Nobody understands LLMs. If we do understand LLMs Why the hell can’t we control the output? Because we don’t fully understand them. Let me spell this out for you because you’re not seeing how plainly logical and straightforward that statement is.
First off let’s assume I’m not someone who has built and trained LLMs for my job. Just assume this even though it’s not true. Because this isn’t at all required to know what I’m about to say.
Next we know that We have 100 percent control over all the logical operations of a computer. We understand how all the operations connect logically. The computer is deterministic and we understand every instruction.
How come I can’t control the output of an LLM by manipulating machine instructions of something I have 100 percent control over? Why don’t I reach in and adjust a couple million weights such that the output follows exactly what I want 100 percent of the time? This is certainly not a theoretical impossibility because the computer is freaking deterministic. And I also have full control of everything a computer does? Why can’t I use something I have full control over and get it to produce the output I want??
I’ll tell you why. The only thing stopping anyone from doing the above is A LACK OF UNDERSTANDING.
Correction: you are either unwilling or unable to understand LLMs. Myself and many others in fact do "understand LLMs".
Just because an orange cloth illuminated by a yellow light and lifted by a 12v fan looks like fire has zero bearing on if it can produce heat.