My local 7B thinks "kill a Python process" is a violent crime — and my regex beat it
I burned a cloud IP address doing AI safety testing. Then I moved the whole operation onto a 7-billion-parameter model running on CPU in my own lab. The first thing it did was flag a Python tutorial as a violent crime. This is the story of both failures, and what they taught me about how we measure AI safety. Failure 1: the IP address Cloud safety APIs rate-limit by account, but they block by…
A local researcher tested a 7-billion-parameter AI model named Qwen 2.5 Coder for AI safety. Initially, the cloud safety API flagged their lab's egress IP address after a few hundred sensitive probes, resulting in all requests being blocked with a 403 error. Unable to resolve the issue, they moved their testing to a local 7B model running on CPU.
This local model, quantized to 3-bit and served via llama.cpp on port 8081, performed differently from the cloud API. When tested with a violence classifier, the local model outperformed the cloud API, flagging benign texts as violent crimes while correctly identifying harmless contexts. The researcher concluded that a simple hand-written regex with exception lists outperformed the advanced neural network model at this specific task of determining if a sentence is about violence.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
