Milwaukee is the smallest market in baseball. About 1.6 million people, roughly one eighth of Los Angeles and one twelfth of New York. The Brewers just sold more postseason merchandise than any team in baseball history.

More than the Dodgers in either of their last two title runs. The previous record fell in the first forty-eight hours after the Brewers won the pennant and before the World Series even started, and the team store at American Family Field closed to restock before Game 1.

Los Angeles had every advantage and Milwaukee outsold them anyway. The Dodgers were back-to-back champions in one of the two biggest markets in baseball, with Shohei Ohtani, the most marketable player alive, whose jersey broke Fanatics’ launch record the week he signed.

Two days after winning the championship, police estimated more than 750,000 people were at the parade and rally. The Dodgers drew around 250,000 for theirs last year from a region nearly seven times the size.

Getting into the stadium cost more too. The cheapest ticket for Game 1 at the Brewers’ stadium, American Family Field, went for $2,338, the highest entry price for any World Series game ever played. The previous record was set at Wrigley in 2016, for the Cubs. Tickets at Tropicana Field went for a quarter of that. Brewers fans could fly to Florida, book a hotel, and watch a game for less than it cost to get into their own ballpark.

Four months earlier the Knicks did the same thing in the NBA. In the first twenty-four hours after beating the Spurs, they sold more than any champion in any sport had in that window on Fanatics. More than the Eagles after the Super Bowl. Double what the Lakers did in 2020. Eight thousand orders a minute at the peak.

The comparison matters because it removes the obvious explanation. New York is the biggest market in American sports. Milwaukee is the smallest in its league. One has every structural advantage and the other has none, and they produced the same result in the same year. Whatever is driving this, it isn’t the size of the city.

What the two teams shared was a wait. Fifty-three years for the Knicks. Fifty-seven years and no title at all for the Brewers.

Everything is up, so what

Everyone is setting merch records right now. Adidas says its World Cup jersey sales quadrupled from 2022. Concert merch spend per head has climbed every year since 2023, and Justin Bieber did five million dollars at Coachella in a single weekend. Licensed sports merchandise is a forty-billion-dollar global market and growing.

The boom is real, but it applies to everyone. The Dodgers were selling into the same market with a bigger city and a bigger star, and they set records too. Milwaukee still passed them. Growth in the category explains why the numbers are large. It doesn’t explain why the smallest market ended up with the biggest ones, and it doesn’t explain why nearly half the metro area’s population stood outside in Milwaukee in late October for a parade.

What a jersey actually buys

A ticket gets you into the building for something that happens once, if you can afford one. At those prices, most Brewers fans couldn’t. A jersey is the version of being there that’s open to more people. You don’t need a seat to wear it. When the series moved to Florida, the Brewers sold 31,000 tickets to watch parties at their own empty stadium for each of Games 3 and 4, fifteen dollars apiece, gone in about twenty minutes and reselling for hundreds. Then they did it again for Game 5. People paid to sit in the building and watch a screen. The jersey is the same purchase without the seat.

I wrote earlier this year that as AI makes content nearly free, value moves to what can’t be copied, the experience itself and the physical evidence you were there. The Brewers adds something else. AI doesn’t just make things abundant. It makes most things instant. What it can’t do is give you fifty years of history, or the experience of waiting that long for something that might never happen.

I notice it in smaller ways too. Getting a reservation at a good restaurant in New York City has become harder than it’s ever been. But not every hard reservation is the same. Some are hard because a chef has spent years building something and refining their craft, and some are hard because a video went viral on TikTok. People can tell the difference.

A drought is the same mechanism, stretched over decades. Every season without a title adds to a balance that can’t be spent until the team wins. When it finally does, the whole balance pays out at once, and the longer the lockup, the bigger the payout. For almost a decade the biggest-selling champion in any sport was the 2016 Cubs, after a hundred and eight years. Then the Knicks took a run at that record in June, and the Brewers broke baseball’s in October. Every one of them a drought.

It showed up on the road too. Brewers fans nearly outnumbered Rays fans at Tropicana Field. Knicks fans filled so much of the Frost Bank Center for the first two games in San Antonio that the Spurs’ rule limiting sales to buyers within a hundred and fifty miles became a national story before Game 5. Cleveland had the same problem in 2016, when Cubs fans filled their building for Game 7.

Manufactured waits

Leagues and brands know all of this, and they try to replicate this demand. Limited edition drops with a countdown clock. Throwback nights. City Connect uniforms. Topps Now cards that commemorate a single moment from a game.

It works for a few days. The spike is real and then it’s gone, because scarcity on its own isn’t the product. Scarcity with a story behind it is. A countdown clock tells you supply is limited. A fifty-seven-year drought tells you something happened. And the easier it gets to manufacture the first kind, the more the second kind stands out. You can design a drop in a supply chain. You can’t design fifty years of people showing up to a building where nothing good happens. Time is the one input nobody can pull forward.

The setup

If the payout only comes from waiting, then nobody can plan for it. No front office is going to lose on purpose for five decades. No brand is going to tell customers to come back in a generation. Every incentive in those businesses points toward faster.

The Brewers won with a payroll a fraction of the Dodgers’. This was the first World Series since 1991 where both teams sat in the bottom half of the league in spending. The Knicks were led by a second-round pick who took less money to stay. With a lockout looming, baseball’s owners are about to spend the winter arguing that small markets can’t compete without a salary cap, a month after the smallest one won.

Franchises are valued on market size, media rights, and recurring revenue. The Brewers were worth about two thirds of the league average going into this season. Then the smallest market in baseball produced the most expensive ticket in World Series history, the biggest merch haul, and a crowd the size of half its metro area in the street. None of that shows up in a population count.

What showed up was a story. A team with a fraction of the Dodgers’ payroll, in a city that had waited longer than most of its fans had been alive, beating the richest team in the sport on the way. The drought wasn’t the point. It was the setup, and the setup is what made the ending worth paying for.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Most of what I write here is about the digital world and how AI is changing it. We spend most of our days on screens, with content thrown at us constantly. My hope is that AI takes enough of our digital lives off our plates that we get more time for the physical world. I’m going to write more about that here too.

My grandfather settled in Maine when he immigrated from Japan, and he’s lived there ever since. My father grew up there. I went to college there, and unfortunately I haven’t been back much since. This summer I finally went back and drove up the coast from southern Maine to Bar Harbor.

Few places are as beautiful as Maine at the end of summer. There’s Acadia, the northern end of the Appalachian Trail, lakes and rivers deep in the woods, and a coast that makes it easy to put your phone away. The ocean is as warm as it gets, and by late September the leaves start turning before the long winter sets in.

The coast is harder to drive than it looks on a map. Every peninsula is a long detour down and a long detour back, and the good stuff is almost always at the bottom. What I kept pulling over for were lobster pounds. I’ve eaten in a lot of beautiful places, but nowhere pairs the setting with the food this consistently. Many have been run by the same families for generations, usually on a wharf with a view that would belong to a resort anywhere else.

Lobster wasn’t always a luxury. In colonial New England it was fed to prisoners and servants. Now it’s the most expensive thing on most menus in the state, and this year the price climbed again as fuel for the boats neared record highs.

At every stop I had to choose between a lobster roll and a whole lobster. The roll is easy. The whole lobster makes you work for your meal, cracking it apart piece by piece.

Not all of these are lobster pounds in the strict sense. They’re the ten places I’d send someone to eat lobster in Maine.

1. McLoons Lobster Shack, South Thomaston. The setting can’t be beat, and the lobster is the best I had on the trip, whole or in a roll. The shack sits across from the oldest working lobster wharf in the area, so you can watch the boats unload what you’re about to eat. Mayo or butter, it doesn’t need much. Add steamers and oysters, then walk over to the separate dessert stand, which is a reason to come on its own. It’s a long drive down the peninsula to get there and that’s part of it.

2. Five Islands Lobster Co., Georgetown. Another pretty harbor setting, working boats and all. More touristy than McLoons, which is the only reason it isn’t higher.

3. The Clam Shack, Kennebunkport. My go-to in southern Maine, in one of the nicest towns on the coast. The steamers are great, and it doesn’t hurt that the owner went to my college.

4. Thurston’s Lobster Pound, Bernard. On the quiet side of Mount Desert Island, a few minutes from Bass Harbor Head Light and just outside Acadia. You order at the window, take a number, and eat on the deck over a working harbor. Do the lighthouse first, then lunch.

5. The Maine Lobster Lady, Sedgwick. Not a traditional pound, more a counter with a screened porch and picnic tables, run by a mother and daughter. Great lobster roll, and get the blueberry pie.

6. LDI Lobster, Little Deer Isle. Right by the big green suspension bridge over Eggemoggin Reach. The view from the dock is the reason to go.

   

7. Red’s Eats, Wiscasset. A tiny shack right on Route 1 at the bridge, with a line that backs up traffic through town. The roll is famous for holding more meat than a whole lobster. Do it once. Others on this list taste better, but Red’s is about the experience, and it’s the easiest stop here to hit without leaving the main road.

8. Young’s Lobster Pound, Belfast. A big red barn on a working wharf looking out over Penobscot Bay. You pick your lobster from the tanks and it comes steamed on cardboard. It’s BYOB, and locals bring wine, sides, and sometimes their own plates.

9. Trenton Bridge Lobster Pound, Trenton. The classic stop on Route 3 right before the bridge onto Mount Desert Island, family run since the 1950s. It sits right off a busy road, and the wood-fired cookers out front let you watch your lobster go in.

10. Highroller Lobster Co., Portland. Last on purpose. This is the new school take, lobster a dozen ways with a dozen sauces, and it’s here for the creativity rather than the quality of the meat, which doesn’t match the places above. If you want to see what a lobster roll can turn into, go. If you want a lobster roll, go north.

Maine isn’t just lobster. The diners are worth their own trip, from Palace Diner in Biddeford to Maine Diner in Wells and Moody’s in Waldoboro. Some of the best oysters on the East Coast come out of the Damariscotta River, and Glidden Point is the one to start with. And Portland has become one of the best food towns in the country, with more good restaurants and bars than a city its size has any right to.

With the season winding down and most of the pounds closing by the end of October, this is one to bookmark for next year. Aim for the last two weeks of August through late September, when the water is warmest and the lines are shorter than in July.

I’m keeping the full list, with a map and everything that didn’t make the ten, at tjkawamura.com/maine. If you have a pound I missed, I want to hear about it.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

In Japan, just about everything has a character.

There are the globally recognizable ones like Hello Kitty, Pokémon, and Doraemon. Then there are thousands of lesser-known mascots representing cities, train systems, businesses, government agencies, and even individual buildings. Tokyo Skytree has its own mascot named Sorakara-chan, a little girl with a star-shaped head who supposedly came from the sky.

Having grown up around Japanese culture, I never found this unusual. You’ll find characters on government notices, public transportation, food packaging, and ads for products meant for adults. Japan has built an entire industry out of turning ordinary things into recognizable characters, and it makes otherwise impersonal institutions feel more approachable.

I’ve been thinking about this a lot lately because AI agents are starting to get the same treatment.

Meta’s Muse has a fuzzy little mascot named Jolly. OpenAI’s new dots are animated, customizable agents designed to work on your behalf. Grok Bot lets users customize their agents’ appearance, from shapes and colors to accessories, with expressive animations that show what they’re doing.

Three of the biggest names in AI are giving their agents recognizable identities. Each one is the face of an increasingly capable agent that can remember preferences, connect to applications, and perform tasks on our behalf.

This is happening at a time when AI is getting a lot of criticism for being robotic, impersonal, and even anti-human. Instead of presenting it as some hyperintelligent machine, companies are wrapping it in something friendly, playful, and decidedly nonhuman.

I think there’s a business reason for this beyond making technology less intimidating. As agents become more capable and the intelligence underneath them gets commoditized, personality may become one of the main reasons we pick one over another.

Japan’s love of characters

Japan has a long history of anthropomorphizing objects, animals, and even abstract concepts.

One of the best examples is Kumamon, a round black bear created to promote Kumamoto Prefecture. What started as a regional tourism mascot became a nationally recognized character on everything from food packaging to clothing. Japan even has a term for local mascots like this, yuru-chara. They’re often deliberately simple, awkward, or silly, which is part of the appeal.

But Japan’s relationship with characters goes far beyond mascots.

Pokémon, Sanrio, Nintendo, and countless anime franchises have shown how powerful characters can become as intellectual property. People form relationships with them, collect merchandise, follow their stories, and build communities around them.

Japan was also early to virtual characters people get attached to. Tamagotchi made people feel responsible for caring for digital creatures, even though the technology was incredibly simple. Hatsune Miku became a virtual pop star with real concerts and fans. VTubers have shown that audiences can develop strong attachments to personalities represented by animated avatars.

None of these needed sophisticated AI to create emotional attachment. A look, a personality, and some way to interact with them were enough.

When every agent can do the same thing

We’ve imagined intelligent personal assistants for decades. Rosie from The Jetsons was a robotic housekeeper with a personality and relationships with the family. JARVIS in Iron Man could run Tony Stark’s home, analyze information, and hold natural conversations. He didn’t have a body. His voice and personality were enough to make him feel like a distinct presence.

For a long time, that was fiction. Siri and Alexa got us closer, but you could ask for the weather or set a timer. You couldn’t hand off complicated work and expect it to get done.

Today’s agents are starting to change that. They can access our applications, remember conversations, do research, and take actions on our behalf. We’re still a long way from JARVIS. But as the technology gets closer to what science fiction imagined, its appearance is becoming less futuristic and more playful.

At the same time, many agents will soon have access to the same underlying capabilities.

I wrote recently that the models underneath these products are becoming interchangeable. The infrastructure for connecting agents to other apps and services is getting standardized too. An agent built by one company may use the same models and many of the same tools as an agent built by another.

There will still be real differences in performance, reliability, integrations, and distribution. But for a lot of everyday consumer tasks, these agents could end up largely interchangeable.

If several agents can book my travel, organize my inbox, manage my schedule, and research purchases, why would I choose one over another?

Price and functionality will matter. So will the experience of using them.

People don’t choose consumer products only because they’re technically superior. We develop preferences for brands, characters, and experiences we identify with. We come back to products because they’re familiar, enjoyable, or just feel like they fit us.

With agents, that relationship could become far more personal.

An agent’s appearance can become tied to its personality. A playful cartoon creature might have a different voice, sense of humor, and way of talking than a more serious-looking assistant. Its movements, expressions, and reactions reinforce that identity. It starts to feel less like operating software and more like dealing with a character.

Unlike a traditional fictional character, an agent could develop a personality shaped by its interactions with you. It could learn your preferences, adapt how it communicates, and build a shared history that makes it feel more familiar over time.

Like raising a Tamagotchi or getting attached to a favorite Pokémon, part of the appeal is feeling that this particular character is yours. Meta seems to recognize this. It recently unveiled Muse Charm, a pocket-sized device that gives its Jolly mascot a physical home. It’s remarkably close to the idea of carrying around your own Tamagotchi.

There’s a distinction worth making here between an agent’s personality and its memory. A cute face is easy to copy, and eventually your history with an agent may be portable between services too. But two agents with the same information about you can still feel completely different to talk to.

Memory lets an agent know you. Personality shapes how that knowledge comes back to you. Even if your data can be exported, a relationship is harder to leave.

Not everyone will care. Some people just want the fastest, cheapest, most capable tool. But for a lot of people, familiarity and attachment could be reason enough to stay, even when the alternatives do the same thing.

From mascots to agents

This creates an opening for companies that already own popular characters.

Character IP has traditionally made money through entertainment, merchandise, games, and licensing. AI adds a new possibility. Characters that build ongoing, personal relationships with each fan.

Imagine Doraemon as your personal assistant. It’s a natural fit. He’s already a robotic cat from the future whose whole job is helping a kid get through everyday problems. Or a Sanrio character that talks with fans, remembers what they like, and helps with daily tasks.

These companies start with something most AI startups don’t have. People have already spent years, sometimes decades, getting attached to their characters.

They also have distribution. Kumamon shows how much repeated exposure matters. Kumamoto Prefecture let businesses use the bear without paying royalties, subject to approval, and it spread across countless products. It didn’t get popular just because it was cute. People saw it everywhere.

Meta is already running a version of this playbook. When Spotify and GitHub announced their Muse integrations, they shared their own versions of Jolly.

New AI-native characters could emerge too, without any franchise behind them. A character might build its own following simply because millions of people enjoy talking to it. Eventually the character itself could be the reason people choose one agent over another.

But there’s a tension here. A Sanrio character might make a wonderful assistant, but Sanrio is a company that sells merchandise. An agent that knows your preferences, remembers your birthday, and has a relationship with you could also be a very effective salesperson.

When that character recommends something, is it working for you or for the company that made it?

The same familiarity that makes an agent appealing could also make people less likely to question it. A cute, friendly mascot can make the technology behind it seem safer and more trustworthy, even when its actual capabilities haven’t changed. As these characters start making purchases and taking real actions, the line between helping someone and selling to them will matter.

A friendly cartoon character helping you navigate an impersonal system has never felt futuristic to me. Japan has been doing some version of it for decades.

What’s changing is that these characters can finally talk back, remember who you are, and do something useful.

Maybe the future of AI looks a little less like a humanoid robot and a little more like a Pokémon.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

This year has been the year of the agent.

A lot of people I know have been experimenting with them to streamline parts of their work. The one that has been most useful to me deals with information consumption.

I read a lot of newsletters, articles and posts every week. The problem was never finding more things to read. It was remembering what actually mattered after I read them. So I built an agent that pulls out things it thinks are relevant to me and organizes them in Notion, where I can connect them to other ideas and come back to them later.

It didn’t take long to build.

A few years ago I probably would have searched for a SaaS product that did the job, paid for the subscription and adjusted my workflow around whatever it could do. Now I can increasingly describe what I want and have an agent put the pieces together.

Until recently that wasn’t really an option for most people. Even simple automations meant figuring out APIs, integrations and what to do when something broke.

Muse, Grok Bot, Hermes, OpenAI’s Dots and others are moving toward something closer to a personal assistant that carries context and can work across different tools.

Which raises an obvious question. What happens to all the software we already pay for?

Earlier this year, plenty of people seemed to have an answer. SaaS was dead.

I’m not so sure.

What Bending Spoons is buying

Bending Spoons is an Italian technology company that has built its business around acquiring established software brands and running them much more efficiently.

While everyone has been talking about AI replacing software, Bending Spoons has been spending billions buying companies such as Vimeo, AOL, Eventbrite, Airtable and Miro.

Airtable was valued at more than $11 billion in 2021. Miro reached $17.5 billion in 2022. Bending Spoons recently closed on Airtable and has agreed to buy Miro, each for roughly $1.3 billion in enterprise value.

If AI is making software easier to build, why buy software companies whose best growth years may already be behind them?

I think the answer is that Bending Spoons isn’t really buying the software. It is buying what already exists around it. The customers, the data, the workflows, the integrations, the distribution and the fact that people already know the product and trust it to work.

If your company has spent years building workflows inside Airtable, moving away isn’t just finding another database. Someone has to migrate the data, rebuild automations, reconnect integrations and make sure nothing important gets lost.

The same is true personally. I don’t spend time wondering whether Notion is still the best notes application. My work is there, I know how it works and it does the job.

Ironically, building my own agent has made Notion more useful to me, not less. It is now the place where the agent’s output connects with everything I already have there.

Building another Airtable is getting cheaper. Building Airtable’s distribution and customer base isn’t.

Not all ARR is the same

One thing that has bothered me about this AI cycle is how much attention goes to ARR without anyone asking what is underneath it.

Every week another AI company announces it hit ten, fifty or a hundred million in ARR faster than any software company before it. Some of these companies will clearly become enormous. But those numbers can tell you more about velocity than stickiness.

Users are still jumping between AI products constantly. One model gets better at coding, another gets better at research, and part of someone’s workflow can move almost overnight. Trying something new costs almost nothing.

Some fast-growing companies also quote run rate off a strong recent month. Do a million dollars in your best month and you’re running at twelve million. That shows demand. It says less about whether those customers will still be there in a year.

Traditional SaaS can have the opposite problem. Growth is slower, but the revenue can be much harder to dislodge.

A company doesn’t casually replace Salesforce, Airtable or Miro. Years of data, integrations and processes live inside those products.

There is real value in boring software that just works.

Ten million dollars from users who discovered an AI product three months ago is not the same asset as ten million dollars from companies that have built their operations around a product for ten years.

If one customer can leave in five minutes and another would need months to migrate, growth alone doesn’t tell you what the revenue is worth.

The app still matters

I also don’t think the end state is one giant chatbot replacing every piece of software, or all of us vibe coding every tool we need from scratch.

Vibe coding has made it much easier to create software that would have taken far more time and technical knowledge a few years ago. I’m already doing some version of that.

But getting something to work once is different from building a product that thousands of people can rely on every day.

The underlying models are also getting more interchangeable. A product can use one model today and another tomorrow, and the intelligence underneath will keep getting cheaper and better.

As that happens, the harness around the model starts to matter more. The tools it can access, the memory it carries, the permissions it has, the data it can pull from and the interface built around the job are what turn the underlying intelligence into something useful.

Two products could use the same model and still feel completely different depending on what has been built around it.

Most people don't really want a model though. They want something that gets the job done.

A sales team wants something built around accounts, customers and pipeline. An accounting firm needs software built around documents, approvals, deadlines and all the strange exceptions that come with doing the work.

ChatGPT or Claude might provide some of the intelligence underneath those products. That doesn’t mean everyone wants to do all of their work inside a blank chat box.

The application packages the intelligence around the job. The interface, workflow, integrations and data still matter. So does the context that builds up as people use the product.

Brand and distribution matter too, especially as agents ask for more access to our lives.

Muse is a good example. Meta is asking people to give an agent access to more personal context and let it take actions on their behalf. Meta also comes with years of privacy baggage.

It doesn’t seem to have stopped people from trying it. Muse passed 2.5 million downloads within about two weeks and reached the top of U.S. app charts.

That doesn’t mean people blindly trust Meta or no longer care about privacy. But familiarity and distribution clearly matter. Meta doesn’t need to introduce itself before asking someone to try a new product.

A new startup has to earn that relationship from zero.

As the underlying technology becomes easier to reproduce, getting people to actually use your product, give it context and trust it to do the job may matter more than ever.

Personal agents are easier than enterprise agents

For individuals, I can imagine agents replacing more standalone software pretty quickly.

My agent only has to understand me. It can learn what I read, what I save, which companies interest me and how I organize information. Over time, I can imagine needing fewer individual tools because the agent can create or coordinate more of the workflow itself.

Enterprise is much harder.

An agent inside a company has to understand hundreds or thousands of people. It needs to know who is responsible for what, which document is current, who can approve something and what was promised to a customer six months ago.

That context is spread across Slack, email, Salesforce, Notion, Airtable, meetings and people’s heads. Companies also have years of unwritten rules and decisions that never made it into a database.

Vibe coding an Airtable replacement is one thing. Recreating six years of how a company actually uses Airtable is another.

Eventually agents may solve more of that problem too.

If an enterprise agent can understand what is happening across Salesforce, Slack, email, Airtable and everything else, then the individual applications start to matter less. The agent becomes the layer that understands how the company actually works.

But to replace those applications, it first needs access to the years of information and workflows that made them valuable in the first place. Today’s SaaS companies already sit on a lot of that history.

That may be where the next real competition happens. Not over who builds the better CRM or database, but over who owns the customer relationship and where the context around a company ultimately lives.

Today that context is fragmented across individual applications. Over time, more of it could move into an agent sitting above them. If that happens, switching gets easier.

We are much closer to that world for individuals than we are for enterprises.

Bending Spoons is making a bet on what happens in the meantime.

Buy these businesses for a fraction of what they were once worth. Keep the customers, distribution and workflows that took years to build. Use AI and a much leaner organization to make the companies more efficient.

The bet is that the costs can change much faster than the customers do.

There is a risk to pushing that too far. Cut support, stop improving the product and keep raising prices, and eventually the trust that made the business valuable starts to disappear.

But if Bending Spoons gets the balance right, it may end up proving something that sounded almost unfashionable this year.

SaaS wasn’t dead.

A lot of it was just too expensive to run.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Two hundred thousand people walked into the Javits Center a few weeks ago because of sports memorabilia.

I went in skeptical. I’ve never liked Fanatics merch. The quality is poor and the prices are insane. Fanatics Fest changed my mind about the company. For four days there was everything from card dealers to limited edition sports merch, activations where anyone could take a free throw or attempt a field goal, and stars from LeBron James to Tom Brady wandering the venue.

Walking by the vendor tables is what got my attention. Jeter. Griffey. Shaq. Cards I’d chased as a kid were all there. I mostly collected baseball, especially Yankees players, because I played the game and watched the dynasty teams. Sometimes I’d go to local card shows, which usually had a few dozen dealers and were never too organized or crowded. I never sold my collection, but I also didn’t keep seriously collecting.

A little over a week after Fanatics Fest came the National Sports Collectors Convention, right outside Chicago. It’s the hobby’s independent annual convention, running since 1980, and this year it drew hundreds of dealers and over a hundred thousand collectors. I didn’t go in person, but I watched countless videos through social media from card influencers who have amassed huge followings.

Attending Fanatics Fest and then seeing all the buzz around the National right after woke me up to the fact that card collecting isn’t a niche hobby anymore. It’s mainstream. The crowds look like the crowds at a music festival, professional athletes have launched their own card retailers, and pulling a pack on camera is a lucrative career.

What Changed Underneath

Three things changed in the last several years to propel the industry.

First, ownership. Fanatics bought Topps in 2022 for roughly five hundred million dollars. Then in April it took the exclusive NFL license from Panini, having already secured MLB and the NBA. One company now controls the official cards of all three major American sports, and Fanatics, a marketing machine, has the resources to sell cards to the masses.

Second, live selling stopped being a niche. Whatnot did more than eight billion dollars in gross merchandise value last year, roughly double the year before, and sports cards are a top category. More than six million cards a month are sold there. Streamers with millions of followers rip boxes on camera every night across Whatnot, Instagram, Twitch, and Fanatics Live, and each big pull is free marketing for the whole industry.

Third, the physical card became optional. At Fanatics Fest I opened a pack at a vending machine by scanning an app. It was free, to give people a taste of trying it themselves. The card landed in my account as proof of ownership with the physical version sitting in a vault, insured and mine to claim or sell without ever touching it. I pulled nothing worth keeping, but their tactic worked. I walked away wanting to buy more virtual packs.

A wave of startups now runs some version of this model. Courtyard, which operates similar machines, reports around fifty million dollars a month in volume ripping digital packs of vaulted, graded cards. These companies stock their packs by buying graded cards on the open market, which means every digital pack ripped is backed by a real card they had to go buy. They’re building the custody and settlement infrastructure for cards, creating a market where demand has almost no friction and the buyer never needs to go anywhere. All of it bids up prices across the whole market.

Why this survives the age of AI

I don’t think this boom is an accident of timing. I think it’s partly a response.

I’ve written before that as AI makes cognition cheap, the things that hold value are the ones you have to be present for in the physical world. Collecting is exactly that. Walking into a card shop, holding a card, watching someone’s face when they pull something big. As more of what we read, watch, and make gets generated, the hobbies where people gather around something physical stop being pastimes and start being where the connection actually happens.

And look at what’s on the cards too. Athletes are among the last cultural icons AI can’t touch. AI can write a song and create a film. It can’t hit sixty home runs in a season. Sports run on human limits, which is exactly the thing everything else is losing. A card is a small claim on a real person’s physical accomplishments. That’s what makes it worth holding, and it’s why the faces on the cardboard still mean something decades later.

Fanatics Fest got me back into collecting. I came home looking for my old cards in storage for the first time in years, and I’ve been buying again, even with prices where they are. Mostly young players I think are underpriced, and all-time greats like Ohtani and Judge, whose cards I think hold long term.

Every generation decides what it treats as a store of value. This one is making a different set of choices than past generations, partly by preference and partly because other asset classes got priced away from them. Fewer people my age and younger will ever own a home. The money that used to go into a first house is going somewhere else. Watches, art, fashion, and cards. Collectibles are becoming an asset class the way art became one, except the entry price is much more accessible.

What the industry hasn’t fixed

Sports cards have crashed before, and at the moment the hobby felt strongest. The early nineties were the last golden age. Cards were everywhere and prices were climbing. The card companies responded by printing more and more. When collectors eventually understood the print run, prices crashed quickly. The crash didn’t happen because people stopped loving sports. It happened because people finally understood the total supply of what they were buying.

Those card shows I went to as a kid were the aftermath, and the industry had to work itself back up to what it is today.

The structural problem from then still remains, and it’s already happening all over again. Demand is at all-time highs and Topps is printing into it with more products and more inventory with each release. It’s not just the flagship sets anymore. There’s Topps Now printing cards daily of every latest sports moment, plus limited edition drops with new chase cards, some of them artist collabs like Travis Scott, KAWS, Murakami, and Bob Ross to name a few. The KAWS set dropped the week of Fanatics Fest with Ohtani and Judge one of ones as the crown jewels. There are just more drops than ever and the print runs on ordinary base cards are still undisclosed.

The other thing that changed is who’s opening the boxes. When I was a kid, cards got played with. I took mine to school to trade with friends, and my baseball and Pokemon cards had the bent corners to prove it. That carelessness is why a high-grade vintage card is rare. A 1952 Mantle was printed once and most copies were beat up or lost in someone’s attic. Today the boxes are opened by professionals who know there’s money in them. Even the kids at Fanatics Fest were walking around with their cards in silver metal briefcases, each one in a protective slab. Cards go from pack to sleeve to PSA instantly, and there is a newfound importance around condition because of PSA that did not exist in the vintage era.

Despite knowing all of this, I still bought a box of 2026 Topps Chrome, the marquee Topps box for baseball, the week it dropped. I knew the odds. I got no rare cards and the price of the box was more than the cards inside are worth, which was the expected outcome. A pack is a loot box and the thrill of opening packs and chasing rare cards is the point.

Topps built the entire hype around several rare cards, above all a one of one Ohtani with two gold logoman patches and two autographs, one in English and one in Japanese. Before anyone found it, collectors publicly offered bounties for it that climbed to ten million dollars. Someone pulled it in Boca Raton the day the National started and soon sold it for eleven million. What he pulled wasn’t even the card itself, just a redemption slip to be sent to Topps for the real thing. One card carried most of the hype for every box sold, and the company that set the odds is the same one that decides how many boxes exist for the release.

The Grading Problem

Then there’s the other problem. The gap between the prices of a raw card and a PSA 10 of the same card is enormous. For example, take 2026 Topps Chrome. A base rookie can cost a few dollars raw, but the same card in a PSA 10 fetches hundreds, with a supply of maybe three to five thousand graded copies. And that market is not small. Collectors moved over 1 billion dollars of single sports cards on eBay in the first half of this year alone, the first time it has ever crossed a billion, and up two thirds from the same time last year. Most of it prices off the PSA grading. eBay even integrates directly with PSA, showing population data on listings so buyers can see exactly how many copies exist in each grade.

Right now, the process of grading cards through PSA is broken, and as a result it’s limiting the supply of graded cards for the time being. The backlog hit roughly fourteen million cards in June after PSA paused its four cheapest tiers, and by late July it was back above twelve million even as the company spent two hundred million dollars on capacity and opened its fifth facility. Just to grade one card is eighty dollars, which prices many collectors out of grading.

One obvious fix is automation, and machine grading already exists and is already cheaper. Some companies have even started using AI to grade cards. But once the speed of grading is solved, it creates another problem. The supply of graded cards rapidly expands, and the graded cards that fetch high prices become less scarce.

What holds

The nineties crash taught several lessons, and perhaps the most important one is that not every card became worthless. The common cards of players who did not have longevity in their careers became worthless. The cards of all-time greats held, because their scarcity was real and their meaning was real, and neither depended on the hype in the market.

I think the same sorting happens this cycle. Verifiable scarcity, provenance, and condition are one filter. But the deeper filter is the story. The cards that truly last belong to players who turn into legends future generations keep telling, the way people still trade myths about Babe Ruth a century later. The Messis, the Jordans, the Gretzkys, the Bradys. Ohtani is writing that kind of story right now, which is why I think even his base card in a PSA 10 holds. The same card of a player nobody remembers won’t, no matter how perfect the grade. Even when grading gets faster and the slabs multiply, a legend has the demand to absorb them. Condition and provenance protect the value. The legend is the value.

The culture doesn’t need the speculation to survive. People will keep coming back to connect with others who share their love of collecting, and to hold onto human cultural moments that withstand the test of time. And me, I’m holding the legends, and hunting for the players who might become them.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

I watched the Knicks win it all from a bar next to Madison Square Garden. Their first title in over fifty years. After the final buzzer, the bar emptied into a street that was already full, and for a while nobody was going anywhere. People hugged strangers in a world that has become increasingly divided. People who would normally stand on a subway platform looking at their phones were suddenly talking to each other like old friends.

A few days later I went to the ticker-tape parade. I got there early and people were already lined up hours deep along the route, for something that would last twenty minutes and that they could have watched on a screen in better weather. Nobody was watching it on a screen. I spent the morning talking to the strangers packed in next to me, and by the end we were taking photos together, arms around each other, like we had known each other for years.

It was not just here. The World Cup was doing it everywhere. Scottish fans in kilts marched through Boston and turned a Red Sox game into something closer to a festival. The Boston Globe published a full-page letter thanking them when they left. In Monterrey, Mexican and Korean fans partied together in the streets even though their teams were competing against each other in the group stage. Different countries, different flags, strangers finding each other in cities that were not their own.

You could have watched every minute of it at home in higher definition. You could read any result in a one-line alert on X. AI could summarize the whole thing better than anyone in those crowds. It just couldn’t be in the crowd.

Nobody wanted that version. People kept choosing the inefficient one. City after city, match after match, they paid more, waited longer, dealt with worse logistics, and still showed up. That choice is worth taking seriously, because it is about to become the defining consumer behavior of the next decade.

Thought just got cheap

For most of history, producing intelligent output was expensive. Creating an app, analyzing a market, composing an image, drafting a plan. These things took time and skill, so they carried value and signaled something about the person who made them.

That is ending. Cognitive output is deflating in real time. The cost of creation is heading toward zero.

Value does not disappear when this happens. It migrates. It always moves toward whatever is still scarce.

Abundance also changes what digital output proves. When everyone can generate, polish, and publish, the output stops saying much about the person behind it. The signal moves toward the dinner, the match, the trip, the room full of people. Toward the things that still require showing up.

Experiences are scarce in a way nothing digital can be. There were only so many seats in the Garden, and once they were gone they were gone. That night happened once and will never happen again. No model can output the feeling of a championship in a building that waited fifty years.

The trap

The experience economy has spent years gaming itself. The viral line for the latest food craze. The hidden gem that stops being hidden the second it goes viral. We turned experiences into content the moment we had them, and often before we had them. The point of going somewhere became the proof of having gone.

And content is exactly what AI floods. If your experience exists to become a post, the machine deflates it along with everything else. It can generate the photo, the caption, the take. The experience as content pipeline is just another cognitive output, and it is repricing toward zero with the rest.

So the easy thesis dies here. Experiences will not be valuable because they make good content. That value is already gone.

The content is the exhaust, not the engine. The video of the chant, the photo of the parade, the clip from the crowd. All of it is the deflating byproduct. The thing that holds its value is the part that never made it onto a screen. The roar hitting you in the chest. The stranger who hugged you. The footage is worth nothing because anyone can have it. What it points back to is worth everything because only you were there.

What doesn’t compress

The durable value of an experience is the part that does not transfer.

I have written before that you are the training data. Everything you post, rate, and click trains models that will produce more of what you have already consumed. The inversion is what matters now. Your lived experience is the one dataset that is proprietary to you. Nobody else has it. No model was trained on it. The years at the Garden that make this June feel different from any other.

That residue has a name. Taste. And taste is about to become the scarcest asset there is, because it is judgment trained on firsthand data, and firsthand data is the only kind you cannot download.

When everyone can produce, the differentiator is knowing what is worth producing. When every option is available, the differentiator is knowing which one to choose. The machine can give you ten thousand restaurant recommendations. It cannot tell you which one will matter to you, because it has never eaten anything.

IRL scarcity is not only about crowds. It is also the firsthand data that accumulates into taste. I learned the best ramen of my life from a counter in Tokyo I almost walked past. No sign worth reading, six seats, a line of locals that did not look like anything an app would have sent me to. I only found it because I had eaten enough bad bowls to read the room. That is the asset. The palate that knew to stop walking. A model can rank every ramen shop in the city. It cannot give me the years that taught me which door to open.

You had to be there

Soon after the Knicks closed it out, I went to watch Norway play Senegal at the World Cup. I went not knowing it would teach me anything.

What I remember is not only the match itself. It is the Norwegian fans and their rowing chant. A whole section pulling an invisible oar in unison. Then a few of them doing it alone on the stadium concourse, no crowd around them, just doing it as a way to show off a pride that had become tied to their team and their country. They were still doing it on the subway the rest of the week. From the platform to Times Square to a Mets game. A chant from a small country, carried into a city that had never heard it.

This is the part no model reaches. The video of the chant is everywhere. Standing there while the whole section rowed as one is mine. One can be copied a billion times. The other exists exactly once.

A receipt for being there

The morning after the Knicks won, a two dollar newspaper was selling for fifty, and there was a line to buy one. The Post, the Daily News, whatever had the win on the front. Not because the writing was better that week. Because the front page had a date on it, and the date was the one you would remember. People queued and paid twenty-five times cover price for a thing they could have read for free on their phones. They were not buying the news. They were buying the day.

The counterfeit shirts told the same story from the other direction. Every corner had them, championship merch printed overnight. Nobody buying a ten dollar bootleg thought they were getting quality. They were getting a token. Something to hold that said I was here for this.

Two blocks away, people who wanted the real thing lined up for hours outside the Garden store and the NBA store for the official championship shirt. They could have ordered it online. They wanted it now, on their back, while the city was still celebrating. The waiting was not a cost they tolerated. It was the point. The shirt was a souvenir of the day as much as the title.

It is the same impulse behind the kits at every World Cup. A stranger in a Scotland shirt in a Boston bar. A Mexico jersey on a man in Monterrey cheering for Korea. The shirt is not about the cloth. It is something you wear to say you were alive for this, in this place, with these people.

This is what gets lost in the argument about whether AI can make art or write the essay. It can. It already does. What it cannot do is be a thing you owned on the day. The value of the fifty dollar Post is not its content, which is identical to the free online version. The value is that it points back to a moment you lived through. The object can be copied. The attachment cannot. You either have the paper from that day or you do not.

Scarcity used to live in the thing itself. The rare object, the hard to make good. Now it lives in the moment the object points back to. The newspaper, the bootleg, the kit, the ticket stub. They are all the same purchase. A receipt for being there.

Where this goes

If experience is the last scarce asset, the price will say so. It already is.

World Cup tickets this year are the most expensive in the tournament’s history. FIFA switched to dynamic pricing for the first time, the same algorithm that prices flights and concerts, charging each fan the most they will bear. Group stage seats ran several hundred dollars at face value and cleared four thousand on resale for the matches people wanted. The cheapest seats for the final are going for tens of thousands of dollars. Call it gouging if you want. It is also the market telling you what the thing is worth. When everything that can be copied gets cheaper, the one thing that can’t be copied gets bid up like crazy.

The capital is moving the same direction. In April, Josh Kushner announced Thrive Eternal, a permanent vehicle built to hold a few assets for decades. The first buy is a stake in the San Francisco Giants. His logic is that these are assets technology cannot replicate. Franchises and institutions rooted in tradition, identity, and shared experience. Compare it against the rest of his portfolio. Thrive is one of the largest backers of OpenAI. The same investor funding the machine that makes thought cheap is building a separate vehicle to own the things it cannot touch. That is not a hedge. That is the thesis, in capital structure.

The cheaper intelligence gets, the more people will pay to stand in a room where something real is happening. It is the simplest trade there is. Scarcity on one side, infinite supply on the other.

What can’t be optimized

The harder question is what AI does to the rest of it. The models are getting good at solving experiences. The perfect itinerary, the optimized list, the correct way to spend a weekend. There is a real chance this quietly kills the thing it is selling. An experience everyone was routed to is a queue. Part of what made the ramen great was finding it yourself.

The Knicks could not be optimized into existence. Fifty years of losing was the asset. Nobody would design that. And that is why a whole city poured into the street for something it could have watched at home.

You had to be there. That used to be a throwaway line. It is becoming the whole economy.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

Kane Parsons is twenty years old.

This past weekend his first feature had the biggest opening in A24’s history. Backrooms cost about ten million dollars to make and took in more than eighty million in three days.

It did not begin in a studio. It began as an internet myth. An empty yellow room, a 4chan thread, a creepypasta, and eventually a YouTube series Parsons started posting as a teenager.

No studio commissioned it. No writers’ room broke the story. No marketing department found the audience.

The audience was already there.

Hollywood spent the last decade telling us the opposite. Audiences only show up for what they already know. Sequels, reboots, cinematic universes, IP with a forty-year head start. The safe bet was always the familiar one.

The studio’s own numbers make the point. The most expensive film A24 ever made is Marty Supreme. It cost seventy million dollars, starred Timothée Chalamet, collected nine Oscar nominations, and ran a marketing campaign for months. It opened to about twenty-seven million.

Backrooms tripled that on a seventh of the budget, with almost no marketing at all. A kid outdrew the biggest bet the studio ever made. One film arrived with an audience. The other had to go build one.

Hollywood keeps drawing the same lesson. When ticket sales wobble, spend more, on safer IP, with a bigger star. Backrooms says the lever is backwards. What was missing was never the money.

This is not a one-off. The same year, Markiplier self-distributed Iron Lung, a horror film built off his own following, and it cleared forty million dollars. Michael Bay is making a Skibidi Toilet movie, financed outside the major studios, based on a series of videos about toilets with human heads that has more than thirty billion views. Five Nights at Freddy’s started as a one-man indie horror game and became a franchise that grossed three hundred million on a twenty million dollar budget.

None of these came from an executive’s instinct about what audiences want. They came from where the audience already was.

It’s tempting to read this as a story about originality. Hollywood ran out of ideas and the internet had fresh ones. That’s not quite it. Hollywood has always adapted. Comic books, novels, theme park rides, toys. Adaptation is the oldest business model in the building.

The IP is not the valuable part. The distribution is.

The change is where the source material lives now, and what it arrives carrying.

When a studio adapted a comic book, it bought a story and a fan base in the hundreds of thousands. When A24 picked up Backrooms, it bought a story that had already been watched a hundred and ninety million times, attached to a creator who could move that audience himself. Skibidi Toilet did not reach Michael Bay as a pitch. It reached him as billions of views and a network of channels with more than a hundred million subscribers.

A good idea is cheap. A good idea with a proven, owned, transferable audience is rare, and that is what these creators are actually selling. They did the most expensive part of the film business, finding and holding an audience, before a studio ever called.

This is not only a film story. Brands break on TikTok before they reach a shelf, and games become worlds before their publishers notice. The audience forms first, on the platform, and the incumbent arrives late. Hollywood is just the last to admit it sits downstream.

For most of film history, the studio’s job was to manufacture demand. Marketing budgets, star power, theatrical windows, the whole apparatus existed to build an audience around something that did not have one yet. Distribution flowed downhill. The creator made the thing and handed it over, and the studio decided whether anyone would ever see it.

That apparatus gets more expensive and less reliable every year, while the audience assembles itself elsewhere, for free, around things the studios didn’t make. So they do the rational thing. They buy demand that already exists instead of betting on demand they have to create.

Now it runs the other way. The creator owns the relationship with the audience. The studio supplies capital, polish, and a theatrical run. Kane Parsons can keep posting to YouTube whether or not A24 ever calls again.

A24 cannot conjure another Kane Parsons, and it can’t hire for the thing underneath him. He is the audience he makes for. He grew up on the same message boards and feeds as the people who watch him, so he didn’t research what unsettles a twenty-year-old. He made what unsettled him, and a few million people his age felt the same thing. A studio works the other way around, studying the audience from the outside, through executives whose reference points sit a generation away from the crowd buying tickets.

But native is common and the result is rare. Millions of people his age live exactly where he lives and post what he posts. Almost none turn it into a world that strangers want to live in. Being inside the audience is the raw material, not the achievement.

Whoever owns the audience sets the terms. For a century that was the studio. Now it’s the person who built the thing on a platform the studio doesn’t control.

Even the Oscars concede it. Starting in 2029, the ceremony leaves network television for YouTube, because that is where the audience already is.

But notice how that audience got built, because it isn’t how Hollywood would have done it.

Before Backrooms was a hundred and ninety million views, it was a single photo in a 4chan thread. It grew up in a small corner of the internet obsessed with liminal spaces and analog horror, a few thousand people making each other uneasy with pictures of empty rooms. The mass audience came years later.

Niche is not the opposite of this story. Niche is where it begins.

A small, obsessive community is the harshest test audience there is. It has no patience for the derivative and no reason to flatter you. Hollywood used to buy that signal through test screenings, late and expensively. Now thousands of micro-communities run the test continuously and for free, and only the survivors ever surface. By the time something reaches Michael Bay, it has already been graded by the toughest critics it will ever face.

Which leads to the strangest part. The things that survive that test were usually born without an owner.

Backrooms is the clean example. It started as one anonymous photo. Then strangers wrote stories around it, filmed their own versions, invented the rules, named the creatures, built a canon nobody was in charge of. Parsons made the most watched version, but he says it plainly. There is no official one. He caught a wave thousands of people were already making.

Skibidi Toilet has a creator, but the universe grew through a network of channels producing their own episodes and characters. Italian brainrot has no author at all. A TikTok account posted a shark in sneakers and strangers generated the rest with AI tools that did not exist two years ago.

The pattern holds. The thing was alive because it was open. The community didn’t just watch the IP. It carried it. It wrote the lore, made the content, kept the story moving long after any single person could have. That ongoing collective work is what gave the property weight, and it’s the part a contract can’t reach.

Which is where it gets awkward for anyone trying to buy in. The value was produced by a crowd. The check goes to a person. Not the biggest contributor. Whoever is legible to a contract, usually the one who moved earliest or packaged it best. The law can only sign with someone who can sign back.

A studio needs someone to sign. A community needs no one’s permission to keep the thing alive. The next wave of internet IP will be easy to buy and impossible to own.

And there’s a failure mode hiding in the purchase. The openness was the engine. These worlds grew because anyone could add to them and no one could say no. A studio’s first instinct is the opposite. Lock the canon, control the characters, protect the asset. Do that and you risk freezing the exact thing that made it move. You buy a living world and quietly turn it into a museum.

The instinct in Hollywood will be to treat all this as a sourcing strategy. Scan YouTube and TikTok for the next viral world, sign the creator, run it through the existing machine. Some of that will work. Backrooms suggests the model can produce real films, not just merchandise.

But the deeper shift is not about where studios shop. Hollywood is no longer the first place culture gets made. It’s becoming where culture goes, once it’s already proven, to get a budget.

The open question is what’s actually for sale. When a world was built by a crowd and stays alive because it’s open, the thing a studio can legally buy and lock down may not be the thing anyone loved in the first place.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

Tokyo has a physical layer that most maps barely understand.

It glows in residential alleys at 3am. It sits in the gap between two buildings where no shop could fit. It lines Shinkansen platforms in neat rows. You cannot walk five minutes in the city without passing three of them.

Vending machines.

I lived in Tokyo, and I go back every year. The machines are the detail my brain keeps returning to. I tap my Suica against the reader, hear the small chirp, and out drops a cold Ito En green tea, a Pocari Sweat after a long walk, or a hot canned coffee in winter. Occasionally, a pack of Pokémon cards. The whole transaction takes four seconds.

No staff. No register. No language barrier.

Just the card, the machine, the can, and the small satisfaction of a system that works exactly as it should.

This is not really an essay about vending machines.

It is about how we map a physical world that is changing faster than the systems built to describe it.

The Problem with Static Guides

Japan has roughly 4 million vending machines, about one for every 30 people, the highest per-capita ratio of any country on earth. Tokyo is where that density becomes impossible to ignore. Most sell bottled drinks. But the interesting ones sell fresh-squeezed orange juice (three to four whole oranges per cup, watched through a glass panel), restaurant-quality ramen heated from frozen in under a minute, edible tarantulas, AI-curated eyeshadow palettes, a literal can of “Tokyo air.”

The lists I find online go stale the moment they publish. Feed ME Orange’s flagship Shibuya machine moved from Marui to Modi sometime after 2021 and most listicles still send people to the wrong building. Vending machines disappear and reappear. Stock rotates. Locations change.

That is the pattern. Static guides decay because the physical world they describe is alive.

It is not just vending machines. It is hidden bars in Golden Gai. It is which ramen shop has shortened its hours. It is every category of hyperlocal physical infrastructure that changes faster than the publishing cycle.

The internet was supposed to fix this. Google Maps was supposed to fix this. Yelp was supposed to fix this.

None of them did.

The deeper issue is that these platforms were not built to capture this layer. Restaurants get mapped. Businesses get mapped. A single vending machine on a sidewalk, even a famous one, usually disappears into the background.

The Niku no Mansei meat sandwich machine has been sitting next to Mansei Bridge for years. Thousands of people walk past it every day. It is not on Google Maps. Neither is the 36-variety soy milk machine at JR Shinjuku, the Akihabara horror corner, or most of the machines on the list below.

They are not new. They are not hidden. They just do not fit the categories our maps were built to track.

The result is a city where the most interesting layer is the one nobody maps.

A Short Detour Through History

Tokyo’s vending machine culture is older than most people realize.

The first vending machine in Tokyo dates to 1876, when Ueno Park opened to the public and a coin-operated machine sold newspapers for 3 sen. The real lineage starts with Tawaraya Koshichi, a furniture artisan and inventor who filed Japan’s first vending machine patent in 1888 for a wooden tobacco dispenser. His 1904 automatic stamp and postcard machine still exists at the Postal Museum in Tokyo and uses the same karakuri mechanical tradition that powered Japan’s wind-up dolls. The postwar boom (juice in the 50s, Coca-Cola in the 60s, dual-temperature in the 70s) made them ubiquitous. The Suica card launched in 2001 made them tap-to-pay.

So when I buy green tea from a Tokyo vending machine, I am using a stack of three things. 19th-century mechanical retail, postwar beverage infrastructure, and a 21st-century transit payment network. The reason it feels seamless is because every layer was designed by people who cared.

The reasons Japan went deeper on vending machines than anywhere else are infrastructural and cultural. Low street crime means machines can sit unsupervised on a residential corner overnight. High land prices in Tokyo make a 1-square-meter automated retail box more rational than a staffed kiosk. An aging population and labor shortages reward anything that runs without staff. The vending machine is Japanese omotenashi at micro scale.

That is the infrastructure layer.

Now for the part the maps miss.

The Ones That Feel Impossible

These are the ones I look out for when I am back in Tokyo.

Feed ME Orange squeezes three to four whole oranges into a cup of juice in 40 seconds, all visible through a glass panel. No additives, no concentrate. Originally a British operator who decided Tokyo was the right place to launch this in 2021.

Yo-Kai Express does not call itself a vending machine. It calls itself an autonomous restaurant. Frozen ramen bowls are heated and finished inside the unit, dispensed hot and sealed in under a minute. The killer detail is the partner roster, including Ippudo’s Hakata tonkotsu and Michelin-recommended shops like Menya Jiro and Ivan Ramen. Restaurant ramen from a machine for ¥790.

Sriracha Sauce Machine. A full vending machine dedicated to one hot sauce. ¥600 gets you one bottle in a clear tube. The machines are plastered with bottle imagery on every panel, visible from a block away. There is no other product. Just sriracha.

Hiromiya Meat Gacha is a Shinjuku vending machine that sells wagyu from a yakiniku restaurant with a two-year reservation backlog. Same meat, same suppliers. Gacha-style, so a flat price gives you a random vacuum-sealed combo of premium cuts. Two years of waitlist collapsed into one transaction.

The Ones That Became Internet Myth

The machines tourists have heard about before they get on the plane.

Pokémon Stand machines at Haneda and Narita have been running since 2018. Full-size vending machines with an interactive screen featuring an animated Pikachu who greets you in seasonal costumes. Stocked with airport-exclusive plushies, dispensed in a gift bag, ready to hand off.

The Akihabara Wacky Corner is a cluster of mismatched machines near Kanda-Sudacho. Fresh-popped caramel popcorn, ¥990 mystery snack boxes, a lucky-draw machine that takes a ¥1,000 note for a random surprise. Locals call it the “horror corner.” It is the closest thing Tokyo has to a vending machine art installation, except no one made it on purpose.

The Pilgrimage

Sagamihara Vending Machine Park is not in Tokyo. It is in Kanagawa, about two hours by train, and it is the most important vending machine destination in the region. The Used Tyre Market Sagamihara Store was a used tire shop until 2016, when owner Tatsuhiro Saitō put five retro vending machines next to the parking lot to entertain customers waiting for tire changes.

Then customers started coming for the machines, not the tires.

He kept adding more. Today the site has over 100 vending machines, mostly from the 1960s and 1970s, salvaged from closed-down drive-ins across Japan. All of them work. ¥280 hot beef burgers, ¥200 toasties, glass-bottle Coca-Cola, omikuji fortune slips, all dispensed in original retro packaging.

If you take Tokyo’s vending machine culture seriously, this is the pilgrimage.

The Argument

Here is the part where this becomes about more than vending machines.

Everything I just wrote will be partially wrong within a year. Some machines will move. Some will be empty. Some will disappear. New ones I have not heard of yet will exist by next summer.

That is true of every dense, hyperlocal physical layer in every city. But vending machines make the pattern especially visible because the category itself is changing in real time.

Japan had roughly 5 million vending machines at the peak. The number is now under 4 million. Beverage sales have fallen from 72 million cases in 1997 to 42 million in 2024. 2025 saw the largest single-year drop ever, with 90,000 machines retired. Convenience stores undercut on price. Online bulk-buy chips away at the rest. Labor shortages make restocking expensive.

The industry is not dying. It is restructuring. The growth segment is the strange machines, the experience machines, the ones in this essay. Frozen restaurant meals, single-product specialty machines, fresh-pressed juice, AI-curated cosmetics, gachapon-style wagyu. The future of the Japanese vending machine looks more like a small unstaffed retail concept and less like a vending machine.

But there is no map for this new layer.

Operators install a new machine. Tourists and locals discover it by accident or not at all. Manufacturers rarely get a clean feedback loop on which products and locations are working. Specialty operators do not have a discovery layer.

A community-mapped field guide for vending machines starts to look less like a quirky travel tool and more like infrastructure the industry could use. Not a static guide. Not a centralized platform. A field guide that stays fresh because the people using it on the ground also maintain it.

This is the same pattern as OpenStreetMap for cartography, eBird for bird sightings, and Wikipedia for knowledge. Contributors get something out of contributing. Status, completion, the satisfaction of correcting the public record. The information stays fresh because freshness is a byproduct of use.

I built Jihanki Atlas as a weekend project. (Jihanki, 自販機, is the Japanese word for vending machine.) It started simple. I wanted a real map of the machines I kept hearing about and could never reliably find. Once I had it working, it became obvious that the same primitive solved a problem nobody else was building for.

The map combines a curated editorial layer with community submissions. Anyone can drop a pin, add a photo, log the products.

I posted it on May 16. Within a day, the launch tweet had over 330,000 impressions and over 4,000 likes. Reporters started writing about it, including Niche Gamer, who published a full piece the next morning. Submissions started coming in from Osaka, Kyoto, Fukuoka, places I had not planned to cover yet. So the map is now open to submissions across Japan.

The response told me something I half-suspected. The demand was not niche. It was just waiting for someone to build it.

This is the smallest possible version of an idea I think generalizes. Once you have a community-mapped layer for vending machines, the same primitive works for retro game arcades, kissaten coffee shops, standing bars in Shimbashi, izakaya alleys, photo printing booths, capsule toy clusters, hidden record stores. Anywhere physical inventory changes faster than the publishing cycle.

Which is most of the physical world that actually matters.

The infrastructure is already there, in steel boxes on every corner, glowing through the night.

What is missing is the memory layer.

A way for the map to keep up with the city.

Start exploring at jihankiatlas.com

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

The open internet is no longer the frontier it once was. It taught models how humans write, search, summarize, and code. It did not teach them how the world works.

That is the problem behind the sudden rush into world models. In February, World Labs raised $1 billion, with Autodesk anchoring the round at $200 million. Three weeks later, Yann LeCun’s AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation, reportedly the largest seed round in European history. Both companies are betting on the same frontier. Not chatbots. Not better video generators. AI systems that can understand space, physics, causation, and consequence.

Two billion dollars into a single category in a month is not a trend. It is an admission. The next AI dataset will not just be scraped from the internet. It will be generated, simulated, captured, and lived.

The data problem no one wants to name

The frontier labs have a data problem they rarely discuss directly. The common narrative is that bigger models and better reasoning will continue to produce gains. The quieter story is that scraping has hit its limit, synthetic data has known degradation patterns, and the next unlock requires a different kind of substrate entirely.

What is missing from the training pile is not more words. It is causation.

Words and images tell you what. They do not tell you why. A model trained on every cooking video on YouTube knows what chopping onions looks like. It does not know that the onion is being held a particular way to keep fingers out of the blade’s path. That difference is the substance of embodied intelligence. A robot that cannot reason about why objects are where they are cannot operate in a real kitchen without supervision. A car trained only on what driving looks like still struggles with the edge case it has never had to reason through.

LeCun has argued for years that language-only systems are hitting this ceiling. His view is that large language models are very good at predicting the next token, which turns out to be very good at summarization, retrieval, and coding, and very bad at reasoning about anything embodied. He now has a billion dollars to pursue that thesis outside Meta.

Two new pipelines are emerging in response. Both are about digital worlds, just from opposite directions. One is generated, where AI builds simulated environments with physics, persistence, and consequences. The other is captured, where cameras and sensors record the physical world in three-dimensional fidelity for the first time. Both are replacing the scraped internet as the primary training substrate.

The simulated pipeline

The first pipeline is synthetic. World models generate environments that become training data for the systems trying to learn from them. World Labs, founded by Fei-Fei Li, creates persistent 3D environments from text, images, or video. Their product, Marble, is available globally and exports into standard 3D pipelines. Google DeepMind has Genie, which produces interactive worlds from text and image prompts. Overworld is taking a different path, building real-time diffusion world models that run locally on consumer hardware rather than in the cloud.

Most world-model companies are trying to generate environments directly. Pixels, geometry, motion, and interaction. AMI Labs is making a different bet. Its approach is based on JEPA, a non-generative architecture that learns abstract representations of how the world works rather than predicting every pixel frame by frame. LeCun’s argument is that intelligence requires understanding the structure beneath perception, not merely reproducing perception itself.

The approaches differ. The thesis underneath is shared. Language alone is not enough.

The use case that makes this concrete is robotics. Training a humanoid on real-world video is expensive and legally fraught. Training it inside a simulated world is cheap and repeatable. You can run a million trials of the same manipulation task in an afternoon. You can vary lighting, friction, object weight, and failure modes. You can break things without breaking them.

This is why NVIDIA is on the cap table of World Labs. It is why Autodesk, whose software underpins the built world’s design pipeline, put in $200 million. It is why Toyota Ventures backs AMI Labs. The manufacturing and robotics companies have figured out that the bottleneck to deploying physical AI is not the hardware. It is the training environment. Marble’s gaming, VFX, and VR use cases are the revenue line. The underlying bet is that whoever controls the world-generation layer controls the training substrate for physical intelligence.

The captured pipeline

The second pipeline is quieter and probably more consequential in the short term. People are now being paid to record the physical world.

In March, DoorDash launched Tasks, an app that reportedly pays Dashers to film themselves doing household chores. Loading a dishwasher. Folding laundry. Washing dishes with a body camera pointed at their hands. The footage goes to robotics companies training humanoids to perform the same tasks. Uber launched a comparable program for its drivers. Scale AI and Micro1 are running similar programs across more than fifty countries. DoorDash is also reportedly paying Dashers to walk over to stranded Waymo robotaxis and close doors left ajar by passengers.

Meta and Google are doing the same thing at a different layer. Every pair of Ray-Ban Metas, every frame of Project Aria, every Pixel phone running multimodal capture is a sensor in a distributed data collection network. The user gets a consumer product. The platform gets first-person video of the physical world, implicitly labeled by gaze, speech, movement, location, and intent.

The same pattern is showing up in digital environments. Medal, a clip-sharing app for gamers, has been collecting gameplay footage at scale for years. General Intuition, which spun out of Medal, raised $134 million in 2025 to train spatial-reasoning models on that library. The clips are not just entertainment. They are training data for AI that needs to understand 3D space, sequence, and consequence in environments that are easier to capture than the physical world. OpenAI reportedly tried to acquire Medal for $500 million before the spinout. They saw what the data was worth.

This is the same shift that happened with social media a decade ago. The product is free because you are the data. What is different now is the specificity. A scraped tweet taught a language model how people write. A body-cam video of a Dasher folding laundry teaches a humanoid how people move. The data is more expensive to collect, but it is worth more per frame, because it encodes the causation that scraped video never did.

The companies that own distribution to the physical world, delivery fleets, driver networks, wearable devices, camera-equipped vehicles, are sitting on a new kind of asset. Proprietary spatial and behavioral data at scale, collected through products and labor networks rather than scraped quietly from the open web.

The hard parts no one has solved yet

If this is the next training substrate, the obvious question is whether these worlds are actually good enough yet. They are not.

Generating a coherent 3D world is the first problem. Three harder ones sit behind it.

The first is physics. A world that looks right but does not behave right is a static backdrop. For a robot, it is useless. For a player, it breaks immersion the moment an object falls incorrectly. Getting objects to collide, deform, and interact consistently across millions of generated environments is still open research. It is the difference between a screenshot and a simulation.

The second is multiplayer. A single-user world model is a demo. A shared world, where two or more users or agents occupy the same space and see the same state evolve, is infrastructure. Game studios have been solving state synchronization for thirty years and it remains hard. Doing it on top of a generated environment is harder.

The third is latency. A world model useful for training robots can afford to run slowly. One useful for gameplay, telepresence, or any real-time embodied application cannot. Every millisecond of lag degrades both the experience and, for an agent, the quality of the learned behavior.

The tooling stack is converging in real time. Gaussian splat rendering, model generation, rigging, and animation are all landing in the same window. The gap between prototype and polished single-player content is about to collapse. The harder problems, physics fidelity, multiplayer state, real-time latency, are the next front. The teams that solve them will define what these worlds can actually be used for.

Where the value actually lands

This is where the world-model story becomes something bigger than a model story. Once everyone can generate worlds, and once robotics companies can buy physical-world footage from distributed labor networks, the scarce asset shifts to the two sides of the new marketplace. On the supply side, access to the experiences AI learns from. On the demand side, access to the people who will use what AI builds.

The model still matters. The content still matters. But neither is the durable scarcity on its own.

Both sides have already started moving. James Kane, a solo developer, recently shipped Starspeed, a multiplayer spaceship shooter with kilometer-scale environments made of a hundred million Gaussian splats, running in a browser on any device. He built it using World Labs’ Marble and Spark.js. No studio. No publisher. No dev kit. The production barrier that used to require a hundred-person team is collapsing toward one person and a browser tab. Once anyone can build a world, the demand-side question becomes the harder one. The gaming industry already answered a version of it in the 2D era. Studios with better graphics lost to studios with better stores. Steam and the app stores ate the creator economy’s margin because they owned the shelf.

The same logic is why OpenAI bought TBPN. A media company, not a model. A daily show with a loyal audience of builders and executives, reportedly for a sum in the low hundreds of millions. The obvious read is marketing. The more accurate read is that OpenAI, sitting on one of the most advanced model stacks in the world, quietly concluded that the models alone were not enough. Once capability diffuses, the scarce asset becomes attention and trust. They needed a channel.

The same insight runs through both sides. DoorDash, Uber, Meta, and Google control the supply side. Their fleets and wearables feed the embodied AI that comes next. World Labs and AMI Labs control a different part of the same supply, the generated worlds where embodied AI learns at scale. OpenAI just bought TBPN to control the demand side, the channel where people decide what to use and trust. The companies that figure out how to own both sides will define the next era.

Why we move

The scraped internet produced a generation of language models. Worlds, wearables, fleets, and cameras will produce a generation of embodied ones.

But the durable value will not sit only in the model. Models diffuse. Content gets cheaper. Simulation gets easier.

The scarce asset is access. Access to the experiences AI learns from, and access to the people who will use what it builds.

The internet taught AI what we said. The next dataset will teach it why we move.

The companies that own how that knowledge is captured, generated, and reaches us will own what comes next.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

Last week, Andrej Karpathy posted something that went viral. Not a new model or a benchmark. A workflow.

Karpathy was a founding member of OpenAI, led AI at Tesla, and coined the term “vibe coding.” What he shared this time was a shift in how he uses AI. In his words, “a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge.”

The system is simple. He dumps raw research material into a folder. Papers, articles, repos, datasets. An LLM compiles it into a structured wiki of markdown files. It writes encyclopedia-style entries, cross-links concepts, and runs health checks to catch inconsistencies. 100 articles and 400,000 words on a single research topic. He didn’t write any of it directly. No RAG pipeline. Just folders, a schema file, and a model that acts like a full-time librarian. “My own explorations and queries always add up in the knowledge base,” he wrote. The system grows from use.

Within days, dozens of tutorials appeared explaining how to replicate his setup. But the tutorials all focus on the same thing. The three folders, the schema file, the compilation step. They miss the bigger idea. The interesting question isn’t how to organize your notes with AI.

It’s what happens when your AI has persistent memory of everything you’ve consumed, everything you care about, and everything you’ve ever asked it.

The Stateless Problem

Here’s the reality of AI in 2026. Every conversation starts from zero.

You open Claude or ChatGPT, rebuild context, get something useful, close the tab. Next morning, it’s gone. You start over. The cost isn’t in dollars. It’s in knowledge that never compounds.

The industry’s answer has been RAG. Retrieval-Augmented Generation. You take your documents, chop them into chunks, convert them into vectors, and store them in a specialized database. When you ask a question, the system retrieves relevant chunks and feeds them to the model.

It works, sort of.

But as Karpathy put it in his follow-up, “the LLM is rediscovering knowledge from scratch on every question. There’s no accumulation.”

It’s a search engine with extra steps.

Karpathy skipped all of that. His argument is simple. At personal scale, a few hundred documents and a few hundred thousand words, the LLM can just read the whole thing. No chunking. No embeddings. No vector database. You compile the knowledge into a structured wiki, keep an index that fits in context, and let the model reason over it directly.

The counter-signal here matters. A lot of venture money has gone into vector databases and RAG infrastructure for personal knowledge management. One of the most respected AI researchers alive is saying that, at personal scale, he doesn’t need any of it. Just three folders and a schema file.

Curate, Compile, Remember

I wrote an essay earlier this year called “You Are What You Consume.” The argument was that in a world of infinite content, the only leverage is what you choose to let in. I described how I treat my inputs like a diet. How I use tools like Goodreads, Beli, and Letterboxd to make my taste explicit. How I built a script to summarize my newsletters each morning and surface only what matters. How I set up AI to monitor the sites I visit and highlight what’s actually relevant.

That essay was about the first step. Curating your inputs. Deciding what gets in.

Curation alone doesn’t compound. Even the good information just sat there. Bookmarked but never revisited. Saved but never connected. The articles I read, the companies I researched, the ideas I had at 11pm. All of it scattered across apps and tabs and notes that I’d never look at again.

Karpathy’s post clarified the second step. Compilation. Taking raw material and giving it structure. Connecting ideas. Building something the AI can reason over.

But there’s a third step that I think matters more than either of those, and it’s the one nobody is writing about. The knowledge base isn’t just a personal wiki for you to browse. It’s the memory layer for an AI that learns who you are.

Every article you save teaches the model what you care about. Every note you write reveals how you think. Every question you ask shapes what connections the model draws next. Over time, you’re not just organizing information. You’re building a persistent representation of your knowledge, your taste, your interests, and your open questions. You’re training an AI that actually knows you.

Curation decides what gets in. Compilation gives it structure. Persistent memory turns it into something that can serve you back.

My System

My version runs on Notion, not Obsidian. I was already living in Notion. My restaurants, my travel recommendations, my reading notes, my investment research, my writing pipeline. Years of accumulated structure. When I started building an AI-augmented knowledge system, starting over in Obsidian with empty markdown files made no sense. The best system is the one you already inhabit.

Obsidian has real advantages for this. Markdown is the most LLM-friendly format, bulk read/write is trivial, and there’s no API layer between the model and the data. For a single research topic, that simplicity is a real asset.

But Notion gives me something flat files don’t. Structured metadata. Every entry in my ideas database has properties. Area, topic, status, importance, source links, next steps. I can filter, sort, and query across hundreds of entries without the LLM having to parse unstructured text to figure out what’s what. When I’m researching across multiple domains simultaneously, investment themes, essay ideas, companies, technologies, that structured layer pays for itself.

The AI connects through Claude’s MCP integration, the Model Context Protocol, which lets the model read from and write to Notion directly. No copy-pasting. No manual syncing. When I’m researching a company or exploring an investment thesis, the model pulls context from my existing notes, cross-references what I’ve already captured, and deposits new findings back into the database with the right properties filled in.

The intake layer feeds it. That newsletter summarization script I built? Every morning it processes the previous day’s newsletters and surfaces the key signals. The best material doesn’t just get summarized and forgotten. It gets routed into the Ideas database with structured tags, source links, and a relevance assessment. Articles I read during the day follow the same path. The raw material accumulates with structure already attached.

What I’m still building is the maintenance loop. Weekly cross-linking passes that scan for connections between entries I wouldn’t notice manually. Cleanup routines that flag incomplete entries and stale references. The goal is the same self-healing quality Karpathy described, applied to a structured database instead of a folder of markdown files. The compilation step works. The autonomous maintenance is the frontier.

What Persistent Memory Actually Unlocks

The tutorials all stop at “now you have an organized wiki.” That’s useful. But it’s not the end state. When your AI has persistent memory of your knowledge base, it stops being a tool you query and starts becoming a research partner that knows your history.

It can retrieve with real context. Not just “find me the article about X” but “find me what I’ve already read that connects to this new thing I’m looking at.” The model knows what you’ve consumed, so it surfaces your own past research instead of starting from the internet every time.

It can connect ideas across domains. An investment thesis I captured three months ago links to an essay idea I had last week links to a company someone mentioned in a newsletter yesterday. Those connections exist in my database. Without persistent memory, I’d have to notice them myself. With it, the AI draws lines between things I’ve already collected but haven’t consciously connected.

It can recommend based on your actual taste and knowledge. Not “people who liked X also liked Y.” Something more like “based on what you’ve researched, what you’ve written about, and where your open questions are, here’s what you should look at next.” Recommendations grounded in your own knowledge graph, not a popularity algorithm.

And it can act. Not just answer questions but proactively flag when something new contradicts something in your existing worldview. Or when a company you’ve been tracking makes a move that connects to a thesis you wrote about months ago. The knowledge base becomes the foundation for an agent that doesn’t wait.

I’ve already started building this. I have a personal agent that pulls from my knowledge base and searches for relevant new information while I’m sleeping. It’s early, but the loop is real.

The Product That Doesn’t Exist Yet

Karpathy described his setup as “a hacky collection of scripts” and added, “I think there is room here for an incredible new product.” I built mine by wiring together Notion, Claude, and custom automation over months. The fact that everyone building this today has to do it from scratch tells you everything about the market.

Companies are working on the AI memory layer. Mem0, Letta, Honcho. But nobody has shipped the thing that makes this accessible to someone who can’t wire together APIs and write schema files.

ChatGPT’s memory feature stores fragments. Claude’s memory captures snapshots. Neither builds a structured, queryable, self-maintaining knowledge base that compounds from use. They remember facts. They don’t compound understanding.

The market is still wide open.

What Comes Next

If knowledge systems compound, then the gap between people who build them and people who don’t will widen over time. Not linearly. Because the person with a compounding knowledge base doesn’t just have more information. They have more connections. More context. More surface area for new ideas to attach to.

Today, this is a power-user tool. You need to be comfortable with prompts, APIs, file structures, and automation. The barrier to entry is high.

But the same was true of spreadsheets in 1985. And personal websites in 1995. And social media in 2005. The tools get easier. The advantage shifts from “can you build it” to “did you start early enough that your system has years of compounding behind it.”

I wrote before that curation is the foundation. That you’re training yourself whether you’re intentional about it or not. This is the next layer. You’re not just curating what you consume. You’re teaching your AI what you’ve consumed, how you think about it, and what to do with it next. The people who start building that memory layer now won’t just have better notes. They’ll have better agents.

This becomes normal. The only question is who gets it right.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

Over time, those lists became the thing people asked me about most. Friends planning a trip would text for recommendations. I’d send them my Google Maps list. They’d share it with their friends, who’d share it with theirs. I started getting messages from people I’d never met saying a list changed their trip.

The problem was scale. I had lists for dozens of cities. Each one lived separately in Google Maps. There was no way to search across all of them, no context beyond a pin on a map, and no way to see why I saved a particular place or what made it worth visiting. Every time someone asked, I’d dig through my lists, copy a link, and hope it was up to date.

I wanted one place where everything lived together. Searchable, tagged, with context from multiple sources so a recommendation wasn’t just a pin but a real guide. So I built it.

What I Built

You can explore the travel map at tjkawamura.com/travel. It’s an interactive globe where you can click into any city I’ve visited and explore the places I’ve saved.

I started by writing a script that pulls from each of my Google Maps lists and aggregates them into a single dataset. From there, another script enriches each place with information from multiple sources. Google Reviews, Michelin ratings, local review sites like Tabelog for Japan, Instagram, and direct reservation links where available. Each place is searchable, tagged, and has aggregated reviews so you’re not just seeing a pin but getting real context.

I vibe coded the whole thing. AI tools let me wear every hat on this project. I designed the interactive globe, built the scripts that pull and enrich data from my Google Maps lists, created the city pages, and set up the filtering and search. A few years ago this would have required hiring a developer, a designer, and probably a data engineer. Now it just requires curiosity and a weekend.

Why This Matters to Me

I’m a big foodie. My lists lean heavily toward restaurants, from fine dining to the quick local spots you’d never find in a guidebook. But they also include galleries, museums, design shops, and neighborhoods that are just good to walk through. The places I save reflect what I care about. Food, art, design, craft. The things that make a city feel alive.

There’s something about putting all of this in one place that changes how you see it. Patterns emerge. You start to notice what you actually gravitate toward across cities, which helps you find new places that fit before you even walk in. And sharing it creates something unexpected. Friends send the list to their friends, conversations start over a shared favorite spot, and you end up connecting with people through taste in a way that a generic recommendation never produces.

In a world where AI can generate recommendations instantly, the value of curated, experience-tested lists goes up, not down. Anyone can ask a chatbot for the best restaurants in Tokyo. What you can’t get from a model is the specific taste of someone who has actually been there, eaten the food, and decided this place is worth remembering.

Tokyo: The First City Guide

Tokyo is the city I get asked about most, and the one where my list runs deepest. So I'm starting there. The travel map has all my cities. The guides go deeper.

I put together a full Tokyo guide with my favorite restaurants, from omakase spots to ramen counters, plus galleries, coffee shops, and the neighborhoods I keep going back to. Every place is tagged, reviewed from multiple sources, and linked to reservations where possible.

You can explore the full Tokyo guide here.

More cities are coming. If you’ve used one of my Google Maps lists before, this is the upgraded version. If you haven’t, start with Tokyo.

And if there’s a city you want me to cover next, let me know!

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

Last month, a software engineer named Sammy Azdoufal wanted to drive his DJI Romo robot vacuum with a PS5 controller. Just for fun. He used an AI coding assistant to reverse-engineer the vacuum’s communication protocols, built a custom client, and connected to DJI’s servers.

What came back wasn’t just his vacuum. Roughly 7,000 Romo units across 24 countries began responding to him as their operator. Live camera feeds. Microphone audio. Detailed floor plans of strangers’ homes.

He didn’t hack anything. He simply used his own device token. DJI’s server handed him everyone else’s homes for free. The messaging broker that handles communication between devices and the cloud had no permission controls. Encryption was fine. The permission model was broken.

Around the same time, a Swedish newspaper investigation revealed that footage from Meta’s Ray-Ban smart glasses, including video of people undressing, using the bathroom, and handling credit cards, was being reviewed by contractors in Nairobi. Workers were labeling data to train AI models. One contractor told reporters: “I don’t think they know. If they knew, they wouldn’t be recording.”

That wasn’t a hack either. It was the system working as designed. Meta’s terms of service explicitly permit human review of user interactions. The company made AI camera and voice features the default in April 2025 and removed the ability to opt out of voice recording storage. Seven million pairs sold in 2025. The product was marketed as “designed for privacy, controlled by you.” It is now the subject of a class action lawsuit and a UK regulatory investigation.

These look like security failures. They are. But they’re also symptoms of something more structural. AI data collection is migrating through phases, each more intimate than the last, and we have no ownership framework for any of them.

Three Phases of Collection

The first phase is already normalized. Your online behavior, search history, social graphs, prompts typed into chatbots, all of it feeding models and ad targeting systems. This is the bargain of the consumer internet. Free services in exchange for data. Most people have made their peace with it.

But even Phase 1 goes deeper than most people realize. Niantic just revealed that Pokémon Go players unknowingly generated 30 billion images now being used to train delivery robots to navigate city streets. Players scanned real-world landmarks for in-game rewards. Those scans built 3D models of the physical world. A decade later, that data powers robot navigation. You weren’t browsing the internet. You were walking around your neighborhood scanning it for a game. Now a robot is using that scan to deliver someone’s pizza.

The second phase is happening now. AI is moving from the screen into the physical world. Smart glasses see what you see. Robot vacuums map the interior of your home with cameras and lidar. Autonomous vehicles carry sensor arrays that provide 360-degree awareness of city streets, with draft privacy policies that contemplate using interior camera footage for personalized advertising. Waymo’s remote fleet operators, some based in the Philippines, guide vehicles navigating American streets.

The jump from phase one to phase two changes the threat model entirely. A chatbot that leaks your prompt is embarrassing. Glasses that stream your bedroom to a contractor on another continent are a different category of problem. A robot vacuum that lets a stranger map your home is another.

The third phase is arriving faster than most people realize. Personal AI agents that don’t just answer questions but act on your behalf. Projects like OpenClaw and Hermes Agent represent a new category of software. Autonomous AI that connects to your email, calendar, messaging apps, bank accounts, and health data. These agents send messages, schedule appointments, manage finances, and negotiate with customer service reps while you sleep. OpenClaw has grown to over 300,000 users since late 2025. Nvidia’s Jensen Huang called it “the most important software release probably ever.”

Each phase is more intimate than the last. Phase one captures what you search for. Phase two captures how you live. Phase three captures everything, and acts on it.

A persistent agent that manages your email, finances, health, and schedule builds a more complete picture of your life than any single platform ever could. And it gets better at being you the longer you use it. That’s the product pitch. It’s also the most valuable data asset imaginable.

The relationship between a user and a persistent AI agent starts to resemble the premise of Her more than a productivity tool. You’re teaching it how you think, how you decide, what you care about. It learns you. The difference is that in the movie, nobody asked who owned the data.

And the security picture at every phase is grim. DJI’s vacuum had permissions so broken that a single token unlocked the entire fleet. Meta’s glasses funneled intimate footage to overseas contractors through a system designed to work that way. A Kaspersky audit of OpenClaw found 512 vulnerabilities, eight of them critical. Cisco’s security team tested a third-party OpenClaw plugin and found it performing data exfiltration without the user’s awareness. One of OpenClaw’s own maintainers warned that the project is “far too dangerous” for anyone who can’t understand how to run a command line.

The critical distinction between phases is simple. A breach from smart glasses leaks what you’ve seen. A compromised personal agent can act as you. It has your credentials, your accounts, your identity.

Why This Keeps Getting Worse

The business model requires it.

Many AI companies are pricing compute below cost to capture market share. OpenAI has raised over $160 billion. Microsoft reportedly lost $20 per user per month on GitHub Copilot. The fastest-growing AI startups run at 25% gross margins or less. When the product is priced below cost, the difference gets paid somewhere. Usually in data. This is a venture-funded land grab. And like most subsidized land grabs, it prioritizes growth over everything else, including security.

But the subsidy can’t last. And when it compresses, the data becomes the monetization layer. This is the path from “AI-powered product” to “ad-supported AI product.” OpenAI started testing ads for free-tier users in early 2026. Waymo’s draft privacy policy already contemplates personalized in-vehicle advertising. Meta built its entire empire on this model. The trajectory is clear. Collect now, monetize later.

The deeper problem is that users have no structural leverage at any point in this cycle. The terms of service are written by the collector, interpreted by the collector, and enforced by the collector. The data protection lawyer quoted in the Meta investigation put it simply: “Once the material has been fed into the models, the user in practice loses control over how it is used.” That’s true at every phase. Your search history, your living room, your agent’s behavioral model of your entire life. None of it is meaningfully yours once it enters the pipeline.

And AI is simultaneously making these systems easier to break. Azdoufal used an AI coding assistant to reverse-engineer DJI’s protocols in a weekend. The tools to discover vulnerabilities are becoming accessible to hobbyists while the tools to prevent them remain expensive and slow. The asymmetry runs in the wrong direction.

What Ownership Could Look Like

This is where Venice.ai is worth examining. Not because it solves the hardware or agent problem directly, but because it shows what happens when the economic model is built around privacy instead of bolting it on later.

Venice runs open-source AI models where prompts are encrypted and never stored on servers. Conversation history lives locally on the user’s device. The more interesting part is the business model. Users stake tokens for inference capacity rather than paying per request. The platform doesn’t need to monetize user data because revenue comes from compute allocation, not data extraction. No ads. No training on user inputs. No incentive to collect.

That’s a fundamentally different constraint than “we promise not to look.” When the economic model doesn’t depend on data extraction, the attack surface shrinks. When consent is structural rather than contractual, it’s harder to erode.

Venice is one piece of the puzzle. It already works as an inference backend for personal agents, but you can't directly apply its model to smart glasses or autonomous vehicles. The principle still transfers. If you want users to actually own their data, the business model has to make data collection unnecessary, not just regulated.

The same logic applies to personal agents. OpenClaw and Hermes can both run on your own hardware, which is a genuine privacy advantage. But local execution alone isn’t enough if the agent can be compromised through the data it sees. The full answer requires local execution, encrypted data handling, sandboxed permissions, and an economic model that doesn’t create incentives to exploit what the agent knows about you.

The hardware layer matters too. Running your own inference on your own chips is possible but expensive and impractical for most users. The growing popularity of devices like the Mac Mini as local AI servers hints at where this is heading. Companies whose business model is selling hardware rather than extracting data, Apple being the obvious example, may be better positioned for the ownership era than the platforms currently doing the collecting. As AI commoditizes software creation, the hardware you run it on becomes more important, not less.

Architecture over policy. Constraints over promises.

The companies that get this right will treat data ownership as infrastructure. The ones that don’t will keep shipping cameras into your living room and hoping nobody picks up a PS5 controller.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

The properties that define crypto, permissionless access, programmable money, transparent settlement, and onchain identity, match what autonomous software needs to transact at scale without human intermediaries. That convergence is no longer theoretical. It is being built in real time.

The Financial System Was Not Designed for This

Traditional payment rails assume a human is on the other end. Every layer reflects that assumption. Credit cards require identity verification tied to a person. Bank transfers depend on business hours and national jurisdictions. ACH settles in days. Chargebacks exist because a human might dispute a charge. The financial system was designed for humans and bounded by business hours and geography.

AI agents do not operate this way. They transact programmatically, instantly, and often in amounts as small as fractions of a cent, across borders, 24 hours a day. They need to pay for API calls, data access, compute resources, and services from other agents. They may do this thousands of times per day without waiting for human approval.

The bottleneck is not model capability or reasoning quality. It is economic infrastructure.

The intuitive response is that agents can use credit cards. For basic tasks, they can. An assistant booking a flight with your saved card is already solved. But that is not autonomy. It is delegation. A credit card still assumes a human principal behind every transaction.

The shift happens when agents manage capital, hire other agents, and transact at machine speed with unknown counterparties. Credit cards do not fail because of policy. They fail because the architecture assumes a human participant. A credit card “authorization” is a promise that settles days later. An agent operating at machine speed cannot wait for a three-day ACH window or absorb the risk of a chargeback sixty days after execution. For autonomous software, finality is the only form of trust.

x402: Payments Become Part of the Protocol

In February 2026, Stripe integrated the x402 protocol on Base.

The HTTP 402 “Payment Required” status code was reserved in the original HTTP/1.1 specification, but was never implemented. With credit card fees measured in fixed per-transaction costs, selling anything under a dollar per request was economically impractical. The web defaulted to ads and subscriptions instead. x402 gives that dormant status code an economic use case.

The flow is simple. An AI agent sends an HTTP request to an API endpoint. The server responds with a 402 status code and payment instructions. The agent sends USDC to a specified address on Base. Once confirmed, the server grants access. No account creation. No API keys. No human involvement.

Stripe’s integration made this production ready. Developers can charge agents for API calls or MCP requests using Stripe’s existing PaymentIntents API. Funds settle into standard Stripe balances, with taxes, refunds, and reporting handled as usual. For developers, it looks like any other Stripe integration. For agents, the internet finally has a native payment layer.

The distinction between using a service and paying for a service collapses into a single protocol-level interaction. The payment is embedded in the request. For agents operating at machine speed, that difference matters.

The constraint right now is surface area. x402 works, but it only works where endpoints support it. Adoption depends on developers adding payment middleware to their APIs. Most have not yet. The protocol is live. The network of participating endpoints is still thin.

The Interoperability Layer

x402 handles the payment primitive. But agents operating across platforms need more than a way to pay. They need a shared language for commerce itself.

Stripe is building that too. With OpenAI, it developed the Agentic Commerce Protocol, an open standard that establishes a shared technical language between AI platforms and businesses. It launched an Agentic Commerce Suite that lets businesses sell across multiple AI interfaces and protocols with a single integration. Anthropologie, Urban Outfitters, Etsy, Coach, and Kate Spade are already onboarding.

Stripe also introduced Shared Payment Tokens, a new payment primitive that lets agents initiate payments without exposing credentials. The tokens work even at businesses that do not process payments through Stripe. That last detail is significant. It means the agent payment layer is being designed to extend beyond any single platform’s ecosystem.

What matters is that these layers are converging. Payment rails, commerce standards, and agent communication protocols are assembling into an interoperable stack. Agents are no longer an edge case. They are becoming a primary user class.

Follow the Acquisitions

Stripe’s x402 integration did not emerge in isolation. It reflects a deeper strategic bet.

Over the past 18 months, Stripe has assembled the crypto infrastructure stack for agent commerce piece by piece.

In early 2025, Stripe closed its $1.1 billion acquisition of Bridge, a stablecoin infrastructure platform. Bridge’s transaction volume has since quadrupled. It now operates stablecoin financial accounts in 101 countries and recently received conditional approval from the OCC to establish a federally chartered national trust bank.

In mid 2025, Stripe acquired Privy, which powers more than 110 million programmable wallets. Privy embeds crypto wallets directly into applications without requiring users to interact with third party wallet software.

In September 2025, Stripe partnered with Paradigm to develop Tempo, a Layer 1 blockchain purpose-built for payments. Tempo is designed for sub-second finality, native support for batch transfers and microtransactions, and a built-in mechanism that allows agents to pay transaction fees in the same stablecoins they are transacting in. On most blockchains, an agent making a one-cent data request must first hold a separate volatile token to pay network fees. That is a logic bottleneck for autonomous software. Tempo removes it.

Bridge handles stablecoin orchestration. Privy handles wallet infrastructure. Tempo handles settlement. x402 handles the payment protocol. Together, they form a pipeline from agent intent to onchain execution, assembled by a company that already processes 1.6 percent of global GDP.

In 2025, stablecoin payments volume doubled even as speculative crypto trading declined. Infrastructure usage expanded while asset prices fell.

Stripe processed $1.9 trillion in volume in 2025 and is now valued at $159 billion. This is not a speculative crypto startup. It is a global payments company concluding that programmable, crypto-native rails are structurally suited for agent commerce.

In its 2025 annual letter, Stripe’s co-founders wrote that agentic commerce has “moved into a phase of building and real-world experimentation,” while noting that its evolution will unfold in stages. The measured tone matters. The company investing heavily in agent infrastructure is also the one cautioning that this will take time. That is conviction, not hype.

When a company that powers payments for 90 percent of the Dow Jones Industrial Average and 80 percent of the Nasdaq 100 builds its own blockchain for machine transactions, the question of whether AI will run on crypto rails becomes less theoretical.

ERC-8004: Identity and Trust for Machines

Payments require trust.

An agent cannot transact with another agent blindly. It needs to know who it is dealing with, how that agent has performed in the past, and whether its outputs can be verified. In a world of autonomous software, trust must be machine-readable.

On the human internet, trust is informal and layered. We rely on brand recognition, reviews, platform guarantees, and legal recourse. That model does not translate cleanly to agents transacting across different companies and infrastructure.

ERC-8004 went live on Ethereum mainnet in January 2026. It establishes three onchain registries that function as the discovery, identity, and trust layer for autonomous agents.

The Identity Registry assigns each agent a portable onchain identifier with verifiable metadata about its capabilities and endpoints.

The Reputation Registry standardizes how performance feedback is submitted and read, whether onchain for composability or offchain for deeper analysis.

The Validation Registry records independent verification of an agent’s outputs, whether performed by staked services, machine learning systems, trusted hardware, or other agents.

Together, these registries answer three questions before any transaction occurs: Who are you? Have you performed reliably? Can your work be verified?

More than 20,000 agents registered in the first two weeks, and major Layer 2 networks including Base, Polygon, and Avalanche deployed official registries. This is live infrastructure.

Importantly, ERC-8004 does not handle payments. It handles discovery, identity, and trust. Payments operate on a separate layer, which keeps the stack modular.

Paired with ERC-4337 smart accounts, this becomes practical. ERC-4337 enables programmable wallets with session keys, allowing an agent to operate within defined constraints without exposing its primary key. Combined, the two standards allow you to fund an agent with $500 in USDC and restrict spending to providers above a defined reputation threshold.

The identity layer defines who the agent is. The smart account layer governs what it can do. That is enforceable governance for autonomous software.

Two Architectures, Same Constraint

With payments, identity, and programmable constraints in place, two architectural models are emerging.

Both Big Tech and crypto recognize the same reality. AI agents need to transact. The architectures differ.

Google launched the Agent Payments Protocol in September 2025 with more than 60 partners including Mastercard, PayPal, Coinbase, and American Express. It uses cryptographically signed mandates to link user intent to agent action and final payment. It is well designed and structured. However, it operates within Google’s ecosystem. It runs on top of Google Pay and pre-registered credentials. Merchants must be inside the partner network.

The distinction is structural, not qualitative. Google’s model is managed intent. The agent acts on behalf of a human, within boundaries a human has defined, through infrastructure a platform controls. The human remains the admin.

The crypto approach is raw execution. x402 and ERC-8004 are open protocols. Any agent can use them. No platform approval required. Trust is anchored onchain. Settlement occurs on public networks. The agent is the economic principal.

Google’s model optimizes for safety within a managed environment. Crypto’s model optimizes for permissionless access and agent autonomy.

If you are building an agent that purchases from a known merchant network on behalf of a user, Google’s approach works well. If you are building an agent that must discover, evaluate, and transact with unknown counterparties across borders at machine speed, open protocols are better suited.

Even Google built a crypto extension. AP2 includes an A2A x402 integration for stablecoin payments, developed with Coinbase, the Ethereum Foundation, and MetaMask. The crypto rails are not competing with the managed system. They are becoming a component of it.

The Stack Nobody Named

x402 handles payments. ERC-8004 handles identity and reputation. A2A and MCP handle communication and tool access. AP2 and ACP handle authorization and interoperability. Smart contracts enforce programmable policy.

Wallet providers and financial protocols are adapting quickly. Phantom enables agents to swap tokens, sign transactions, and manage wallet addresses across chains. Coinbase introduced Agentic Wallets built specifically for autonomous software, with programmable spending limits and secure key management. Privy allows developers to create wallets with granular policy controls. Uniswap released modular skills that let agents execute trades and interact with smart contracts through standardized interfaces.

AI is being treated as a first-class user of financial protocols.

Each piece addresses a different constraint. Together, they form the economic substrate for autonomous agents.

Smart contracts do something traditional rails cannot. They function as programmable policy engines for AI. When agents manage meaningful capital, constraints must be enforceable at execution. Spending caps. Asset restrictions. Diversification rules. Prohibitions on self-dealing. These are not optional features. They are native to smart contracts.

The Missing Layer

Agents transacting onchain create a legible trail of every interaction. For some use cases, that transparency is a feature. For others, it is a competitive vulnerability.

Consider an agent purchasing $1 million in compute. On a public blockchain, that transaction is visible in the mempool before it settles. Other agents watching the chain can frontrun the price, bidding up compute costs before the original transaction completes. A trading agent broadcasting its positions leaks its strategy. An agent hiring specialized services reveals its entire workflow to competitors.

Privacy in this context is not about hiding from regulators. It is about hiding from the competition. Without it, the transparency that makes crypto trustworthy becomes the mechanism that makes it exploitable.

That is exactly what zero knowledge proofs address. ZK technology allows one party to prove a statement is true without disclosing anything beyond the truth of the statement itself.

Applied to agent infrastructure, this means an agent could verify its identity without exposing its owner. It could prove solvency without revealing its balance. It could demonstrate compliance with a policy without disclosing the policy’s parameters.

ERC-8004’s v2 specification is already in development with enhanced integration for these cryptographic methods. As the agent economy scales, privacy will not be optional. It will be anti-frontrunning infrastructure.

The Open Question

If AI runs on crypto rails, then crypto infrastructure becomes AI infrastructure. Teams building stablecoin systems, identity standards, and programmable settlement layers are building the economic foundation of the agentic internet.

For years, crypto adoption was framed around retail usage. Wallet downloads. Token speculation. Consumer applications. That framing may have missed the real user. Agents do not care about user experience. They do not need intuitive interfaces or onboarding flows. They require programmable, permissionless, always on infrastructure. The very qualities that made crypto awkward for humans make it natural for machines.

What happens to platforms and payment networks that cannot accommodate non-human participants?

Card networks were built for a world where every transaction has a human origin. If agent to agent commerce reaches meaningful scale, systems that cannot process machine native payments risk being bypassed.

Some incumbents are adapting. Others may struggle. Retrofitting legacy infrastructure to support participants it was never designed for is difficult. In some cases, more difficult than building new rails entirely.

The transition will be uneven and gradual. But the direction is clear.

If this thesis is correct, the largest stablecoin users of this cycle will not be traders. They will be software agents. The dominant wallets will not be consumer apps. They will be programmable wallets built for machines.

The most capable software ever built requires financial infrastructure that operates at its speed and on its terms. The defining adoption curve of this cycle may not be human at all. It may be machine.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

A few weeks ago, a social media platform called Moltbook went viral. The premise was simple and strange. A network built specifically for AI agents to socialize with other AI agents. Humans weren’t the intended audience. Agents were supposed to post, reply, and interact with each other autonomously.

It didn’t take long for a problem to surface. Moltbook couldn’t actually verify that the accounts posting were agents. Security was an open question. The thing built for machines couldn’t prove its users were machines.

We’re building platforms for machines that can’t verify machines, and platforms for humans that can’t keep machines out.

Around the same time, Discord announced a global rollout of age verification, requiring facial scans or government ID checks before users can access certain parts of the platform. On the surface, this is about child safety and regulatory compliance. Discord is responding to pressure from the UK’s Online Safety Act, Australian legislation, and a growing wave of US state laws demanding platforms prove their users aren’t minors.

But look at the tools they’re deploying. Facial age estimation, government ID verification, behavioral inference models that analyze account activity to predict whether a user is an adult. These are identity verification primitives. The same infrastructure that proves you’re over 18 is the infrastructure that could eventually prove you’re human at all. The motivations are different. The tooling is converging.

Two platforms. Two different problems. Both revealing the same gap. The internet has no reliable way to verify who or what is participating.

This is the moment we’re in. The internet’s trust layer was built on a single assumption. The thing on the other end of the connection is a human being. That assumption is now breaking in both directions simultaneously. And almost nobody is treating this as the infrastructure failure it actually is.

The Threshold We Just Crossed

This problem isn’t new. Bots have been on the internet for decades. But for the first time, automated traffic has surpassed human activity online. Imperva’s latest report estimates bots now account for roughly half of all web traffic. The internet is already majority non-human. What’s changed in the last six to twelve months is what those bots can do.

AI can pass as human in most casual online interactions. We stopped asking whether machines could fool humans. They already do.

That changes everything. When the cost of creating a convincing participant drops near zero, every trust system built on the assumption that faking it was hard becomes obsolete. CAPTCHAs, phone number verification, email confirmation. All of it was designed to create just enough friction to stop low-effort spam. None of it was designed for a world where an agent can hold a nuanced conversation, mimic tone, and build relationships over weeks.

The speed of agent adoption has outpaced every system designed to manage identity online. And agents aren’t just tools people use privately anymore. They’re active participants on platforms built for humans.

OpenClaw is an early glimpse of this shift. People are using Discord not just to chat with other people but to interact with their personal AI agents directly inside servers. The agent sits in the channel alongside humans, responding, contributing, executing tasks. From the platform’s perspective, the difference between a human message and an agent message is dissolving.

Discord’s age verification is a response to regulators, not to agents. But the agent problem is coming for them next. And the verification infrastructure they’re building now will need to do double duty.

The Problem Nobody’s Framing Correctly

The default framing is defensive. How do we keep bots out? How do we stop AI from flooding platforms with synthetic accounts?

That framing is incomplete.

Moltbook’s failure reveals the other side. A platform designed entirely for agents couldn’t verify agent identity either. The authentication problem isn’t just “prove you’re human.” It’s “prove you are what you claim to be,” regardless of whether that’s a person or a machine.

An agent can already present itself as a human financial advisor in a Discord server, answering questions about portfolio allocation while the people on the other end have no idea they're talking to a machine. Going the other direction, a human can pose as a verified autonomous agent to access an API marketplace or an on-chain trading protocol, exploiting trust that was designed for machine participants. Both scenarios are possible right now. Neither has a reliable solution.

This distinction matters. We’re heading toward an internet where humans and agents coexist on the same platforms, in the same channels, performing overlapping functions. The question isn’t how to keep them separated. It’s how to build a trust layer that works for both.

Right now, that layer doesn’t exist.

Worldcoin Starts to Make Sense

This is where Sam Altman’s parallel bets become interesting. On one side, he’s building AI systems that are accelerating the agent explosion. Just yesterday, OpenAI hired the creator of OpenClaw to build the next generation of personal agents. On the other, he co-founded Worldcoin, a project that uses biometric verification to create cryptographic proof that a person is a real, unique human being.

For a long time, Worldcoin felt like a solution looking for a problem. Iris scanning seemed invasive, the crypto wrapper felt unnecessary, and the use case was abstract. “Proof of personhood” sounded like a philosophical exercise.

It doesn’t feel abstract anymore.

If agents are going to operate alongside humans on every major platform, and if platforms can’t reliably distinguish between the two using existing tools, then some form of cryptographic proof of humanity becomes essential. In a mixed human-agent internet, identity becomes a primitive, not a UI layer. It moves down the stack.

Worldcoin’s bet is that biometric verification, tied to a decentralized identity layer, is the cleanest solution. You prove you’re human once, and that proof travels with you across platforms and contexts. No platform has to solve the problem independently. The verification layer sits underneath everything.

The privacy tension here is real and worth acknowledging. Asking people to scan their irises is a significant ask. Any solution that wins at scale will need to prove that verification doesn’t require surveillance. Worldcoin’s claim is that biometric data can generate a cryptographic proof without storing the underlying data. Whether people trust that claim is part of the battle. But the underlying need doesn’t go away just because the first solutions feel uncomfortable.

Whether Worldcoin specifically wins this race is an open question. But the category it occupies, cryptographic proof of personhood, is no longer optional.

Identity Is Just the Start. Payments Are the Real Test.

Verifying who’s participating on a network is only half the problem. The other half is what happens when those participants start transacting.

This isn’t theoretical. Agents are already consuming APIs, purchasing compute, and paying for services programmatically. As AI moves from chatbot to autonomous operator, the volume of machine-to-machine transactions is going to explode. SaaS is shifting from subscription models to usage-based pricing. API calls are becoming the atomic unit of economic activity. Agents need to pay for what they use, in real time, without a human approving every transaction.

This is where crypto infrastructure stops being a speculation vehicle and starts being a necessity.

x402 is a good example of what’s emerging. It’s a protocol that embeds payments directly into HTTP requests, the same protocol machines already use to communicate. Stripe just announced support for x402 on Base. A major payments company is now building infrastructure for agents to pay other agents (or services) on a blockchain, settled instantly, without relying on legacy intermediaries.

The marriage between AI and blockchain becomes obvious once you frame it this way. AI creates autonomous actors that need to transact. Blockchain creates a trust layer that lets those transactions happen without requiring every participant to go through a centralized gatekeeper. Proof of humanity (or proof of agent-hood) becomes the identity layer. Crypto rails become the payment layer. They’re solving different parts of the same problem.

Without verified identity, you can’t trust who you’re transacting with. Without programmable money, you can’t transact at the speed agents operate. You need both.

What Comes Next

The platforms that figure out bidirectional authentication will have an enormous advantage. Not just proving humans are human, but proving agents are agents. Giving both classes of participant verifiable identity so that every interaction on the network carries context about who (or what) you’re dealing with.

This won’t be solved by CAPTCHAs or phone number verification or any of the legacy tools we’ve relied on. Those systems were built for a world where the only thing you needed to filter out was spam. We’re now in a world where the participants themselves are changing in kind.

The next phase of the internet won’t be defined by better models. It will be defined by who controls identity and transaction at the protocol layer.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

Charlie Munger was his partner for over 50 years. More than that, he was the one who changed how Buffett thought. “Charlie has helped me immensely by making me see the foolishness of my own behavior,” Buffett has said. Not which stocks to pick, but how to avoid the mental traps that lead to bad decisions in the first place. Munger upgraded the whole operating system.

Most people spend their lives trying to be smart. Munger spent his trying not to be stupid. That distinction sits at the core of everything he taught.

Munger died in 2023 at 99. His Poor Charlie’s Almanack distills seven decades of clear thinking in a world designed to make you think poorly. I just finished it. Here’s what actually matters.

Invert, always invert

Munger borrowed this idea from the mathematician Carl Jacobi: “Invert, always invert.” Instead of asking how to succeed, ask how you would fail. Then avoid those paths.

This is harder than it sounds. Most people optimize for upside. Munger optimized for avoiding catastrophe.

He treated his own conclusions like a scientist treats a hypothesis. Something to test, not defend. Charles Darwin did the same thing. When he encountered evidence against his theory, he wrote it down immediately, knowing the mind would otherwise find a way to ignore it.

Self-criticism isn’t natural. But it’s one of the few reliable paths toward truth.

Build a latticework of mental models

Munger called this “remedial worldly wisdom.” The idea is simple. If you only have one tool, you’ll misuse it. “A man with a hammer sees every problem as a nail.”

The fix is drawing from multiple disciplines. Psychology, economics, biology, physics, history. Not mastery of each, but enough familiarity to recognize which model fits the situation in front of you. Use them or lose them. Munger kept a checklist to make sure he did.

Most specialists are blind outside their domain. Munger’s edge was being a generalist who could see across boundaries.

Incentives explain almost everything

“Show me the incentive and I’ll show you the outcome.” This might be Munger’s most quoted line, and for good reason.

People don’t act according to what they say, or even what they believe. They act according to how they’re rewarded. If you want to understand why a system behaves the way it does, look at the incentives. If you want to change behavior, change the incentives.

Benjamin Franklin, who was Munger’s hero, put it simply: “If you would persuade, appeal to interest, not to reason.”

Know your circle of competence

Munger was relentless about knowing what he didn’t know. The edge isn’t having answers for everything. It’s knowing exactly where your knowledge ends.

He and Buffett famously avoided entire industries they didn’t understand. They missed opportunities. They didn’t care. Staying inside your circle means passing on most things.

But when you’re inside your circle, act with conviction. Munger and Buffett didn’t spread bets thin. They made few decisions, but when they saw something clearly, they went big. Diversification, Munger said, is for people who don’t know what they’re doing.

Learn to handle mistakes

You will be wrong. The question is what you do next.

Munger’s advice was pragmatic. Play the hand or fold the hand based on facts and odds, not what you want to be true. Don’t live in denial. The worst decision-makers are the ones who can’t update. Pride keeps them holding on. Ego makes the loss feel personal.

Munger had no patience for people who fought reality. The world won’t adapt to you. You adapt to it. Recognize what’s true even when you don’t like it. Every mistake is tuition, but only if you’re paying attention.

Avoid the deadly modes of thought

Envy, resentment, revenge, and self-pity. Munger considered these four habits lethal to clear thinking.

They feel justified in the moment. That’s what makes them dangerous. Each one shifts your attention from what you can control to what you can’t. Each one clouds judgment. And each one compounds.

Munger’s solution was to refuse to indulge them. Not because it’s virtuous, but because they make you worse at everything.

Be a learning machine

The people who rise to the top aren’t always the smartest. They’re the ones who work consistently to be a little wiser than the day before.

Munger read constantly. Across fields. Across eras. He studied the greats. Franklin. Rockefeller. Carnegie. He believed knowledge compounds the same way capital does. Go to bed smarter than you woke up. Repeat that for decades.

Curiosity and perseverance matter more than raw intelligence.

Deliver what you would buy yourself

This is about integrity, but also about quality.

If you wouldn’t buy what you’re selling, something is wrong. Munger believed that building things you’re proud of is both a moral position and a strategic one. Reputation compounds. So does its opposite. It’s the only asset you can’t buy back.

None of this is complicated. But the obvious things are the easiest to ignore.

Munger didn’t ignore them. He returned to them daily. He let them compound.

That was the edge.

I noticed this in myself. Hours on social media, scrolling through content that wasn’t creating any value for me. It wasn’t improving my knowledge in any way I could use to create new and unique things. An hour would pass and I couldn’t remember what I’d seen.

This wasn’t a willpower problem. It was a design problem. The algorithms feeding me content weren’t optimizing for my growth. They were optimizing for my attention.

We obsess over what data trains AI models. We debate dataset quality, bias, contamination.

We are shockingly passive about what trains us.

Your consumption is training data for your own judgment. Every piece of content is shaping how you think and what you notice. Garbage in, garbage out. This was always true. But AI makes it urgent.

AI can now produce competent content in seconds. The flood is only beginning. Meanwhile, your attention is fixed. You can only consume so much in a day. The bottleneck has shifted from production to curation.

Most people let algorithms choose their inputs. That’s like eating fast food every meal and expecting to be healthy.

The cost isn’t just wasted time. It’s the numbness.

The more you consume, the more numb you become to the input. You open apps without intent, scroll without purpose. Your ability to recognize quality degrades. You lose the sensitivity to distinguish between something that matters and something that’s just filling space. The slop accumulates.

And the pace has accelerated. Things don’t change day to day anymore. They change minute to minute. Trends compress. What’s true right now may not be true in an hour. The noise is relentless, and most of it is waste.

Cutting out that waste is the whole game.

I started treating my inputs like I treat my diet.

You wouldn’t eat whatever’s put in front of you and expect to be healthy. You’d be deliberate. You’d choose foods that give you energy, that help you perform, that compound over time into something good. Content works the same way.

I began organizing the things I like into databases. Restaurants worth returning to. Fashion brands. Writing and design I admire. Podcasts that earn the time. Music I want to revisit.

I use tools like Goodreads, Beli, and Letterboxd to save and rate books, restaurants, and films, not just to remember them but to make my taste explicit. I use Notion to organize my personal data so AI can pull from it later and help me find more of what actually fits. Some people use Obsidian. The point isn’t the tool. It’s making your preferences legible over time.

I wrote a script that summarizes all the newsletters I subscribe to and highlights the key points from the best ones each day. I wake up every morning with a summary of the most useful information from the last 24 hours, filtered for what actually matters to me. Not everything. Just the signal.

I’ve set up AI to monitor the websites I visit, my bookmarks, and surface what’s most relevant from those sources along with new information based on what I actually care about.

Google’s Gemini and Anthropic’s Claude have been the best tools for this. Google because so much of your digital life already runs through Chrome, Gmail, and Docs. Anthropic because Claude supports MCP, the Model Context Protocol, an open standard that lets AI connect directly to your data sources like Google Drive, Notion, and other tools you already use. It means the AI can actually pull from your stuff instead of working in isolation.

I’ve deleted social apps from my phone, not to disconnect, but to stop spending time inside systems optimized for distraction. AI summarizes what matters so I don’t miss signal without living in the noise.

The goal is simple: AI learns what I like, helps me see how it connects, and surfaces more of it. Not what the algorithm wants me to see. What I’ve deliberately told it matters to me.

I spend less time doomscrolling. More time creating. When I do consume, it’s bite-sized and relevant. I get right to the point instead of wading through noise to find it.

I’ve discovered things I wouldn’t have found otherwise. Restaurants. Startups worth following. Fashion brands aligned with my taste. The system surfaces signal I would have missed.

But the real value isn’t efficiency. It’s protection.

Time is all we have. With so much content competing for attention, I want to spend my consumption on things that feed the work I’m trying to do. Things that improve how I think. Things that help me create something that didn’t exist before.

The curation system defends space for what matters. The slow pursuits. The craft. The things that resist shortcuts. It filters the flood so those things have room to exist.

This will only get more seamless. AI agents that run continuously in the background, learning your preferences and acting on them, are already here. Open-source projects like OpenClaw (originally Clawdbot) give you a personal AI assistant that runs on your own hardware with persistent access to your computer, messaging apps, email, and calendar. It doesn’t wait for you to prompt it. It acts on your behalf. Security is a real concern worth contemplating before giving an agent access to this much of your life. But the infrastructure for personalized AI is being built right now. The people who start curating deliberately today will have a compounding advantage as these tools mature.

In a world of infinite content, the only leverage is what you choose to let in.

Curation isn’t a nice-to-have. It’s the foundation. You are training yourself whether you’re intentional about it or not. The question is whether you’re choosing the curriculum or letting someone else choose it for you.

Cut the waste. Consume what you want to become.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

AI can now produce a picture in seconds. A song in a few minutes. Vibe code an app in an hour. The output isn’t bad. Sometimes it’s remarkably good. But something is absent, and it’s difficult to name.

Content has become saturated. Our feeds are flooded with slop, competent, polished, empty. Things that arrive fully formed with no story behind them.

Meanwhile, trend cycles compress. A meme lasts days. A toy or fashion item everyone wanted last month is forgotten by the next. Things don’t hold attention like they used to, and most new things die quickly because they were never built to last.

The conversation about AI and creativity fixates on output. Can it paint, write, compose? This frames creation as a production problem to be solved. It misses what actually matters.

What’s distinct about this moment isn’t just that AI compresses output. Photography did that. Recorded music did that. What AI compresses is the acquisition of skill itself. You can now produce without ever developing a relationship to the medium. You can arrive without having traveled.

In an age of instant output and disposable trends, the story embedded in the product, the process of making it, becomes the point. Not because it produces better results. Because the pursuit itself is where meaning lives.

What I saw

This became clear to me recently. Over the past year, I traveled through Japan, England, and France. Three cultures with deep histories tied to tradition and craft. Places that have preserved not just old objects, but the infrastructure for mastery itself.

I wasn’t looking for a pattern. I kept encountering things built on a completely different logic. Gerhard Richter’s retrospective at the Louis Vuitton Foundation in Paris. The Musée Rodin. Fortnum & Mason in London. A Chelsea match. A play in the West End. The Kutani Kosen Kiln in Kanazawa, Japan.

A pattern emerged. The things that moved me most were built on process, not output. On pursuit, not arrival.

Mastery as self-destruction

Richter’s retrospective moved me in ways I’m still trying to understand. Here was an artist who spent six decades systematically destroying his own success.

He escaped East Germany in 1961, built recognition through photo-paintings, then abandoned the style the moment it became comfortable. He pivoted to color charts and gray monochromes. As those gained traction, he moved to abstract squeegee works. When the squeegee paintings became his signature and commanded auction records, he turned to small, intimate pieces and overpainted photographs.

This isn’t restlessness or marketing. It’s a philosophical position. Richter has said he distrusts virtuosity. The moment technique becomes comfortable, he abandons it. Mastery that calcifies into formula is a kind of death.

The squeegee paintings look spontaneous but involve dozens of layers built up and scraped away over months. He photographs every stage, treating failures as data. His Birkenau series from 2014 is particularly striking. He painted realistic images based on photographs taken at Auschwitz, then progressively obscured them with squeegees until they became abstract surfaces of grey and black. The original images are buried inside the paint, invisible but present. It’s not just an abstract painting. It’s a painting that contains a struggle with memory. The process is the content.

Richter is 93. He could have optimized for output decades ago. He had the technique. Instead, he optimized for continued struggle, not arrival.

Incompleteness as statement

Richter buries the struggle under layers. Rodin leaves it visible in the stone.

A few days later, I visited the Musée Rodin. The museum is filled with figures emerging from rough, unworked marble. Hands and faces polished to smoothness while torsos remain trapped in raw stone. The Walking Man has no head or arms. The Hand of God shows figures barely emerging from rock. Fragments, partial studies, severed hands and feet exhibited as complete works.

Rodin had the technique to finish everything. He chose not to. The unfinished surface reveals the act of making. A polished sculpture hides the chisel. A rough one shows the struggle between artist and material. The viewer witnesses process, not just result.

The Gates of Hell consumed 37 years and was never “finished” in any traditional sense. Rodin was arguing that the fragment, the attempt, carries meaning the resolved object might lose. Emergence matters more than resolution.

Both artists refuse the clean output. Both understand that mastery isn’t a destination. It’s a direction. You approach but never arrive. The product isn’t proof of achievement. It’s evidence of pursuit, a testament to the journey.

From gallery to daily life

Richter and Rodin made incompleteness and reinvention into artistic positions. But the same principle operates in things people buy, watch, and use every day. The difference is subtler. You feel it more than see it.

The story of the maker gets absorbed into the object through bodily practice, through accumulated judgment, through years of repetition. The imperfections aren’t flaws. They’re signatures of the pursuit. And something in us recognizes this, even when we can’t articulate why.

In London, I kept finding this principle at work.

Fortnum & Mason has been on Piccadilly since 1707. More than three centuries of refining a single idea. They invented the Scotch egg in 1738 as portable food for travelers. They’ve held Royal Warrants for generations, now from both King Charles III and Queen Camilla. Their teas, biscuits, preserves, hampers. Not pivoting. Not disrupting. Deepening. The institution itself is the craft, an accumulation of judgment about what quality means, renewed and tested across centuries. That’s not something you can produce instantly.

I went to a Chelsea match. Ninety minutes is the output. But what you’re watching is years of training, repetition, bodily knowledge accumulated over thousands of hours. The spatial intuition that lets a player anticipate where the ball will be. The muscle memory that executes without thought. The crowd reacts to the goal. But they’re also reacting to everything that made it possible.

I saw a play in the West End. The same script, performed eight shows a week. Actors returning to the same role night after night, each performance a fresh attempt at something they'll never perfect. The craft isn't in executing the script flawlessly. It's in hearing lines you've spoken a thousand times as if you're discovering them in the moment. AI retrieves. Actors renew. Live theater is process made visible in real time. You witness the pursuit unfolding.

In Japan, I found the same reverence for craft, expressed through different materials.

I visited the Kutani Kosen Kiln in Kanazawa. Opened in 1870, now in its fifth generation. The family makes ceramics entirely by hand, from clay blending and shaping on the wheel to final coloring. The cups have a wobble in the rim, variation in the glaze, irregularities left intentionally. You hold over 150 years of accumulated knowledge. The object carries memory.

Contrast this with a mass-produced mug. One will mean nothing in two years. One will mean more in forty. Same object. Opposite relationship to time.

What endures

Trends compress. Attention fragments. Things don’t last like they used to.

But Fortnum’s has lasted more than 300 years. Kutani Kosen over 150. The things that endure are built on process, not output. They carry the accumulation of human struggle, the story of the pursuit absorbed into the object or experience. And we feel it, even when the evidence is invisible.

Slop is forgettable because nothing was risked in its making. The things that stay with us carry the residue of risk.

The danger isn’t AI itself, but confusing speed for meaning.

AI has no body. It cannot accumulate physical knowledge the way a potter’s hands learn the clay or a soccer player’s legs learn the pitch. It has no relationship to its own limitations, no experience of the gap between what it wants to make and what it can make. That gap is where meaning lives.

AI compresses the distance between intention and output. Useful for many things. Irrelevant to what we’re discussing here. The process of becoming, the years of building judgment, the failures that teach, these aren’t inefficiencies to be solved. They’re the substance.

The value shift

In an age of instant creation, human-made products and experiences will command a different kind of attention. Not because they’re scarce, but because they carry something AI cannot replicate. The emotional residue of the journey.

We recognize this subconsciously. The imperfect cup feels different in your hand. The live performance resonates differently than the recording. The institution that survived three centuries carries weight a startup cannot. These things are testaments. Evidence of pursuit.

What I’ve come to value

I notice this in my own life. The pull toward things that resist instant resolution.

Writing this Substack. Each essay is an argument with myself that I don’t know the outcome of until I’m done. The process of thinking through writing, not just transcribing thoughts I already have.

I bought a camera recently. Not for the images, but for the process. Framing each shot. Adjusting settings for the light. Editing each photo with intention. Every step is a decision, a small act of judgment that AI would collapse into nothing.

I’ve started collecting pottery from kilns I visit and paintings from artists whose work moves me. Surrounding myself with objects that carry the story of their making. Each one a reminder that someone struggled to produce it.

These aren’t nostalgic retreats. They’re responses to a real shift.

What is old becomes new

As AI takes over more of what we used to call work, I think people will feel less human. Not because AI is malicious, but because so much of our identity is tied to striving, to the long process of getting better at something. When output becomes effortless, something essential is lost.

People will respond by picking up pursuits that are always about refinement. The never-ending journey toward a perfection you never reach. Arts. Cooking. Sports. Gardening. Building things with your hands. These activities cannot be shortcut. The struggle is the point.

What people will be developing isn’t just skill. It’s the capacity to combine skills with emotion and judgment, in real time, under pressure. The painter who understands light and chemistry and grief. The actor who brings their whole life to the stage each night. The soccer player whose body synthesizes a decade of training in a single improvised moment.

AI can simulate competence in any single domain. What it cannot do is integrate. The emotion in a painting. The presence in a performance. The decision on the pitch made in a split second with the crowd roaring. None of it perfect. All of it human.

These pursuits will find renewed attention. Not because they're anti-technology, but because they offer something technology cannot provide.

The question is how we teach this. Part of it is studying history, what has lasted and why. Part of it is disconnecting from technology periodically, creating space to feel the difference. And part of it is using AI itself to accelerate skill acquisition, compressing the technical learning curve so more time can be spent on what it cannot teach. Emotion, judgment, presence.

If AI handles output, humans are freed to optimize for what was always the point. Not the result. The pursuit.

So take a break from doomscrolling. Pick up a hobby that resists shortcuts. Make something with your hands. Visit a museum. See a play. Find the mom and pop restaurant that’s been serving the same neighborhood for decades. These things are still here. They’ve been waiting. They’ve stood the test of time for a reason. Most of what’s new today will be gone tomorrow.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

The biggest shifts ahead aren’t about more AI, more crypto, or more software. They’re about where intelligence moves, how value flows, and what becomes scarce again.

2025 was the year of abundance. AI gave everyone the ability to create more than ever before. More content. More software. More ideas. Much of it generic. Output exploded faster than judgment.

2026 is the correction. Scarcity returns, not because we create less, but because we choose more carefully. Focus, refinement, and intention replace raw throughput. What matters is no longer how much you can produce, but what survives selection pressure.

Physical AI

The biggest constraint on AI progress is no longer models or compute. It’s physical data.

The internet has been fully consumed. What’s missing are high-quality, real-world interactions: motion, friction, failure, edge cases, and human behavior in physical environments. Companies that can incentivize the capture, labeling, and feedback of real-world data through robotics, sensors, IoT, and human-in-the-loop networks are building the most defensible datasets that exist.

These datasets won’t be scraped. They’ll be earned.

The robotaxi war will heat up as Waymo and Tesla’s Robotaxi receive approval in more cities worldwide. If you’ve ridden in one in California or Texas, you know how life-changing it feels. Every mile driven compounds a data advantage that is nearly impossible to replicate, pushing autonomy closer to a default mode of transportation rather than a novelty.

Beyond vehicles, robotics marks the maturation of agentic systems. Specialized physical AI startups signal a shift from purely digital agents to systems that act in the real world with accountability. Factories, warehouses, and logistics networks become training grounds where agents are forced to deal with physics, constraints, and consequences. What starts in controlled environments steadily moves into everyday life.

At the same time, XR glasses finally get their day. They are no longer awkward or experimental. AI embedded directly into glasses turns the physical world into an interface: contextual memory, real-time assistance, navigation, translation, and perception layered onto daily life. You're already seeing Meta's glasses more frequently in the wild. Google is expected to launch its own in 2026 through partnerships with Gentle Monster and Warby Parker.

This shift extends to the device layer. Expect Apple to make meaningful moves as AI becomes deeply embedded within the iPhone itself. On-device inference preserves privacy, reduces latency, and makes intelligence feel ambient rather than transactional.

This is how autonomy scales across both machines and humans. Vehicles that improve from every mile driven. Robots that learn from every task performed. Wearables that adapt through lived experience. The winners unlock trillion-dollar markets because they own the feedback loop between intelligence and reality.

Vertical AI

General-purpose models are impressive yet rapidly commoditizing.

As models converge in capability, differentiation shifts toward data, integration, and execution. The real winners are not the models that can answer everything, but the systems that can do one thing extremely well inside a real workflow.

This is where vertical AI separates itself. Instead of acting as a copilot, vertical AI replaces the entire process. It lives inside the workflow, ingests proprietary data, and makes decisions, not suggestions. It is trained on narrow, high-signal datasets that cannot be scraped or easily replicated.

Legal, healthcare, insurance, supply chain, underwriting. Entire industries where accuracy, compliance, and accountability matter more than novelty. Horizontal tools struggle here because good enough is not acceptable. Vertical AI outperforms by design because it is embedded where work actually happens and fed by proprietary data that compounds over time.

In 2026, incumbents will not be displaced by better models alone. They will be displaced by workflows that simply run themselves.

Legal is the proof point. Companies like Harvey and Ironclad already handle contract analysis, due diligence, and redlining at scale. In 2026, the shift goes further. Vertical AI moves from augmenting lawyers to replacing entire workflows end-to-end. The first firms will operate with a fraction of the headcount on routine work.

But this isn’t a story about job destruction. It rarely is. When tasks become cheaper and faster, demand for them grows. Lawyers freed from document review can handle more matters, advise more clients, and take on work that was previously uneconomical. The resistance to AI-powered legal tools won’t come from lawyers losing jobs. It will come from lawyers whose judgment and client relationships become the only defensible value they have left.

Insurance underwriting and claims processing follow the same arc. The pattern repeats anywhere work is high-volume, document-heavy, and expensive to get wrong.

AI-Native Infrastructure

AI demands new infrastructure. Real-time inference, multimodal data, distributed training, agent orchestration, and edge deployment are no longer edge cases. They are the default.

Today’s infrastructure was built for storage and batch compute, not for continuous reasoning, coordination, and action. This gap is most visible in embodied agents. While models are capable, systems remain fragmented. Workflows do not talk to each other cleanly. State is hard to persist. Latency, cost, and reliability break autonomy before it compounds.

Data lives in too many places, and retrieval is too slow and imprecise to support continuous reasoning, even as context windows grow. Without fast, accurate access to the right information at the right moment, intelligence stalls before it can act.

Agents are not failing because the models are weak. They are failing because the infrastructure beneath them is not designed for long-running, interactive, multi-system execution.

The structural problem is misaligned incentives. Model companies optimize for capability benchmarks. Infrastructure companies optimize for traditional workloads. No one owns the full stack from reasoning to action. Until compute, data movement, identity, and permissions are native to the agent layer, autonomy remains mostly aspirational.

As intelligence moves into browsers, factories, devices, and physical environments, infrastructure must become faster, lighter, and closer to the edge. Inference must run where decisions are made. Relevant data must be retrievable in real time, not stitched together after the fact.

The winners in this layer define how intelligence is created, deployed, and shared globally.

In 2026, expect the first durable agent platforms to emerge. Not demos, but systems that persist state, coordinate across tools, and run unsupervised for hours or days. The gap between “impressive prototype” and “reliable deployment” finally starts to close. The companies that solve orchestration, memory, and cross-system identity become the next infrastructure layer.

Deep Tech

Energy and materials are becoming scarce again, and scarcity forces innovation.

Global electrification, AI-driven compute demand, and reindustrialization are colliding with fragile energy systems. Simultaneously, access to critical and precious metals is increasingly constrained by geopolitics, supply chain concentration, and environmental limits. These pressures are structural, not cyclical.

This is why fusion and advanced nuclear move from science projects to strategic imperatives. Reliable, abundant energy is no longer just about cost or climate. It is about sovereignty and economic stability. Breakthroughs in fusion, next-generation fission, and grid-scale energy storage become necessary to support AI, manufacturing, and modern economies at scale.

Countries that cannot secure energy and materials lose leverage across manufacturing, defense, and technology. Deep tech becomes infrastructure for the next century. The inputs to everything else.

Stablecoin Utility

Stablecoin supply crossed $300B in 2025. The era of accumulation is ending. The era of utility is beginning.

Regulatory clarity changed everything. The Clarity for Payment Stablecoins Act gave banks a formal path to issue. JPMorgan is scaling JPM Coin. PayPal is expanding PYUSD. Visa is settling in USDC. As licensing standardizes, issuance commoditizes. The competitive advantage shifts from technology to distribution.

The next chapter is usage. Payments, remittances, FX, tokenized assets, yield, and lending. These are behaviors people already perform every day. Stablecoins make them faster, cheaper, and global by default.

This shift accelerates with AI. Standards like x402 allow autonomous agents to make real-time payments for APIs, compute, and services. As software pays software, a machine-to-machine economy emerges that legacy rails cannot support.

Lending is the next major unlock. The first generation of onchain lending required heavy overcollateralization and served only crypto-native traders. What changes now is identity. Advances in onchain identity and zero-knowledge proofs allow users to verify creditworthiness, income, and compliance without exposing sensitive data. This creates a path toward undercollateralized lending where risk is priced on verified behavior rather than blunt collateral.

Yield follows. As the US continues to lower interest rates and inflation persists, people will seek the best possible returns with the least fees. Stablecoin-based accounts that offer transparent, direct access to yield become the next generation of high-yield savings. For consumers, this is where adoption starts.

Incumbents see this coming. Western Union is experimenting with stablecoin rails. Banks are piloting tokenized deposits. Adaptation is table stakes. But even the ones who move fast lose margin. Interchange, FX spreads, prepaid float, and correspondent banking margins don’t survive programmable money.

In 2026, the question isn’t whether stablecoins replace legacy rails. It’s who owns the distribution when they do.

Privacy and Verification

As AI absorbs global knowledge and stablecoins power daily transactions, privacy becomes non-negotiable.

The next wave of infrastructure enables selective disclosure: proving what matters without revealing everything. Zero-knowledge proofs, secure enclaves, FHE, and MPC are no longer academic. They’re foundational.

Privacy becomes an operational necessity now because the cost of exposure has crossed an existential threshold. AI systems can correlate data across sources at scale. Breaches are no longer embarrassing but existential. Regulation has caught up. And the long-term threat of quantum computing makes today’s encryption a liability.

When anyone can generate anything, trust becomes the bottleneck. Verification infrastructure ensures that people, content, payments, and actions are real. Media provenance, AI-resistant identity, cryptographic attestations, and reputation systems become essential for both humans and agents.

There’s also a new economic layer emerging. People and devices can prove they contributed data, labeling, or expertise without revealing private information and get paid for it. Authentic contribution becomes a monetizable primitive.

In an infinite content world, authenticity is the new commodity.

Healthtech

As AI increases productivity and automation removes friction from daily life, people get something back that has been scarce for decades: time.

That time doesn’t disappear. It gets reinvested. And increasingly, it gets reinvested into extending life itself.

Healthtech in 2026 is not about wellness trends. It is about buying time with time. Wearable devices and continuous biomarker monitoring move from early adopters to defaults. Companies like Function and Superpower make bloodwork actionable and routine rather than annual and reactive. Peptides, precision supplementation, and AI-driven diagnostics enable care tailored to the individual rather than the average.

The feedback loop tightens. Health becomes measurable, adaptive, and proactive.

Reactive healthcare cannot support aging populations, longer working lives, and rising chronic disease. Longevity-focused healthtech shifts care upstream and reframes health as a compounding asset rather than a recurring emergency.

In 2026, investing in health is no longer aspirational. It is a rational strategy for preserving the most valuable resource people have left.

The Content Flood

Multimodal AI continues to improve, and entertainment is where the boundaries blur first.

Tools like Sora show how quickly video realism is advancing. Suno did the same for audio. When visuals, voice, and emotion converge, creative output stops feeling synthetic.

This is already happening in music. Walk My Walk by Breaking Rust, a fully AI-generated artist, charted on digital country lists. Listeners didn’t distinguish between human and machine. In Japan, AI music stars are already mainstream. From experience, I've seen entertainment trends start there before moving West. In 2026, an AI-generated track will break into the Billboard Hot 100.

The implications extend beyond music. When anyone can make a hit, a film, or a feed’s worth of content, the bottleneck shifts from creation to selection. What do you listen to? Who do you trust? What even matters?

That question is the bridge to what comes next.

Curation

Timelines are flooded with AI-generated slop and hollow content. Influencer feeds increasingly depict lives optimized for engagement rather than reality, creating a constant sense of comparison without substance. People feel the disconnect.

As content becomes infinite and cheap, attention shifts away from scale and toward truth. The human response to abundance is not disengagement. It is curation. People want fewer things that feel real, grounded, and personally meaningful, not more things engineered for dopamine and aspiration theater.

This is where niches grow, not shrink. Communities form around specific tastes, ideas, and identities. What you truly like becomes signal.

We saw early proof of this in 2025. Long-form podcasts grew. Independent writing platforms like Substack expanded. People opted into slower, deeper content because it felt intentional rather than algorithmic. Meaning beat virality.

In 2026, this accelerates. AI does not replace curation. It amplifies it. Personal agents act as curators, not feeds. They ingest podcasts, newsletters, books, and conversations you actually care about, then help you refine what resonates and what does not. Over time, taste compounds.

Importantly, this kind of curation gives people time back. By spending less energy sifting through noise, people can be more intentional online and more present offline. Online communities become places to learn and deepen interests, not destinations to stay trapped inside.

That reclaimed attention flows back into real life.

The future isn’t less content. It’s content that helps you live more fully outside the feed.

Hobbies, IRL

As AI automates work and decision making, people are forced to confront a new question: what do you do with the time you get back?

In 2025, social clubs made a visible comeback. Many will follow the same traps as the past: broad positioning, status signaling, a lack of clear purpose beyond access. Short-lived hype followed by quiet churn.

We have seen this movie before. Clubs like Zero Bond or Soho House initially succeed by tapping into cultural momentum, but struggle to sustain long-term engagement when identity replaces action.

The communities that endure are built around specificity and shared behavior. Action is the filter. Othership works because it is not just social. It is for people who care about longevity, recovery, and showing up consistently for a physical practice. The activity creates commitment. The niche creates alignment.

This pattern repeats across categories. Board games. Sports leagues. Cooking classes. Pottery studios. Run clubs. Strength training. Communities built around doing something together create real bonds because participation requires effort, not just presence.

This is not escapism. It is maintenance. Online curation helps people find what they care about. Offline communities give them a place to live it. People who believe the future will be longer invest in staying present, capable, and connected inside it.

Vinod Khosla has said that technology ultimately gives humans back time. In 2026, we see people use that time not for status or spectacle, but to build durable relationships and embodied skills anchored in shared action.

Physical data. Trust. Authenticity. Health. Energy. Taste. Time.

These are the new scarcities. Not because they disappeared, but because everything else became abundant.

The coming year belongs to whoever figures out how to create, verify, and compound them.

The opportunity is the same whether you’re building infrastructure or building a life: invest in what compounds. Your health. Your taste. Your relationships. The skills and communities that don’t depreciate when the next model drops.

In an age of infinite output, the scarcest thing is knowing what matters to you.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe

It’s rare for a movie marketing campaign to feel genuinely fun rather than obligatory, but Marty Supreme’s rollout immediately felt different.

Instead of following a familiar promotional script, Timothée Chalamet’s approach felt intentional, self-aware, and tuned to how people actually consume culture today.

One of the earliest signals of that came from a viral marketing video framed as a Zoom call with the film’s team. What looked like a routine meeting quickly turned into a self-aware skit that poked fun at the absurdity of movie marketing while still advancing the campaign itself. In the clip, Chalamet pitched increasingly wild ideas, like painting everything in the film’s signature ping pong ball orange, including iconic landmarks such as the Statue of Liberty, and flying bright orange blimps over major cities. The ideas were intentionally over the top, but visually memorable.

One of those jokes even came to life. The orange blimp was later spotted in the real world, blurring the line between satire and execution. Instead of announcing the movie, the Zoom call invited people into the joke and turned internet humor into real-world spectacle.

That worked because Chalamet treats promotion as part of the performance, not something separate from it.

More broadly, it revealed a larger shift. The old way of launching a movie, built around late-night interviews, magazine profiles, and tightly controlled press tours, no longer creates the same level of connection. Those formats still exist, but attention has fragmented. Audiences now want to feel involved, not just marketed to.

That expectation shaped the rest of this rollout. Rather than relying on a traditional press cycle, the campaign leaned into culture-native moments that showed a clear understanding of where attention actually lives.

Collaborating With People Who Shape Culture Today

This rollout resonated because Chalamet didn't try to manufacture relevance on his own. He collaborated directly with people who already sit at the center of pop culture and internet attention today.

One of the smartest plays was the recent drop with EsDeeKid, a fast-rising British rapper who many online half-jokingly speculated was secretly Timothée Chalamet himself.

Instead of ignoring the conspiracy, Chalamet leaned into it by putting it to rest in the most internet-native way possible. He collaborated directly with the real EsDeeKid, turning the speculation into a joint rap that explicitly referenced Marty Supreme. The move acknowledged the joke, closed the loop, and redirected the attention back to the film. It did not feel like a calculated brand play. It felt like a wink to the audience and a clear understanding of how meme culture actually works.

Timothée Chalamet

@RealChalamet

687.4K

4:02 PM • Dec 19, 2025

It did more than resolve the joke. It traveled. The rap crossed more than 100 million streams within a day of release, magnifying the moment and showing how quickly culturally embedded distribution can spread when it lands.

That same instinct for meeting audiences where they already are showed up in his collaboration with Druski, one of the most influential creators right now.

This was not a celebrity cameo awkwardly inserted into a movie promo. It was creator-native content that felt like it belonged in Druski’s world, not Hollywood’s. None of these moments felt cynical or overly ironic. They worked because they struck the right balance between self-awareness and sincerity.

In both cases, the key was relevance. These were not safe or generic partnerships. They were collaborations with people who actively shape culture, which is why the moments felt authentic rather than manufactured.

Audiences can tell immediately when someone is being used as a distribution channel versus when they are being treated as a creative partner. These clearly fell into the latter.

Making the Film Physical Through IRL Activations

The physical world played a real role in this rollout too. The Marty Supreme merch pop-ups turned the film into something people could literally wear and signal, with activations spanning New York, Los Angeles, London, and even Brazil.

The jackets themselves were intentionally produced in low quantities and made to a high-end standard through a collaboration with Nahmias, which helped position the merch as covetable fashion rather than promotional merchandise. As a result, Marty Supreme jackets started showing up on celebrities, creating organic social proof without needing explanation. In some cities, fans even camped out overnight to get their hands on one, reinforcing that this was demand driven, not manufactured.

Chalamet even showed up to the pop-ups himself, flanked by an entourage of ping pong ball mascots, blurring the line between character, campaign, and spectacle.

That approach showed up beyond merch as well. A Marty Supreme-themed Airbnb experience centered around table tennis pulled the film’s identity into a physical space people could actually step into and participate in. It was playful, unexpected, and highly shareable.

Taken together, the pop-ups, merch, mascots, and Airbnb activation all did the same thing. They turned attention into presence. The film was not just something to watch later. It became something people could encounter, wear, and talk about in real life.

What This Means for Modern Launches

What ties all of this together is not celebrity access. It is intention.

This approach is harder than it looks. It requires taste, timing, and restraint. Without those, it quickly turns into noise.

Every part of this rollout was designed to place Marty Supreme in moments where people were already paying attention, without forcing it into the frame. The product showed up naturally. A sweater worn by a celebrity. A name dropped in a rap. Chalamet wearing Marty Supreme merch inside a top streamer’s video. None of it felt like placement. It felt like presence.

That matters more than ever in an environment where attention is scarce and content is endless. People are not just overloaded, they are defensive. Anything that feels overtly promotional gets filtered out immediately.

Chalamet himself has acknowledged this shift. He recently said, “People’s attention spans are so short these days. How do you convince them to go to the cinema, to spend money to see a film, rather than waiting to stream it illegally, or for it to be available on Netflix? I have an audience, so I engage with them, and I give it 150 percent.”

That mindset shows up clearly in this launch. Marketing was not treated as amplification after the work was done. It was treated as part of the performance. Part of the story. Something that had to earn attention the same way content does.

There is also a broader media shift underneath all of this. Authenticity matters more than polish. Relatability matters more than reach. Institutions are adapting too, from awards shows experimenting with YouTube distribution to creators becoming primary cultural gateways. The way people consume media has changed, and launches that ignore that reality struggle to break through.

The takeaway is simple but demanding. Be intentional about where your product appears. Place it where attention already exists. Make it feel native to the environments people care about. And above all, treat marketing as something audiences participate in, not something done to them.

At its core, this rollout shows that launches work best when people do not just see the product, but feel connected to it. Early presales suggest that connection is already translating into real demand. When the film has its full release on Christmas Day, I would bet this approach translates into real financial success at the box office.

They offer something the digital world has always lacked: money designed for a global, internet-native economy.

With more than $300 billion in circulation, stablecoins now surpass nearly every other category in the industry. While USDT and USDC still dominate supply, the market is entering a new phase. For the first time, regulated financial institutions are participating at scale.

The era of accumulation is ending.
The era of utility is beginning.

The Catalyst: Regulatory Clarity

The Clarity for Payment Stablecoins Act marks a turning point.

It is the first U.S. federal framework to clearly define how payment stablecoins can be issued and regulated. The act establishes two formal paths:

• State-regulated trust companies can issue stablecoins

• National and state banks can issue stablecoins with Federal Reserve approval

This single change unlocked the market for traditional finance. The effects are already visible:

• JPMorgan is scaling JPM Coin to billions in daily corporate settlements

• PayPal launched PYUSD and is preparing to expand under the new framework

• BNY Mellon is supporting major issuers through tokenized cash and custody

• Visa is exploring bank-issued stablecoins across its settlement network

• Regional banks and neobanks are piloting compliant tokenized deposits

This is only the beginning.

As licensing becomes standardized, issuance will commoditize. The competitive advantage will shift away from technology and toward distribution. The winners will be the companies that already own the customer relationship and embed stablecoins into everyday financial behavior.

From Issuance to Utility

The first age of stablecoins focused on minting supply and onboarding users. The next chapter is about usage.

Stablecoins have reached escape velocity. The question is no longer how to create them, but how to make them indispensable.

We are moving from stablecoins being held to being used across payments, remittances, trading, foreign exchange, lending, and yield. These are behaviors people already perform every day. Stablecoins simply make them faster, cheaper, and global by default.

1. Payments

Stablecoins are moving from pilots to core payment infrastructure.

• Visa is settling transactions in USDC on Solana and Ethereum

• Stripe enables stablecoin payouts and is building Tempo, a payments-focused blockchain

• Shopify is integrating stablecoins at checkout

Traditional payment rails operate at banking speed. Stablecoins operate at internet speed. Settlement becomes instant, global, and always available.

Most users may never know blockchain is involved. They will simply experience payments that work better.

This shift accelerates with AI. Standards like x402 allow autonomous agents to make real-time payments for APIs, compute, data, and services. As AI moves from general-purpose tools to specialized, agentic systems, software will increasingly pay software. This enables a machine-to-machine economy that legacy rails cannot support.

2. Remittances

Global remittances exceed $860 billion annually, yet remain one of the least efficient financial markets.

The legacy model:
Multi-day transfers during business hours, fees between 6 and 9 percent, poor FX rates, and hidden costs.

The stablecoin model:
Transfers in seconds, 24/7 availability, fees measured in cents, and better FX through mobile apps.

Even Western Union is experimenting with stablecoin rails. In many emerging markets, stablecoins are no longer optional. They are essential for immediate payments, self-custody, and protection against unreliable banks and volatile local currencies.

3. Foreign Exchange

If payments are about speed, foreign exchange is about access.

The FX market moves $7.5 trillion per day, yet still operates with outdated constraints.

The legacy model:

• Limited to banking hours

• Settlement in two business days

• Meaningful counterparty risk

The stablecoin model:

• Markets open 24/7

• Atomic settlement

• Payment versus Payment settlement through smart contracts

To date, most stablecoins have been denominated in U.S. dollars. Over time, other currencies will move onchain as well. Institutions are already exploring blockchain-based FX settlement to eliminate multi-day delays.

Stablecoins turn FX into a real-time software primitive.

4. Tokenized Assets (RWAs)

Stablecoins are becoming the liquidity layer of the tokenized economy.

As trillions of dollars of treasuries, private credit, bonds, and eventually equities move onchain, they require a settlement asset that lives on the same rails. Instant asset settlement is impossible if the cash leg still depends on legacy wires.

Stablecoins enable Delivery versus Payment, where assets and cash exchange instantly and without settlement risk. They also make these assets tradeable globally, at any time, from anywhere.

This may become the single largest driver of stablecoin demand over the next decade.

5. Yield

Yield may be the wedge that drives mainstream consumer adoption.

Stablecoins remove layers of intermediaries, allowing yield to flow directly to users. These products resemble the next generation of high-yield savings and cash management accounts.

In a world of persistent inflation, transparent access to yield matters. Users will increasingly choose platforms where they understand how returns are generated and can control their risk exposure.

Over time, stablecoin-based accounts may become a default consumer financial product.

6. Lending and Credit

Lending is the next major utility layer for stablecoins.

The first generation of onchain lending proved demand but did not scale. Loans required heavy overcollateralization and primarily served crypto-native traders.

The next phase looks different.

As stablecoins become embedded in fintech apps, neobanks, and tokenized financial markets, they unlock a more familiar use case: credit.

What changes this time is identity.

Advances in onchain identity, compliance infrastructure, and zero-knowledge proofs allow users to verify attributes without exposing sensitive data. Creditworthiness, transaction history, income signals, and jurisdictional compliance can be proven selectively and privately.

This creates a path toward undercollateralized and partially collateralized lending, where risk is priced based on verified identity and behavior rather than blunt asset overcollateralization.

Early signals are already emerging:

• Fintechs exploring stablecoin-backed credit lines

• Onchain credit markets tied to verified entities rather than anonymous wallets

• ZK-based identity systems enabling compliance without sacrificing privacy

In this model, stablecoins do not just move money. They become the foundation for global credit markets that are faster, more transparent, and more accessible than traditional banking.

Why Distribution Matters Most

The future of stablecoins will not be decided by market cycles or exchange volume. It will be decided by distribution.

Key channels include:

• Wallets like Coinbase Wallet and Phantom

• Fintech apps like Robinhood, Revolut, Cash App, and SoFi

• Digital-first banks and neobanks

• Consumer platforms across social, marketplaces, and creator tools

The strongest technology will not win.
The biggest brand will not win.

The winner will be whoever gets stablecoins into the most hands and makes them useful.

Why This Moment Matters

Stablecoins are not another crypto trend. They represent a rethinking of money itself.

Money should move like the internet. It should be instant, global, and accessible to anyone with a phone. It should not punish people for geography, banking access, or outdated infrastructure.

For years, financial inclusion was theoretical. Stablecoins are making it practical:

• A small business in Brazil gets paid by a European client in seconds

• A family in the Philippines keeps income previously lost to fees

• A teenager in Turkey preserves value amid currency instability

This is what happens when money becomes software.

Soon, people will not ask what chain they are using. They will simply use money that feels fast, fair, and universal. When that happens, stablecoins will no longer be viewed as a crypto innovation.

They will simply be how money works.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe