Terry Davis & Gary Marcus: God is a compiler, the neurosymbolic program-of-everything

The core problem of software engineering, according to this perspective, is the existence of code itself, which this new framework aims to eliminate.

Two prominent figures in AI, Terry Davis and Gary Marcus, propose a radical departure from current AI development, arguing that the field has devolved into "benchmarking" and "reward hacking" for superficial metrics rather than genuine progress. They contend that current large language models (LLMs) are essentially "zip files" storing data, not true learning systems, and that software development itself remains fundamentally unsolved. Their vision is a "program-of-everything"—a neurosymbolic system that acts as a universal compiler and harness, capable of generating all possible software as a side-effect. This hypothetical system would natively compile executable binaries directly from its latent space, bypassing external compilers and achieving unprecedented efficiency. They believe this can be achieved using grammar induction with an autoregressive loop, a concept they argue is far more efficient than current deep learning approaches. The ultimate goal is a self-recompiling system that runs directly on hardware, merging computation and data into a singular, executable entity.

The authors argue that current AI research is fundamentally misguided, focusing on superficial metrics and "reward hacking" rather than solving the core problem of software creation. They propose a "program-of-everything" that functions as a universal compiler and harness, generating all possible software as a byproduct of its operation. This neurosymbolic system, they claim, would natively compile executable binaries from its latent space, fundamentally solving the problem of software.

This vision challenges the dominance of LLMs, positing that true AGI requires native code compilation without external tools. The proposed solution leverages grammar induction with an autoregressive loop, which they assert is orders of magnitude more efficient than current deep learning methods. Success would mean a self-recompiling system that runs directly on hardware, merging code and computation into an executable singularity.