Reflex swaps text for CHIP-8 opcodes

This is a strong demo of the “understanding vs. actuation” split: the LLM can still encode intent, but the execution layer becomes a specialized controller instead of a text parser. That makes the project intellectually useful, even if it is nowhere near a general software-control architecture.

• –The cross-attention setup is the most interesting part: instructions query machine state directly, which is a cleaner pattern than generating intermediate text and parsing it back into actions.
• –CHIP-8 is the right sandbox here because it is small enough to prove the concept for arithmetic, branching, and subroutines without pretending to solve arbitrary code generation.
• –The `two plus three` failure is the key limitation: the model seems to lose the latent bridge from natural language variants to executable structure once the decoder path is removed.
• –This reads more like a research prototype than a product, but the architectural idea could matter for VMs, embedded control, or other structured action spaces.
• –The repo suggests the real bottleneck is data coverage and representation alignment, not opcode synthesis itself.