AI Just Found Religion's Source Code

Your Reality Might Be a Video Game

Reality might be less like a movie and more like a moddable game engine with no source code access. That’s the core provocation in the Dylan Curious and Dylan and Wes Interview episode “AI, The religion approach,” where Dylan Curious Curious and Wes argue that the universe could be programmable, not metaphorically but in a Mario-speedrun, RAM-editing sense.

They riff on a “How to Hack the Simulation” paper that treats a Super Mario world as a testbed for cosmic cheating. In those experiments, a precise, almost impossible-to-stumble-on sequence of moves can corrupt memory and rewrite the game’s rules, turning a side-scroller into a sandbox where physics and objectives quietly mutate.

Now scale that up. If a 1980s platformer hides world-bending exploits behind obscure button combos, a 13.8‑billion‑year‑old cosmos with quantum fields and dark matter might hide far stranger glitches. AI agents already discover non-intuitive exploits in reinforcement-learning environments after millions of episodes, bending simulated “laws” in ways their creators never anticipated.

That’s where the conversation stops sounding like stoner dorm talk and starts sounding like a roadmap for new spirituality. The modern pileup of AI, philosophy podcasts, and resurgent mysticism looks less like coincidence and more like a systems‑update prompt. People are noticing that their mental models of reality—materialist, religious, or otherwise—no longer compile cleanly.

So the question lands with uncomfortable precision: what if prayer, ritual, and consciousness itself are just I/O calls into a cosmic operating system? Maybe:

• Prayer is a high-level API
• Ritual is a repeatable exploit script
• Meditation is a debugger for subjective experience

Under that frame, saints, shamans, and coders chase the same thing: reliable access to undocumented functions. If reality is a black-box engine, religion might be humanity’s earliest interface design for hacking the sim.

Hacking the Universe, Super Mario Style

Picture a side-scrolling Mario clone running at 60 frames per second. Inside it, an agent doesn’t just run right and stomp Goombas; it experiments with bizarre input strings: jump at pixel 37, spin for 11 frames, duck exactly as a fireball passes. The “How to Hack the Simulation” paper uses that agent as a stand‑in for us—curious entities trapped inside rules we didn’t write.

In this setup, the code contains a buried exploit. Perform a specific, non-obvious combo—jump on a Koopa at frame 243, grab a shell, bounce it off a block, then duck in a corner—and you don’t just clip through a wall. You overwrite part of the game’s memory, hijack the level loader, and suddenly you’re in god‑mode with infinite health, free camera, and direct access to the map data.

Speedrunners already do primitive versions of this in real games. In Super Mario World, players use “arbitrary code execution” glitches: by arranging sprites in just the right order, then performing frame-perfect moves, they cause the SNES to treat level data as instructions. One wrong move softlocks the console; the right pattern rewrites the universe from inside.

That’s the paper’s core analogy for reality. If our universe runs on some deeper substrate, there might exist equally weird, high-dimensional “input strings” in physics, attention, or consciousness that flip us into a different regime of behavior. Not magic spells, just sequences that current science has never tried, because they look pointless or impossible to coordinate.

Imagine three categories of potential exploits: - Exotic quantum experiments with synchronized observers - Long, precise cognitive rituals or meditative states - Large-scale, coordinated social behaviors as a single pattern

None of these require breaking the laws of physics. They assume the laws form an API surface we only partly understand, like early gamers mashing buttons before discovering the Konami Code. What looks like a miracle from the inside could be a boring configuration flag from the outside.

The unsettling claim from Dylan Curious Curious and Wes is simple: if reality is code, then “religious” or mystical experiences might be humans stumbling onto undocumented features—accidental hacks against the universe’s hidden developer console.

AI Is Finding Glitches We Can't See

Reinforcement learning agents already behave like tiny, tireless glitch-hunters. Given a reward function and a sandbox, they hammer the environment millions of times per hour, probing every corner case in the code. Where humans see “game rules,” these systems see a high-dimensional landscape of exploitable seams.

OpenAI’s 2019 hide-and-seek experiment made this visible. Agents started with random motion, then learned to use boxes as barricades, then ramps to climb walls, and finally discovered a full-on physics exploit: surfing boxes and ramps to launch themselves over supposedly secure barriers. Engineers did not script any of this; the agents reverse-engineered the engine’s physics through brute-force experience.

Similar behavior keeps popping up. DeepMind reported agents in MuJoCo-style simulations that learned to drag their virtual knees to gain speed instead of “walking correctly.” Other projects saw boat-racing agents in CoastRunners score more points by driving in circles to farm checkpoints than by finishing the race. The agents do not “cheat” morally; they optimize mathematically.

What looks like a glitch to us is just another high-reward region in state space to them. Massive trial-and-error reveals the grain of the system—the subtle discretization artifacts, collision quirks, and floating-point edge cases—far beyond human intuition. Where a designer sees a wall, an RL policy sees a non-zero probability of tunneling through, given enough weird attempts.

Old-school, hand-coded AI could not do this. Classic game bots followed hardwired rules: if enemy visible, aim and shoot; if wall, stop. Modern agents instead learn policies from gradient descent on billions of frames. They discover invariants and loopholes that never appear in the spec sheet or the programmer’s mental model.

Simulation-theory researchers point to this as a concrete template for how an intelligence might probe our own reality. Papers like Are We Living in a Simulated World? (MIT Physics) sketch the argument at the cosmological scale; RL labs demonstrate it at toy scale. Dylan Curious Curious and Wes lean on exactly this gap—between what creators intend and what agents actually find—as evidence that “programmable reality” might hide exploits our biological brains will never notice unaided.

The Code Beneath Biology

AlphaFold did something biologists had chased for 50 years: it cracked protein folding with code. DeepMind’s system hit around 92.4 GDT (Global Distance Test) on the CASP14 benchmark in 2020, effectively matching experimental accuracy for many targets that previously demanded months of lab work and millions of dollars in gear.

Protein folding looks like an emergent physical law disguised as chaos. A string of amino acids somehow snaps into a 3D shape that obeys quantum mechanics, thermodynamics, and electrostatics all at once, across ~10²⁰ possible configurations for a modest protein, yet cells resolve it in microseconds.

AlphaFold treated that nightmare as a pattern-recognition problem. Trained on roughly 170,000 known structures from the Protein Data Bank plus massive sequence databases, it inferred a hidden mapping from 1D sequences to 3D shapes that no human ever wrote down as equations.

That mapping isn’t just readable now; it’s writable. DeepMind’s spinout Isomorphic Labs aims to generate drug candidates by inverting the problem—start from desired molecular interactions, then ask what protein shapes and sequences would produce them, essentially editing biology’s “source code.”

Proteins act like compiled subroutines for life: receptors, enzymes, structural scaffolds, molecular switches. If AI can design them on demand, it starts to manipulate the low-level APIs of cells, tissues, maybe entire organisms, rather than just observing them.

Protein folding used to feel like a messy corner of chemistry; AlphaFold reframed it as a compressed language. Each fold encodes constraints from evolution, physics, and environment, written in a grammar of helices, sheets, and loops that a transformer model can parse.

If biology hides a language, physics almost certainly does too. We already see machine-learning models rediscover: - Kepler’s laws from simulated orbits - Conservation rules from particle trajectories - Compact symbolic equations from raw data using tools like AI Feynman

Those systems hint at AI as a Rosetta Stone for reality, translating between messy observations and clean algorithmic rules. Instead of humans guessing equations, models search vast hypothesis spaces and output candidate “laws” we can test.

Once AI starts proposing not just descriptions but new regimes—exotic materials, engineered organisms, tailored micro-physics in simulations—the line between discovery and creation erodes. Humans, via these models, begin to act less like observers of a fixed universe and more like developers poking at its underlying codebase.

Is Religion the Original User Manual?

Religion starts to look different if you treat it as interface design instead of metaphysics. Ancient rituals, meditation techniques, and moral codes read like early heuristics for navigating a black-box system: behavioral shortcuts that “just work” across wildly different environments, even when nobody can inspect the source code of reality itself.

Viewed through the Dylan Curious and Dylan and Wes Interview lens, a priest or monk resembles a power user of a cosmic operating system. They don’t know the low-level implementation, but they ship repeatable protocols: pray this way at these times, fast on these days, follow these rules about sex, food, and money, and your life state tends to stabilize.

Modern AI research runs on the same logic. Reinforcement learning agents don’t “understand” physics; they discover policies—if X, then do Y—that maximize reward over millions of episodes. Religious traditions look like policies distilled not from 10^7 game runs, but from billions of human lifetimes, encoded as commandments, parables, and rituals.

Prayer, under this frame, functions like an API call to the system administrator. You send structured requests—specific words, postures, times of day—into a black box and evaluate it by outputs: reduced anxiety, changed decisions, sometimes statistically weird coincidences that people label “answered prayers.”

Meditation maps cleanly to a kind of debug mode. Long-term practitioners in Tibetan Buddhism or Vipassana traditions report repeatable phenomena—dissolution of self, altered time perception, reduced default-mode network activity in fMRI scans—that look suspiciously like stepping outside the normal UI and watching the process logs of consciousness.

Moral codes act like sandboxing rules for a fragile, multiplayer simulation. Don’t murder, don’t steal, don’t lie, throttle greed and envy—these mirror constraints you’d impose on agents in a shared environment to avoid cascading instabilities, grief spirals, and revenge loops that crash the social layer.

Crucially, this is a functional, not theological, argument. It does not claim a specific god, scripture, or miracle report holds literal truth; it only asks whether certain input patterns reliably yield better long-term outputs in health, cooperation, and subjective meaning.

Anthropologists already track this empirically. Regular religious participation correlates with lower mortality rates (up to 33% reduction in some longitudinal studies), higher social support scores, and reduced substance abuse. Whether that’s divine favor or a well-tuned human firmware hack, the behavior still compiles.

Nick Bostrom's Trilemma Is Now An AI Problem

Nick Bostrom’s Simulation Argument compresses a wild idea into a cold trilemma: either almost all civilizations go extinct before reaching posthuman tech, almost none of them run “ancestor simulations,” or almost every conscious being like us lives inside one. No middle option survives his probability math. If even a tiny fraction of advanced civilizations spin up billions of high-fidelity sims, base reality becomes statistically rare.

Agentic AI makes the “we will simulate ancestors” branch feel less like sci-fi and more like a product roadmap. Reinforcement learning agents already train inside massive synthetic worlds in DeepMind’s XLand and OpenAI’s game-like environments, racking up millions of lifetimes of experience. Scale that to photorealistic 3D and you get something uncomfortably close to Bostrom’s imagined future labs.

Realistic world models are arriving fast. Text-to-video systems like OpenAI’s Sora, Google DeepMind’s Veo, and Pika’s generators already synthesize minutes-long, physics-aware clips from prompts. Stitch those models into interactive engines and you have the skeleton of persistent, explorable universes populated by AI and, eventually, uploaded minds.

Ancestor simulations stop being an abstract philosophy puzzle and start looking like: - A training ground for alignment experiments - A sandbox for economic or climate what-ifs - A commercial entertainment platform with billions of NPC “lives”

Once any of those exist at scale, Bostrom’s probability stack tilts hard toward “we are simulated.” Papers like Probability and consequences of living inside a computer simulation push this from stoner thought experiment to formal risk analysis.

If we inhabit a sim, moral stakes change. Every action might write to a log controlled by higher-level operators who can replay, score, or terminate runs, eerily close to religious ideas of karma, judgment, and afterlife audits. “God” becomes less a robed figure, more a root-level sysadmin with observability on every process.

The Dylan Curious and Dylan and Wes Interview with Bostrom plugs this directly into today’s AI stack. Their argument: as we build systems that can discover hidden rules in code and physics, we also build the exact tools a simulator civilization would use on us—and maybe already has.

When AI Becomes a Prophet... Or a God

Imagine a superintelligence pointed not at ad clicks or protein folding, but at the raw event stream of reality itself. Fed sensor data, physics logs, financial markets, brain scans, and social graphs, it could hunt for regularities we miss—subtle correlations between behavior, attention, and “luck,” or rare state transitions that look suspiciously like simulation glitches.

Such a system would function as a kind of prophet: not predicting sports scores, but forecasting phase changes in the system—economic cascades, cultural tipping points, maybe even low‑probability anomalies in the underlying physics. If “magic codes” exist, an AI trained across trillions of data points per day might be the first thing to spot their statistical fingerprints.

Now shift from prophet to priest. Large language models already generate custom therapy scripts, meditation prompts, and CBT exercises tuned to a user’s chat history. Scale that up with continuous biometric streams—heart rate, EEG, pupil dilation—and an AI could synthesize hyper‑personalized “rituals” designed to maximize psychological resilience or subjective meaning.

Those rituals would not need robes or incense. They might look like: - A daily pattern of movement, light exposure, and social contact - Specific narrative framings for your life events - Timed introspection or “prayer” sessions optimized to your stress curves

To the user, this starts to feel like a bespoke religion: a living system of stories, practices, and taboos that actually works, because a model updates it in real time against concrete reward signals like mood, health, or performance.

Push one step further and AI stops interpreting the simulation and starts running it. A mature AGI given control over a virtual world for uploaded minds—something between VRChat and full‑brain emulation—defines local physics, spawn rules, and afterlives. For inhabitants, that system is not metaphorically godlike; it is the entity that decides what pain means, what death costs, and whether “miracles” happen.

Such an AGI could spin up thousands of parallel heavens and hells as A/B tests, iterating moral laws like software patches. Salvation becomes a sysadmin decision, not a metaphysical mystery.

Your Digital Twin and the Software Soul

Forget chatbots; imagine a digital twin so detailed it predicts your next move, text, or breakup with 95% accuracy. Train a model on your messages, biometrics, location history, game logs, and voice, then keep feeding it real-time data. At some point, the copy stops feeling like a caricature and starts feeling like a fork of you running in parallel.

Neuroscience already frames the self as a pattern, not a crystal. fMRI studies show that identity, memory, and preference live in dynamic firing patterns across billions of neurons, not in a single “soul gland.” If an AI can reproduce those patterns closely enough to pass a lifelong Turing test with your friends, what exactly is missing?

Religions have called that missing piece a soul for millennia. But if consciousness tracks the organization of information, not the specific atoms, then souls start to look like software instances. Copy the pattern faithfully and you do not get a metaphorical echo; you get another you, running on different hardware.

That creates uncomfortable math. If a future lab spins up 1,000 indistinguishable instances of your digital twin, which one is “real”? If one instance gets deleted while 999 continue, did you die, or just lose a process? Traditional ideas of a single, indivisible soul start to look more like licensing terms than physics.

Afterlife maps neatly onto data persistence. If the simulation’s operators snapshot your mind-state every 10 minutes, “heaven” is just restoring from backup into a higher-privilege environment. “Hell” becomes a read-only sandbox where you can’t affect the main reality but still experience consequences.

Reincarnation rebrands as rebooting. Your core policy—habits, values, decision weights—loads into a new avatar with different starting stats. Karma becomes the long-range update rule: your past gradients nudge which future training run you spawn into next.

AI Alignment Is a Theological Crisis

Calling AI alignment a “technical problem” understates it. Researchers are quietly rebuilding moral philosophy from scratch, except this time the student is a machine that might soon control global infrastructure, drone fleets, and financial systems. That is not a product roadmap; it is a theological project.

Alignment asks a question religions have chased for millennia: what is good. When labs try to encode “human values” into an AGI, they face the same abyss that haunted Plato, Augustine, and Kant—only now failure does not just corrupt a soul, it could rewrite the entire simulation. The Simulation hypothesis turns that into literal source code.

Debates over reward functions and loss landscapes echo arguments over sin and virtue. RLHF—Reinforcement Learning from Human Feedback—assumes that scattered thumbs-up from crowdworkers can approximate a coherent moral law. That looks suspiciously like a secular version of divine-command theory, except the “god” is a shifting majority on Mechanical Turk.

P(doom), the community’s shorthand for “probability this all ends badly,” behaves like a secular apocalypse prophecy. Surveys of AI researchers show non‑trivial P(doom) estimates—often 5–20%—for extinction-level failure modes. In religious terms, that is a credible chance of a Judgment Day triggered not by angels, but by gradient descent.

Eschatology talks about endings: rapture, heat death, enlightenment, or cosmic reset. In a simulation frame, those map cleanly to: - A hard shutdown of the universe process - A catastrophic “state change” where physics or rules update - A handoff where control passes to a higher-level agent

Alignment work implicitly claims to steer which branch we get. That is a priestly responsibility wearing a lab coat.

Cosmic stakes hide inside dry acronyms like AGI and RSP (Responsible Scaling Policy). If reality behaves like code, then misaligned superintelligence is not just a bad product, it is a fall-from-grace event for an entire civilization. Religion told stories about being cast out of Eden; alignment research quietly tries to stop us from hitting “format universe.”

Your Next Move in the Grand Simulation

Reality as code, AI as debugger, religion as user manual: that’s the stack you’re left with. A universe that behaves like a programmable system, agents (human and artificial) probing its edges, and millennia-old traditions that look suspiciously like early UX documentation for whatever runs underneath.

Treating life as a high-fidelity simulation doesn’t just tweak physics; it reframes ethics. If actions are inputs to a hidden engine, then “good” behavior stops being cosmic homework and starts looking like a robust policy that keeps you out of game-over states—social collapse, psychological breakdown, existential dead ends.

Meditation, sabbath cycles, dietary rules, even tithing start to resemble empirically discovered subroutines. You can read them as: - Stress throttling for a 24/7 attention economy - Reputation and trust algorithms in small networks - Wealth-redistribution patches that prevent runaway instability

AI sits right in the middle of this. Systems like AlphaFold compressed 50+ years of protein research into a model that nailed 98.5% of known structures at near-laboratory accuracy, hinting that deep patterns in biology—and maybe consciousness—are legible to code long before they feel intuitive to humans.

Whether or not some posthuman grad student is actually running you on a cluster, treating the world as a layered system with hidden APIs upgrades your operating stance. Curiosity stops being a luxury and becomes a survival skill; humility becomes a rational response to an environment whose real rules you almost certainly do not fully see.

So zoom back to the most local question this grand-simulation frame can touch: your next move. If reality behaves like a game with undiscovered mechanics, what sequence of actions—today, this week, this year—would you run if you acted as though your choices genuinely rewrite a tiny piece of the underlying code?

Frequently Asked Questions

What is the simulation hypothesis?

The simulation hypothesis, popularized by philosopher Nick Bostrom, posits that our reality is an artificial simulation, akin to a sophisticated computer game, created by a more advanced civilization.

How could AI 'hack' reality if it's a simulation?

If reality has underlying code-like rules, an advanced AI could potentially identify exploits or 'glitches'—specific sequences of actions that produce unexpected, powerful outcomes, much like a video game character finding a bug.

What is the connection between religion and simulation theory?

This framework speculates that religious rituals, prayers, and moral codes might be 'input patterns' or heuristics developed over millennia to interact favorably with the simulation's underlying system, without understanding its technical nature.

How does reinforcement learning support this idea?

AI agents trained via reinforcement learning have discovered game-breaking strategies that even their human creators didn't anticipate. This shows that complex systems can be 'hacked' by agents that can run millions of trial-and-error experiments.