Andrej Karpathy's recent AI startup school talk described LLMs as "new computers programmed in English" with properties of operating systems. We built LLMunix to explore these concepts practically.

  Core concept: Everything is defined in markdown documents that Claude interprets as functional system components - no traditional code, just natural language specifications.

  Key experiments:
  - Pure markdown OS where .md files become agents, tools, and system processes
  - Sentient state architecture with behavioral constraints that evolve dynamically
  - Memory-driven learning where each execution contributes to behavioral intelligence
  - Agent/tool duality for autonomous workflows

  Example: llmunix execute: "Monitor 5 tech sources and generate intelligence briefing" - the system adapts its approach based on API limitations, user sentiment, and historical patterns.

  This is early-stage research into how far we can push natural language computing. The results suggest Karpathy's vision of LLM operating systems might be closer than expected.

  Links:
  - Karpathy's talk: https://x.com/karpathy/status/1935518272667217925
  - LLMunix repo: https://github.com/EvolvingAgentsLabs/llmunix

  Looking for feedback from the HN community on this approach to LLM-native computing.