[ben_w]

> No, it’s not. You’re making my statement abstract for the sake of arguing.

> I’m not a cook, doctor, or a lawyer. I can’t prepare meals for a party of more than 2.

They are demonstrating how over-broad your own statement was with an *equivalent* statement to show how it only passes on an unhelpful technicality.

Immediately after your quotation is this:

> you only need to spend 5+ years in medical school + whatever extra it takes to become proficient

LLMs pass the bar exam and the medical exam. These are things which I assume I would be able to do myself if only I were willing to dedicate 5 years of my life to each.

> I can write exactly all code an llm could write.

I can often see many errors in the code that ChatGPT produces. Within my domain, it's just a speed-up, a first draft I have to fix. Outside my domain, it knows what I *can't* Google because I've never heard the keyword that would allow me to.

On legal questions, ChatGPT (despite passing the bar exam) seems to make up cases. I belive this because I can google the cases and fail to find them. Is this because they don't exist, or because they're not indexed on Google? I don't have the legal background necessary to know — and it would take me years to get the knowledge necessary to differentiate "it's worse than first glance" from "it's better than second glance".

[xigoi]

Of course a memorization machine can pass exams that are mostly based on memorization.

[ben_w]

Even if LLMs were "a memorization machine" (they're bad at that), the statement is obviously false because actutal literal memorisation machines (books, video recorders, Google) cannot pass these exams.

LLMs only started to because they could follow the questions.

But even that aside, it doesn't matter why LLMs can do what they can do or what else can also do that, what matters is that it would take most humans several years to get to the level of current LLMs in a subject that human isn't already familiar with.