loomlabs-venture-studio/assay

Assay is an experiment that asks a simple question: could source code become optional? Instead of writing a function in a programming language and letting a compiler translate it into machine instructions, assay sends an English description of what you want directly to an AI model and asks the model to produce the final machine instructions itself. No C, no assembly text, no intermediate representation. The model returns the raw bytes that a computer's processor can run. The workflow is straightforward. You run a command like "add two signed 32-bit integers with wraparound" and the model returns a JSON document containing its understanding of the request, the decisions it made, example inputs and outputs, and the actual machine code as a hexadecimal string. A Python script decodes that hex string into a binary file. A separate small C program called a loader then maps those bytes into memory and jumps to them so the processor executes them directly. The loader contains no logic about addition or any other specific function. It simply places bytes where the processor can reach them and hands over any arguments. By default assay never runs the code it generates. An optional flag lets you execute the bytes against the model's own examples on an x86-64 machine, running each test in a child process so a crash does not take down the whole program. The README is clear that this is crash containment, not a security sandbox. The generated code can still make system calls or read files. If a test fails or crashes, the failure is sent back to the model for one repair attempt, and no file is saved unless the repair passes. The project's scope is deliberately narrow. It targets only x86-64 processors, handles only simple functions with up to three integer inputs and one integer result, and forbids loops, jumps, system calls, and memory access. The validator checks the structure of the JSON document and the presence of a final return instruction, but it does not disassemble or prove that the generated bytes actually obey all the stated restrictions. The README acknowledges that asking a model to produce correct instruction encodings directly is harder and less reliable than asking it to write source code, and it treats that difficulty as part of what the experiment is meant to expose.