Engineering technician Ben Santora tests small language models on Linux, using CPU-only environments like servers and containers. While testing the gemma-3-4b model, he observed that many tasks assigned to AI are already solved efficiently by the OS and its shell. This isn't an argument against AI, but a recognition of the shell's distinct capabilities.

Inside the operating system, Bash excels as an agent for files, logs, and processes. Tasks like log inspection, which can tax CPU cores when handled by an SLM, complete instantly via a simple shell pipeline. This efficiency matters for testing, as it frees model cycles for tasks requiring interpretation and judgment, not just mechanical execution.

Santora argues that modern 'agent' setups often bypass the OS, sending raw system data to a model first. However, Unix tools remain unmatched for their designed purpose: clean pipes, explicit behavior, and no inference cost. For his use cases, the most productive role for language models is assistance—explaining flags or suggesting commands—not replacing the shell itself.