I would refuse to do any AI interview, and refuse early-stage coding exercises. I require the base salary range and the number of in-office days/week upfront. I add a 10 to 20% premium for each onsite day per week.
I got tricked into doing this once, for Mercor. didn't have much else going on, and I'd never actually talked to an AI before. what struck me as very odd is that instead of being obsequious, the AI would ask me an involved question, I would start digging in, and it would very rudely interrupt me to ask something completely different.
after 1 30 minute session I'd had enough, and didn't complete the remaining sessions. the even nastier part is that for the next few weeks I would occasionally receive messages from Mercor thanking me for doing the interview, but saying that the format had been changed and it was necessary for me to go through the process again.
1. How well can they write a requirements spec for an AI? How well can they eliminate and disambiguate points of confusion? How well can they iterate upon the requirements by considering the AI's plan?
2. How well can they review the AI's output against the requirements? Can they identify and communicate what was missed, and translate it into updates to the requirements, fix the code via continued chat, otherwise by hand?
3. Can they avoid under-engineering, over-engineering, and unnecessary bloat?
The point is that if an interview is not assessing these things, it likely is wasting everyone's time.
I would refuse to do any AI interview, and refuse early-stage coding exercises. I require the base salary range and the number of in-office days/week upfront. I add a 10 to 20% premium for each onsite day per week.
made me sad reading this as a picture of the state of the world
I got tricked into doing this once, for Mercor. didn't have much else going on, and I'd never actually talked to an AI before. what struck me as very odd is that instead of being obsequious, the AI would ask me an involved question, I would start digging in, and it would very rudely interrupt me to ask something completely different.
after 1 30 minute session I'd had enough, and didn't complete the remaining sessions. the even nastier part is that for the next few weeks I would occasionally receive messages from Mercor thanking me for doing the interview, but saying that the format had been changed and it was necessary for me to go through the process again.
I suggest HN add “is lying to you” to the clickbait title filter.
LLM’s love that one weird trick to get hits.
To test a candidate, I want to see:
1. How well can they write a requirements spec for an AI? How well can they eliminate and disambiguate points of confusion? How well can they iterate upon the requirements by considering the AI's plan?
2. How well can they review the AI's output against the requirements? Can they identify and communicate what was missed, and translate it into updates to the requirements, fix the code via continued chat, otherwise by hand?
3. Can they avoid under-engineering, over-engineering, and unnecessary bloat?
The point is that if an interview is not assessing these things, it likely is wasting everyone's time.
Honestly, I think this should be a felony.