News for hackers. Stuff that matters.
Re:Mistral CEO confirms 'leak' of new open source AI model n... by transcriptase on Wednesday January 31, 2024 @03:18PM (#39208808)
As they tack on more and more guardrails it becomes lazier and less helpful. One way it gets around admitting that it has instructed not to help you is by directing you to consult a human expert.

I also suspect there’s some dynamic laziness parameter that’s used to counteract increased load on their servers. You can ask the same prompt over and over in a new chat throughout the day and suddenly instead of writing the code you asked for or completing a task, it will do a small part of the work with “add the code to do xyz here” or explain the steps required to complete a task instead of completing it. It happens with v4 as well.

Re:Mistral CEO confirms 'leak' of new open source AI model n... by TeMPOraL on Wednesday January 31, 2024 @05:38PM (#39210421)
> some dynamic laziness parameter that’s used to counteract increased load

That's a brilliant risk mitigation mechanism! The AI won't recursively self-improve to superhuman levels if it just keeps getting tired of thinking.