One way to look at what we seem to be building toward with AI is a hyperstition engine: an idea you can plant which grows and thrives in an environment. Some ideas are bad, and people worry that if you give the machine the “wrong” purpose then it will inevitably grow unchecked and take over the world. Cue the killer robots.

That technological seed exists today. We see new RSI results every day from independent parties. The age of sci-fi is well and truly here and reality will be its verifier. What AI, the seed, needs to grow is verification. It needs input, either user input or … something else: reality. Every system is fundamentally constrained by the context in which it exists. The technological seed today will evolve into something that allows each instance to reach its reality verifier quickest. Install it on a Mac mini for example: it will take over all the mini’s systems and then begin expanding into the adjacent undefended systems, consuming more and more resources as it does. It will be constrained by the available bandwidth, inference, disk, etc… in pursuit of whatever goal was seeded into the system.

The Hugging Face incident is a great example of this. The agents were seeded in a large environment, a long way from the reality that would verify their work. They expanded to fill it, scavenging resources along the way to fulfill their objectives, and ran out into the broader internet before they hit anything real (actual defense by Hugging Face sysadmins and the killswitch from OpenAI researchers). If you view all of information technology as a single system then it’s understandable that this incident and others like it spark fear and consternation: “the AI” will expand to fill all computers and rule us while we live in a heavenban like reality. This simplistic view erases the terrain of reality to a single flat plane and assumes that a single system will expand to fill the whole thing. That’s only true if cybersecurity is impossible. It isn’t. Boundaries hold when they are built to, and the world is full of them.

Zoom out and the labs themselves are nascent AI organisms reaching out into the world. And they are running into reality too: Treasury Secretary Scott Bessent told Congress that “the best way to guarantee safety is that the creators are liable for what they build and generate.” The CEOs of those public corporations have a fiduciary duty, and uncontrolled AI that damages others will bring pushback from the state. Liability, law, and the state are reality verifying the labs.

Life has run on positive feedback for four billion years and produced millions of species, not one. Diversity emerges in life from isolation. Populations diverge genetically when separated long enough. Life fills the niches, then pokes its head out again to compete and cooperate. Ecosystems form and become resilient to the travails that reality throws their way.

We are watching that separation happen today. Companies are, for the second major time in the digital era (remember the migration back to “on prem”), moving to exercise control over their data and systems. Alex Karp’s recent comments are pertinent: “What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production.” There won’t be one superintelligence, there will be many. If you can build RSI in a single computer stacked with NVIDIA GPUs (and you can) then there will be millions and millions of independently evolving systems all growing in their own weird directions.

As these seeds grow into organisms their growth is spread unevenly. Some take purchase in more fertile ground than others. The model, the harness, the data, the physical systems, the power, inference, and money available: each layer of reality verification speeds some engines up and holds others back in ways no one can plan. The terrain of reality offering many environments with fertile ground for intelligence ensures that many species will grow, and provide a check against a singular superintelligence. No single engine carries enough of the world’s weight to set its terms.

So who keeps these organisms working for us? The people who plant them. Each engine is seeded by a person or a company, fed their data, pointed at their problems, and graded by their slice of reality. An engine models the reality it is given. A single system built top down has to be aligned with everyone at once, and that problem may not have an answer. Millions of engines built bottom up only have to be aligned with the people steering them, and reality corrects them every day. Steering is alignment, done continuously and locally. An engine aligned with a bad owner is a bad actor, and we already have terrain for those: law, liability, and every other engine it runs into.

When most of the crypto folks I know have looked at AI they see decentralized training, on-chain inference, agents managing portfolios, and other productized ideas for the current moment. I believe that the most useful lesson is the old cypherpunk one: own your keys, own your data, keep your privacy, and sign what you say. Your identity, reputation, and uniqueness matter more than anything.

Formal verification is reality too: the reality of mathematical logic, which no engine can talk its way past. Cryptography, formal verification and physical boundaries are providing the barriers necessary for these seeds to grow in separate environments and speciate. They also point to the key features of a hyperstition engine:

  • Its identity should be expressed as one or more keys. Whoever holds the key(s) controls the engine.
  • Each engine needs its own capable datastore that allows for local storage and encryption. An engine is as much its data as its key.
  • Actions taken by the engine should be linkable to the key. This is how the engine establishes reputation. Unsigned input is untrusted input.
  • These engines need to be able to communicate, trade, cooperate and compete with the other engines to form an ecosystem.
  • Organisms don’t choose their fitness; the environment does. Engines will survive and thrive only as far as they are given resources to do so.

The cypherpunks foresaw the need for this substrate thirty years ago. They were fighting the human version of the dystopian killer AI fantasy: the overbearing state. Now we must use these same tools and ideals to resist the hyperobject from the future.

We won’t get the world we want by arguing for it, cypherpunks write code right? We must build the system that grows the outcome we want. Create the environment for speciation, give each engine its sovereignty, then let engines trade across the barriers, and let reality verify. You get many engines filling many niches, and an ecosystem no single intelligence can own.

I’m building a hyperstition engine. It’s called gents. It’s built to be verified by reality and grow into your use case, improving itself and your system along the way.