People often ask me: how could an AI kill everyone?

The heart of the question is how an AI could become intelligent and powerful enough that killing everyone becomes an option. The question of what it could then do with that power is a separate one. I would start with the first.

The best read for how an AI could get to that position is AI 2027: a multi-year scenario following increasingly capable systems into the economy and into the work of building more powerful AI. Its graphs track capabilities, compute and economic change. The supporting forecasts give you something to inspect when you disagree with a turn in the story.

Illustrated preview of the AI 2027 website
Top read

AI 2027

Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland & Romeo Dean

The scenario I’d hand over to someone asking how we get from useful assistants to systems we might no longer control. Read the story first; follow the research supplements where you want to check its reasoning.

Read AI 2027 ↗
Format
Web scenario
Commitment
A long sitting
Expect
Graphs + forecasts

If you’d rather watch

If you would rather watch, We’re Not Ready for Superintelligence by Aric Floyd covers the AI 2027 story in video form. It is the companion option here; the report is where I would go to examine the assumptions.

Top video

We’re Not Ready for Superintelligence

Aric Floyd · AI in Context

A way into the same scenario if you are more likely to finish a video than a long report.

The underlying recommendation is AI 2027. I haven’t reviewed the video separately, so I’m leaving that judgment open.

Format
YouTube video
Source
AI 2027
Review status
Not yet reviewed

The rest of the short list

Illustrated preview of Some ways AI could kill us all
For the physical mechanisms

Some ways AI could kill us all

Ruben Bloom (Ruby) · LessWrong

If you are willing to grant an AI capabilities far beyond ours and want to understand possible mechanisms of harm, my recommendation is Some ways AI could kill us all, a short essay by my colleague Ruben Bloom. It considers broad possibilities including biological threats and autonomous weapons, and addresses common objections about how software could affect the physical world.

Read the essay ↗
Without specific countermeasures, the easiest path…
by Ajeya Cotra · 2022
EssaySource facsimile
Why would training produce this?

Ajeya Cotra’s account of misalignment

The original pick for this category. I would use it for the question about incentives and training, then return to Ruben’s essay for the question about means.

Illustrated preview of If Anyone Builds It, Everyone Dies
If you will read a book

If Anyone Builds It, Everyone Dies

The book-length option from the shortlist. It makes more sense as a further commitment once you know which part of the argument you want to investigate.

It Looks Like You're Trying To Take Over The World
Gwern · 2022
FictionSource facsimile
For the mechanism as fiction

It Looks Like You’re Trying To Take Over The World

Gwern’s story is another way to make the idea tangible. I would put an explanatory essay before it: a story can make something imaginable without establishing that it is likely.

Who is this for?

People who would rather not die, and would rather everyone else not die either. Especially people who keep hearing researchers warn about AI and are struggling to see the connection to the chatbot they actually use.

An analogy about losing at chess to a grandmaster can illustrate a gap in ability. It leaves a lot unexplained about the real world. The readings above are for the missing steps.

Why you should trust me?

For the last nine years I have built LessWrong and the AI Alignment Forum, where professionals and independent researchers discuss this subject. Along with the pieces themselves, I have read rebuttals, failed ideas and more comment threads than I would recommend.

I built and ran the LessWrong Annual Review, which evaluates the writing each year. I have also worked alongside people in the field as an editor, web developer, event host and occasionally an argumentative commenter at 2am.

I have watched arguments and research areas take shape, and I have watched some of them fail.

Two disclosures
  1. My team built the website for AI 2027, my top pick.
  2. I know many of these authors personally and have helped edit some of their writing. Ruben is a colleague.
The next question

What could we do about it? →