Transformer becomes a processor: C code is burned directly into model weights

🤖 Transformer becomes a processor: C code is burned directly into model weights

Percepta published an article titled Constructing an LLM-Computer. The system compiles a WebAssembly virtual machine into transformer weights, while an MILP optimizer selects the operation layout across layers. Any C code compiled to WASM is fed as tokens and executed by the model itself; the weights occupy ~1.5 MB.

🌍 This is a mechanism, not a promise: the program is stitched into the weights by construction, without training or an external interpreter. Execution is claimed to be exact integer arithmetic, but this is currently based on Percepta's blog posts and the author's demo, without independent reproduction.

👤 In the browser, you can load a C program (fibonacci, collatz, sudoku), compile it to WASM, and see how the model executes it step by step. The code and weights (~1.5 MB) are open on GitHub — you can run it locally.

Source 1: https://www.percepta.ai/blog/constructing-llm-computer Source 2: https://www.percepta.ai/blog/can-llms-be-computers