Resources for Chapter 11: An Alchemy, Not a Science | If Anyone Builds It, Everyone Dies | If Anyone Builds It, Everyone Dies

Resources for Chapter 11

An Alchemy, Not a Science

This is the online resource for Chapter 11 of If Anyone Builds It, Everyone Dies, which discusses how modern AI labs are approaching the problem of aligning artificial superintelligence. Refer to the book for answers to questions including:

How should we rate the current readiness of AI companies to solve the ASI alignment problem?
Where does “interpretability research” — research to read and understand AI minds — fit into the picture?
Can’t we just direct AI to solve the problem for us?

Below, we cover a grab bag of AI alignment and deployment ideas and proposed reasons for optimism, as well as proposed reasons it could be a good thing to advance the AI frontier even if the situation looks grim.

Won’t we just muddle through, like always?

The world usually muddles through by trial and error. In this case, early errors wouldn’t leave survivors. 1 min read

Do you see alignment as all-or-nothing?

No. But “partial alignment” is still likely to be catastrophic. 2 min read

Won’t the situation get better once governments get more involved?

It depends on how (and how soon) they get involved. 2 min read

Won’t the most reckless companies naturally be the most incompetent, and thus not a threat?

The more cautious companies today are still reckless. 4 min read

Isn’t it important to race ahead because of “hardware overhang”?

That would be suicidal, because we’re too far from an alignment solution. 3 min read

Isn’t it important to race ahead so we can do alignment research?

We strongly recommend against this entire AI paradigm. 3 min read

What if AI companies only deploy their AIs for non-dangerous actions?

Actions that seem benign can still require dangerous capabilities. 9 min read

Why not just read the AI’s thoughts?

Their thoughts are hard to read. 6 min read

What if we made AIs debate, compete with, or oversee each other?

If the AIs get smart enough to matter, they likely collude. 3 min read

What about various other AI alignment plans?

We cover additional alignment proposals in the book. 1 min read

Won’t there be early warnings researchers can use to identify problems?

Warning signs don’t help if you don’t know what to do with them. 8 min read

Your question not answered here?Submit a Question.

Extended Discussion

More on Some of the Plans We Critiqued in the Book

We Know What It Looks Like When a Problem Is Being Treated with Respect, And This Isn’t It

Shutdown Buttons and Corrigibility