individual is one-on-one tutoring. There's a phenomenon called the "Bloom's 2 Sigma effect," where one-on-one tutoring raises a student's performance by two standard deviations, taking a 50th-percentile student to the 99th percentile. Every royal and aristocratic family in history knew this. Alexander the Great was tutored privately by Aristotle.

The problem was economics. Only the wealthiest people in society could afford one-on-one tutoring. Now AI can bring it to everyone. When a kid is curious about something, they can ask an LLM questions endlessly and get instant feedback. Say "teach me this" and it teaches you. Say "explain it more simply" and it lowers the level. Say "now quiz me" and it tests you. This is possible right now.

Andreessen suggested a hybrid approach for parents: kids still go through the traditional education system, and AI tutoring fills in. A new private school called Alpha is apparently already running on this philosophy. It has in-person classes and teachers, but AI and AI tutoring are a core pillar.

8. AI can do everything? Then have it teach you too

This was one of the points Andreessen stressed most. The reason people don't get enough out of AI is that they focus only on having AI do "work."

"Being taught by AI" is just as powerful. "I'm a coder, teach me how to be a PM." AI will happily do that. It gives you assignments, evaluates your results, and points out where you fall short. He said that's as powerful as AI doing work for you.

"Anybody who really wants to advance their career should, in my view, be spending all their free time at this point talking to AI and saying, 'Train me, super-empower me.'"

Lenny shared some tips too. One is to simply watch an AI agent work. Even if you're not an engineer, observing how the agent thinks and makes decisions teaches you architecture naturally. Another is that when you get stuck on a problem and then solve it, ask the AI, "What could I have said differently to avoid this error in the first place?"

9. The three stages of an AI-native startup

Andreessen said the most forward-looking founders are thinking on three levels.

Stage 1 is redefining the product. Is AI a feature added to an existing product, or does it overturn the product category itself? For example, Adobe Photoshop adding AI image editing is one thing. Not editing images at all and generating new ones with AI is another. If generating is easier than editing, the reason for the 40-year-old Photoshop franchise to exist starts to wobble.

Stage 2 is redefining the organization. If a company that needed 100 coders can get the same results with 10 coders super-empowered by AI, what do you do? Or do you keep all 100 and get 10 times the output? That's exactly the question our company was discussing today.

Stage 3 is redefining the company itself. Founders have started imagining a company where one person commands an army of AI bots and does everything. There's a holy grail the industry has chased for a long time: "the one-person billion-dollar company." Bitcoin's Satoshi Nakamoto pulled it off. Ethereum did it with a very small team. Instagram and WhatsApp also achieved huge results with very small teams. In the AI era this is becoming more realistic. He said the most forward-looking founders are already experimenting with Stage 3.

10. Nobody knows where the moat is

On moats in AI, Andreessen was surprisingly humble.

In his experience, when a really big technology shift happens, people make predictions early on with enormous confidence. "This kind of company will win, that kind won't be able to adapt, the moat is here." The media naturally prefers definitive answers. But look back a few years later and almost all of those predictions were pretty badly wrong. He said if you reread the confident predictions about the internet from 1993 to 2010, nearly all of them were wrong.

He said you can look at the defensibility of AI models from both sides. From one side, it takes billions of dollars, scarce engineers, and massive compute, so it seems like two or three companies will dominate. From the other side, within a year and a half of ChatGPT's launch, five companies in the US, five in China, and open source all put out products at nearly the same level. Something that looked like magic got commoditized extremely fast. DeepSeek came out of a Chinese hedge fund, reimplemented the ideas of the big US labs, and added its own innovations on top.

He also brought up Claude Code and Cowork. Anthropic announced that it built Cowork with Claude Code in a week and a half, and Andreessen saw it two ways. One: "It's remarkable that AI built a product in a week and a half." The other: "How big a barrier to entry can something built in a week and a half have?" Every other model company will obviously build agents and products for everyday users. Six months from now, Claude Code could get overtaken the way GitHub Copilot did.

He said that if you have a drink or two with the smartest people in the industry and talk candidly, what you hear is that "there really aren't many true secrets among the big labs." They have the same information and the same knowledge and regularly leapfrog each other, but there isn't much that's proprietary.

Andreessen's conclusion was clear: "I don't think I can predict what the industry structure will look like in five years, which companies will win, or what the killer app will be. Being flexible and adaptive at this point is a much better use of time."

11. Determinate optimism vs. indeterminate optimism

On investment philosophy, Andreessen cited Peter Thiel's framework. Thiel proposed a 2x2 matrix: optimistic or pessimistic, and determinate or indeterminate. In Thiel's view, Silicon Valley has too much "indeterminate optimism," believing the world will get better without being able to explain why. What Thiel wants is "determinate optimism," the attitude of "the world will get better because I'm going to build this specific thing." Elon Musk is the classic example. Electric cars, solar, Mars. Very specific.

Andreessen firmly classified himself as an "indeterminate optimist," and said this is where he and Thiel part ways. As Andreessen sees it, indeterminate optimism isn't just wishful thinking. Silicon Valley is a system where there isn't just one determinate optimist like Elon, but 100, 1,000, 10,000 of them. Individual founders should be determinate optimists. But the people running the system above them should run as many experiments as possible and get as many smart people as possible trying interesting things.

He said history remembers founders, not seed investors. We remember Henry Ford. We don't remember the investor who put seed money into Ford and lost it on ten other car companies. He said 99.9999% of the credit should go to the founders, the determinate optimists.

12. AGI, and an IQ of 200

Andreessen separated two definitions of AGI.

The "cosmic definition" is the singularity: AI enters a loop of improving itself, and human judgment stops mattering. Andreessen said, "I don't think we're lucky enough, or unlucky enough, to live in that world."

The "practical definition" is the point where AI does every economically valuable task as well as a human. He said a co-founder of Anthropic narrowed this down to "a basket of the 10 to 15 most valuable economic tasks." By that standard we've already almost reached it, or will soon.

What Andreessen found more interesting is what lies beyond. Human IQ biologically tops out around 160. That's Einstein level. At 140 you're a world-class researcher or bestselling author, at 130 a really good lawyer, at 110 a good manager, and at 105 an accountant for a small business. The range where humans do intellectually impressive work is roughly 110 to 160.

But AI has no such biological ceiling. Current AI models test at around 130 to 140, and in math they're starting to approach 160. Andreessen said AI at 160, 180, 200, 250, and 300 is coming soon. "Is the world better or worse with more Einsteins? Better, obviously. With machines smarter than Einstein? Better, obviously."

He said the idea of "human equivalent" will just be a footnote. "Oh yeah, we passed that on some Tuesday in 2026." The question after that matters far more: what can we do in a world where we have machines smarter than humans at our fingertips?

13. Media diet: X and old books

Andreessen's media strategy was interesting too.

He called it "an almost perfect barbell strategy." On one end, he uses X (Twitter) to see what's happening right this moment. On the other end, he reads old books, 50 years old or more. They've passed the test of time, so he figures there's something timeless in them.

He said he's skeptical of everything in between, like newspapers and magazines. Reread last Friday's newspaper and you'll see that almost nothing it predicted came true. Magazines are worse. They come out weekly or monthly, so by the time an article runs, the information is already old.

What he thinks is still underrated is "content made directly by practitioners in the field." Substack newsletters and podcasts are exactly that kind of channel. He said there's huge value in someone who actually works in a field explaining it themselves. Sure, there's a bit of selling their own product in it, but fundamentally people want to explain what they do and want others to understand it.

Still a human bottleneck

To be honest, the podcast wasn't the only reason I wrote this post. What was weighing on me wasn't AI. It was global thinking.

I studied in the US too, but after living and working in Korea for a long time, I inevitably became more Korean. No matter how much I try to consume social media and YouTube in English, I speak Korean every day and I'm around Korean coworkers, so my way of thinking inevitably becomes Korean too.

If our company wants to become a unicorn and build a global product, I think that, simply put, we have to become a global organization. But realistically, everyone on our team is Korean. That gap was the biggest bottleneck I felt. Me included.

So last week I went to two meetups. One was an English coffee meetup run by an American founder friend from Silicon Valley who does business between Seoul, Tokyo, and Shanghai. We drink coffee together, but the topic is AI. Through this meetup I easily met global talent in Korea who are running businesses fairly aggressively in AI and speak fluent English. Since everyone came with the same mindset, even though it was our first meeting, the conversation was actually more relaxed and candid.

That evening I went to meet a friend of a friend who had worked as a PM at a publicly listed US financial company in Silicon Valley, recently quit, and now runs a community for PMs, hosting meetups in Korea, Singapore, Japan, and elsewhere. Walking through Gangnam Station, I was surprised at how many foreigners there were, and when I got to the event, I got to see Korean American friends from the tech industry for the first time in a while. There were also people from all sorts of countries beyond the US, like Brazil, Australia, and India.

In the end, here's what I want to say. Realistically, we can't be in the US. But at the moment when I felt that global-thinking bottleneck most strongly, I felt hope that we could build a global community and culture here in Korea that feels a lot like being in the US.

I believe AI can solve just about everything, but unless I plug AI into my brain, AI can't change the way I think. Then it hit me that I could improve that in a pretty easy way. Take part more actively in global communities, and, as Andreessen said, "spend all my free time having AI train me." Those two things were pointing in the same direction.

AI can super-empower me, but first I need the will and a direction to want to be super-empowered. And in the end, I find that direction among people.

Originally written in Korean on Brunch, April 7, 2026.