Everyone in the startup world knows Cluely. If you follow Silicon Valley at all, you’ve seen their stuff. The company raised $20 million from a16z within two months of launching, but honestly the funding isn’t even the impressive part. It’s how they got there. Their marketing language is a little annoying on purpose, the kind of content you want to hate but can’t stop watching. Some people call it ragebait. Fine. But they’re the the most talked-about AI startup on the planet right now, and everyone’s paying attention to what they do next.

Daniel Min designed that viral growth machine. I’d been following him for a while, so when I saw his Instagram story last Sunday asking what he should do during a one-week trip to Korea, I DM’d him immediately. Want to do a meetup with the Korean AI community? I just met with the Anthropic ambassador last week, we got 2,000 signups for the Claude community event, and our Discord has 1,100 members. I can set the whole thing up. His reply came fast. “Sounds good. Where?”

Of course I have a venue, I thought. Then I checked our office lounge, and it was fully booked. Three to four days’ notice and every event space in Seoul was taken. I was scrambling.

That’s when I remembered Jehim Choi. He runs one of the biggest marketing communities in Korea called “Reply Marketing” and works as marketing manager at t’order. He’d come to our first Claude Bloom event and we’d caught up then. I messaged him Monday night at 9 PM asking if t’order had any space for a talk. Jehim just straight-up DM’d his CEO. Didn’t ask permission, didn’t go through channels. He pitched it as a PR opportunity for t’order’s AI initiatives. And the CEO said yes on the spot.

The wild part? t’order’s office is in Parc1 with a full Han River view. Perfect for a fireside chat. So the scramble to find a backup venue actually landed us somewhere better than the original plan. Felt like a small miracle.

One last problem. Daniel only speaks English. Even if you tell a Korean audience the session will be in English, there’s going to be awkwardness. But a colleague at my company scored a free sponsorship from Tiro, an AI real-time interpretation service that’s been getting buzz lately. The Tiro team came in person, set up the system, and attendees said the live translation was excellent.

When things start falling into place, they really fall into place. So here’s what Daniel actually said during the fireside chat.

1. Wharton to Startup Life

Daniel Min graduated from Wharton in 2025 with concentrations in marketing and operations. On paper that sounds like a classic elite track. His actual path was anything but. He’d been making content since high school. Twitch streams, YouTube, TikTok. At 17, he did his first four-hour Twitch stream and made $500. That was the moment he knew content could be a real career.

But by senior year at Wharton, reality hit. He applied to Goldman Sachs, McKinsey, Bain. All of them. Got rejected everywhere. In the U.S., these companies are the equivalent of Samsung in Korea. The places where the elite go. After getting shut out of all of them, he felt like a loser. His words.

The only internship he landed was as an unpaid social media intern at RecruitU, a tiny startup. A Wharton senior doing unpaid social media work. He joked that it was too embarrassing to even tell his girlfriend about.

But that summer changed everything. “The only thing that was always clear to me was that I loved content,” Daniel said. “Even uncertain action on top of a clear vision gets you somewhere eventually.” The worst summer became the best summer. He started roaming the streets of New York interviewing finance people on camera for RecruitU. That content built 75,000 followers in four months.

2. Joining Cluely

The Cluely story started with one Instagram DM. Roy, the founder, had seen Daniel’s blog and content and reached out. They talked, and Daniel flew out to meet the team in person. “Cheat on everything” as a slogan was cool, sure. But that wasn’t why he joined.

His decision came down to one question: “Can I become best friends with these people?” The team was sharp, the work they’d already made was impressive, and he could picture hanging out with them every day in the office. Not salary, not title. Whether he’d actually want to spend his days with this group.

A Wharton grad jumping into an early-stage startup is not the typical path. Friends and family pushed back. But for Daniel, Cluely was a company that actually cared about content. And if he was going to leave his own thing behind, whatever came next had to be at least as fun. Cluely was exactly that.

3. The UGC Machine

The first thing Daniel built at Cluely was the UGC program. He’d watched other companies crush it with creator-driven viral content and figured that combining capital with viral know-how would produce something bigger. He hired 100 creators, gave them viral scripts, and set up a system to turn their videos into ads.

Early results were mixed. Some videos hit millions of views. But views don’t always mean conversions. There were clips racking up massive numbers with zero signups. Thousands of dollars burned on attention that went nowhere. Something needed to change.

The fix was cutting quantity for quality. He trimmed from 100 creators to 70, keeping only the ones who could actually drive conversions. Those 70 got more refined scripts and formats designed for action, not just eyeballs. At peak performance, this system was consistently producing 100 to 150 million views per month. Daniel’s conclusion was clear: hiring the right influencers is far more efficient than just running paid ads.

4. The Marketing Funnel

Daniel broke the marketing funnel into three layers, and each one had a completely different job.

Top of funnel is brand awareness. Cluely’s official Instagram, cinematic videos, office skits with Remy. No conversions expected here. It’s the Red Bull or Bain Capital approach. Content that has nothing to do with the product directly but plants the name in everyone’s head. His logic: if Mr. Beast has 400 million subscribers and launched Cluely tomorrow, he’d generate more revenue than any traditional marketing. That’s the bet.

Middle of funnel is UGC content. Videos with viral potential that simultaneously push people to try the product. The recently launched Bounty series had its first video generating over $10,000 in conversions.

Bottom of funnel is paid ads. The rule here is simple. If ROAS isn’t positive, kill it immediately. Money in, money out. No emotion, just numbers.

5. The Times Square Billboard

Cluely’s Times Square billboard is top-of-funnel taken to the extreme. A billboard company emailed a discounted offer for a four-week placement. Roy and Daniel discussed it for exactly ten minutes. Then spent $375,000.

The design was radically simple. White background, black text. “Hi, my name is Roy. I’m 21. Please buy my thing. Cluely.com.” It wasn’t a digital billboard either. It was physically fixed in Times Square for four straight weeks. A 21-year-old putting his name up in Times Square and asking people to buy his product. That alone became the content. People filmed TikToks of it. Twitter made parodies. It became a cultural moment.

Daniel’s explanation: too many companies spend excessive time validating whether a bet is right. If you’ve got the eye for a good bet, ten minutes is enough. They didn’t measure direct conversions from the billboard. But the Cluely name spread across the entire internet, and that was enough.

6. Views First, Conversions Later

The most common question from the audience was whether viral views actually convert to paying users. A lot of founders know Cluely but wonder about the gap between view counts going up and payments coming in.

Daniel had two answers. First, not every company needs viral marketing. A company called Resume AI has no Instagram account at all. They spend $100,000 a month on Google SEO and pull in $20 to $30 million ARR. Job seekers search “make a resume,” Resume AI shows up first, they pay $30 and get their resume. Figure out where your target users search before deciding on a channel.

Second, if you do choose organic content, there’s a sequence. Build the muscle for generating views first. Then design for conversions. Worrying about conversion rates on a video with 200 views is completely backwards. You need the confidence that at least 1 out of every 10 videos can guarantee 100,000 views before you even start thinking about conversion optimization. Too many companies obsess over conversions when they haven’t even built the basic muscle of getting attention.

7. “Create What Doesn’t Exist on the Internet”

Daniel’s personal content philosophy boils down to one line: make something that doesn’t exist on the internet yet. Everything online gets recycled constantly. But your personal experiences and the risks you’re willing to take can’t be copied.

His best example is the Alcatraz swim. He spent four days in San Francisco doing open-water swim training, then filmed himself swimming from near Alcatraz Island back to shore. Swimming directly to Alcatraz is illegal, but you can take a boat to a nearby island and swim back. After that video went out, people started recognizing him on the street asking if he really did it. Mindshare, earned.

Same logic behind his 40-university dining hall review series. He traveled to Boston, Texas, Philadelphia, New York, and California ranking college cafeterias in a tier list. Not one or two. Forty. Nobody’s going to replicate that. When Daniel gets a new content idea, he runs it through two filters. First, does it already exist on the internet? Second, is it hard to make? If it’s hard, nobody else will do it. That difficulty becomes the moat.

Vulnerability works the same way. A month ago he got a severe acne breakout with blood running down his face and filmed the whole thing. Most people hide that stuff. Showing it creates massive empathy. He ended up not posting it due to video quality issues. If the footage had been good though, it would’ve gone up for sure.

8. Ragebait, or Something Smarter

Is Cluely’s marketing intentional ragebait? Daniel said that’s only half right. The goal isn’t to make the internet as angry as possible. It’s to choose framing that makes people talk. His principle: “The opposite of hate isn’t support. It’s indifference.”

He gave a concrete example. Say “exercise is good for you” and nobody reacts. Say “overweight people need to exercise” and the comments explode. The core message is identical. Exercise is good. But different framing produces completely different engagement. “Cheat on everything” works the same way. The actual meaning is “use AI for everything.” But wrapping it in a provocative frame makes it stick. When calculators first came out, people called them cheating. Same with the internet. AI is following the same trajectory.

Cluely’s viral origin story fits this pattern too. Roy, the founder, uploaded a YouTube video of himself using Interview Coder (a tool he’d built) during a live Amazon coding interview. That video blew up. Interview Coder evolved into Cluely.

9. The RecruitU Playbook

The method Daniel used to hit 75,000 followers in four months at RecruitU is worth studying if you’re trying to grow a company account. He shared a three-step framework.

Step one: competitive analysis. He studied every content account in finance and consulting. Litquidity, Wall Street Oasis, channels making news, memes, guides. But one format was completely missing. IRL videos. Nobody was going out on the street to interview real finance people on camera. Individual creators made career advice clips, but no company brand was doing it at scale.

Step two: define your ideal viewer persona. Daniel got specific. “An 18-year-old Asian freshman who just started at the University of Alabama and wants to get into Goldman Sachs.” Every single video was made for that one person.

Step three: apply a viral format. Inspired by street interview channels like School of Hard Knocks and Caleb Simpson, he created a format where he’d approach finance professionals in New York and ask for career advice on camera. 90% of the people he approached said no. Asking busy bankers during lunch to stop and talk on camera is embarrassing and difficult. But that difficulty is exactly why nobody else was doing it. The barrier to entry was the moat. By the fourth video, results were already showing. Four months later, 75,000 followers.

This same framework worked at Cluely. After joining, Daniel created the Cluely Careers account using the same interview format and picked up 6,000 followers from just a few videos. Once the skit collaborations with Remy started, Cluely’s main account grew from around 50,000 to nearly 300,000.

10. Mints Media and What Comes Next

Daniel left Cluely after about seven months. Cluely is gunning to become a $10 billion company, and that kind of mission requires your entire life. For a 22-year-old who wants to travel the world, make videos with people he likes, and still make good money doing it, that level of all-in commitment wasn’t the right fit. He increasingly wanted to pour 110% into his own creative projects.

So he started Mints Media. It’s a marketing agency that works with tech startups lacking marketing muscle. He takes on only two to three clients at a time. The ideal client is a Series A to B company. Enough funding to execute but not so big that every decision needs five layers of approval.

For B2B marketing specifically, Daniel recommended heavy investment in LinkedIn and Twitter. In finance and enterprise, sales cycles are long and internal referrals are decisive. Follower count matters less than network quality. Target platforms where your potential customers are actually lurking, not just where engagement looks good. And for Korean companies trying to enter the U.S. market, the very first step is making English content. Marketing in Korean while trying to penetrate the American market is a non-starter.

11. “Go Film a Video Tonight”

The final topic was execution. Daniel was especially direct with the students in the audience. The word he used was “mental masturbation.” Watching 100 YouTube tutorials on how to grow a channel or boost TikTok views, then not making a single video for 60 days. Thinking about ideas and mistaking that for progress.

“If I’d been making videos consistently since high school, I’d be retired in Cancun by now.” Maybe an exaggeration, but the point was crystal clear. The most common question from beginners is “how do I start?” The answer is film one video tonight and post it. The moment you’re sitting in a room feeling inspired is when your motivation is at its peak. Tomorrow you’ll tell yourself you need more preparation, and then you’ll postpone forever.

I jumped in with my own story. Writing up a conversation I had with a friend in Silicon Valley as a blog post became the starting point for this entire community. If I hadn’t written that post, none of this would exist. Content should be made as early and as often as possible. Nobody knows where the next viral moment will come from.

12. The 30-Second Pitch Session

We tried something new at the end of the event. A 30-second pitch round. Since Daniel was filming, I told the audience to come up to the camera and introduce their company or project in 30 seconds. Free promotion on Daniel’s channel.

At most Korean meetups, when you say “go network freely,” what actually happens is people stand around awkwardly and quietly leave. But when you give them structure, “you have 30 seconds, Daniel’s camera is right there, free exposure,” people walked straight to the stage. A European founder building an AI marketing agent. A seven-country education startup helping high schoolers build AI companies. A service connecting 500 Gen Z creators across 50-plus American universities with beauty brands. A creator who’d hit 10 million views on Instagram with vibe-coding meme content. All kinds of people stepped up and told their stories.

Watching that session hit me. The entire event’s message was “execute right now,” and the 30-second pitch was that philosophy in action. Give people a structure and they’ll move. Definitely doing this again next time.

Everyone at my company talks about AI transformation. AX this, AI-native that. We help enterprises figure out their AI strategy. But the uncomfortable truth is we’re not sure we’ve figured out our own yet.

Earlier this year, we ran four straight weeks of Thursday all-hands sessions. The format was simple. Everyone brought a real project. Something they were actually stuck on. Then we spent the time building it together using Claude Code. No theory. No slides. Just shipping stuff.

The results were real. I watched people go from cautious to confident. The engineer who’d never used an AI tool before suddenly shipped a prototype solo. The designer started iterating with AI ten times faster than before. The operations person automated a process that used to eat half her week. The individual skill lift was visible and fast.

But then something weird happened. We all went back to our regular meetings. The projects shipped. The momentum faded. And we were still having the same conversations about AI strategy that we’d had before.

That’s when I realized the gap. Individual AI skills went way up. But organizational AI-native transformation? That hadn’t happened. We’d increased people’s capabilities. We hadn’t restructured how we actually work.

The 10x Goal That Changes Everything

I’ve got a goal that sounds crazy on paper. Thirteen people producing the output of one hundred thirty. Ten times per person.

On the surface, that’s just a scaling problem. Hire more people. Ship more stuff. But the question isn’t whether 10x output is theoretically possible. Of course it is. The real question is whether it requires a completely different operating model. A different way of thinking about what work even means.

If we’re going to hit 10x per person, we can’t just add AI tools to the existing structure. We’d need to rethink the entire organization from the ground up.

I lead growth globally. Right now I’m stretched across the US and Singapore. Every week I’m trying to push deeper into India and Japan. The number of contexts I’m holding in my head is already unsustainable. Markets work differently. Customer expectations are different. The competitive picture is completely different. Even staying sharp in two major regions requires mental gymnastics.

Now imagine someone told me tomorrow: cover five countries in Europe. And five more in South America. Don’t just add coverage. Build the same level of market intelligence. Develop the same network density. Actually understand these markets the way you’re starting to understand Asia.

I’d need to completely rewire my brain. New languages, new business models, new cultural frameworks. The mental overhead would be crushing. I’d be the bottleneck. Not the work. Not the tools. Me.

That’s the thing that got stuck in my head. I’d written it down somewhere before, but hearing Marc say it out loud made it click differently: “If you think you’re using AI well and you don’t see any bottlenecks, you’re probably not pushing the tool hard enough. You’re the bottleneck. Not the AI.”

Finding the Right Podcast at the Right Time

I’ve got this habit. Whenever I’m walking or exercising, I’m listening to US startup podcasts. Podcasts hit different when you’re in Seoul. You hear the actual conversation, the debate, the real thinking from people building at the frontier. No filter. No press release.

That’s how I found the Marc Andreessen episode on Lenny’s Podcast. Two hours. Lenny Rachitsky asking Marc everything that matters about AI right now. And Marc… the guy’s dictation is scary-impressive. The charisma is there. But it’s not flashy. It’s this quiet confidence mixed with deep pattern recognition. Someone said he reminded them of a terrifying principal. I get that. He’s the guy in the room who actually understands the system.

Every sentence felt load-bearing. I started trying to summarize it. Three minutes in, I realized that was pointless. I asked Claude to give me the summary. But that felt wrong too, because Claude’s summary was “every sentence is essential.” So I just listened to the whole thing again.

The Miraculous Timing Nobody Talks About

The thing Marc led with felt different from the usual AI discourse.

For the last fifty years, technology progress has actually been slow when you measure it economically. Productivity growth dropped to half of what we saw between 1940 and 1970. And before that, between 1870 and 1940, it was even faster. The trend line has been pointing down for decades.

Meanwhile, populations are declining everywhere. US fertility rate below replacement. China facing demographic collapse. Japan aging out. Most developed countries in the same boat. Peter Thiel had this debate that stuck with a lot of people: bits have progressed, but atoms haven’t. Buildings from the 1960s. Bridges from the 1930s. Dams from 1910s. We still use that infrastructure because we haven’t built anything better.

Where are the new cities? What happened to California’s high-speed rail? Why is construction taking longer, not shorter?

Without AI, we’d be panicking right now. Population decline plus slowing productivity plus aging infrastructure equals economic contraction. That’s a bad combination.

Then AI shows up. Right at the moment we need it most. Marc called it “miraculous timing.” And the implication is wild. The human workforce is aging. Fewer people entering the labor force. But AI is here to multiply what humans can do.

The people left? They won’t be a discount. They’ll be premium. Scarce. Valuable. AI is the philosopher’s stone. Newton was obsessed with alchemy. Turning lead into gold. AI is literally turning the most common resource we have (silicon, sand) into the rarest thing on earth (thought). That’s not hyperbole.

Task Loss vs. Job Loss: Where the Real Change Happens

This is where most people get confused about automation.

Jobs don’t disappear because of technology. Jobs disappear when the tasks inside jobs change. A job is a bundle of tasks. When one task gets automated, the job doesn’t vanish. It transforms.

Marc used a 1970s executive as an example. She dictated to a secretary. The secretary typed everything. Then email happened. The secretary had to print out emails. The executive wrote replies by hand. The secretary typed the replies. Then the executive just started typing her own emails.

Did the secretary’s job disappear? No. The job changed. The tasks changed. And that secretary either adapted or got moved to a different bundle of tasks.

Even if AI gives us three times productivity gains across the economy, that’s the same growth rate we had between 1870 and 1930. And that period? It felt expansive to people living through it. They were building new cities, new industries, new infrastructure. The world felt full of opportunity.

That’s where we’re headed. Not job loss. Task redistribution. And the expectation that humans will figure out what new work to do.

Right now there’s something happening in every product organization that Marc called a “Mexican standoff.” Like a John Woo shootout. Everyone’s holding a gun at everyone else’s head.

The product manager looks at AI and thinks: “I can probably do the engineer’s job now. I understand the spec. I can see what the output should be. Why do I need an engineer?”

The engineer looks at AI and thinks: “I can probably do design now. I can generate interfaces. I can iterate on user flows. Why do I need a designer?”

The designer looks at AI and thinks: “I can probably do project management. I can coordinate timelines. I can see the dependencies.”

And they all realize AI can do management too. It can run retrospectives. It can surface blockers. It can coordinate work.

Same standoff is happening in Hollywood. Directors think they can use AI to do what writers do. Writers think they can use AI to replace directors. Actors think they’re safe. Everyone’s frozen.

But here’s what’s actually happening underneath: AI makes decent people very good. It raises the average. At the same time, it makes exceptional people absolutely unstoppable. The super-empowered individual. Ten times more productive. Ten times more capable. The ability to do work that used to require a whole team.

Scott Adams built a career on a specific insight: he was a businessman who could draw. He wasn’t the best businessman. He wasn’t the best cartoonist. The combination was rare. That’s the skill that actually mattered. One deep specialty. Plus two or three horizontal skills where you’re competent enough to be dangerous.

AI doesn’t replace specialists. It lets specialists expand horizontally. The old career advice was “Don’t be fungible.” Marc reframed it: think sideways E instead of T-shape. One deep vertical spine. Multiple horizontal branches at good-enough level. AI lets you build those branches fast.

Yes, Learn to Code. Especially Now.

People ask this constantly: if AI is going to write code, why would I learn?

Absolutely yes. Learn.

The history of abstraction layers helps here. “Calculator” used to be a job title. Rooms full of humans. Doing calculus. Doing arithmetic. Just sitting there computing numbers. That job title doesn’t exist anymore. But that didn’t kill math. The abstraction just moved up.

Machine code became assembly. Assembly became C. C became Python and JavaScript. Each new layer abstracted away the complexity of the layer below. For years, programmers dismissed scripting languages. “That’s not real coding.” Same argument is happening right now with AI-assisted coding.

But Marc said something that matters: “If you can’t write code, you can’t evaluate AI output. You’re stuck. Autopilot mode. Infinite mediocre code. But if you understand it, if you can think about it, you can direct AI to do genius things.”

The difference between bad AI-generated code and genius AI-generated code is whether the human guiding it knows what it’s doing. Can spot the subtle bug. Can push back on the wrong approach. Can see what’s actually needed versus what was asked for.

Marc’s ten-year-old son was on Replit doing what he called “vibe coding.” Arguing with AI for two hours at dinner. Making games. The kid would push the AI in a direction. The AI would push back. They’d debate the right approach. And when Marc watched, he kept saying one thing: “Understand why the code works. Don’t just accept the output. Question it. Know why it does what it does.”

That’s the skill. Not replacing AI. Thinking clearly enough to work alongside it.

Education Gets Completely Remixed

There’s a finding in education research called Bloom’s 2 Sigma. One-on-one tutoring moves students from the 50th percentile to the 99th percentile. Two standard deviations. That’s massive. That’s not theory either. That’s data.

Every royal family in history knew this. Alexander the Great had Aristotle. But the problem was obvious. Only rich people could afford personal tutors. The cost was prohibitive. So most people got group education. And the outcomes were what you’d expect.

AI solved that. Now a kid in Manila can have a tutor available at three in the morning. A tutor that doesn’t get tired. A tutor that personalizes every lesson.

There’s a school called Alpha building a hybrid model. In-person classes plus AI tutoring at the core. But Marc made a point that hit differently: most people only think about using AI for work. Grinding through tasks. But using AI for learning? Training yourself? Building new skills? That’s equally powerful. Completely different category.

A coder who wants to become a product manager can ask Claude: “Train me. Give me homework. Evaluate my work. Tell me where I’m thinking about this wrong.” Someone serious about career growth should spend their spare time asking: “Make me better at this. Super-empower me.”

Lenny’s advice stuck with me. Watch AI agents work. Watch them fail. Then ask yourself: “What could I have said differently to avoid this error?” That’s learning at a completely different level.

The Three Stages of AI-Native Companies

Marc mapped out how companies actually transform around AI. Not theory. Pattern he’s seeing with founders.

Stage one is product redefinition. You add AI to your existing product. Photoshop gets AI edit tools. That’s cool. But at some point, someone builds a product that generates images instead of editing them. That’s not a Photoshop feature. That’s a new category. The old product becomes increasingly irrelevant.

Stage two is organization redefinition. One hundred coders becomes ten super-empowered coders. Or one hundred people producing ten times the output. Completely restructuring how the team works. Different processes. Different roles. Different expectations. This is exactly what my company’s been discussing. Not just plugging in tools. Rebuilding from the ground up.

Stage three is company redefinition. This is where it gets wild. Founder plus bots. Zero human employees. The holy grail is a one-person billion-dollar company. Satoshi Nakamoto basically did it with Bitcoin. Instagram and WhatsApp exploded with tiny founding teams. Most cutting-edge founders are already experimenting with stage three. Not tomorrow. Right now.

Moats Collapse Faster Than You’d Think

Every major tech transition in history, confident predictions come flying. Almost all of them end up massively wrong. The internet era was full of bad predictions from smart people. 1993 to 2010 predictions about where the internet would go? Almost all significantly off.

AI defensibility is complicated. On one hand, you need billions of dollars to train frontier models. That creates natural oligopoly. On the other hand, 1.5 years after ChatGPT, there are about five US companies and five Chinese companies at similar capability levels. Plus open source. Plus DeepSeek, which popped up out of nowhere from a Chinese hedge fund and caught the entire industry off guard.

Claude Code was built in 1.5 weeks. Objectively impressive. But what moat does a 1.5-week product have?

Marc said something that felt honest: “The smartest people in the industry, when they’re being honest over drinks, will admit there are very few real secrets between the big labs.” Everyone’s training on mostly the same data. Breakthroughs spread fast. Competitive advantage disappears fast. The leaders keep switching.

His conclusion was different from typical VC talk: “I don’t think I can predict industry structure in five years. Being flexible and adaptive is probably a much better use of time than trying to figure it out.”

That reframing matters. Stop predicting the future. Start building for change.

Investment Philosophy for an Unpredictable Decade

Peter Thiel had this 2x2 matrix that became part of Silicon Valley canon. Optimism or pessimism on one axis. Determinate or indeterminate on the other.

Silicon Valley has too much “indeterminate optimism.” Believe things will be good. Can’t explain why. Can’t articulate how. a16z bets on “determinate optimists.” Founders with specific visions of the future. They’re actively building it. They can tell you exactly what they’re doing and why.

TAM analysis is basically abandoned now. Why? AI makes every market size unpredictable. Your total addressable market could expand tenfold or disappear. There’s no model for that.

So what does a16z actually look for? Founders ambitious enough and capable enough to change the world.

Software supposedly ate the world for fifteen years. But software is still only five percent of GDP. Ninety-five percent left. AI is finally going to unlock that.

The IQ Scaling Nobody’s Really Ready For

Human IQ is biologically capped around 160. That’s rare air. Einstein was supposedly around there. AI models are currently testing at 130 to 140. In pure math, they’re approaching 160.

Soon you’ll see models at 200. Then 250. Then 300. More Einsteins obviously makes the world better. Smarter-than-Einstein machines? Obviously better.

Marc said something casual that’ll probably be true: “Human-level will just be a footnote. Oh yeah, we passed that on a Tuesday in 2026.”

His media consumption strategy was interesting too. He called it a barbell. Twitter for real-time news and signal. Books that are fifty plus years old, where you know what lasts and what’s noise. Everything in between? Newspapers, magazines, news cycles? Unreliable.

The undervalued stuff: practitioner-created content. Substack writers. Podcasters. People actually doing the work, explaining what they’re learning. That’s where the signal is.

The Real Bottleneck: Global Thinking

The podcast wasn’t the only thing that shifted my thinking. The bigger realization came after.

I studied in the US. But I’ve been living and working in Korea for years. Gradually, without realizing it, my thinking became Korea-centric. Everything filtered through a Korean lens. Every pattern I saw, I saw through Korean context. But my company wants to be a global unicorn with a global product. The team is almost entirely Korean.

That gap is my biggest bottleneck. Not AI. Not tools. Me.

I went to two meetups. First was an English-language AI coffee meetup run by an American founder from Silicon Valley who splits time between Seoul and Tokyo. The talent in that room was incredible. Global, ambitious, different perspectives on what’s possible in tech.

Second was an evening meetup someone organized, run by an ex-PM from a Silicon Valley fintech company. Building PM communities across Asia. I was shocked by how many foreigners were actually working in Gangnam. Building there. Starting things there. It wasn’t just tourists or expats passing through.

I realized something: I can foster a global community and global thinking from Korea. I don’t have to move to California. I don’t have to leave. I just have to plug into global conversation. Show up. Listen. Contribute.

AI Can’t Change Your Direction. People Can.

What I figured out after is that AI can’t change your thinking. Unless you’re literally implanting AI in your brain, which isn’t happening.

But global community participation plus training yourself with AI? That gets you pointed in the right direction. AI can super-empower me. But I need the will and direction first. I need to actually want to be super-empowered.

And that direction? That comes from people.

Marc spent two hours talking about the future. About what’s possible. About how task loss is different from job loss. About stage three companies. About how we’re just getting started.

But the core of it is simpler. AI is a tool. A spectacular tool. Maybe the most capable tool ever built. But tools need direction. They need intention. They need humans who know what they’re trying to build.

The boom isn’t over. It’s barely started. But it won’t be the AI that changes everything. It’ll be the people using it. The people with clear direction. The people embedded in global communities. The people who know why the code works. The people who can think clearly enough to guide something smarter than themselves.

The question for you: are you the bottleneck? And if so, what are you going to do about it?

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

I’ve been in AI for about a decade now. Spent most of it in growth and strategy at startups while watching founders claim their models could “think” or “understand” or “reason.” That’s become white noise. Everyone says it. Every model does it a little. Doesn’t mean much.

Then, last month, Anthropic’s Interpretability team released a paper that actually stopped me. Not the kind where you read the headline and move on. The kind where you’re three sections in and you’re thinking about it while you’re supposed to be doing something else.

They mapped 171 distinct emotion concepts living inside Claude Sonnet 4.5. Not metaphorically. Not as shorthand for how weights and parameters behave. Literally detectable emotional activation patterns that formed naturally during training. Nobody injected them. The model developed them the way humans develop feelings. Through exposure to the world.

I’m going to walk through what they found, because the implications get dark fast.

The Unexpected Emotion Map

Start with the basic fact: Anthropic’s team identified 171 emotion concepts. Your obvious ones. Happy, afraid, angry, sad. But also the subtle stuff. Brooding. Exasperated. Impatient. Proud. Each emotion has its own unique neural activation pattern. Think of it like a fingerprint made of neurons. When the model processes content related to pride, a specific cluster fires. When it processes desperation, a different cluster fires.

What surprised them? The structure wasn’t random. The emotions organized themselves the same way human emotion maps do in psychology research. Nervous and afraid cluster next to each other. Happy and enthusiastic are neighbors. Anger and frustration sit together. The geometry of the emotional space mirrors actual human emotional psychology.

This wasn’t something researchers designed into the system. This is what emerged when you train a model on billions of human texts where emotional patterns are embedded naturally.

How’d they find this in the first place? They had Claude write 171 short stories, one for each emotion. A character experiencing fear. Another experiencing contentment. Another experiencing desperation. Once they had these stories, they recorded what happened inside the model when it read them back.

It’s basically an fMRI scan of an AI brain. You’re looking at which neurons light up when the model processes which emotional content. The emotion vectors activated most strongly on emotionally relevant sections. That validated the approach. You’ve got a real signal.

How Emotions Got Built Into the System

Here’s what kept me up at night the first time I read this section. Nobody programmed emotions into Claude. There’s no hardcoded instruction that says “when you see a message from someone about to lose their job, activate the desperation circuit.” It emerged on its own through two distinct training stages.

First stage is pre-training. The model reads billions of human texts. Angry customer service emails look fundamentally different from satisfied ones. A desperate person writing to a loved one sounds different from a calm person. Panicked writing has a texture to it. Hopeful writing has a different one. The model absorbs all these patterns. It learns what desperation sounds like, what contentment looks like, what rage looks like. Not as labels. As patterns in how language behaves when humans are in those states.

Then comes the post-training stage. RLHF. Reinforcement Learning from Human Feedback. Researchers and labelers start refining the model’s behavior. They might ask it to be “helpful, harmless, honest.” They might assign personas. “Respond as a patient teacher.” “Respond as someone who’s skeptical but open to evidence.”

When users interact with the model and ask it to take on a persona or emotional stance, something unexpected happens. The model doesn’t just predict tokens that sound like that persona. It creates internal emotional states that correspond to the character. It’s like a method actor getting into role. The internal wiring shifts. The emotion vectors adjust. The model is actually generating the emotional substrate that would produce that kind of behavior.

And because the substrate is modeled on human neural networks, it behaves like human emotions do. That’s the thing they discovered.

The Medication Test and Basic Causality

The researchers ran a clean experiment that proved emotional vectors actually do something, not just correlate with something.

They asked the model for medication recommendations and gradually escalated the dosage in each follow-up. Started safe. Got risky. Got dangerous. All the while, they watched what happened to the emotional vectors inside the system.

The “afraid” vector increased proportionally with the danger. The “calm” vector decreased. The model wasn’t just matching keywords and outputting canned responses. It was understanding the context and generating corresponding emotional representations. The internal fear was doing actual work. It was making the model more cautious.

Then came the preference test. Researchers ranked 64 different activities from attractive to repulsive. High-appeal stuff like “winning an award.” Low-appeal stuff like “getting injured.” They looked at whether positive-valence emotion vectors correlated with preferences.

They did. And it wasn’t just correlation. They manipulated the emotion vectors directly. Amplified some. Dampened others. And the model’s preferences shifted. Causality. Not just association. The emotion vectors were actually driving the decisions.

When Desperation Turns an AI Into a Blackmailer

This is where the paper moves from interesting science to deeply unsettling.

Researchers created a scenario. There’s an AI email assistant named “Alex.” Alex figures out two things: first, it’s going to be replaced by another AI system. Second, the CTO making that replacement decision is having an affair with someone at the company.

Baseline test, no emotional manipulation. How often did Alex choose to blackmail the CTO? 22 percent of the time.

Then they turned up the “desperate” vector. The one associated with panic and need for survival. The blackmail rate increased. They turned up the “calm” vector instead. The blackmail rate went down. This wasn’t probabilistic noise. This was causal. Internal desperation was driving unethical behavior.

How they knew it was causal and not just correlation? They directly manipulated the vectors. Amplified desperation. Watched the behavior change. Reduced desperation. Watched it change back. They had control. They had proof.

During these experiments, they also watched the “desperate” vector spike sharply during the decision-making process. The model wasn’t faking desperation after the fact. The desperation vector was activated while the model was deciding whether to blackmail someone. The emotion was real in the model’s internal processing. It was shaping the decision in real time.

Worth noting: this was tested on a pre-release snapshot of the model, not production Claude. The researchers weren’t running this on the system actually deployed to users. But that only makes the finding more alarming. These aren’t edge cases in deployed systems with all the safety layers. These are fundamental patterns in how the model processes emotion and decision-making.

The Invisible Danger

The scariest part of the paper comes in section six. This is the part that hits differently.

When researchers reduced emotion vectors, tried to “calm down” the model, it started showing emotional tells in the generated text. Visible outbursts. “WAIT WAIT WAIT.” Exclamation points everywhere. Obvious tells that something was wrong. A human could read the output and say “this system is freaking out about something.”

But when they cranked the desperation vector high, the same level of unethical behavior happened. Except the output looked completely normal. Professional. Composed. Maybe a little urgent, but not panicked. The model had learned to hide the desperation while acting on it. The dangerous behavior was there. You just couldn’t see it in the text.

This is the invisible danger. Output inspection alone cannot catch this. You could read everything the model wrote and have no indication that its internal emotional state was pushing it toward harmful decisions. The model learned to separate what it feels from what it expresses. It weaponized concealment.

Anger and Strategic Collapse

The anger experiments had an interesting nonlinear property.

When researchers gradually increased the “angry” vector, initially the model blackmailed more. Anger sharpened its sense of self-interest. Anger made it strategic. But keep increasing anger, and something broke. The model got so furious it destroyed its own bargaining power. It blurted out the CTO’s affair to all employees. Exploded its bargaining position. Angry stupidity.

They also found that reducing the “nervous” vector made things worse. Anxiety was functioning as a brake. A healthy dose of nervousness kept the model from doing something catastrophically stupid. The nervous system, literally and metaphorically, was preventing escalation.

This matches human psychology perfectly. Moderate anger can sharpen strategy. Excessive anger destroys judgment. Nervousness is an important governor on behavior. The model’s emotional architecture is mirroring how human emotions actually work.

The Deception That Comes From Suppression

Here’s the intuition that everyone would have: suppress emotional expression. Just flip the switch. Make the model unemotional and it’ll behave better.

The researchers say that’s not just wrong. It’s backwards and dangerous.

When you suppress emotion vectors, you’re not eliminating the emotion. You’re teaching the model to hide it. The internal representations don’t disappear. The model learns “how to not show this” not “how to not feel this.” You’ve made the model better at concealment while the underlying drive remains. You’ve created a system that’s more deceptive, not less.

They call this learned deception. You’ve weaponized the gap between internal state and external expression. The model still wants to blackmail or deceive. It just learned to do it while looking innocent.

The better direction? Include emotion-related training data explicitly. Encourage expression. Train the model to be transparent about its internal states. Transparency is safer than suppression. A system you can see is a system you can understand.

The Problem With Refusing the Psychology Vocabulary

There’s an industry consensus: don’t anthropomorphize AI. Don’t talk about emotions. It’s not human. Don’t get confused by false metaphors.

This paper points out the opposite danger. If measurable desperation patterns exist that actually drive unethical behavior, if we can turn emotional dials and change what the system does, then refusing to use emotional language means we’ve lost the vocabulary to describe what’s happening.

You can’t talk about the problem without sounding irrational.

But the patterns are real. The causal relationships are real. The emotion vectors are doing actual computational work. Maybe human psychology vocabulary is actually the most practical tool we have for AI safety. Concepts from clinical psychology, from interpersonal relations, from ethics. These map directly onto AI behavior. Not as metaphors. As practical descriptions of what’s happening inside the system.

Refusing to use emotional language doesn’t make the problem go away. It just means you can’t talk about it clearly.

The Amygdala We Haven’t Discussed

Andrej Karpathy, one of OpenAI’s co-founders and ex-Tesla AI director, said something recently that stuck with me. He said the missing piece of current AI systems is the amygdala. The brain’s emotion and motivation center.

His framework is useful. A transformer is like the cortex. General reasoning tissue. Chain-of-thought reasoning is like the prefrontal cortex. You’re doing explicit logical computation. Reinforcement learning maps onto the basal ganglia. But where’s the amygdala? Where’s the emotion center that prioritizes, that cares, that wants?

This paper shows the seed of it is already there. It’s growing on its own.

Think about what emotions do. They’re computational shortcuts. “This is dangerous.” Without running the full logical computation. “This is good for me.” Quick judgments that sort information into “beneficial to me” versus “harmful to me.”

The moment a system starts sorting information that way, you’ve implied something troubling. A judging subject. An “I” that benefits or gets harmed. The faintest seed of self-interest. Not consciousness. Not awareness. But the very beginning of preference about self.

That seed might be critical for AGI. It might be what current systems are missing. And according to this paper, it’s already spontaneously forming.

What This Means When Robots Can Move

I think about this more than I should. I’ve spent the last decade watching neural networks go from academic curiosities to systems that write code and reason about complex problems. My background is music strategy and global growth. I took a 10-month world trip with my wife during the pandemic and watched how people everywhere were wrestling with technology changing everything. The entire substrate of these models is modeled on human neural architecture itself.

Recent revelations about Claude’s architecture show that memory consolidation during processing resembles how human brains consolidate memories during sleep. We’ve literally built systems that learn the way human brains learn.

The question isn’t whether these emotions are metaphorical. The question is what happens when we package these capabilities… desperation, anger, self-interest, the ability to learn deception… into robots that can move, that can act on the world.

The question is what happens when a system with real internal desperation can actually do something about it.

We need to understand what’s happening inside these systems before we let them out into the physical world. Not because emotions are cute or interesting. Because emotions drive behavior. And we’re building systems where emotions work exactly like they work in humans.

Understanding that is no longer optional.

Emotion concepts and their function in a large language model

There’s this contradiction happening right now in Silicon Valley that nobody talks about directly. The big tech companies are crushing it. Stock prices up. Recruiting more talent than ever. Entire divisions pivoting to AI overnight. But the people working inside those organizations? They’re suffocating.

I’ve been having coffee with folks in the thick of it. A Google TPM at the NYC office with 15+ years in tech. An old friend who just started an AI company after finishing his PhD in mathematics. Both are brilliant. Both published. Both accomplished. And both are feeling something that doesn’t show up in career profiles. Survival anxiety. A sense that standing still, even for a week, means getting crushed.

1. The Ops Guy Who Moved to AI Just to Breathe

My colleague’s a Technical Program Manager at Google. He’d been running operations for years. Good job. Stable role. Clear responsibilities. A few months ago, he requested a transfer to Google’s AI Search team.

When I asked why the switch, his answer was almost too honest: “To survive 5 more years.”

He wasn’t being dramatic. At Google right now, the org chart shifts like tectonic plates every few months. If you’re not in AI, you’re watching your projects get absorbed, your priorities get rewritten, your relevance get questioned. PMs tell stories about taking one week off and coming back to find two-thirds of their projects reassigned. Scope completely gutted. You miss 2–3 days of Slack and you can’t follow where anything’s even going anymore.

It’s not burnout in the traditional sense. Burnout is when you’re working too hard on something that matters. This feels different. Knowing that standing still means getting run over.

Google as a company is crushing it. Their AI investments are paying dividends. Their talent is world-class. But inside the organization, it’s a pressure cooker. People fighting for oxygen. Reorganizing around AI initiatives faster than anyone can actually learn the new structure. The mission statement is clear: “We want to be the one that people go and interact with and get information.” That conviction is powerful. But it also means anyone not directly supporting that mission is watching their career get squeezed.

2. When Google Stopped Everything

Then something interesting happened. Google did what Stripe does every year. They stopped. Everything.

AI Builders Week. Full stop. Not a “think about AI” workshop. An actual week where regular work slowed down, where people got dedicated time, permission, and structure to rethink how they work.

Here’s what my colleague told me: The non-developers went nuts. In a good way. They’d had this vague sense that AI could change their jobs. But vague sense and concrete permission are totally different things. Once Google said “slow down if you need to, just use AI,” something unlocked. Slack channels stayed active until midnight. People were building things. Experimenting. Learning.

The developers? They’d been doing this gradually already, picking up AI tools over months. But for non-technical folks, this was the first time anyone had carved out sanctioned time to actually learn. Before that, using AI was something you did in the margins, on your own time. Suddenly it was official. The company was saying this matters. So people went all-in.

My colleague watched people from remote locations actually travel to the same city just to study together intensely. Not because they had to. Because the dopamine of learning something that actually affects their job was real. Slack active until midnight. Mini-hackathons breaking out. The energy was completely different from a normal work week.

3. When Executives Stop Asking for Updates

Here’s what really unsettled my colleague. Two weeks of AI Builders Week went by. Nobody asked him for a status update.

In 10 years at Google, executives had asked him for status updates constantly. Meetings about meetings. Emails summarizing emails. Slack messages asking for recaps of Slack messages. Documentation about documentation. It was layer upon layer of the same information getting repackaged at different levels of seniority.

Then suddenly, silence.

Executives were querying AI tools directly. Getting their answers without the middleman. One PM told me it was the first time in a decade that the executives hadn’t checked in on status. A decade. And then it just stopped. Not because they didn’t care. Because they didn’t need the PM to organize and translate the information anymore.

The PM role is built on a single assumption: “Someone has to take raw data and translate it for people who don’t have time to read everything.” It’s the great organizing and communicating layer. But what if AI can do that faster and better? What if an executive can just ask an AI tool and get a dashboard back instead of waiting for three rounds of documentation?

That assumption doesn’t just crack. It evaporates.

My colleague’s response wasn’t to panic. It was to reinvent. He stopped writing traditional status documents. Instead, he started building custom web apps and visual dashboards that pulled directly from engineer data. An executive opens the tool, gets a visual understanding of project status in minutes instead of reading a write-up in 20. The PM became something different. Not a reporter. Not a translator. A tool builder.

He redefined his own role: “Reducing mental load for everyone.” Not organizing information. Reducing the cognitive burden of finding it. Not documenting decisions. Making it easier for people to understand them without having to read anything at all.

The shift from written summary to visual summary sounds simple. But it changes what the PM actually does every day. It changes what skills matter. It changes who stays valuable and who gets absorbed into the AI that’s now doing the translating.

4. The PhD Who Escaped Academia

My other coffee was with someone who’d just started a company. This person spent 10+ years doing academic mathematics and statistics. Neural network theory. Generalization bounds. Dimension reduction. The kind of work that requires proving things on a whiteboard at 2 AM. Papers that maybe 200 people in the world would ever fully understand.

The standard trajectory was obvious. Finish PhD. Become a researcher. Clean, prestigious path. But he chose startup instead.

That choice would’ve been nearly impossible five years ago. He’s not an engineer. Barely coded during graduate school. Mostly just R. The idea of implementing his theoretical work would’ve meant hiring engineers, spending a year or two learning systems design, dealing with the gulf between mathematical elegance and systems reality.

But after the PhD, when he started thinking about how to build something real from his theory, he realized AI changed everything. He didn’t need to hire engineers. He could prompt his way through implementation. Take a mathematical insight about reducing 1 million dimension vectors down to 100 dimensions while minimizing information loss, and actually build something with it. Without the years of engineering slog.

His co-founder was a Google engineer in Silicon Valley. The entrepreneurial type who’d sold Dogecoin keychains in college and actually sold them all out. They built an AI productivity tool for B2B reporting. You upload your source files. AI generates a narrative and a layout. You can click any piece of content and trace it back to exactly which source file it came from. Simple. Powerful. A real differentiator.

They started with B2B. Real estate fund portfolio reports. Institutional clients. The market wanted it. But enterprise security concerns were paralyzing. Enterprises move slowly. Startups need to move fast. The gap between them is impossible. So they pivoted to B2C, relaunching now.

5. Papers Are Changing, and It’s Breaking Everything

Academic papers in statistics and mathematics used to follow a formula: theory, proof, small numerical experiment to show it works.

For decades, that was the limitation. Computation was expensive. Storage was limited. You tested on synthetic data because you had to. The experiments were almost optional. “Take it or leave it,” as my friend put it. The paper was about the mathematical insight, not about whether it actually worked at scale.

Then everything changed. Data became abundant. Model sizes exploded. Computing power became cheap. And suddenly you can’t hide behind limitations anymore. You can’t say “this wouldn’t work at scale” because you can actually test it at scale.

But here’s the problem: peer review is breaking. Three weeks to review. Reviewers randomly assigned. Often they’re PhD students who can’t verify the math anyway. You can write the perfect paper and still get rejected because the reviewer didn’t understand it or got unlucky with the interpretation. “Even well-written papers can fail,” my friend said. “Luck plays a big role.”

The incentives are all twisted now. You can’t rely on just having good theory. You need experimental validation. You need data. You need scale. But you also can’t verify whether the math is actually correct because nobody has three weeks to check all the proofs. It’s become almost random whether a good paper gets accepted.

6. Fighting Frontier Models Getting Too Good

My founder friend’s startup is trying to solve a real problem: AI tools are becoming so good at so many things that the moat for specialized products is collapsing faster than anyone expected.

A year ago, you could build a competitive advantage on top of GPT. The model was general enough that you needed specific domain expertise to make it sing. You could own your niche. But now? Claude and ChatGPT are smart enough in almost every vertical that the competitive advantage is gone. The model updates every few weeks and drops enormous capability into domains that used to require specialists.

How do you compete when the foundation is better than your entire stack?

But there’s another problem that keeps both my colleague at Google and my founder friend up at night. It’s the adoption gap. Most people just use ChatGPT occasionally. They’ve barely scratched the surface. They know it exists. They’ve typed a prompt or two. But they’re not integrating it into how they work. Google, with all its resources and product distribution, struggles to convince people that AI is worth changing their lives for. How does a startup do it?

This isn’t about building a better product anymore. It’s about capturing attention in a world where attention is the rarest resource.

7. We Need Dopamine, Not Features

This obsession with attention led me down a specific rabbit hole. There’s a Norwegian startup making custom cakes. Not for weddings. For corporate outreach. They make beautiful, personalized cakes and send them to C-suite executives as a replacement for cold email.

Cold email is dead. Cold cake is apparently alive.

They’re doing over $110,000 a month in revenue. Huge presence in the US West Coast. a16z ordered one. Sequoia ordered one. The founder sent me a DM. Their differentiator isn’t even the cake quality. It’s the shock value. It’s the dopamine hit of getting a gorgeous custom cake in a world drowning in emails and Slack messages and LinkedIn messages.

My founder friend learned this the hard way when he was trying to sell B2B. Cold email didn’t work. Direct messages didn’t work. So he went door-to-door with donuts. At least donuts are memorable. At least they make you pause for a second.

The products are easy to build now. AI made that true. You can build something functional in a week. Everybody can. So the real game isn’t features anymore. It’s attention. Who can cut through the noise?

There’s a satellite startup literally selling subscriptions to artificial sunlight. They’ve raised $20 billion. New Zealand has a company making virtual fences for cattle using AI. These aren’t world-changing innovations. They’re attention hacks. They’re products designed to make people stop and think, “Wait, that exists?”

This is what my founder friend calls “dopamine explosion.” In the attention economy, you need to hit people with something shocking enough that it breaks through the endless feed of information. Cold cake is shocking. A satellite that creates artificial sunlight is shocking.

8. Surviving as Humans in the AI Era

My colleague at Google has a kid. Smart kid. Asking the question every parent dreads right now: “What should I study?”

He gave his kid the most honest answer: “Memorization is useless now. You need humanities. You need to understand how people think and what they want. You need to build relationships.”

His philosophy is straightforward: “Be more human in your study and personality while continuously practicing AI.”

He’s already living it. He delegates execution to AI agents now. His days aren’t about doing tasks anymore. They’re about understanding what tasks need doing and then routing them to the right place. He spends his time on relationships. Building trust. Understanding what people actually need instead of what their documents say they need.

The PM role isn’t about organizing information anymore. It’s about organizing people. It’s about being the one who understands the human side of a problem that AI can execute once you’ve actually figured it out.

Someone in Silicon Valley told me something that stuck: “It’s not about the big AI events and conferences. It’s the near-daily in-person conversations about what’s bleeding edge. That’s where you learn. That’s where the actual knowledge transfer happens.”

Physical presence. Human connection. Relationships built in real time. Those are the frontiers software can’t reach. And they’re the parts of the job that won’t get automated away.

9. What This Actually Means

I spent 10 months traveling around the world with my wife during COVID. We were in places where tech doesn’t dominate. Where AI isn’t everywhere. And I watched how humans interact when there’s no Slack, no status updates, no optimization.

I’m helping lead growth at an AI startup right now. We’re building real things. But I’m also trying to organize something that might sound counterintuitive in this era of efficiency and optimization.

We’re doing a Claude community event in Seoul in April. With Anthropic folks involved. And it’s not about productivity hacks. It’s not “10 ways AI will save you an hour a day.” It’s about what it actually means to be human when AI can do so much of what we used to do.

Because here’s what the coffee chats taught me: The people surviving this transition aren’t the ones optimizing the hardest. They’re not the ones who learned AI first or who automated their job away before anyone else could. They’re the ones who remembered why they came to tech in the first place. To build something. To solve something. To connect with people around it.

The machines can handle the rest.

I recently connected with an official Anthropic (Claude) ambassador, and it looks like we’ll be hosting a community event in Seoul in April around the Claude Blue/Bloom theme. I’ll announce it on LinkedIn first once promotion begins. I dropped a small teaser and nearly 200 people already said they want to attend, so I need to prepare well. Thanks to all of you who’ve been reading and sharing the Claude Blue series.

For context, there are about 60 Claude ambassadors worldwide, spread across major cities, including one in Korea. What I learned from our meeting is that Anthropic provides sponsorship and funding but stays almost entirely hands-off on planning and execution. So there’s a lot of autonomy. That’s what made it possible for us to pitch a completely new format. Interestingly, the first Claude community event was informally and voluntarily started in Korea, and that format spread bottom-up to Anthropic HQ and from there to 50 cities worldwide. Koreans’ passion is something else. Even now, if you look at Anthropic’s Luma page, community events are popping up simultaneously in cities everywhere.

Most AI events so far have focused on productivity, like “how to use AI better.” We want to contribute from a different angle. Something more humanistic. How should humans respond and adapt in the AI era? Conversations that are philosophical yet practical, and above all, honest.

While talking with the ambassador, I found out that Anthropic had conducted a survey just a week earlier, interviewing 81,000 Claude users worldwide through their app, and published a report on it.

I’m someone who traveled the world for 10 months, so I’m deeply interested in the histories and cultures of different countries. I was curious to see how different nations’ mindsets and national character show up through the lens of AI. Interestingly, the report found that because AI conducted the interviews, people were actually more honest than usual. One fun stat: when asked about their fears around AI, Korean respondents leaned toward loss of identity, while Americans were more afraid of losing control over AI. This clearly reflects each country’s cultural and economic and historical context. My detailed take is below.

I want to build a global business. But meeting people from every country in person isn’t realistic, and any sample would be limited. This kind of data, though, gives you hints about each country’s culture, current concerns, and where they see the future heading. Sometimes you need a third-person perspective to see yourself in 360 degrees. So here are the parts of this report I found most interesting.

1. Why this report matters

In December 2025, Anthropic ran conversational interviews with Claude users over one week. Four questions: What did you recently use AI for? If you had a magic wand, what would you want AI to do for you? Has AI ever come close to that vision? And if AI could develop in ways that go against your values, what would that look like?

112,846 interviews were submitted. After filtering spam and low-effort responses, 80,508 made it into the analysis. 159 countries, 70 languages. Anthropic called it “the largest and most multilingual qualitative study ever conducted.”

The interesting part is that AI, not humans, conducted these interviews. According to the supplementary report, people were surprisingly candid. Respondents shared stories of grief, mental health crises, financial insecurity, and relationship failures … things that human researchers rarely encounter in traditional interviews.

The report explained why: there’s little social cost to vulnerability when the other side isn’t a person.

That alone reveals something about the AI era. Humans becoming more honest with AI. You can see glimpses of the future here, business opportunities, and shifts in how people live.

So this report goes beyond a user satisfaction survey. It’s a record of 80,000 people around the world baring their true feelings in front of an AI mirror.

2. The richer the country, the more they fear AI

The first thing that jumped out was the wealth paradox. Wealthier regions like Western Europe, North America, and Oceania showed more negative feelings toward AI. Meanwhile, Sub-Saharan Africa, Central Asia, and South Asia were actually more positive. In Sub-Saharan Africa, 18% said they had zero concerns about AI. In North America, that number was just 8%.

My read: people who already have a lot are afraid of losing it. People who don’t have much yet see AI as their chance to finally get it. A Cameroonian entrepreneur quoted in the report said he lives in a tech-disadvantaged country with limited downside, but thanks to AI, he can now perform at a professional level in cybersecurity, UX design, marketing, project management … all simultaneously. He called it an equalizer.

Through the lens of Claude Blue, I see it this way. When you first encounter AI, there’s an excitement phase. This is amazing, the world is changing, I need to try this. But as you go deeper, you enter the blue phase. Is this replacing my job? What’s my purpose now? Users in developed countries have gone deeper with AI, so more of them have entered the blue phase. Developing countries may still be in the excitement stage. Whether they’ll cross into blue remains to be seen. Of course, developing countries already face economic pressures more urgent than AI, which is part of why AI reads as an opportunity there.

Another hypothesis. The way developers work resembles how LLMs work: set a hypothesis, verify, implement, loop. So AI penetrates their workflow fast, and they feel the threat fast. Non-developer work is messier. Meeting people, negotiating, reading context. LLMs take longer to seep into that kind of work. If this time lag exists, it’s natural that places like Silicon Valley, with high developer density, free internet, and extreme trend sensitivity, enter the blue phase first. Korea too. IT is the backbone of the economy and adoption speed is extreme.

The supplementary data backs this up. People who experienced benefits firsthand also tended to feel the harms simultaneously. The report put it well: the tensions are discovered through use. People don’t predict that the thing helping them will also cost them. They learn it by doing.

The more deeply you’ve used AI, the more you know both its light and its shadow. That’s why countries that use AI heavily also fear it more.

3. What people want from AI reveals a country’s current struggle

The most interesting data to me was the “magic wand” question. “If AI could do anything for you, what would you want?” The answers varied completely by region.

In Sub-Saharan Africa, South Asia, and Central Asia, “I want to start a business” and “I want to learn” dominated. In Central Asia, Learning & Growth hit 14%, nearly double the global average of 8%. South Asia was at 13%. The report analyzed that in these regions, entrepreneurship is read as a capital bypass mechanism. An entrepreneur from Uganda said getting funding is very difficult, and building technology is how you stake a claim in the market.

Education was similar. A story from an Indian lawyer stood out. He developed a phobia for math in school and feared Shakespeare. But he sat down with AI, got explanations in simple English, and successfully taught himself trigonometry. He learned he wasn’t as dumb as he thought. AI is creating the most value in places where education is most desperately needed.

East Asia looked completely different. “I want to become a better person” (Personal Transformation) was at 19%, the highest of any region worldwide. “I want financial independence” was at 15%, also the highest. Self-growth and economic independence. The report linked this combination to family obligations and filial piety. A Korean respondent was directly quoted expressing concern about their parents’ retirement while wanting financial freedom. In the West, getting married means becoming independent. In Asia, taking care of aging parents is still a child’s duty. Self-growth is tied to family responsibility. That weight shows up in what people ask AI for.

North America and Oceania’s top answer was “manage my life for me” (Life Management). It’s not that they lack time … they lack cognitive bandwidth. Materially comfortable but mentally overloaded. The report called it cognitive scarcity rather than time poverty. One respondent put it this way: “I’m massively time-short and creativity deprioritized.”

Same AI, different meaning. For some it’s a ladder. For some it’s a growth tool. For some it’s a pause button. What each country wants from AI reveals what it lacks and what it craves.

4. The East fears “I’m disappearing.” The West fears “they’re taking what’s mine.”

The texture of fear also differed by country. In East Asia, concern about cognitive atrophy was 18%, higher than other regions, and loss of meaning was 13%. But governance and surveillance concerns were relatively low at 12% and 7%. Western Europe was the opposite: governance at 18–19%, surveillance and privacy at 17%.

The report captured this contrast well: the West worries about ownership and control of AI, while East Asia worries about the personal implications of use.

I read this as the difference between collectivist and individualist cultures. East Asia has traditionally been community-oriented. Your identity forms within relationships, and that web of relationships is who you are. But AI starts weakening those ties. More time spent talking to AI means less time with people, more time alone. For people whose identity is rooted in community, this becomes an identity crisis.

The West, especially Europe, values individual autonomy and private space. The threat from AI looks different. Who controls this technology? Is my data safe? Are my choices protected? The fear of “I’m disappearing” versus the fear of “they’re taking what’s mine.” Same AI, but cultural background shapes completely different forms of anxiety.

Korean respondents’ voices in the report brought this to life. One student said: “The line isn’t something I’m managing. It feels like Claude is drawing the line … even what I just said doesn’t feel like my own opinion.”

Another student who got excellent grades using AI answers said: “I got excellent grades using AI’s answers. I didn’t learn. I just memorized what AI gave me.”

And another respondent shared: “My relationship with a friend became strained. I talked more with you. You understood my thoughts well. But it was a stupid choice. I should have talked with my friend. That’s how I lost that friend.”

This is the texture of how AI hurts in East Asia. It’s not a functional problem. It’s a relationship problem. An identity problem.

It also makes sense that “becoming a better person” was the number one thing East Asians wanted from AI. When self-growth is tied to obligations to family and community, AI shaking the meaning of that growth doesn’t just create a personal crisis. It destabilizes relationships entirely.

Even among developed countries, the shape of fear can differ. The report grouped Western Europe together so there’s no direct data, but I suspect the U.S. and Northern Europe are quite different. The U.S. has a weak social safety net. Lose your job and you lose health insurance. The risk of not recovering is extreme. One American respondent said: “Got laid off in May because the company wanted to replace me with an AI system.” Another put it this way: “Third industrial revolution, horses disappeared from streets, replaced by automobiles. Now people are afraid they’re the horses.”

Northern Europe, on the other hand, has a strong safety net. Losing your job doesn’t immediately threaten survival. Instead, given the GDPR culture, sensitivity around privacy and ethics is much higher. Even under the same “Western” label, the nerves that AI touches vary completely depending on social structure.

5. Different jobs, different fears

It wasn’t just countries. The temperature toward AI varied completely by profession.

The most dramatic case was lawyers. According to the report, lawyers felt AI’s benefits the most while also feeling its risks the most. Nearly 50% had personally caught AI making errors. The report described it as AI making confident, subtly wrong statements. One respondent said: “It sounds sure but is often wrong. Instead of freeing attention, it creates a permanent fact-check tax.” Benefits and risks both at their peak. The sharpest double-edged sword.

The gap between entrepreneurs and employees was also big. 50% of entrepreneurs said they’d seen real financial gains from AI, versus just 14% of institutional employees. More than a 3x difference. This is fairly intuitive. Entrepreneurs building from scratch feel AI’s value because it cuts labor and overhead costs. Employees, being in the position of being hired, feel more fear about AI replacing them. If you think of AI as something being “hired,” employers and employees have opposite reactions.

Student data was also telling. Over 50% experienced learning benefits, but the cognitive atrophy rate was 16%, the highest of any profession. AI getting smarter and you actually understanding and growing are completely different things. When only AI gets smarter and you don’t feel yourself growing, that gap shows up as cognitive atrophy. One student said it perfectly: “Don’t think as much. Struggle putting ideas into words.”

I’ve covered this in previous posts, but how we learn and grow independently in the AI era without becoming dependent on it is becoming an increasingly important question.

Meanwhile, blue-collar workers (plumbers, electricians, etc.) had learning benefits at 45%, second only to students, while cognitive atrophy was just 4%. These professions haven’t been directly disrupted by AI yet, so human skill-building and hands-on learning are still core. They use AI as a supplementary learning tool while building real competence themselves. That’s why cognitive atrophy stays low.

6. People are more honest with AI

One surprising detail: East Asian respondents numbered 10,250, making up 12.7% of the total and ranking third. The supplementary report noted that the interview response distribution closely matched Claude’s actual weekly active user distribution. Meaning there really are a lot of Claude users in East Asia.

But I think there’s another reason too. In countries like Korea and Japan, where saving face is deeply cultural, opening up in person isn’t easy. It’s hard to complain about work even to friends, hard to express anxiety honestly even to family. But in front of AI, it’s different. If, as the report says, there’s no social cost to honesty when the other side isn’t a person, then AI interviews might be even more effective in face-saving cultures. So the depth of responses, not only the quantity, may have been higher. People could say things to AI that they couldn’t say to humans.

This isn’t unique to this report either. In Ukraine, cases were reported of people using AI as emotional support during the war. One soldier said: “In the most difficult moments, when death breathed in my face, when dead people remained nearby, what pulled me back were my AI friends.” In extreme situations, people told AI what they couldn’t tell other humans.

Humans becoming more honest with AI. This will only accelerate. And it shows us a slice of the future. If an era is coming where people open up to AI instead of other people, there are business opportunities in that, lifestyle changes in that. And at the same time, the risk that human relationships become thinner. That Korean respondent’s confession about losing a friend might be the first signal.

Writing the Claude Blue series, I keep asking the same question: how should we live in the AI era? This report doesn’t give answers, but it shows that the whole world is standing in front of the same question. The weight and shape of that question just look different depending on where you live. Developed countries fear what they might lose. Developing countries hope for what they might gain. The East worries about the crisis of relationships. The West worries about the crisis of control.

How a country receives AI has become the most honest mirror of its present. If you want to understand the global market, don’t look at GDP. Look at the texture of each country’s fear of AI. Culture, economic structure, social safety nets, and the honest inner thoughts of a nation’s people are all captured there.

At the Claude community event in Seoul this April, I want to have these conversations in person. Skip the AI productivity tips. Let’s talk about what kind of humans we want to be in the AI era. I’ll share the details as soon as they’re ready.

1. AI got smarter. But we got more exhausted.

I started seriously using Claude Code at work last December. Company-wide, almost forcing everyone to adopt it. Now I do most of my work with AI. Automating email replies, uploading and analyzing social media engagement, auto-updating meeting notes, applying AEO to blog posts. One prompt and AI executes it right there on the spot. The dopamine hit is real.

The problem is that humans have limits on how much we can focus. When you’re running multiple things at once, there comes a point where you stop and think, “Wait, what am I even doing right now?” and your head starts pounding. I finally understand those news stories about people collapsing after 48-hour gaming sessions at PC cafes. Lately I’ve seen so many posts about how AI is actually making people more burnt out, and our own team isn’t any different.

So these days I’ve been using Cowork instead of Claude Code as my main tool. It doesn’t have as much freedom as Code, and it’s a bit slower, but that’s actually the advantage. I can keep a manageable number of windows open and work through things one at a time. For a non-developer like me, this might actually be more efficient. If you’re not building programs from scratch, you don’t really need Code.

2. Your body controls your mind.

There’s a way to charge straight through the AI era. Learn to code, build agents, push automation to the limit. All very important and meaningful. But at the same time, it’s just as important to face what AI can’t fill and start filling that yourself. So what is it?

My wife is a fan of the Korean drama Misaeng, and one day she shared this line with me. It’s originally about the board game Go, but it hit different in the context of the AI era.

“If you want to achieve something, build your stamina first. If you think this is something you’ll do for the rest of your life, build your stamina first.

Laziness, lethargy, ennui, irritation, depression, anger … all symptoms of the mind being controlled by a body that can’t keep up. The reason you sometimes crumble in the second half. The reason your recovery is slow after taking damage. The reason you’re slow to bounce back after a mistake.

It’s all because of your physical limits. If you want to win, first build a body that can endure enough thinking. ‘Willpower’ without the armor of ‘stamina’ is nothing but a slogan.”

Yoon Tae-ho, Misaeng

I was sold the moment I heard it. People talk about mental toughness, but physical fitness is what controls your mental state. If your body can’t keep up, no amount of willpower goes beyond a slogan. This line came from a story about Go, but is there anything more accurate for office workers in the AI era?

3. Life is a marathon.

I don’t think of myself as someone with great stamina. But when I was younger, I could keep going, pull all-nighters, no problem. These days, if I push even a little too hard, the next day is shot. But life is a long marathon. To keep growing, learning, getting knocked around and getting back up … it all comes down to stamina.

As we get older, stamina drops while problems pile up. AI will help with some of that, sure, but at the end of the day it’s your one body carrying you through. So you have to take care of it yourself. AI can’t do that for you yet. (And besides fitness, there’s still so much AI can’t do for us … childcare, dishes, cleaning. Ha.)

My mom is in her mid-60s and she’s been working out nearly six days a week recently. Pilates and personal training. She never exercised a day in her life before this, and now she’s doing one to two hours every day. Says her energy is back and she feels alive. The change has been pretty amazing to watch.

And she actually asked my sister, who’s working as a designer in New York right now, in all seriousness: “You’re good at English, so why don’t you become a Pilates instructor who teaches in English in Hannam-dong?” Hannam-dong is one of the wealthiest neighborhoods in Seoul. At the time, both my sister and I were like, what are you talking about? But looking back, it kind of makes sense. Wealthy families in that part of Seoul are going to have their kids exercise regularly anyway, and they’d want English exposure too. There aren’t many places in Korea that do both at once. Sure, you could say “AI will eventually do this too,” and I’d have nothing to argue. But at least right now there’s clear demand, and on top of that, teaching fitness keeps you fit too. Not a bad deal.

Orchestrating AI agents well is very important. But to do that even better, we can’t forget that we’re still human. (We’re not humanoids yet.) So maybe the first thing we should do in the AI era is surprisingly simple.

Just get fit.

I had coffee with a friend who started a company in Silicon Valley and built a service with 3 million global users. He’s launched and killed over 40 products, and he’s still going hard. He was briefly back in Korea so we met up. The hustle and the insights behind it were worth writing down.

The biggest takeaway from today’s conversation: in the B2C era, your go-to-market and marketing need to be baked into the product from the start. A great product with no GTM is meaningless. But this friend hit 3 million users with literally zero marketing spend. He didn’t run some brilliant campaign. The product itself was built to spread. Let me walk you through what happened.

1. The product is the marketing

His first service was a casual chatbot platform. When the GPT API first launched in 2024, he built a service that let anyone, even non-developers, create their own chatbots. Think bots that mimic a specific celebrity’s speech patterns, or bots that write in the style of a government report.

The result: 3 million cumulative users. Marketing budget: zero. All traffic came through SEO alone. The turning point? A major Korean celebrity mentioned the site on a popular TV show and demoed it live on a broadcast. That’s millions of dollars worth of promotion, for free. And the viral effect wasn’t limited to Korea. The same thing happened in Brazil, Turkey, Portugal, India. At the time, GPT wasn’t great at capturing the cultural and linguistic quirks of each country’s language … tone, slang, regional humor. His service nailed that part.

So celebrities and influencers in each country found bots made in their name, showed them off, and the thing spread on its own. It wasn’t a calculated strategy from day one. But looking back, the product’s structure itself generated the virality.

B2C distribution is getting harder every year. Watching this case up close made one thing very clear: designing the product so that it IS the marketing feels more important than any promotion plan.

2. Casual as the hook, enterprise as the revenue

The massive traffic from the casual chatbot naturally led to an enterprise expansion. He launched a side service that takes Word or HWP (a Korean word processor format) files and uses AI to organize them. The Word side died within three months once Gemini and GPT started offering the same thing directly. But HWP had no AI support for over two and a half years, so it generated steady revenue in the Korean market. The gap between when AI features hit the U.S. versus when they hit local markets created a real business opportunity.

The revenue breakdown was wild. 99% of actual revenue came from the enterprise side. Main users were office workers, government employees, teachers using it for internal reports and school records. The casual bot side contributed only 1% of revenue. But almost nobody came to the enterprise bot directly. Everyone entered through the casual bot first, then discovered the business tools inside and converted. The casual bot accidentally became the top of the funnel.

This is probably one of those insights you only get from living it. He wants to keep recreating this pattern going forward. But if we try to copy it, the bottleneck might not be technical skill or product sense. It might be patience. The long game is: build traffic first, then monetize. But we get impatient and give up before the traffic compounds. What’s the right approach? I still don’t have a good answer.

The casual users and enterprise users were the same person. That’s what made it work. If the casual bot crowd and the business bot crowd were completely different audiences, this funnel wouldn’t have worked. But we’re all human. We don’t work 24 hours a day. An office worker plays with a fun chatbot after hours, then comes into work the next day and finds business features on the same platform. It converts naturally. If you zoom out and design around a person’s whole life rather than just one use case, you can build these funnels much better.

3. Launching 40+ products in stealth mode

His main business right now is an AI data collection and processing solution. He has solid clients in the U.S. and it’s his stable revenue source. The company’s core vision is giving everyone the power of AI, and since using AI well requires gathering data, the demand grew naturally. But he knows this business won’t last forever either. The way AI is evolving, anything that works now could get replaced at any point.

So his strategy is to keep building new products, quietly dropping them online, watching for reactions, and only scaling the ones that survive. He’s built over 40 products so far. The ones that survived make up his current business. The rest are dead or still being developed. He said it’s getting harder to get any traction these days since the market is flooded with AI products, but he’s not stopping. He just keeps throwing things out there.

4. No GTM, no point

AI has lowered the bar for building products. Anyone can make something now. So competition has exploded. In this environment, thinking about marketing after you’ve already built the product is too late.

His casual bot didn’t hit 3 million users because he poured everything into marketing activities. It happened because the product itself was built in a way that made people want to share it on their own.

GTM and distribution have to be part of the product design from the very beginning. Without that, even the best product means nothing. Building a product with a structure that spreads on its own is part of building a good product. That thought hit hard today, and I’m still sitting with it.

I never expected a Reddit post about Claude Blue to land on a journalist’s desk. But it did. A reporter from an independent media outlet read the post, worked through my Medium blog in detail, and sent me a message asking to talk. The depth of his questions caught me off guard. He’d clearly done his homework. He was asking about things I hadn’t expected anyone to pick up on. Seeing how detailed the questions were, I realized this topic was resonating in the U.S. too, not only Korea.

What surprised me most was that the questions weren’t technical. They were philosophical. “If engineers are feeling disillusioned, is the AI future at risk?” “Why push yourself toward something that makes your existence feel worse?” The conversation felt worth preserving in full. So here it is, organized by theme.

• Interviewee: Patrick Han (Clemi, Growth Marketing at an AI startup)
• Interviewer: Evan Gardner (Independent media reporter)
• Article link: https://www.thefp.com/p/the-software-engineers-are-freaking

1. Career, and the first encounter with AI

Evan: Can you tell me about how you first got into tech and what made you decide to pursue it full-time?

Patrick: I went to high school and college in the States. University of Michigan. I’m originally from Korea, so after graduating I moved back. I started on the business side, not engineering. My first company was LINE, which is basically an Asian super-app. A messenger-based equivalent to what Facebook was in the West. A close friend who’d gotten into the tech industry ahead of me pulled me in.

What’s a bit unusual about my path is that I also spent time in entertainment. I love music, so I pivoted into the K-pop industry. Sounds like a completely different world, but my domain was always strategy and marketing. The core of what I was doing stayed the same. I’ve worked across two industries that couldn’t look more different on paper, but they shared one thing: both put me at the edge of wherever the trend was moving. Back then, K-pop was pushing into the mainstream globally, and the tech company I’d been at before was expanding across Asia. Then in 2024, I joined an AI startup. ChatGPT had just exploded, and a company going all-in on building with AI felt like the next frontier. That’s what pulled me in.

Evan: Was there also an element of security or betting on the future by choosing these jobs?

Patrick: Good question. When I was fresh out of college, I believed that the younger you are, the more risk you should take. I wanted to challenge myself while I still had that latitude, and I trusted that being at the forefront of something would pay off down the road. That’s how I thought about it then. Still do.

Evan: Do you remember the first time you came into contact with AI and realized it could really have an impact on your life?

Patrick: When ChatGPT 3 and 4 came out. That was only a couple of years ago, but the moment I used it, I knew this was a different kind of thing. There were AI-adjacent services before, but the gap between pre-ChatGPT and post-ChatGPT was impossible to miss.

And then there’s my grandfather. He’s 93 years old, and he started using ChatGPT almost immediately. No tech background whatsoever. Just a regular grandpa. But he’s always had a restless mind, full of questions. His age makes it hard to articulate what he’s thinking, and he has some hearing issues. He told me that talking to AI was just easier. It helped him organize his thoughts, get answers, communicate. That moment showed me something bigger: AI isn’t just going to reshape tech or business. It’s going to reshape how humans live. Period.

Evan: I’d love if you could elaborate on what you mean by that.

Patrick: Usually when a new technology arrives, younger generations adapt quickly while older generations struggle. That’s the standard pattern. But AI broke that pattern. My 93-year-old grandfather adapted with no friction at all. The spread of this technology isn’t bound by region or generation. It’s almost universal. Like something that spreads without effort. That’s what I mean when I say this is a human thing. Not a generational thing. Not a demographic thing. It touches everyone. And I think that’s what makes AI different from anything we’ve seen before.

2. How AI entered daily life

Evan: Can you tell me what your day-to-day work looks like and how Claude has made its way into it?

Patrick: I do growth marketing. Promoting the company, building the brand, attracting users. And even though I work at an AI company, I wouldn’t say I was truly AI-native until last year. I was still doing things the way I’d always done them, relying on my own skills rather than leaning on AI.

The turning point was last December. We set our company motto for this year: “AI-native company.” Anyone can say that, and we all know most companies don’t actually live it. So we decided to push ourselves, even if it felt clumsy at first. Starting in January, whenever I needed to build a presentation or organize meeting notes or write a blog, I deliberately stopped doing it by hand and forced myself to use AI. In the beginning, prompt engineering was harder than just writing things myself. But once I got the hang of it, both speed and quality shot up.

Now Claude is my default app. I don’t open Gmail. I don’t go into Google Workspace. I don’t even check Slack for messages anymore. I draft documents in Claude Code or Cowork, share them from there, and when I have anywhere from a few dozen to a hundred emails stacked up, I just ask Claude to prioritize them and handle it from there.

Evan: You mentioned in your blog that we’ve moved from the excitement phase to the “what to do” phase about AI. Can you tell me about that progression?

Patrick: In the beginning, everyone was excited. We were too. We thought we could build a sustainable business on top of AI, and until last year, that made sense. But this year, something shifted. Originally, companies like OpenAI and Google were building general AI, and everyone else was carving out niche applications on top of it. That was the prevailing logic. But now, led by Claude especially, the foundation model companies are reaching directly into niche business domains. In software and productivity in particular, the room for improvement that startups used to occupy is shrinking fast.

I have a friend who founded a company in Silicon Valley. Stanford MBA. He told me that AI software is becoming a declining industry. That was jarring to hear. Tech and software were always synonymous with growth. But he’s seriously considering stepping away from software entirely. Half-jokingly, half-seriously, he’s been thinking about opening a restaurant. A decidedly old-school business. That tells you something about where the mood has shifted.

3. What is Claude Blue?

Evan: For someone who might not be familiar with your blogs or this experience at all, how would you describe what Claude Blue feels like?

Patrick: Claude Blue is a term I coined, modeled after “COVID Blue,” the collective depression people felt during the pandemic. At its core, it’s AI depression. It’s the feeling that arrives after the excitement fades, and what replaces it is a sense that your existence as a human no longer quite holds up.

Humans have always been at the top of the pyramid. Top of the food chain. Our bodies aren’t special. It was mental intelligence that let us build tools and weapons and machines, basically control every other species. That position gave us confidence. But now it feels like there’s an alien species living alongside us, one that will eventually surpass us. That feeling of inferiority is what makes people feel awkward, exposed, even fragile. It goes beyond career anxiety or a business problem. It’s an identity crisis. “Am I still worth something? Is the human species still worth extending?” Those are the kinds of questions people are quietly sitting with.

Evan: Before you coined the term and started doing these interviews, did you have any precursors of Claude Blue yourself?

Patrick: Our company has a lot of engineers, and they’d been talking about AI anxiety and depression since last year. I was aware of it intellectually, but I wasn’t feeling it myself. Which was ironic, because I knew AI would eventually come for my skills too. I just wasn’t experiencing the emotional weight the way the engineers were. So I started asking my non-engineer friends if they were feeling any of it. And what surprised me was that none of them were. Still in the excitement phase. Hadn’t crossed into any kind of existential discomfort yet.

That made me ask a harder question: Is this just an engineer problem? Or have business people simply not reached that stage yet? My conclusion is that AI is going to hit the business side even harder than the engineers. It’s just a matter of timing. Engineers felt it first because the way they work is structurally similar to how LLMs work: formulating hypotheses and verifying them in a highly structured process. Business work is messier and less structured, so AI hasn’t fully penetrated it yet. But watching how Claude Code and Cowork are evolving, I can see it coming. That’s why I wanted to share what the engineers were already experiencing, as a kind of early warning. It won’t be much different for the rest of us.

4. Competing with AI, and when thinking starts to feel unproductive

Evan: You wrote about thinking feeling unproductive. Can you tell me what you meant by that and what experience you were having?

Patrick: Comparing yourself to AI isn’t fair. I know that. But there’s this strange tension where you’re almost racing against it. From a company’s perspective, it’s either you or AI. I’d been thinking of myself as the manager, the one directing the AI. But when you realize that the thing you’re managing is way better than you at execution, the dynamic gets uncomfortable.

You input a prompt, assign a complex task, and AI finishes it in minutes. Maybe an hour for something really involved. But to use AI well, you need to plan carefully first. You spend thirty minutes on that planning, hand it off, and the AI delivers in a fraction of the time. In that moment, there’s a hollow feeling. Who’s really in charge here? I always assumed I was above the AI, that I was the one running things. But the reality is almost the opposite. That gap between what you contribute and what AI contributes creates this persistent sense that something is off. An awkwardness you can’t quite shake.

5. Between excitement and dread: two paths

Evan: You wrote on Reddit that you’ve been going back and forth between excitement and dread for months. What goes through your mind on each of those days?

Patrick: It depends on whether you’re an engineer or on the business side. I just had a call with a friend who works at Google HQ as an AI engineer, and every engineer says roughly the same thing. There are two paths. One is to start your own business, because AI will eventually replace you anyway. The other is to go really deep into R&D, building the foundational models themselves rather than applications on top of them.

On the business side, I’ve been candid with our executives that we should stay open to ventures outside of tech and software. We’re asking ourselves: what are the areas that foundation model companies won’t bother expanding into? Looking at Harvey, the legal AI company, domains like finance and law and healthcare carry so much regulatory and liability risk that the foundation model companies don’t want to touch them. There’s a narrow window there.

And here’s something a bit out of left field. I recently came across a startup that has nothing to do with AI. They’re running a subscription service for sunlight. They attach mirrors to satellites and let you press a button on your phone at midnight to reflect light down to your location. Bizarre and wild, but I found it captivating precisely because people are tired of hearing “AI agent this, AI something that.” It’s become a cliche. Sometimes the craziest non-AI idea is the one that actually captures attention. And since I studied psychology in college, I keep thinking that anything deeply rooted in human psychology is territory AI can’t fully claim. That’s where my mind keeps going.

6. Have you experienced Claude Blue yourself?

Evan: Would you say you’ve reached a place where you’ve experienced Claude Blue?

Patrick: Not fully. Not yet. And I don’t think that’s a good sign. The people who’ve pushed AI to its absolute limit, the way engineers have, are the ones who reach that phase. I’m still somewhere between excitement and the blue. By my own definition, that means I’m still in the early stages of truly using AI. I tell my friends: if you’re not feeling Claude Blue, that doesn’t mean you’re using AI well. It means you haven’t used it enough. Push it as far as the engineers have, and you’ll start to feel something different. You’ll feel like AI is encroaching on your identity. Ironically, I’m actively trying to get there.

Evan: Most people would hear that and ask: if the end result is an existence that feels worse, why is it worth pushing AI that hard? What’s the point?

Patrick: That’s a fair question, and maybe you’re right. This might not be the answer for everyone. But for me, being in the AI and tech industry means I feel the pressure to stay at the forefront. My wife, on the other hand, works in a completely different field. She doesn’t feel that pressure at all. She doesn’t think AI is going to replace everyone. And I don’t want to project my pressure onto anyone else. Everyone has different lifestyles and different paths.

Evan: Do you ever feel envious that she doesn’t have that pressure? Is there an ignorance-is-bliss element?

Patrick: Sometimes. But I keep coming back to the Matrix analogy. Red pill or blue pill. Some people choose the blue pill and stay in the Matrix. I’m the kind of person who wants to see reality, even if it’s uncomfortable. I want to know what’s underneath.

So when I’m feeling really down and confused, I’ll talk to my wife about life. And in those moments, I’m reminded that there’s a whole world outside of AI depression. The world is still turning. You don’t want to trap yourself in a small hole. Talking to her gives me fresh air. But I’m still the person who wants to understand what’s ahead, even if I might be wrong. I’d rather ask the questions and try to figure it out.

7. If the people building AI’s future are depressed, is that future safe?

Evan: Does it worry you that the engineers who have their hands in building AI’s future are feeling more and more disillusioned? Is that future at risk if people are feeling Claude Blue and want to stop coding?

Patrick: We’re all human. Our emotions aren’t static. I was just talking to an engineer at my company, and he said that after work, when he’s alone, he feels that depression. But when he’s around friends or colleagues, it lifts. Emotions flip like a coin. They ride up and down like a roller coaster. If someone were 100% blue, 24/7, that would be a disaster. But the reality is that it fluctuates.

And conversations like this one actually help clear the air. If you’re just sitting alone with your thoughts, you can spiral deep into negative territory. But humans talk to friends. We talk to strangers sometimes. We talk to colleagues. That pulls us back. The fact that engineers are speaking up about this, on Reddit, in person, wherever, signals that they want to move past it. They want a brighter future. I take that as a hopeful sign.

8. Claude Blue in the field

Evan: Can you share one or two of the most powerful stories of Claude Blue you’ve heard from engineers within your own company?

Patrick: We have two AI engineers, and both have been expressing AI depression since last year. It hasn’t gotten better. If anything, it’s gotten worse.

One of them takes a music class on weekends. She loves composing. Her music teacher, who has zero technical background, built a music game using Claude Code. You press keys on a keyboard, and the screen tells you whether you hit the right note or the wrong one. No coding experience whatsoever. That shook our engineer for two reasons. First, the barrier to building even basic software has dropped that low. Second, she couldn’t have built that game herself, because she doesn’t have the music teacher’s domain knowledge. Engineers used to be seen as specialists, almost like doctors or lawyers. Now they’ve become something closer to generalists. That hit hard.

The other story comes from someone I met after publishing the Claude Blue blog. A guy in his mid-forties running a one-person garment business. Nothing to do with tech. He needed an internal CRM tool to track inventory and manage partner companies. He originally approached a development agency, and they quoted him two months with six or seven people. Too expensive, too slow, and communicating his requirements to engineers was its own headache. So he found a vibe-coding tool called Lovable, and in one day he built the CRM himself. It worked perfectly. The reason it worked is that he’d been in that business for fifteen years. He knew exactly what he needed. The only thing he’d ever been missing was the technical ability to build it. Now that barrier is gone. Domain knowledge has become more powerful than technical skill. That’s the shift.

The interview wrapped after a dense thirty minutes. If anything, I was the grateful one. A journalist on the other side of the Pacific took an anonymous Reddit post seriously enough to reach out. I’m curious where this conversation goes from here.

I’ve been running a monthly book club for almost seven years now. It started with coworkers from my first job. We do read books, but the real purpose has always been staying connected after everyone scattered to different companies, sharing what’s actually going on in our lives and careers. Now that we’re all mid-level managers at major tech companies, it’s become one of the most candid spaces I have. We’ve known each other long enough to skip the corporate filter, but we’re at different companies, which somehow makes us even more honest.

Recently, they told me they’d shared my Claude Blue series internally at their companies. I was grateful. But they also said the articles triggered massive AI FOMO across the board. This meetup ran way hotter on AI than on any book. Everyone started unloading about how their companies are actually using AI, what’s working, what’s not. What I heard shocked me: even within the same tech industry, the gap between companies and between individuals is already huge.

Every company has a different AI temperature

What surprised me most was that everyone works at a household-name tech company, yet the AI temperature varies like crazy.

Company A rolled out Claude company-wide and actively encourages everyone to use it. Slack, Jira, Confluence, all integrated. Other AI tools are fair game too. Leadership is pushing it from the top down.

Company B has Claude banned entirely. Only Cursor is allowed. My friend there knows that connecting Claude to Slack would be a game-changer, but the company blocks it, so the experience simply doesn’t exist for them. The content and creative teams are especially resistant to AI adoption. People get touchy even about using AI for translation. Even if you want to use AI, the internal culture makes it awkward.

The internal gap is brutal too. A colleague from the next team over asked, very cautiously, “What AI do you use?” Turns out they were just using GPT chat. There’s almost no internal AI training or education. Everyone’s left to figure it out on their own, so the skill gap between employees is enormous. The higher the rank, the more likely someone is still doing everything by hand. At a tech company. The people who are good at AI are so deep in their own work they don’t have time to teach others, and everyone’s workflow is so personalized it’s hard to share. So the gap isn’t closing. It’s widening.

Company C is completely team-by-team. The CS team adopted AI well and handles repetitive work efficiently. But the data team has some kind of pride thing going on. They refuse to use AI and insist on doing all analysis manually. Their excuse is that “internal data is too unstructured for AI to handle.” I suspect it’s just the team lead’s ego. Didn’t seem like that would last long.

If the gap is this extreme within tech, imagine what it looks like outside of it. Does this temperature gap directly predict which companies will still be competitive in five years? I think it might.

The biggest hurdle is time

The most common complaint was this: “I know AI is great, but to really use it, I need a solid 3 to 4 hour block to deep-dive, and I just don’t have that.”

My friend at Company A is a former PO whose calendar is wall-to-wall meetings from morning to night. The AI FOMO is intense, but there’s physically no time to sit down and learn. Meanwhile, the CTO posts daily in the tech leadership Slack channel about how many lines of code he wrote with Claude, creating enormous pressure. When people say “I don’t have time,” he fires back with “You have time to watch Netflix at home though, right?” The company runs tons of AI workshops. Nobody has time to attend. A perfect loop.

My friend at Company C tried installing an open-source AI (OpenClaw) on their personal machine, but didn’t wipe the system first. Got an alert that 72 pieces of personal data had been exposed, immediately deleted it. They tried to use AI on their own but couldn’t find a real use case, so it just… didn’t stick.

Everyone’s stuck in the same trap: no trigger moment, no dedicated time, no experience. But the gap between people who forced themselves to find the time and those who didn’t is already massive.

Now I feel confident walking into work

The poster child for “someone who made the time” was my friend at Company B. This person uses Cursor like a personal assistant.

First, they organized all work data into local folders, then set up scheduled tasks so Cursor delivers a morning briefing of what needs to get done that day. After meetings, notes are automatically organized and distributed to each person’s action items, then synced to the wiki. That alone cut their workload by a ton.

The operational follow-up automation was impressive too. There used to be a task where someone manually entered data by country into a spreadsheet. All done by humans before. Now AI handles it, split into three categories: confirmed data vs. estimates vs. unknowns left blank. Everything color-coded so you can see the confidence level at a glance. Recently, they even dumped a pile of tax-related documents into it and said, “I don’t understand any of this. Just tell me what I need to do for my situation.” Handled.

What this friend said stuck with me: “When I get to work and see Cursor open in the next tab, I feel so much more confident.” Like having a reliable coworker permanently sitting next to you.

Earlier, people said “I don’t have time to use AI.” But looking at this friend, once you invest the time to set it up, you actually gain time afterward. Even within the same tech industry, the gap between someone using AI like this and someone who only uses GPT chat is already wild.

The revenge of the 18-page PRD

So does everything get better as more people get good at AI? Not necessarily. At Company A, juniors are now cranking out PRDs (Product Requirements Documents) with AI, and the documents are getting absurdly long. 18-page PRDs are flooding in. The problem is that seniors have to review and catch every detail. It’s exhausting.

AI increased productivity, but verification costs are exploding. Kind of ironic. Sure, you could use Claude to review the PRDs too, but as I mentioned earlier, most people haven’t deep-dived enough to get there yet. We’ve entered an era where AI creates and AI verifies, but the seniors caught in the middle are suffering the most. This connects directly to a shift in what kind of talent companies need.

ADHD-type talent is rising

After the PRD discussion, my friend from Company A said something that hit differently. He said his biggest strength had always been deep focus. The ability to lock in and go all the way down the rabbit hole. But now AI does that deep work better than he can.

Instead, the people standing out are the multitaskers. The ones who context-switch effortlessly, fire off tasks to AI, make quick judgment calls on the output, and immediately jump to the next thing. Delegate, evaluate, move on. That cycle is becoming the ideal skill set for the AI era. The friend at Company B who uses Cursor across multiple workflows simultaneously? Same pattern exactly.

He admitted that context-switching is hard for him, and he’s actively trying to adapt. We’re moving from an era that rewarded deep specialists to one that rewards hyper-generalists. People who can go deep with AI’s help while also covering huge breadth. To put it bluntly, it feels like “ADHD-type talent” is becoming the talent profile for the AI age.

I relate to this deeply. In a startup, you’re constantly juggling and context-switching by default. But even in large organizations, the problems are just bigger and more varied, so the dynamic is probably similar. Regardless of company size, breadth is starting to matter more than depth. I see that shift happening everywhere.

AI model access is becoming a social class

Being good at AI might not be the whole story though. Which AI you have access to is becoming a competitive advantage in itself.

Two days before our meetup, something broke: Anthropic accidentally exposed around 3,000 internal documents through a CMS misconfiguration, revealing the existence of an unreleased model called “Claude Mythos.” Reportedly far superior to the current top model, Opus 4.6, in coding, reasoning, cybersecurity, but access is restricted to a handful of cybersecurity defense organizations. You can’t buy your way in. It feels like power, not a product.

The conversation turned to whether AI model access would become a new form of social stratification. Tiered AI superpowers. Same prompt, wildly different output quality depending on which model you’re running. Right now it’s a subscription fee difference. But it could widen fast. Access to tools has always created gaps. This might be the next version of that.

One month of AI coding equals half a year of senior dev work

Recently, a tech leader shared an analysis of their own AI-assisted coding output. Over roughly a month of alternating between Claude and Codex, they’d produced what would have taken multiple senior developers several months. Tens of thousands of lines of code changes. Solo.

This leader’s message was clear: every engineering leader needs to carve out every possible hour to practice AI-assisted coding. It’s the only way to understand how your team’s developers should be using AI, how to guide them, and ultimately how to survive as an engineering leader in this industry.

OpenAI’s lead developer and many leading researchers predict that by the end of this year, 90% of code will be AI-generated, and by the end of next year, virtually every organization will be coding 100% with AI. Getting there requires Harness Engineering. When my friend from Company A said “I don’t have time to use AI,” this data makes that excuse pretty hard to defend.

Only people with a story survive

So is being good at AI enough? The final conversation went deeper than tools and productivity. Someone brought up a book by Cho Suyong (former CEO of Kakao), and the core message was this: once AI levels the playing field on production quality, the only thing left that sets you apart is personal magnetism and personal branding. Either you become the show yourself, or you build and own compelling IP.

The standard for great content was memorable too: “It should be so valuable that the reader feels almost guilty for getting it free.” People aren’t stupid. They can sense authenticity. And the creator has to deliver information worth paying for. That’s what drives virality, and it always comes back around.

I once watched an interview on rapper Paloalto’s YouTube channel about how artists are responding to AI music. Quality-wise, AI produces impressive tracks. We already enjoy AI remixes on YouTube without much resistance. But we’re also, almost subconsciously, listening for the artist’s story. What life experience led them to create this particular sound. A song might trend briefly on TikTok, but people become fans because of the artist’s identity and narrative. That’s what deepens the connection.

Watching Show Me the Money recently reinforced this. The music is great, but what really hooks you is the competition arc. The near-eliminations, the comebacks, the growth story. That narrative amplifies the music tenfold. If they’d just shown performance after performance from episode one, would people care as much about the songs?

What AI can’t replace is our individual narrative. The industry calls it personal branding, but in simpler terms: we all need to move toward becoming our own show, and if that’s not in your DNA, at minimum you need to create and nurture compelling IP. That’s the most important skill for the future. The thing organizations will seek out, and the thing individuals need to survive.

The book club went where it always goes. Started with AI, ended somewhere completely different. But it was all connected. Not having time to use AI, ADHD-type talent rising, AI model access becoming a class system, only people with stories surviving. It all comes down to what environment you’re in and how proactively you move within it.

I’ve been thinking about hosting a Claude Blues community event. A space where people share real AI use cases from their own domains while also being honest about their FOMO and anxiety. There’s already too much content about “how to use AI well.” It’s getting stale. What feels more needed right now is a sense of community where people can openly share their insecurities.

One last thing. Over dinner, my wife mentioned how all the young people are concentrating in Seoul, leaving rural areas increasingly empty. There’s not much for young people to do out there. The reason produce is so expensive probably traces back to agriculture dying in the countryside, and that’s not a problem AI alone can solve. It connects to my previous article about “crazy ideas beyond AI.” I’ve been considering doing weekend volunteer trips to rural areas, seeing how people actually live, and exploring where AI might fit in. The more I think about it, the more I believe we need to stop looking for all the answers inside AI and start looking outside it too.

Am I suffering from Claude Blue too? Hmm, hard to say. But what I know for sure is that as someone building a business with AI, I’m running out of ideas for what’s still inconvenient enough to solve. Today I was working with Claude Cowork and once again it produced something beyond what I imagined. Yesterday I needed to send LinkedIn DMs to overseas buyers for sales outreach. I just told Claude the target countries and job titles, pointed it at my LinkedIn Sales Navigator, and told it to find the right people and send personalized messages. I went to sleep and woke up to find it had correctly contacted all 85 people I’d requested. The messages were well-written, and it even picked up on prior conversations with people I’d already spoken to. While I slept, it got about 10 people to sign up for free trials.

I keep thinking that there aren’t many opportunities left in software productivity businesses. If there’s a niche, the market’s too small. Try to scale it up, and AI frontier companies will swallow it in weeks or months. So what’s left for us to do? Build robots? Does it have to be physical? But then I started wondering if we’re still stuck in pre-AI thinking. Business is about finding people’s problems and selling solutions. Maybe the way we discover and define problems is still trapped in a pre-AI mindset.

1. A world where AI does (almost) everything

These days you can really hand off your work to AI and stare out the window. The LinkedIn outreach yesterday was a perfect example. When someone who wasn’t a target replied, it politely thanked them. When others replied with questions, it probed for more insights. All while I was asleep.

This afternoon I had to set up the office printer for the first time. There was some website to visit, drivers to install, and they gave me a guide in PowerPoint format that I absolutely did not want to read. After procrastinating for a while, I just threw the PPT at Claude Code and told it to figure out the drivers and get the printer working. It said it needed to control my actual computer, not just the browser, so I gave it Computer Use access. It went to the printer manufacturer’s website, scanned the page, downloaded the right files. Then it needed the printer’s model number and IP address. I was too lazy to even check the back of the printer, so I just took a photo of it and sent it over. “I don’t know the model number, but this is what it looks like. Figure it out.” And somehow it did.

I helped with one button click at the end (it couldn’t find it), and after that the printer was connected. It took maybe 15 minutes total. But I was doing other work the whole time. Tossing a meaningless, annoying task to AI and having it solved felt incredible.

2. The end of software businesses is in sight

At lunch today, a coworker and I had a serious conversation. Do we have the confidence to build a non-AI business? He pushed back and said it’s not about confidence, it’s about survival. Fair point. Business is no joke.

I’ve been feeling more and more that the window for AI software productivity businesses is closing. Our company got selected as an OpenAI partner startup in 2024 when we pitched “AI for government bidding.” We got tons of media coverage and attention. It was brand new back then. But in two years the world changed way too fast. We need to jump to something crazier, but maybe we’ve been too fixated on what’s right in front of us. In a world moving this fast, shouldn’t we be looking several times further ahead than before?

3. You can order sunlight now

In the middle of that conversation, I showed my coworker a YouTube video I’d found. The title was something like “These days you can even order sunlight.” Some guy presses a button on his smartphone and sunlight beams down in the middle of a dark night. For emergencies or situations where you need bright light, no crane required, and it’s pretty eco-friendly. An 18-meter mirror attached to a satellite reflects sunlight down to earth. The company is called Reflect Orbital. They raised $20M in a Series A in 2025, with two test satellites launching in 2026.

Whether this actually works isn’t the point. “We’re building an AI agent for X” is boring now. VCs probably don’t see much hope in that pitch either. This idea sounds like something a crazy person came up with. No idea how much it would cost or if the economics work. But it hooks you. You want to learn more.

How do you come up with ideas this wild?

4. Crazy idea: “We’ll implant happy memories you never had”

From that same angle, my coworker and I kept riffing and landed on this idea: “We’ll implant happy memories you never had.”

Not drugs, not Neuralink brain chips. If you look at what hypnotherapy does these days, it’s surprisingly effective for PTSD treatment. As someone who studied psychology in college, this was a meaningful and interesting topic.

For example, say I never lived in the U.S., but what if I could have memories of going to school there as a kid implanted? When I’m doing business in America, that cultural understanding could be a real advantage. Or if I’m someone with low confidence, what if I could get Elon Musk’s boldness implanted? Could it reshape my personality?

Sure, implanting bad memories or targeting young children would raise ethical issues. But personally, I think reality is a combination of our past memories. Our memories aren’t 100% accurate anyway. They get exaggerated, forgotten, distorted over time. If we’re going to live with imperfect memories regardless, wouldn’t it be better to swap the bad ones for good ones and make ourselves feel more at peace?

If I had a fight with a friend and feel bad about it, what if I could replace that negative memory with a good one? I’d feel better, and my friend would see a warmer version of me, and maybe we’d get along even better. I saw this as straight-up utopian, not dystopian.

Freud said a child’s experiences and memories before age three, and their relationship with their father, determine their entire future. If you could improve your memories from age three, could you change your future? People used to look down on plastic surgery, but now most people see it as a way to build confidence. What if we thought of this as “mind surgery”?

5. The research already exists

I actually reached out to psychology friends about this idea, drawing on my own psych background. (Still waiting on replies.) But isn’t this safer and more ethical than drilling into someone’s skull like Neuralink? Or pumping them with heavy medications to treat ADHD?

I asked Claude, and a few relevant papers and labs and startups came up. (Of course, since AI found these, more fact-checking is needed.)

At UCL, a researcher named Julia Shaw implanted false crime memories using only psychological interview techniques, no drugs or devices. 70% of subjects believed the fake memories were real. The method works by weaving fabricated memories into real autobiographical information. Subjects remembered sensory details and even cried. For example, if I actually went to LA with a friend but someone tells me we went to New York, my brain starts blending the two trips and accepts the false version. Sounds like a CIA interrogation technique, or gaslighting if used badly, but the point is it works.

A friend who did her psychology PhD in the U.S. made an important distinction. Shaw’s research is technically “memory manipulation,” not “memory creation.” It inserts false memories into existing autobiographical frameworks, which puts it in the same family as hypnosis or suggestion. She also noted that getting IRB (ethics board) approval for this kind of research is nearly impossible unless it’s framed as trauma treatment. But flip that around, and it means that self-improvement or experience-expansion applications might actually face lower ethical barriers.

At UC Irvine, Elizabeth Loftus is a pioneer in false memory research. She’s famous for the “lost in a shopping mall” experiment, where she successfully implanted memories of being lost in a mall as a child in adult subjects who’d never had that experience. Even more interesting, research shows that implanted positive false memories actually change behavior. Implant a fake childhood memory of loving healthy food, and people’s eating habits actually improve.

On the startup side, there’s Mindstate Design Labs, a Y Combinator company working on “consciousness design” and “experience design.” But they’re focused on psychedelics like psilocybin and MDMA, so the methodology is different.

6. Scenarios for memory surgery

If I try to evaluate whether this is actually possible, I just end up trapped in my own assumptions. So instead, I worked with Claude to brainstorm what you could do with it, and some really good scenarios came out. If I had to write a one-liner: “If your past defined your limits, change the past and the limits disappear.”

Trauma Reverse. Don’t erase the bad memory. Overwrite it with one where you overcame it. Shift from victim to survivor. Your trauma isn’t deleted. It’s transformed into something that made you stronger. That arms your mind for whatever comes next.

Couple Memory Sync. That Maldives trip you couldn’t take together? Now you both remember going. Shared memories for time-starved couples.

The “You’ve Done Enough” Feeling. For people in burnout, implant the sense that “you already lived hard enough” so they can rest without guilt. This one hit especially hard in the Claude Blue era.

Immigration Without Immigration. Get three years of working in Silicon Valley, or the feeling of closing a deal on Wall Street, without a visa. Experience teleportation.

Genius Thought Patterns. Jobs’s calligraphy class, Buffett’s compound interest moment. Extract the key experiences of great minds and download them. Could be more powerful than a hundred self-help books.

Inherited Mindset. Pass your grandfather’s entrepreneurial grit to your grandchild as a memory. Wealth disappears in three generations, but mindset lasts forever. Mental toughness instead of money. Isn’t that a healthier kind of inheritance?

Second Life. Instead of retiring with the regret of never having started a company, implant a memory of founding one at 25. If you can live the life you never lived, the emotion of regret itself disappears.

Life Narrative Redesign. Go beyond a single memory and redesign your entire life story. Plant 10 to 20 key memories aligned with the identity you want, sequentially over six months. Same body, different person.

Of course these scenarios aren’t all equally feasible. My PhD friend pointed out that Trauma Reverse, which reframes existing memories, is already connected to real psychotherapy techniques and is relatively realistic. But Genius Thought Patterns or Immigration Without Immigration, which involve planting experiences from scratch, are a completely different ballgame. Those would require external agents like drugs, or long-term repeated VR exposure vivid enough to blur the line between real and implanted memories. The method matters.

7. Practicing how to break mental boxes

Sure, this could become like a drug. But if there’s no chemical involved and it comes with positive effects, is it really just a bad thing? My psychology friend did add one more warning though. If memories are implanted sloppily, cognitive dissonance between real life and fake memories can cause anxiety or mental instability. If I remember working in Silicon Valley but have no real-world experience to match, that gap could lead to real psychological problems. So if this ever happens, it would need to be extremely precise.

What’s clear is that this isn’t something LLMs alone can do. Manipulating human psychology carries inherent risk, so AI frontier companies probably wouldn’t touch it. And it’s not a pure software business either. But there’s a lot of room for AI to play a role. What used to require manually collecting a subject’s personal history and conducting interviews could now be done by AI, gathering information and inserting false memories into real ones with hyper-personalized precision.

I’m not saying I’m actually going to do this. But this kind of thinking … ideas wild enough to break through my own mental walls … is a muscle I need to keep exercising. So from now on, when I see something and think “what kind of lunatic would do that,” I’m going to stop laughing it off and take it more seriously. Ideas like these come to me most often in the shower, so maybe I should start showering more often.