You ask a rote question and you'll get a rote answer while the interviewee is busy looking at a fixed point on the screen.
You then ask a pointed question about something they know or care about, and suddenly their face lights up, they're animated, and they are looking around.
It's a huge tell.
I just tried and it's: hard. It feels like being ask to keep one's breath like writing something.
I need to focus too much on keeping my eyes closed, I don't have enough bandwith left to thing about anything relevant.
The coding challenge is supposed to be solved with AI. We can no longer afford not to use LLMs for engineering, as it's that much of a productivity boost when used right, so candidates should show how they use LLMs. They need to be able to explain the code of course, and answer questions about it, but for us it's a negative mark of a candidate proclaims that they don't use LLMs.
Do you state this upfront or is it some hidden requirement? Generally I'd expect an interview coding exercise to not be done with AI, but if it's a hidden requirement that the interviewer does not disclose, it is unfair to be penalized for not reading their minds.
I am personally of the view you should be able to use search engines, AI, anything you want, as the task should be representative of doing the task in person. The key focus has to be the programmer's knowledge and why they did what they did.
They also take this approach of "whatever tool works," but their coding test is "here's some symptoms of the SVG generator misbehaving, figure out what happened and fix it," which requires digging into the commit history, issues, actually looking at the SVG output, etc.
Once you've figured out how the system architecture works, and the most likely component to be causing the problem, you have to convert part of the code to use a newer, undocumented API exposed by a RPC server that speaks a serialization format that no LLM has ever seen before. Doing this is actually way faster and accurate using an AI, if you know how to centaur with it and make sure the output is tested to be correct.
This is a much more representative test of how someone's going to handle doing actual work knocking issues out.
It's not a hidden requirement per se to use LLM assistance, but the candidate should have a good answer ready why they didn't use an LLM to solve the challenge.
Also, what is a good answer for not using one? Will you provide access to one during the course of the interview? Or I am just expected to be paying for one?
We are providing an API key for LLM inference, as implementing the challenge requires this as well.
And I haven't heard a good answer yet for not using one, ideally the candidate knows how to mitigate the drawbacks of LLMs while benefiting from their utility regardless.
Again, what would be a good answer? Or are you just saying there isn’t one?
"I considered various approaches for solving this problem. Initially, I thought about using an LLM, as it's great for natural language processing and generating text-based solutions. However, for this particular challenge, I felt that a more algorithmic or structured approach was more appropriate, given the problem's nature (e.g., the need for performance optimization, a specific coding pattern, or better control over the output). While LLMs are powerful tools, they may not always provide the precision and control required for highly specific, performance-critical tasks, so I chose to solve the problem through a more traditional method. That said, if the problem had been more open-ended or involved unstructured data like text generation, I would definitely consider leveraging an LLM."
This answer reflects the candidate's ability to critically assess the problem and use the right tools for the job, showing maturity and sound judgment.
- GP, probably
"LLM-esque AI, especially in my industry, is under heavy scrutiny and I want to wait for the dust to settle before exploring options with such tools."
I was never asked as such, but I do have an answer to that.
Frankly, if an interviewer told me this, I would genuinely wonder why what they're building is such a simple toy product that an LLM can understand it well enough to be productive.