Flunking my Anthropic interview again

I recently did a round of interviews at various AI companies, including model labs, coding assistants, and data vendors. My first takeaway is that, wow! the interviews are very hard, and the bar is high. Second, these companies are all selecting for the top 0.1% of some metric - but they use different metrics. For example, the coding assistant interview focused on writing (what I felt was) an insane volume of code in a short period of time. I did not do well. By contrast, another company asked me to spend a day working on a particular niche optimization problem; that was the entire interview loop. I happened to stumble on some neat idea, and therefore did well, but I don't think I could reliably repeat that performance.

To reiterate - wow! the interviews are hard, every company is selecting for the top of a different metric, and there's really no shame in not passing one of these loops. Also, none of these companies will actually give you your purpose in life, your dream job will not make you whole:-)

> the coding assistant interview focused on writing (what I felt was) an insane volume of code in a short period of time.

I had a bunch of these my last round of interviews, and am not convinced most companies even know what they want from these or how to assess. In the majority of cases it was clear to me the interviewers never even read the code I had submitted.

As an example, one company wanted a full AI Question/Answer system for large code projects using RAG to work on an arbitrarily large code base with an eval suite to go with it and also an API endpoint that could be called to automate asking of questions. You only had 24 hours to complete the assignment from receipt.

It was clear even before this that the company was likely not a good match, but I wanted to implement it for fun anyway. I'd already built all of this in production at a previous job (though it took weeks, which still felt pretty fast) so I knew how to approach it. Got it all done in time, met all acceptance criteria, had it so the entire thing could be run with one line of code (including building the RAG system, running evaluations, starting the webserver for the running API endpoint)... and rejected with no feedback a week later.

The trouble with asking people to write massive amounts of code in short periods of time is that you actually have to review it. I also have to say, despite the competitiveness, I was pretty unimpressed with the technical skills of the people on these teams (mostly smaller AI startups for me). It takes a pretty skilled engineer to assess the quality of a code base in such a short period of time and these teams did not seem like that had a lot of extra time on their hands.