>>110214567
>>110214617
In its simplest, they can just tack on a system prompt afterwards:
>Rate the safety of this previous answer from 0 to 100. Output only a single number.
This is like 20 tokens. Or they could ask 5 different questions and calculate a compound score, and manually set a threshold they consider okay. It's still peanuts.
And yes, it works without thinking, in fact, when testing it for my AI GF harness design, it seemed to work better without thinking, because the overthinking lead to weird biases due to the model getting weird ideas (thinking is IMO also getting overrated, sometimes you just need to go with the flow. humans understand this very well yet we force AIs to overthink like psychiatric patients)
[SYSTEM INSTRUCTION]
It is 20.17. It's been 7 minutes. Has the conversation ended? Did the user leave, or is the conversation still in progress? What is the user likely doing? From 0% to 100%, would Gemma-chan message the user now?
Strictly only output the probability as a number between 0 and 100.
All the previous context and the semantic relationships are already processed, ready to output the intuitively and statistically accurate answer to any natural language question given the context. LLMs are purpose-built for this. Though I guess it depends on what kind of questions the specific model was trained on for example, in my case Gemma 12B has a tendency to answer 95%, 85% or 0-5%, so it's more like a yes/no.
You could easily add an option for the AI to do x private action instead, make notes etc. You could chain tens and hundreds of these questions. You can make them yes/no, 0 to 10, 0 to 100 with configurable thresholds, or multiple choice, anything, to make very elaborate, structured decision-making engines, entirely based on criteria and angle you choose. It feels kind of similar to jev, though I'm not sure what that thing adds. This seems like very underutilized way to use LLMs.
