Hello, a couple of days ago one of my assistants started answering with mash of randomness, the situation persists even for the simplest of tasks. Here’s an example of assistant who’s job is to respond with random greeting in specified language:
The instructions are simple: “given the ISO language code, answer with a short and creative greeting message” - every time, after some successful tries it goes haywire and starts streaming nonsense. Any clues what causes it, or if it is indeed a bug, where else to report it?
Thank you for sharing, unfortunately, this is not the solution - neither closing and starting the browser again (with emptying caches along the way), nor full restart of the computer made any difference - after few correct answers the assistant streams its mash again.
Short prompts are bad practice when dealing with AI.
Especially when a conversation history is prepended.
The AI has a very high chance of generating gibberish if there are too few input tokens.
The only way to mitigate this is to lower the temperature, increasing the amount of input tokens or starting a new chat every time so the conversation history won’t get prepended.
This is because the AI (in its training) gets punished for repetition.
It notices “Hey, I’ve been generating the same thing for the last 4 responses, let’s switch it up!”
Which in this case you do not want.
It’s probably better to solve this programmatically anyway and not utilize AI in the first place.
