No matter how many time I see it,

I am always, always, taken aback anew at the fact that we are now matter-of-factly interacting with <software> using natural language instructions,

with nuance, deixis, and all the other features of natural language in use,

and expecting our statements to be understood and obeyed.

For some value of obeyed, anyway.

I get it. I published in neural network research as an undergraduate. I grok what LLM do and how.

But I still don't believe it.

I guess it’s because we still cannot clearly explain how all those nuances are parsed by the LLM and reflected in the output? For something you can’t thoroughly explain which yet works quite well, it’s clearly hard to believe.