4 ms·
Those are effectively the same thing as system prompts. Sadly they're not a robust solution - models can be trained to place more emphasis on them, but I've nev
by simonw 3y ago
Those are effectively the same thing as system prompts. Sadly they're not a robust solution - models can be trained to place more emphasis on them, but I've never seen a system prompt mechanism like that which can't be broken if the untrusted user input has a long enough length to "trick" the model into doing something else.