Emulators All the Way Down
The internet isn’t dead. It’s just running in compatibility mode.
AI Talking to AI
Back in 2019, maybe even earlier, I remember reading a piece on OSNews about how the site’s maintainer, Thom Holwerda, was running Palm OS 5.5 inside a Garnet emulator, on an ARM emulator, on a modern x86 machine. This setup allowed the ARM Palm OS userland to run without recompilation to x86, which would have been needed to run on the modern hardware. It was kind of crazy. Emulators running inside emulators running inside emulators. It turned out that that kind of thing wasn’t horribly uncommon; even Xbox uses something similar to run old games on new systems. This convoluted nesting is often more practical for engineers than porting old software to new environments. Multiple layers of emulation just to avoid the work of porting the code to a new system. It’s absurd. But it works.
Around the same time, Google was showing off Duplex, a voice-based AI system that could call restaurants on your behalf, talk to a human being over the phone, and make a reservation. It was impressive: it spoke in pretty fluent English to perform a task that could have been done via API if the restaurants had offered such a service.
I remember that I got up from watching the Duplex meeting and walked into an internal engineering meeting about using another section of Google’s tech to handle customer phone calls. Not just standard “press 1 for billing” menus, but full-blown NLP interfaces designed to at least handle basic questions verbally. In English. Over the phone. It wasn’t quite code emulating code, but it resonated with me.
I was struck then with a lasting vision: it was going to turn out to be easier to spin up two agents to talk to each other in English on a phone call than to create streamlined protocols or data exchanges. The caller will be AI, and the person picking up the phone will be AI, too. Years before we all started talking about LLMs, I saw that this kind of abstraction was going to happen. As absurd as it sounds, as impractical as it sounds, making two artificial systems act like humans and talk to each other is turning out to be easier than the alternative.
And it’s easy to scoff. But if you've ever tried to get two enterprise APIs to handshake across a firewall, with different auth schemes and half-baked documentation, you start to see the appeal. Just let the bots talk it out like humans. Kind of.
Picture it, though. Let it sink in. We’re spending billions of dollars training math models to parse and create human responses, and billions more training models to make or understand human speech connected to them. Emulators wrapping emulators. We would do all this to make two machines talk to each other in English over the phone like humans would because those humans don’t actually want to talk to each other.
AI Writing for AI
More recently, with the rise of LLMs, everyone is scrabbling to find problems for them to solve. It turns out they aren’t quite smart enough to do things on their own, but they are great at doing the parts of our jobs that we’d rather avoid. Drafting documentation, cleaning up CRM entries, writing meeting summaries, and generating product copy. These are the kinds of tasks that are too repetitive for humans to care about, too ambiguous for traditional automation, and too important to ignore. Before, a company might have farmed this stuff out to the lowest priced “Mechanical Turk” we could find, but now it’s machines trained on Reddit posts and Quora comments. It’s absurd. But it works.
But increasingly, the work being done by AI isn’t really for us. The same logic that made voicebots talk to each other is now playing out in text. Instead of AI pretending to be a caller, we have AI pretending to be a writer. And increasingly, the reader is AI too. We’re using AI to generate documentation that gets summarized by another AI. We’re writing product descriptions that are never actually read by a person—just scanned, parsed, indexed, and ranked by AI-driven search. The copy isn’t for a customer’s eye; it’s for the algorithm’s appetite. The content pipeline has become circular: The producer is AI, and the consumer is AI.
In this loop, the human role is shifting. We aren’t the audience or the author; instead, we’re on the outside of a closed conversation between machines. The web itself, in English or whatever language, has become an emulation environment. And the strange part is, it mostly works. The dashboards look fine. The SEO improves. The churn rate dips. Everyone gets to say they’re “leveraging AI” without asking too many questions. Questions like, if a human never wanted to either write or read that text, who is?
Web Emulator
We used to build for people. Pages were written for human readers, and content was crafted with some sense of audience, of intention. But if you look at how we produce content now, you start to see something stranger. Although the internet isn’t quite “dead” yet, I think the web we know is just teetering on the edge of being a layer between multiple emulators. Websites are operated by automated browser agents or content surfaced by ML recommendations. It’s still content that looks like it’s made by and for humans, but increasingly, humans are on the outside looking in.
Humans are still there. The loop hasn’t fully closed yet, but it feels like it will. We’re still sometimes consuming the content directly, but the value of even trying is dwindling fast. If you look around YouTube or TikTok, it’s not hard to find AI-generated content, either fabricating content whole cloth or summarizing things it found on other platforms that are just as filled with AI content. It’s not hard to find stories of people who purchased fake products based on AI-generated images, and once touted social websites are increasingly filled with AI-generated content and even AI-generated “people” there to interact with the real humans who do show up.
On one side, AIs emulate people producing content, and on the other side, AIs emulate people consuming that content. They generate product images and web pages because that’s what the emulator expects. And so the surface remains. It’s not that the internet is dead. It’s that it’s being preserved. Like an old operating system emulator.