LLM as CPU: Executing Pseudo-Code Directly Without Code Generation

Ok-Lab-7347 · reddit · 2026-08-30

Explores a paradigm where the LLM acts as a CPU/interpreter executing pseudo-code programs directly, bypassing traditional code generation. While probabilistic nature poses reliability challenges—errors compound as program steps increase—the concept is gaining traction. The post surveys prior art including CoRE (LLM interpreter), LATM (amortized tool synthesis), ToolGate (verifiable execution via preconditions), and Agent JIT (compiling tasks to executable plans). The author is conducting a fuller study using Agent Bayes connected via MCP.

Original post →

More from coding & agent

coding & agent channel →