An instruction in an LLM prompt is not a control. It's an input the model weighs against everything else it has read, in...

An instruction in an LLM prompt is not a control. It's an input the model weighs against everything else it has read, including whatever the user typed. If a rule matters (spending limits, what an agent may promise, which data it may reveal), it needs a deterministic check outside the model, plus a record of when that check fired. That's ordinary software engineering. It's strange how often AI projects skip it. #LLM #AIAgents #SoftwareEngineering

Read Original

Related