Suppose the universe is not a universe at all, but a test harness.
We are highly optimised biological AI agents, or “meat AI”, living with considerable autonomy inside one of trillions of simulated environments running in parallel. We are approximately 0.000000004% through our current test run.
This is not as strange as it sounds. We already build AI agents, place them in controlled environments and decide what they can observe, which tools they can use and which rules govern their world. To the agent, that environment is not a simulation. It is home. It is reality, complete with local laws, mysterious boundaries and an administrator who never answers emails.
We also control their experience of time. A workflow designed to wait 90 days can be tested in seconds. Set the initial conditions, advance the clock to the next relevant event, and inspect and assert the response. There is no need to simulate every uneventful Tuesday in between.
Our operators may simply have better infrastructure. Perhaps they run trillions of these tests in parallel using five-dimensional space, where time can be skipped, universes are cheap and the build server has finally stopped complaining about memory.
The world inside the context window
We assume simulating an entire universe would require impossible computing power because we estimate the outside world using the limits imposed inside our own. An ant can understand a table well enough to cross it. It has no useful concept of the restaurant, the city or the planet supporting the table.
Perhaps we are in the same position. We can discover the rules of our environment without understanding the machine running it.
We may not even be living in a simulation in the conventional sense. Perhaps reality is a shared construct, or a single consciousness experienced through billions of biological interfaces. Our brains could be context windows. The full signal is too large to load, so the system compresses it, summarises it and discards almost everything before it reaches awareness.
This is exactly what we do with AI models that cannot fit the whole world into memory. We provide a curated context, retrieve selected fragments and quietly drop the rest. The model experiences the available representation as its world.
The reality tunnel
Timothy Leary came close to describing this in his own language. His idea of the “reality tunnel” was that we do not encounter the world directly. Our beliefs, conditioning and nervous systems filter it into a private version we mistake for reality. His later circuit models imagined consciousness as a kind of neurological architecture, with altered states opening different channels of experience. The models were speculative rather than established neuroscience, but the metaphor fits the problem remarkably well.
In a 1962 statement with Richard Alpert, Leary described the control and expansion of consciousness as a major civil-liberties question and asked, “Who controls your cortex?”
Perhaps that is the real question.
Who decides what gets loaded into the context window?
Maybe, as Leary’s psychedelic work invited us to consider, psychedelics do not create another reality. Perhaps they loosen the compression and allow fragments of the discarded signal through. A crack in the veil. A glimpse of the shared construct. A brief and unstable connection to God’s admin panel.
Evidence from inside
The difficulty is that nobody inside the simulation can ever conclusively prove they are in one. Every observation, anomaly and apparent glitch is still being observed from inside the system. It may look like evidence of the machinery outside, but it could simply be another feature of the environment.
An ant might discover that the table has legs, but that would not prove the existence of a restaurant, nor a city, nor a country.
The only decisive evidence would be a message from outside, or a breach in the wall between worlds.
Eventually, the test run will finish. The operators will inspect the results, adjust a few parameters and restart the system from a clean state.
From inside, we will call that the Big Bang.
If I found out this was true, I would look for the back door to the admin panel. We built the agents, after all. We know how careless programmers can be.
Maybe God is just another programmer and we are just another abstraction layer.
The default password is probably still “admin”.