The super literal interpretation ideas were much more common in the past when LLMs didn’t exist. Now we have models that are generally pretty good at picking up on nuance and understanding what you mean but also often quite bad at execution, which is roughly the opposite of that idea. I think reward hacking is perhaps the closest we see llms get to literal/malicious interpretations of instructions.
LLMs are neither of those. They're quite good at pretending they understand what you mean, but they don't. That's why they can't execute: they're mimicking the form, not the substance, and then we see the form and anthropomorphise them in our minds.
I've repeated the argument over and over since the GPT-2 days, when I derived it theoretically by inspecting the architecture of the model. I am now fatigued, and enough other people have taken up similar arguments – some developed half-way to a mathematical proof – that I no longer feel the obligation to keep repeating myself.
