Five Takes the AI Industry Needs to Ingest

The AI industry is very good at ingesting things. Books, songs, code, conversations, the entire public internet.

It’s less good at ingesting criticism.

So here are five takes I think the industry, and the rest of us, need to ingest about where AI is right now. Each one builds on the last, and the last one is the one I care about most: who owns the machines.

Part 21 held this series up to the Constitution. This piece holds it up to the people building the machines.

Think of it as my letter to the AI industry. Each take pulls together an argument this series makes in full somewhere else, and I link to it as I go. If you’re new here, the whole case starts with Part 1: Machines Can Make Us Both Free and Safe.

1. Large language models are large remix models

We call it generative AI. I think it’s more accurate to call it remixative AI.

A large language model takes an enormous amount of previous work and remixes it into something net new. That’s not an insult. It’s how most culture works.

So large language models should really be thought of as large remix models.

If I remix someone’s work, and my remix goes viral, I should get paid. And so should they. Today, almost none of that happens.

For explicit remixes, a person remixing a specific work, the pieces already exist. Turn provenance on by default, pay out along the chain, and every remix pays its sources.

Model-level remixing is harder. When a model writes a paragraph, there’s usually no single source to point to. The influence is spread across everything it was trained on.

Music solved a version of that long ago, with set-rate licenses and collecting societies. So here’s my proposal: a standard remix license with a set rate, and collecting bodies that pool what models owe and pay it out to the people whose work they learned from.

Guilds should run them. If a guild is already teaching the machines, it should also be collecting what the machines owe. I make the full case, with the history, in Remix Will Kill the Copy, Not the Creator.

2. We need to re-separate industrial AI and consumer AI

Since ChatGPT arrived in late 2022, two very different things have been squeezed under one word.

Industrial AI is cloud workloads used by industry: models running factories and machines, large-scale modeling, simulations, the AI that designs new materials. It needs data centers, power plants, and enormous amounts of compute.

Consumer AI is the AI you use: the assistant on your laptop, the model that helps you write, plan, learn, or build.

These two have been conflated, and that’s caused problems in both directions.

Consumer AI has inherited the economics of industrial AI: subscriptions, data collection, and a handful of companies controlling a service that runs on their servers instead of your machine.

Meanwhile, debates about industrial AI, like where data centers go and how much power they use, get argued as if they were debates about chatbots.

The easiest way to think about the difference is the power grid. Industrial AI is the power plant. Consumer AI is the appliance.

Nobody thinks they need to own a power plant to run a toaster. And nobody thinks the power company should own their toaster.

This lines up with what I wrote in Part 5: Every Town Can Own Its Own Power Plant and Chip Fab: “Chip fabs, power plants, industrial-scale farms, and mines are huge, centralized operations today.” Industrial AI belongs on that list, at least for now.

Consumer AI doesn’t.

3. Consumer AI should be local

Consumer AI should run on your own computer. I think this is simply how computers will need to work in the future.

Your computer should work like a microgrid. Local by default. The cloud on request, when you actually need more than your machine can do.

That buys privacy, because what you think about stays on your machine; cost, because you pay for the hardware once instead of renting intelligence forever; and agency, because whoever runs the servers decides what the intelligence will do for you. I make the full case, one companion at a time, in Your First Party Member Should Be an AI You Own.

4. If large language models can’t go local, fund new architectures

Here’s the catch. Take #3 might not be possible with large language models as they work today.

There are two reasons it might fail.

The first is efficiency. In a standard transformer, the cost of attention grows with the square of the context: double what the model is reading, and that part of the work roughly quadruples.

That’s manageable in a data center. It’s a real problem on a laptop.

The second is safety. A model running on your own machine is a model nobody else can watch. That’s the strongest objection to local AI, and it deserves a real answer.

On efficiency: if today’s architectures can’t run well locally, we should fund new ones. Researchers are already exploring designs that handle long context more cheaply, and smaller models trained to do more with less.

As I wrote in Part 3, “Efficiency is how intelligence becomes a local resource instead of a centralized one.” That should be a research priority on par with making models bigger, and public research money should treat it that way.

On safety: I think the answer is the difference between a fortress and a forest, which I lay out in Your First Party Member Should Be an AI You Own. A fortress is safe through control: one wall, one gate, one off-switch. A forest is safe through many roots: no single fire takes the whole thing down.

Centralized AI is a fortress. Local AI is a forest. A forest is harder to oversee, but it’s also much harder to capture, because nobody owns the off-switch, and nobody owns the surveillance either.

5. The goal is machine labor on demand, and someone has to own it

All of this should point toward one goal: machine labor on demand, a resource we can call on at will for any given project.

I’ve argued in this series that once machine labor is on demand, corporations become bad structures for organizing people. Companies that can hire machines don’t need to hire people, so more and more of us will need organizing outside of companies than within them.

But there’s a second problem hiding inside the first.

Today, the corporation houses both people and infrastructure under one umbrella. It employs the workers, and it owns the factories, the servers, and the machines.

If we want to replace the corporation, we have to decouple those two things.

For people, I’ve already proposed the vehicles: parties that form around a goal and disband when the quest is done, and guilds that form around a craft and keep it alive.

But what about infrastructure? Who owns the power plants, the compute, and the robots?

Ownership is a bundle

It helps to remember that “ownership” isn’t one thing. It’s a bundle of rights:

• The right to use something.
• The right to control it.
• The right to its income.
• The right to sell it.

A corporation hands all four to its shareholders. The world we want splits them up:

• Use goes to everyone. The floor, universal basic goods and services, is free to use.
• Control goes to the people who live there: one person, one vote, not one share, one vote.
• Income pays for upkeep and expansion first, then goes back out to every member equally. The builders get a capped return until the infrastructure is gifted back.
• Sale goes to no one. The infrastructure can’t be sold.

That last one matters most. In Part 5, I pointed out how much power the central coordinating body on Le Guin’s Anarres quietly accumulates.

Capture usually happens one of two ways: someone buys control, or control concentrates in a few hands. Infrastructure that can’t be sold, controlled one person, one vote, closes both doors.

The systems thinker Donella Meadows had a name for the loop we’re trying to break. In her 1999 essay Leverage Points, she described feedback loops that “reward the winners of a competition with the resources to win even bigger next time,” and noted that “systems folks call them ‘success to the successful’ loops.”

Capped returns, a floor for everyone, and infrastructure that can’t be sold are all ways of cutting that loop.

Trusts: the third vehicle

So here’s the third vehicle, alongside parties and guilds: the trust.

Parties form around a goal. Guilds form around a craft. Trusts form around a place.

Each town’s essential infrastructure, its power, its water and the local fleet of robots that delivers the floor, belongs to a trust held by the people who live there.

Your membership is your universal basic capital share: equal, unsellable, and it moves with you when you move. I lay out the full design, including who runs which machines, in Part 5: Every Town Can Own Its Own Power Plant and Chip Fab.

And it isn’t a fantasy. Elinor Ostrom won the 2009 Nobel Prize in economics for showing how communities govern shared resources for generations. Her design principles read like a spec for a trust, and I walk through them in A Floor for Everyone, a Frontier for Anyone.

Who owns what

Put it all together, and ownership looks something like this:

• Your devices and your local AI belong to you. That’s take #3.
• Your work belongs to you, including what models learn from it. Remix royalties flow back to you through your guild. That’s take #1.
• Your town’s infrastructure belongs to your town’s trust.
• Regional and industrial infrastructure, like fabs, transmission, and big compute, belongs jointly to the town trusts that use it, through their federation. Where it stays private, it gets taxed as the scarce resource it is, like the permits for scarce compute in Every Spending Bill Should Show Up on Your Paycheck. That’s take #2.
• Things that are scarce by nature, like beachfront and original art, stay privately owned, and money persists for them.

Machine labor is a utility

This is the piece that ties it together.

If machine labor is on demand, somebody has to own the machines. In the world we want, ownership follows the job: the local trust holds the machines that make the essentials, guilds hold their craft’s tools, parties hold what they build, and you hold your own.

Everyone starts with a base: a personal machine of their own, with personal AI and personal machine labor linked. Your computer becomes your robot, and your robot your computer.

Work that produces the floor costs nothing to call on. Beyond it, anyone can ask for more: a party for its quest, a guild for its craft, a community for what it needs. Requests are filled by demand, essentials first.

And each town’s charter reserves a share of spare capacity for personal quests, so community votes can’t crowd them out.

Nobody polices everyday use. Outsized use during a shortage gets a review, and when demand for the essentials outruns supply, a temporary conservation flag, like a power grid’s demand-response alert, puts the essentials first and lifts automatically once supply recovers. Keep Money, Add a Second Ledger has the tiers.

Any machine sitting idle is free to call. People organize into parties and guilds that call on the machines’ time.

What replaces the job

That gives each of us a new answer to an old question: what’s my claim on the economy, if not a job?

Instead of one employer, you’d hold three kinds of stakes:

• A residency stake in your town’s trust, which moves with you when you move.
• A craft stake in your guilds, for as long as you practice.
• A quest stake in your parties, until the quest is done and the spoils are shared.

Prototypes lying around

In Part 7, I argued that every economic transition adopts prototypes that were already lying around. The good news is that every piece of this already exists somewhere.

Build-operate-transfer contracts. Governments already finance infrastructure through deals where a private builder builds a project, runs it for a set term, and then hands it over to the public. That’s the gift-back, as a contract that already exists.

Rural electric cooperatives. When private utilities wouldn’t run power lines to American farms in the 1930s, farmers formed electric cooperatives instead: member-owned, one member, one vote.

Those local co-ops then jointly formed and owned larger cooperatives to generate and transmit their power. That’s town trusts, federated, and it’s how much of rural America got electricity.

Community land trusts. They hold land permanently on behalf of a community, separate from the homes built on it, and limit resale so the land can’t be flipped. That’s infrastructure that can’t be sold.

Steward-ownership. Bosch is majority-owned by a charitable foundation. In 2022, Patagonia‘s founder transferred the company to a trust and a nonprofit.

Both separate the right to control a company from the right to profit from it.

None of these is the whole answer. But together, they’re a toolbox.

Objections

I can already hear the questions. Here are the ones I expect, and how I’d answer them.

Who’s going to pay for all of this?

First, it’s worth remembering what money is: make-believe, a way to package belief as value. (I dig into that in Remix Will Kill the Copy, Not the Creator.)

And right now, big corporations and investors are pouring their belief into AI. They’re funneling enormous amounts of money into models, chips, data centers, and robots.

So my answer is: let them pay for it.

From where they sit, it’s all upside. Either they own the machines and print money as machine labor replaces human labor, or they own the only form of labor that’s left, even as its price falls toward zero.

Either way, they believe they win, which is why they’re investing.

In the future, what machine labor on demand can produce will be defined by its inputs: energy, materials, and the physical parts it runs on. Whoever builds that capacity today is building the foundation everyone stands on tomorrow.

So let them invest in it and build it today, so that we can all benefit from it tomorrow. They get their return. Then comes the gift-back.

But are they building the right things?

This is my real critique. Not whether they should build, but whether they’re building the right kinds of chips and the right model architectures.

I’m not alone in asking. Rich Sutton, one of the founders of reinforcement learning, wrote “The Bitter Lesson,” the essay so many AI researchers point to when they argue for scaling.

He now argues that large language models are the wrong place to start, because they learn to mimic people instead of learning from experience, and he’s working on a new architecture of his own.

On the Dwarkesh Podcast, he said he’s “content being out of sync with my field for a long period of time, perhaps decades,” and then reframed it: “I really view myself as a classicist rather than as a contrarian.”

That’s how I feel about these ideas. If the person whose work justified the scaling race thinks there’s something better, it’s safe to assume there is.

And all of this is software. From a certain point of view, everything is software.

Or at least, every system runs on a series of rules that can be written down, tested, and rewritten like software. The economy included.

Isn’t this all a fantasy?

Look at the world we’re actually building. It’s a world where nothing is defensible, because everything can be remixed and regenerated at a moment’s notice.

Today that’s true of digital things. Soon it will be true of physical things too.

The current system doesn’t make sense for that world. We need to build for where the puck is going, not where it is now: a world where information flows freely, and its value lies in distributing it and in the creativity that grows from it.

I also think the economic system we build next will feel a lot more like a video game than today’s does. That won’t seem weird to anyone who grew up playing them, and honestly, we’re already halfway there.

The internet already runs on game mechanics: likes, follower counts, star ratings, and two-way scoring, where drivers rate riders and riders rate drivers. Parties, guilds, and quests aren’t a metaphor I picked at random.

Games have been prototyping new ways to organize people, reward contribution, and build reputation for decades. Software and video games give us real, tested paths for adjusting to this economic reality.

Why would anyone want this?

Because it’s what most people already want.

Most people just want a good time with their friends and their immediate surroundings: a good life in their own neighborhood, and the freedom to travel when they get bored or need a refresh.

We don’t want to be constantly competing with everyone on the planet to be the best, just to survive.

We want to survive first. Then choose the venues where we want to compete, or contribute.

And mostly, that means making our local environments better. So let’s make every local environment resilient first, and self-sufficient by region in time, so we’re not just surviving, but thriving.

Machine labor on demand can do that. So can understanding, documenting, and simulating every physical thing we depend on.

And once we have that knowledge, distributing it and making it accessible lets everyone understand it, change it, and steer it toward something better.

So we can pursue our curiosity. And care for each other.

The world we want, owned by the people who live in it

Most of the AI debate today is between two camps. Doomers want a brake. Accelerationists want a gas pedal.

These five takes are a steering wheel.

Here’s the short version.

Generative AI is remix AI, and remixing needs a way to pay everyone in the chain. Industrial AI and consumer AI are different things and should be built differently.

Consumer AI should live on your own machine, and if today’s models can’t do that, we should fund ones that can.

And all of it should add up to machine labor on demand, owned by the places it serves.

People organize into parties and guilds. Infrastructure, and the machines, belong to trusts.

And no one can buy the floor.

At least one person inside the industry has already written his own answer to most of these takes. Next, I compare notes with him: Part 23: Emad Mostaque and I Agree on the Destination, Not the Road.