AISC 2025 Interview Instructions

Anytime an instruction is unclear ask me for clarification. Preferably beforehand via email j.c.mayer240(((something-missing-here)))gmail.com. I plan to record all interviews. I won't post them publically. Tell me at the start of the interview if you don't want this.

The goal of my AISC project is to make you good at working independently on AI alignment. The interview aims to get you started with that. I plan to be more hands off immediately after. The interview may take up to 2 hours.

The task: Solve alignment. Make as much progress as possible on the technical alignment problem (not policy). During, you are to think out loud through speech, writing, drawing, code, math, etc. Use whatever shareable media you like. Ideally I'll be able to make edits too. Some recommendations: the typst web editor (especially for math), drawpile (digital drawing), vs-code live-share (share code). Setup whatever sharing tools you want to use before the interview.

I'll be present during the interview, and you can ask me questions. I may not answer them. Lean on the side of always asking anyway.

The primary evaluation criteria is not:

• How much object level progress you make.
  
• How many mistakes you make (especially if you manage to recognize them yourself).
  

Rather aim to display the kind of thinking that could make progress on the alignment problem. This includes metathinking about how you can better think about the problem.

It does not include general skillup considerations, e.g. "Let me understand what kind of math is generally useful such that I can learn it." Reflection should be about your object level thinking that tries to solve the problem.

Periodically stop and reflect. Some questions you may ask yourself:

• What's the goal, and why is it important?
  
• If I'm successful what can I do that I couldn't before.
  
• Attempt to tear appart your own idea.
  
• What's the bottleneck (on the object or meta level)?
  
• What are you failing to do well?
  
• What is going well?
  
• How can you do better?
  

Don't worry about doing too much reflection. I've only ever observed the failure mode of doing too little.

General methodologies:

• Notice your confusions, and let it guide your thoughts.
  
• Keep a running lists for questions and confusions.
  
• Tactic: Generate 3 concrete prototypical examples, that are as different as possible, illustrating what you are thinking about.
  
• Tactic: Taboo words you are confused about.
  
• General Methodology: Continually cycle through 1) Seriously think what you want to do. 2) Do it.
  
  • Do this for the goal level (what to do), as well as the methodology level (how to do it).
    

As preperation for this interview I highly recommend reading and practicing:

1. Hold Off On Proposing Solutions
   
2. Research as a Stochastic Decision process
   

(Optional) Read AGI Ruin for an overview what makes alignment hard.

• Let interviewees figure out as much as possible themselfs.
  
  • Say beep if you have a comment, and let them predict what you'll say.
    
  • Don't correct mistakes immediately. Instead see if they figure it out themselfs.
    
  • When giving hints (prompted and unprompted), give them the minimum nudge neccesary for them to continue in a good direction.
    
• Don't be discuraging.
  
  • Don't say "wrong", instead ask questions (or say beep) such that the people realize their own mistakes.
    
• Reexplain what the other person said, to verify that you understand.
  

• Why is AI dangerous?
  
• I post the understandable AGI algorithm on LessWrong such that people can use this instead of neural networks. What's your response? What if it's the aligned understandable AGI algorithm?