Hacker News Reader

Comment by aprilthird2021 | Hacker News Reader

aprilthird2021

No one gets bothered that these weird invocations make the use of AI better? It's like having code that can be obsoleted at any second by the upstream provider, often without them even realizing it

View on HN

My favourite instantiation of this weird invocation is from this AI video generator, where they literally subtract the prompt for 'low quality video' from the input, and it improves the quality. https://youtu.be/iv-5mZ_9CPY?t=2020

I've just migrated my AI product to a different underlying model and had to redo a few of the prompts that the new model was interpreting differently. It's not obseleted, just requires a bit of migration. The improved quality of the new models outweighs any issues around prompting.

It's brittle, for sure. But ultimately I am the API connector so any output goes through me before being actioned on.

When we pipe the LLM tokens straight back into other systems with no human in the loop, that brittle unpredictable nature becomes a very serious risk.

Not really, it's just how they work. Think of them as statistical modellers. You tell them the role they fill and then they give you a statistically probable outcome based on that role. It would be more bothersome if it was less predictable.

You don't "tell them a role", they don't have any specific support for that. You give them a prompt and they complete based on that. If the prompt contains an indication that the counterparty should take on a certain role, the follow-up text will probably contain replies in that role. But there's no special training or part of the API where you specify a role. If the "take on a roll" prompt goes out of the context window, or is superseded by other prompts that push the probability to other styles, it will stop taking effect.

> You give them a prompt and they complete based on that. If the prompt contains an indication that the counterparty should take on a certain role, the follow-up text will probably contain replies in that role.

Or, more succinctly, you give them a role.

If I tell you to roleplay as a wizard then it doesn't matter that you don't have a "role" API does it? We would speak also of asking them questions or giving them instructions even though there's no explicit training or API for that, no?

Yes, if the role goes out of the context window then it will no longer apply to that context, just like anything else that goes out of the context window. I'm not sure how that affects my point. If you want them to behave a certain way then telling them to behave that way is going to help you...

The point is that "having a role" is not a core part of their model. You can also tell them to have a style, or tell them to avoid certain language, or not tell them anything specific but just speak in a way that makes them adopt a certain tone for the responses, etc.

This is similar to how you can ask me to roleplay as a wizard, and I will probably do it, but it's not a requirement for interacting with me. Conversely, an actor or an improviser on a stage would fit your original description better: they are someone who you give a role to, and they act out that role. The role is a core part of that, not an incidental option like it is for an LLM.

That's right. But I'm not sure how that is supposed to lead on from the point I made above.

It's not a requirement for interacting with them, but if you want them to behave a certain way then it's helpful to ask them to behave that way.

It's not necessary to ask a human to behave like a clown but if you want them to behave like one, say for the clown movie you're shooting, then helps to ask them to.

It's not necessary to ask an LLM a question, but if you want an answer then helps to ask one... Etc etc.

Maybe I'm missing your point but it seems a bit absurd, could you clarify?

Those “weird invocations” are called English.

Reply on news.ycombinator.com