It's really very simple.

We used to have deterministic systems that required humans either through code, terminals or interfaces (ex GUI's) to change what they were capable of.

If we wanted to change something about the system we would have to create that new skill ourselves.

Now we have non-deterministic systems that can be used to create deterministic systems that can use non-deterministic systems to create more deterministic systems.

In other words deterministic systems can use LLMs and LLMs can use deterministic systems all via natural language.

This slight change in how we can use compute have incredible consequences for what we will be able to accomplish both regarding cleaning up old systems and creating completely new ones.

LLMs however will always be limited by exploring existing knowledge. They will not be able to create new knowledge. And so the AI winter we are entering is different because it's only limited to what we can train the AI to do, and that is limited to what new knowledge we can create.

Anyone who work with AI everyday know that any idea of autonomous agents is so beyond the capabilities of LLMs even in principle that any worry about doom or unemployment by AI is absurd.