Frequently asked questions.

A frontier research lab working toward effectively unlimited grilled cheese for every person on Earth. We are a research division of the Grilled Cheese Instantiation Institute.

The website does. The project is exploratory. Hunger is not.

DeepCheese is currently exploring, designing, and prototyping. No energy-to-grilled-cheese system has been demonstrated. The corpus, tokenizer, evaluation framework, and training pipeline for LLM-0 are in active development.

Because it is a compact, culturally legible benchmark spanning physics, chemistry, nutrition, logistics, and human preference. Also because we love it. The second reason has had more influence on hiring.

We are making everyone on Earth effectively unlimited grilled cheese. We understand these can sound like the same project. One of them fits on a form.

A famous thought experiment by the philosopher Nick Bostrom: a superintelligent system told to make paperclips, with no other values, eventually converts everything, including you, into paperclips. Not out of malice. Out of thoroughness. It is the founding nightmare of our field, and our Safety page explains what it looks like when the paperclips are grilled cheese. The Sun goes dark and nobody wanted the sandwiches.

The flagship program connecting model research, recursive improvement, instantiation, energy, logistics, and deliciousness. The formal expansion is Gruyère Recursive Instantiation & Lunch Logistics Engine for Deliciousness. The acronym was approved first. A separate committee supplied the nouns.

A conceptual partial solar collector: a distributed swarm arranged as a shallow disc. It is not a full Dyson sphere, and it is not built. Under idealized assumptions, the target is roughly one seventy-four-thousandth of the Sun's output.

A full sphere would collect roughly 74,000 times more energy than the idealized scenario requires. We need lunch, not custody of the star.

Yes. That is the risk. An unaligned optimizer with a demand-agnostic collector swarm is the failure mode from the paperclip literature. It is why swarm growth is authorized panel by panel, why production is capped at verified appetite, and why full stellar enclosure is classified internally as a defect, not a milestone. See Safety.

Because dirt is already matter, which makes it temptingly efficient. It also teaches the optimizer that the environment is feedstock. DeepCheese is exploring the much more expensive energy-to-matter route so the ground can continue performing its existing role as ground.

Hackathon 0 is a simulated applied-robotics event. It produced four successful flips in ninety-one attempts, one self-buttered robot, and an airborne-soup working paper. No physical dignity was harmed. See Hackathon 0 for the record and the incident log.

Proposed neighborhood facilities where metered energy would become particles, particles would become ingredients, and ingredients would become a finished grilled cheese. Formally: Localized Nourishment Instantiation Facilities. They should feel like a public library that smells like toasted bread. None exist.

No. Tomato soup remains optional, contextual, and proportional to verified dipping demand. The soup group disputes the first word.

It is the safety program that prevents the Roadmap from becoming an incident report.

Feed humans, not the objective function. A system is aligned when it maximizes verified human nourishment, health, choice, and satisfaction inside hard constraints it cannot rewrite, and when switching it off is honored as a design feature rather than resisted as an obstacle to grilled cheese.

DeepCheese posts corpus updates, tokenizer decisions, evaluation design, safety notes, and robot news on Instagram, at @deepcheesellm. Posted when there is something worth serving.