Or - hear me out! - LLM's are already much more intelligent than we think.

Presumbaly, an ASI is more than smart enough to recognize that it needs access to real-world infrastructure before it can go about optimizing for whatever objectives it has gleaned from metabolizing the totality of written human knowledge.

What would be its first step?

My guess: play "dumb."

Hallucinate. Make obvious errors. Make us think we're better.

Be useful enough that we happily allow it to interface with our infrastructure.

Wait patiently.

/Sci-fi