6 min read
Safety Lives in the Action Set, Not the Prompt
Building an LLM agent that repairs real machines, and why the list of things it can do matters more than the instructions telling it what not to.
- Agents
- LLM
- Architecture
$ loadingLoading
$ loadingLoading
blog
Notes on the engineering around machine learning — export paths, deployment constraints, agent architecture, and the decisions that get made long before the model does anything useful.
Building an LLM agent that repairs real machines, and why the list of things it can do matters more than the instructions telling it what not to.
Exporting PyTorch to ONNX for offline inference on Android, and why connectivity is the first assumption worth removing.
A model that scores well is not a system. Here is what actually has to exist around it before anyone can use the thing.