Why Every AI System Eventually Becomes an Operating System
Scarce attention, untrusted inputs, competing workloads: the forces that produced Unix are loose in your AI stack, and your postmortems prove it.
Key Takeaways
- Teams building serious AI systems independently converge on operating system architecture - schedulers, quotas, memory hierarchies, privilege levels, preemption - without planning to.
- The convergence is structural: AI stacks recreate the three conditions that produced operating systems - scarce shared resources (attention, tokens, latency), mutually untrusted workloads (agents, tools, retrieved content, user input), and high-stakes arbitration.
- The OS curriculum therefore predicts your postmortems: lost updates demand locking, prompt injection demands protection rings (trust-tiered context with structural action gates), working-set overflow produces context thrashing with textbook remedies.
- The analogy has limits - the substrate is stochastic, premature kernels are overengineering, and providers keep absorbing layers - but as a map of what pressure produces next, it has been the most predictive frame I've used.