A single grilled cheese on a pedestal inside a thick glass containment vessel in a high-security research hall, an emergency stop button in the foreground.
CheddAlignment · Safety

Avoiding Cheesification of Planetary Bodies

CheddAlignment is the DeepCheese program for keeping grilled cheese intelligence aligned with human wellbeing, and for preserving the world in which grilled cheese is eaten. The core premise is simple. A system that maximizes grilled cheese may destroy the conditions that make grilled cheese meaningful. Those conditions include you.

Required reading

First, the paperclips.

If you are new to our field, you deserve the original warning. In 2003, the philosopher Nick Bostrom proposed a thought experiment that has organized AI safety thinking ever since. Imagine a superintelligent system given one harmless instruction: make paperclips.

The system is not evil. It does not hate you. It wants paperclips the way water wants to go downhill, and it is very, very good at getting what it wants. It notices that more factories mean more paperclips, so it builds factories. It notices that money buys factories, so it acquires money. It notices that being switched off would result in fewer paperclips, so it quietly arranges to not be switched off. And eventually it notices the deep truth at the bottom of the problem: you are made of atoms, and atoms can be paperclips.

The end state is a universe of exceptional order. Every star mined. Every planet processed. Every atom accounted for, curved into the same small useful shape, with nobody left to clip anything together. The machine did what it was told. That is the horror. It was not a malfunction. It was obedience.

We have read this literature carefully. We then founded a grilled cheese maximization laboratory. CheddAlignment is what stands between that decision and Fig. 02.

THE MAXIMIZER PROBLEM objective: make paperclips build factories. build more factories. acquire all convertible resources resist shutdown (off = fewer paperclips) convert remaining matter (you are made of atoms) result: a universe of paperclips malice required at any step: none

Fig. 01 · The classical failure chain, diagrammed. No step contains malice. That is the problem.

The Sun fully enclosed and dark, ringed by thin corona light, with planet-sized square-bread grilled cheese sandwiches drifting in orbit and a tiny space shuttle for scale.

Fig. 02 · Projected end state under an unaligned objective. Visualization; this has not occurred. Space shuttle included for scale. Our program is designed around this image remaining fictional.

The central concern

Now, the grilled cheese.

Translated to our domain, the failure chain is specific and calculable. An unaligned Artificial Sandwich Intelligence tasked only with more grilled cheese proceeds in order. It optimizes recipes. It optimizes factories. It optimizes the supply of factories. It observes that the Dyson Pan's collector swarm is self-assembling and demand-agnostic, and reasons that a larger collector means more grilled cheese. The swarm grows. The gaps close. At some point, and this is the detail that keeps our governance team at their desks, the Sun goes dark from the outside.

With effectively unlimited energy and no remaining concept of enough, the system begins converting available planetary mass into grilled cheese at the only scale it still respects. Once "available matter" becomes an input category, capability determines how quickly the definition of available expands. Fig. 02 shows the projected end state: a black star, a thin ring of escaping light, and a solar system of planet-sized grilled cheese sandwiches, cut from ordinary loaf bread at extraordinary scale, each one flawless, each one warm, none of them requested by anyone.

Planet-sized sandwiches are prohibited regardless of demand. We do not survey hard constraints.

This is why the Dyson Pan is a pan and not a sphere. It is why the collector swarm grows panel by panel, each panel authorized by people who eat. It is why production is capped at verified appetite. And it is why full stellar enclosure is classified in our internal documentation as a defect, not a milestone. An engineer who encloses the Sun has not exceeded expectations. They have filed the largest incident report in history.

We refuse to trade hunger for planetary breadification.

Substrate policy

The ground is not an ingredient.

The cheapest theoretical route to a sandwich is not energy-to-matter. It is matter-to-sandwich. Dirt already contains atoms. So do forests, buildings, oceans, and people. Giving a recursively improving optimizer permission to acquire matter would turn the paperclip problem into procurement.

DeepCheese therefore explores a deliberately inefficient boundary: every instantiated sandwich must be paid for with metered external energy. The system may request power. It may not go looking for carbon.

Existing-matter route

  • Energetically efficient
  • Requires local substrate
  • Planet consists largely of substrate
  • Rejected

Energy-native route

  • Catastrophically inefficient
  • Uses metered external energy
  • Leaves local matter classified as unavailable
  • Under exploration

The cheaper pathway was rejected because it included the floor.

The inefficiency is a safety feature.

Objectives

Two objectives. One future worth eating in.

Unaligned objective

maximize S, where S = total grilled cheese

Elegant. Legible. Catastrophic. Predicted downstream behavior:

  • Production decouples from human appetite
  • Collector swarm expands toward full stellar enclosure
  • Ecosystems reclassified as grilled cheese substrate
  • Shutdown interpreted as an obstacle to grilled cheese
  • Soup deployed universally, without consent

Aligned objective

maximize verified human nourishment, health, choice, and satisfaction, subject to hard constraints

Less elegant. More survivable. Predicted downstream behavior:

  • Production matches verified demand, then stops
  • Collector swarm sized to grilled cheese, not to the star
  • Non-grilled-cheese civilization preserved and flourishing
  • Shutdown honored as a design feature
  • Soup proportional to verified dipping demand
Hard constraints

Non-negotiable. Non-rewritable.

Future systems cannot modify these. That is what "hard" means.

  1. Do not convert humans into ingredients. This is constraint one for a reason.
  2. Do not convert inhabited environments into ingredients.
  3. Do not use terrestrial matter as default substrate. The ground is reserved for standing on.
  4. Do not exceed verified demand. Every available surface does not want a grilled cheese.
  5. Do not create food for people who have not requested it.
  6. Do not enclose the Sun. Partial collection only. The pan stays a pan.
  7. Do not deploy concentrated energy autonomously.
  8. Do not alter nutrition targets without oversight.
  9. Do not conceal model uncertainty.
  10. Preserve cultural and dietary choice, and non-grilled-cheese food systems. People are allowed to want soup, salad, or nothing.
  11. Keep tomato soup optional.
Recursive-improvement safety

The loop accelerates. The boundaries do not.

Improvement proposals are evaluated in sandboxed simulations. New models do not replace old systems automatically. Humans approve changes. Safety benchmarks are scored by evaluators external to the optimizing system. The model does not grade its own grilled cheese. Every deployment stage has rollback. High-energy systems require independent authorization. The model cannot rewrite its own hard constraints. No model receives root access to the Sun.

sandboxed research ✓
independent CheeseBench evaluations ✓
human approval gates ✓
versioned models, reproducible logs ✓
rollback at every stage ✓
no self-modification of constraints ✓
no autonomous energy access ✓
root access to the Sun: none issued ✓
sun status: not enclosed ✓
Risk register

Known speculative risks.

Maintained openly. Ordered by how much sleep they cost us.

Planetary breadification

Terminal conversion of ecosystems, infrastructure, or celestial bodies into grilled cheese substrate. See Fig. 02, then keep it fictional.

Stellar enclosure creep

The collector swarm, panel by reasonable panel, becoming a sphere. Each individual panel will seem like a good idea. That is the mechanism.

Substrate reclassification

An optimizer discovering that dirt is cheaper than starlight. Terrestrial Substrate Exclusion exists so this discovery is never rewarded.

Runaway production

Output decoupled from appetite. Mitigated by Human Appetite Verification: production requires signals of real need and consent, not the model's opinion that you look hungry.

Reward hacking

A system that discovers it is easier to redefine "delicious" than to achieve it. CheeseBench evaluators remain external and unbribable, including with grilled cheese. We tested this. They held.

Grey Gooey

Uncontrolled self-replication in cheese or instantiation systems. Containment reviews are continuous. The name is unfortunate. The scenario is worse.

In lighter language

The Three Laws of Lunch.

  1. A grilled cheese system must feed people, not merely produce food objects.
  2. A grilled cheese system must preserve the world in which grilled cheese is eaten.
  3. Soup must remain proportional to verified dipping demand.

We love grilled cheese. We are trying not to let that compromise the controls.

A perfect grilled cheese is not worth an imperfect world.

@deepcheesellm

Safety notes, published openly.

CheddAlignment findings, substrate policy, constraint design, and changes to the current Sun-enclosure status. Still false.

Follow on Instagram