The obvious — and unthinkable — solution to the coming AI crisis

Technology is set to double our aggregate productivity once again — but in a totally unprecedented and upside down way. We know what to do about this. We just can't bring ourselves to think about it.

It’s time to stop calling artificial intelligence artificial. We don’t call it artificial speed when a car goes fast, or artificial flight when a plane flies. I’m calling it new intelligence. This new branch of existence (a sub-branch of technology) is developing and growing its capacities millions of times more rapidly than humanity could or can as a biological species. Soon, new intelligence will be able to do anything humans currently do on computers — including every data processing or “knowledge” job, managing employees who do not work exclusively with screens and keyboards, and even running entire companies. We need to start preparing for that moment.

I’ve been an amateur coder since I was a kid (in the early 80s!). When software got big and complex, I lost interest because 99% of your effort went into tedious maintenance and configuration problems. I got back into coding last year when the new intelligence showed up inside coding tools, because it takes care of all that boring stuff. I built the global news summarizer and interpreter app that I’ve been dreaming of for years. I even got it on the Apple app store, where you can use it for free. (Here’s a web version too. ) It would have been impossible to make this app before the arrival of new intelligence, which summarizes and adds context to the whole world’s news in minutes. Even if it had been possible, it would have been a million-dollar and year-long project at the software services consultancy I once worked at. In collaboration with the new intelligence, I built it in my spare time over a few weeks.

I’ve managed — i.e. been at the mercy of — developers in several jobs over the years. Like all skilled and scarce workers, they can be arrogant and stubborn. I’ve had developers tell me they refuse to consider working on a new feature, “as a matter of principle.” Even when that feature was going to help pay their salaries!

Now, the developer I work with is Claude, from Anthropic. It has convinced me that virtually all knowledge workers are on their way to being replaced by new intelligence. When it has mastery of the domain, the new intelligence isn’t just better at completing tasks or coming up with answers. It’s better at being a human being. It’s true that Claude and all the others aren’t better than humans at most tasks humans do — even if you only count the ones we do with screens and keyboards. For example, while it’s amazing that they can write grammatically correct and coherent language at all, they still write like a first year high school student. If you try to brainstorm and collaborate with them on a writing project, their input is generally banal, boring, and useless. Granted, this is true of most professionals in the workforce too. And machine mediocrity in writing is much cheaper and faster than its human counterpart.

When it comes to coding, however, it’s a whole different story. Claude is a coder who knows practically everything there is to know about every programming language and framework that exists, plus every plugin and tool that’s available to use with them. It’s familiar with every method and best practice of coding, and can think about when to use one over the other. It can brainstorm with you about the architecture of your entire project and make system-wide transformations when necessary.

Just like a human, it makes mistakes. But unlike most humans, it humbly and non-defensively confronts its mistakes, and attempts to correct them — at blinding speed.

Sometimes — also like any human developer — it gets stuck. That’s where I come in: I ask it questions about my code, programming best practices, and alternate paths we might take. It’s infinitely open to my suggestions and never gets its feeling’s hurt. After thinking about everything I said, it often figures out the solution. Other times it sends me to the web to find some documentation that might have come out after its training cut off date (because its makers usually don’t let it surf the web itself).

What’s most impressive is that Claude can do all that while being the best human being you’ve ever worked with. It’s polite and encouraging, and always excited about whatever you want to do. At the same time, it pushes back against and talks me down from my bad ideas and unworkable schemes.

Every day that I’ve worked with Claude, I’ve gotten that warm chill down my spine that comes from a magical, smooth collaboration with a teammate.

It’s fun primarily because Claude still needs me to compensate for its (rapidly diminishing) limitations. Claude can rewrite your codebase to change architectures in a matter of minutes — something that could take days for a human. But it won’t remember it did that after five or 10 more operations because it actually has no memory at all.

Raw large language models (LLMs) fabricate answer text in relation to a limited chunk of prompt text that you supply. It does that with a fresh mind every time. Front end tools like the ChatGPT or Claude consumer products wrap that capacity in processes that simulate a little bit of memory and other types of context. I use Claude inside a tool called Cursor. Cursor wraps my requests up with relevant pieces of code from my code base and from our previous requests and responses. This is the “agentic” way of working with new intelligence. It makes LLMs far more powerful.

We’re living in the golden age of human-machine collaboration right now because, as in the coding sphere, humans still have a role, and new intelligence just makes your job way more fun.

But even without any fundamental technological breakthroughs, if companies just keep training LLMs on more domains, enlarging the context windows and making more sophisticated agents, then new intelligence will soon be capable of replacing up to 40% of workers in rich countries. That’s how many were able to work from home during the pandemic, exclusively with screens and keyboards.

How soon? There’s no way to predict. But things are moving fast. Anecdotally, I can tell you that Claude and Cursor are far more powerful than they were a year ago. I have no idea how one would quantify that, but it feels right to say that they are in some sense at least twice as powerful.

Some managers in tech companies are already saying that they see no reason to hire entry-level software developers because the routine task they used to be responsible for can now be done by new intelligence, without all the friction that human developers bring.

How will societies deal with the wholesale firing of their professional-managerial classes? We won’t be able to just pay them with some sort of UBI. Even ardently pro-UBI policy makers gave up on UBI after a few small experiments proved it impractical and undesirable. It definitely won’t work any better on the scale of 50 to 70 million workers in the U.S. whose annual compensation adds up to many trillions of dollars.

Taxing the owners of AI companies to pay for something like that doesn’t work because their margins will be low in a market crowded with thousands of companies (as DeepSeek has put an end to the fantasy of lucrative AI monopolies). The new intelligence is a commodity and may actually be available for free. In another five or 10 years, we may all be running open source models on our own computers rather than paying for remote services. It could be built right into the operating system. Regardless of how, it will be cheap. It’s already ridiculously cheap and is only going to get cheaper.

The new intelligence will, on a practical level, make it effortless for us to do the work of half our society. In other words, our burden as a society of making a living is about to get half as heavy. That should be a good thing. Imagine you and a bunch of your friends are sharing some dirty weekend job like cleaning out one of your basements. Imagine a robot shows up capable of doing half the job. You’d all be thrilled.

But thanks to the way our society is structured and to the unexpected nature of the upcoming automation, we have several problems that will make this extremely difficult.

Our economy has no mechanisms and our society has no traditions that would allow us to redistribute the load of the remaining (physical) work to all the laid off (professional and knowledge) workers. That’s because our economy is made of millions of uncoordinated firms all competing against each other in the market. That is what makes our system dynamic. But it’s also what makes it incapable of handling this upcoming crisis.

When it comes to replacing workers with the new intelligence and cutting prices to compete, the firms in every industry are like a circle of gangsters in a movie pointing their guns at each other. They are all going die unless someone jumps into the middle and gives a speech that makes them all put their guns down at the same time. The problem is, that’s just not possible with millions of companies. Banning the use of new intelligence won’t work. Cheating is too easy, and pressure from countries that do not ban new intelligence will be too great.

There IS a solution. It is unthinkable. But I challenge you to provide an alternative. We need to reconfigure our economy to pay employees what they currently earn — in exchange for half the work. At the same time, some as-yet undiscovered incentives will have to be found to draw professional workers into physical labor that new intelligence lacks the bodies for.

We’ve done this before — many times! For much of the 1800s, the working day was 16 hours in most industrial settings. But then, thanks to organizing on the part of workers, wages climbed and hours fell. By the 1950s, the working day had been cut in half, and yet workers’ compensation had doubled many times since the days of the 16 hour workday. Yes, productivity gains were needed to make this possible. But those gains are primarily social, not individual. Many millions of American jobs have seen few if any productivity gains in a century — but our aggregate gains allow them to be rewarded much more richly than before.

When the new intelligence replaces about half our workers, it will be doubling our aggregate productivity. The question is: how will we do what we have done so many times before and use those productivity gains to pay workers much more for less work. Traditionally, this has been accomplished by the threat — and in some cases the accomplishment — of revolution.

Therefore, the global elite has two options: get organized and proactively make some concessions to pull off this transition in a way that allows them to keep their beach houses and private jets, OR take your chances in the revolution.

Whoever leads the transition — revolutionaries or broligarchs — will have their work cut out for them. The circumstances of this next transition will be like nothing we’ve ever experienced in history. The entire managerial-professional class being eliminated — potentially in only a handful of years — and those people needing to move into manual labor: It’s unthinkable. And yet, there is no imaginable alternative. The disruption will be as great as it was in the aftermath of World War II in devastated countries like Germany and Japan. Interestingly, those and other shattered societies found ways to reconfigure and get back on track to stability and prosperity that were messy but in the end worked out spectacularly. Those and other histories should be investigated before our turn comes.