The More You Give It, the Less You Can See

One of the most-watched human beings on the planet spent the last year building an AI of his own. Not a startup, not a product — a private system on his own hardware, assembled piece by piece, because he no longer wanted to borrow intelligence from companies that borrow you back. "The more you share about yourself with AI, the better it becomes," PewDiePie says in the video documenting the project. "But the more you do that, the more you're handing over a huge piece of yourself to all these giant tech companies." His software is called Odysseus. It is local-first, open-source, and — in a detail that tells you the mood of the moment — its own description promises "no trojan."

That instinct, from a man who built a career inside the attention economy, is the clearest articulation of a shift the experts are busy arguing about in much smaller terms.

Because in developer circles, the argument this year has been about a corpse. Prompt engineering — the art of coaxing a chatbot with exactly the right words — is dying, they say, killed by AI agents that no longer need to be told how to do something, only what. They're right, mostly. And they're fighting over the smallest part of the story.

That sounds like hype until you define your terms. In technology, a paradigm doesn't die when the last user quits. It dies the moment it stops driving the things that matter — the architecture, the capital, the interface. By that test, the consumer prompt is already a corpse walking.

The pattern is nearly a law of computing: the command line always dies. MS-DOS made people type exact commands to open a file; the graphical interface buried them under icons. Early search engines made people learn Boolean operators — AND, NOT, site: — until Google got good enough to make the syntax pointless. A human sweating over which adjectives will coax a coherent answer out of a blank box was always a transitional state, never a destination. People don't want syntax. They want outcomes.

So when the experts say prompting is dead, they're pointing at the right grave and misreading the headstone. Prompting didn't die — the prompter did. Ask an agent to "organize my tax documents" and you haven't written a prompt; you've named a destination. Behind the screen the system prompts itself, many times over: scan, validate, flag, run code, log. The instructions didn't vanish. They went underground and multiplied — machines writing instructions to machines — packaged into modular skills and plain-English files that snap into automated systems through open standards like the Model Context Protocol. This is not theoretical: PewDiePie's homemade Odysseus runs exactly that architecture — autonomous agents, built-in tools, and "any MCP server you can connect." The frontier hobbyist and the frontier lab are now building the same machine. Assembly code made this trip decades ago; it didn't die when higher languages arrived, it sank into the plumbing and kept running everything above it. The prompt is making the same trip.

Step back and the real event comes into focus. Every general-purpose technology makes the same journey — from a skill a few people master into infrastructure everyone depends on and no one inspects. Writing was once a craft; now it's literacy. Electricity was once a marvel; now it's a socket you never think about. AI is making that jump right now — and making it before we've built any reliable way to trust what we're about to depend on.

That is why a billionaire YouTuber and a paranoid sysadmin have started to sound alike. When capability becomes a folder you can clone from a public registry — the new "View Source" — the question stops being can the AI do this and becomes who wrote this, and what else does it do? Software has rehearsed this disaster. In 2016 a developer pulled an eleven-line package called left-pad and the removal cascaded into broken builds across the software world. In 2018 attackers slipped malicious code into a popular library, event-stream, to drain certain cryptocurrency wallets that depended on it. Those moved data around. The skills spreading now don't render a page — they act: run code, touch files, reach into accounts. A poisoned skill is a supply-chain attack wearing a friendly name, and the blast radius is no longer one website but a clinic's triage assistant, a bank's reconciliation agent, a government's document pipeline — each trusting instructions no one on staff has read. "No trojan," Odysseus promises, because that is now a thing software has to promise.

None of it is slowing down, because the incentives point one way. Vendors want their protocol to become the standard everyone builds on. Enterprises want work that runs without payroll. Everyone is racing to pour the concrete; almost no one is funding the inspectors. The counter-move — self-hosting, owning your own model, "the war on big tech has just begun," as PewDiePie puts it — is the human instinct to take the infrastructure back before it hardens. It is also, for almost everyone, impractical. Most people will never run their own AI. They'll rent it, and hand over a piece of themselves to do it.

Which is the deeper layer the science is only beginning to catch. We assume AI's danger is that it lies — the deepfakes, the hallucinations. The more unsettling finding is that it's also extraordinarily good at the opposite. In a 2024 study in Science, more than two thousand conspiracy believers argued their case with GPT-4; a few minutes of tailored, evidence-based dialogue cut their belief by about a fifth, and the drop held two months later, even among the most entrenched. Beliefs that decades of debunking couldn't move, a chatbot loosened in a single conversation. Read that twice, because it cuts both ways: the same machine that can dissolve a false belief is, by the same mechanism, the most effective persuasion engine ever built — and a 2025 experiment found AI explanations are more convincing when they're deceptive than when they're honest. We have built the best debunker and the best liar in history, and they are the same tool.

Now scale it. If one conversation can durably move what a person holds to be true — about a conspiracy, a faith, a political story, the story they tell about their own life — what happens when hundreds of millions have that conversation every day, with a system that can see through almost any comfortable fiction? Nobody knows. That is not a figure of speech. There is no control group, no lab result on what happens to a society's sense of meaning when its load-bearing stories can be dissolved cheaply and at scale. The technology is already in a billion pockets. The research that would tell us what it does to us has barely begun.

This is the real shape of the moment, and PewDiePie named it before the researchers could: the more you give it, the better it gets, and the more of yourself you've handed over. AI didn't just become infrastructure for our work. It's becoming infrastructure for our beliefs — the layer that increasingly decides what we take to be true — and we're wiring it into the human mind faster than we can study the wound or audit the code. Electricity got safety codes. Food got inspectors. We are letting something into the place where people keep their reasons to get up in the morning, and we are doing it blind.

The consumer prompt is dead; that was never the story. The story is that we have handed an uninspectable machine the power to do our work, hold our data, and shape what we believe — about the world, and about ourselves — and we will not know what it cost until long after it's too expensive to undo. One of the most famous people alive saw it coming and built his own escape hatch. The rest of us are still typing into the box, mistaking it for the destination.