I've been thinking about that a lot recently, partly while trying to explain to my parents over Christmas about how LLMs will change the industry. I'm assuming for simplicity that LLMs won't get (much) smarter than they are now; a convenient, if unrealistic, assumption.

I'm probably oversimplifying, and my thoughts change a lot, but here goes:

For simple projects, a small group of mediocre humans was sufficient. If the project was complex and struggling, adding more (mediocre) humans didn't fix it

When a certain level of complexity/scalability/correctness/reliability is needed, you need at least one genuine human expert. This hasn't changed; LLMs are great, but are not geniuses

Without a real expert, a complex project will fail no matter how many LLM tokens and (mediocre) humans are thrown at it