The quantum state is real. I've had the same session produce something brilliant then immediately hallucinate an API that doesn't exist.

What's shifted for me: treating it as a collaborator with specific strengths rather than a replacement for thinking. It's great at:

- Boilerplate I understand but don't want to write

- Explaining code I'm reading

- First drafts of tests

- Brainstorming approaches

It's terrible at:

- Anything requiring deep system context it doesn't have

- Low-level work where one wrong assumption cascades

- Knowing when it doesn't know

The leverage comes from learning its failure modes. Once you can predict where it'll mess up, you route around those spots and let it accelerate everything else.

And yeah - it's the worst it'll ever be. Which is the most interesting part.