Over the last year and a half I have witnessed something that is truly alarming. I am calling it "The LLM Death Spiral".
This is the case of two people at a company using LLM's extensively for email communications. Over time, the LLM being used by the Boss, has been tuned to interpret emails from the Worker, as being negative, aggressive, inappropriate, etc.
I should add that, before LLM's the Boss was not a writer or reader. In other words, not great at creating or reading written communications. As a result, as soon as LLM's entered the picture, the Boss jumped in head-first and every single email the Boss writes is visibly created by an LLM.
The Death Spiral refers to this effect whereby this situation, through a negative feedback loop, degrades over time. Every Worker email is interpreted in an increasingly negative light, no matter how carefully crafted it might be. The situation got so bad that the Worker got written-up and threatened with firing.
This almost became a perfect embodiment of Plato's Allegory of the Cave. The Boss was, after a few iterations, only seeing the shadows on the wall cast by the LLM effectively trained by the Boss through interaction.
I wonder how much of this might be going on in this brave new paradigm, with the potential of LLM's truly damaging human relations and by, effectively, largely removing the human factor from communications.
This is the case of two people at a company using LLM's extensively for email communications. Over time, the LLM being used by the Boss, has been tuned to interpret emails from the Worker, as being negative, aggressive, inappropriate, etc.
I should add that, before LLM's the Boss was not a writer or reader. In other words, not great at creating or reading written communications. As a result, as soon as LLM's entered the picture, the Boss jumped in head-first and every single email the Boss writes is visibly created by an LLM.
The Death Spiral refers to this effect whereby this situation, through a negative feedback loop, degrades over time. Every Worker email is interpreted in an increasingly negative light, no matter how carefully crafted it might be. The situation got so bad that the Worker got written-up and threatened with firing.
This almost became a perfect embodiment of Plato's Allegory of the Cave. The Boss was, after a few iterations, only seeing the shadows on the wall cast by the LLM effectively trained by the Boss through interaction.
I wonder how much of this might be going on in this brave new paradigm, with the potential of LLM's truly damaging human relations and by, effectively, largely removing the human factor from communications.