What I look for when hiring an ML fresher in India
A year after ChatGPT, the resumes crossing my desk have changed and most of them changed in the same wrong direction. Everyone can now generate a working project. That has not made the good candidates easier to find. It has made the one thing I was always screening for, judgment, the only thing that still tells them apart.
It has been about a year since ChatGPT, and the fresher resumes crossing my desk have changed. The projects are more polished. The code is cleaner. Everyone can now generate a working thing in an afternoon that would have taken a week to figure out before. You would think that makes hiring harder, because the floor rose. It made it easier, because it stripped out the one thing that used to be a signal and left only the thing that actually matters.
Here is what I look for in a fresher now, in India, in a market where a tool can write the project for you.
Can you explain a decision the tool did not make for you
Generating code got cheap this year. Deciding whether it is right did not. So the first thing I probe is whether you understand the thing you submitted or whether you assembled it. I will pick a choice in your project and ask why, and then ask why not the simpler thing. If the tool made the choice and you accepted it, that surfaces in about two questions. If you made the choice and can defend the tradeoff, we are having a real conversation. The polish of the project tells me nothing now. The defence of it tells me everything.
Did you ship it, and did you handle the mess
The three things I have always looked for still hold, and I wrote them down before any of this: did someone who is not you use it, did you handle real messy data, can you explain a tradeoff. What changed is that the first two got easier to fake and the third got harder to fake, so I weight the third more than I used to. A fresher who can walk me through a bug they did not understand at first, and how they figured it out, is worth more than one whose project runs perfectly and who cannot tell me why. The struggle is the evidence. The polish is now free.
Initiative, because I cannot teach it and the tool cannot supply it
The trait I am really hiring for has not changed and will not: did you build this because a problem annoyed you, or because a course told you to. In a year where the tool can produce the assignment for you, the person who went and found their own problem to solve stands out more, not less, because the tool lowered the cost of doing the assigned thing and did nothing for the cost of caring. Initiative is the one input a language model cannot provide, and it is most of what I am screening for.
What I no longer weight
Certificate count, never much. Accuracy numbers, less than ever. And increasingly, raw ability to produce code, because that is the part that just got commoditised. If your pitch is that you can build things fast, so can the tool and so can everyone else. The pitch that works is that you can tell whether the thing you built is right, and explain why, and that you went looking for the problem yourself.
If you are the fresher, this is good news, oddly. The bar moved off the thing that was always a bit of a lottery, whether you had the time and resources to grind out an impressive-looking project, and onto the thing you actually control, whether you understand what you built and whether you cared enough to pick it yourself. Build the real thing, know it cold, and be ready to defend it. The parts of the interview I described back when I wrote about breaking into ML and the interview rounds that decide it still hold. There is just less room now to hide behind a project you cannot explain.