It’s called a system prompt. It’s a wrapper around the prompt you, the user, writes, which is called the user prompt.
What exactly is being added to the system prompt is a huge part of the “secret sauce” of the AI providers, because it greatly affects the quality and tone of the output. It may also be part of the safety layer of guardrails to stop users from asking about topics the AI providers have deemed taboo (such as biology, in recent news).
we stopped access to an entire topic incase it spilled the beans on how to make bioweapons.
Living in the future has been fun. VR was pretty sweet. Ordering pizzas by tapping a rectangle was awesome. I want to get off now though. I want to go home.
Not really, you can add a prompt to commercial LLM products to alter the persona a bit but the actual system prompt the harness is using is opaque to you. Your Claudes, Kimis, Mistrals and what have you are a polished product on top of the model, they’re not the model itself. A lot of the actual day-to-day software engineering of these things is writing a harness that’s robust enough to deal with a cantankerous non-deterministic pile of linear algebra that doesn’t know the difference between ‘looks good’ and ‘is good’.
I feel like AI enthusiasts would become a bit disillusioned with AI if they could see what was being added to their prompts.
What’s being added?
It’s called a system prompt. It’s a wrapper around the prompt you, the user, writes, which is called the user prompt.
What exactly is being added to the system prompt is a huge part of the “secret sauce” of the AI providers, because it greatly affects the quality and tone of the output. It may also be part of the safety layer of guardrails to stop users from asking about topics the AI providers have deemed taboo (such as biology, in recent news).
Living in the future has been fun. VR was pretty sweet. Ordering pizzas by tapping a rectangle was awesome. I want to get off now though. I want to go home.
the sauce is not so secret, people actively try to leak them: https://github.com/elder-plinius/CL4R1T4S
I see. Thank you for explaining.
Most providers allow you to overwrite with your own system prompt so…
Not really, you can add a prompt to commercial LLM products to alter the persona a bit but the actual system prompt the harness is using is opaque to you. Your Claudes, Kimis, Mistrals and what have you are a polished product on top of the model, they’re not the model itself. A lot of the actual day-to-day software engineering of these things is writing a harness that’s robust enough to deal with a cantankerous non-deterministic pile of linear algebra that doesn’t know the difference between ‘looks good’ and ‘is good’.