Use a plain script when the input format is fixed and the logic is deterministic; use an LLM or AI agent when the task needs reasoning over messy, unstructured data. The Qwen 3.8 27B run on Simon Willison's laptop shows how far the second option has shifted - a 17GB local model now handles work that no amount of regex and if-statements could cover.
In Simon Willison's writeup, he reproduces an experiment from a recent paper: feed an LLM the raw floating-point weights of a small policy network and ask which game it was trained to play. Qwen 3.8 27B, running locally via LM Studio, correctly identified the game from the numbers alone. No code execution, no tools - just reasoning over the weights.
What this changes for small business automation
The useful signal is not that your laptop can now reverse-engineer neural nets. It is that a free 17GB model handles reasoning tasks that a year ago needed GPT-4 class cloud calls. For a lot of business flows - invoice classification, contract triage, support ticket routing, OCR post-processing - a local Qwen is now good enough, which kills the cloud API bill and keeps data on the client's machine. That matters when documents cannot be sent to OpenAI for compliance reasons, like in a legal document pipeline I built. Rule of thumb: keep the deterministic parts (file moves, API calls, DB writes) in code, and route only the reasoning step to the local LLM.
Why it matters: If you were waiting for local models to be usable before replacing paid API calls in your automation, the wait is largely over for text-heavy small-business tasks.
Source: Hacker News