G-Research Interview Questions
& Process

After my CV was screened, I got sent a proprietary online assessment with 10 MC questions. Some were unusual, some were tricky, but most were standard quant brainteasers you can pass if you've done a lot of practice. Then, there was an online interview with a quant who asked me a barrage of technical questions. The bar for passing probably depends on your answers and how many you get right. I wasn't prepared for some questions and totally fumbled.

It started with HR, then an online test, followed by two rounds of online interviews. The first online interview was a one-on-one, and the second involved three interviewers covering three different technical topics. After that, there was an onsite interview which included more technical questions and a check for cultural fit.

So the interview process had four rounds. The first round was for an hour. It was split into two parts: 30 minutes for computer science fundamentals and 30 minutes for pair programming where I could use any language I wanted. The second round was two hours long. The first hour was spent solving an algorithmic problem, which wasn't too hard. The second hour involved solving a software architecture problem, focusing on OOP. The third round was also an hour and focused on a high-level system design task. I think it was about designing a chat system, but I can't recall the exact task. Finally, the fourth round was a casual chat with the hiring manager about my career goals and motivations. It wasn't technical at all.

The interview process started with an online test covering math, probability, and ML. After that, I had a triage interview. To be honest, out of all the interviews I've had, I don't respect the G-Research interview. I really think the interviewer was an idiot. I was expecting the interviewer to assess my problem-solving and logical thinking skills, but that didn't happen. He asked me about deriving eigenvalues, and I showed him 'n' equations to solve for 'n'-polynomial equations. Then he asked about the time and memory complexity of an ML model and ways to optimize them. I suggested several tricks like mixed precision training, storing intermediate values on disk (accelerator), and replacing full attention with group attention, etc. The feedback was that he was considering recomputing the activation. So what? My point is that both these questions are just knowledge points, not practical applications or tests of problem-solving and logical thinking. I asked the interviewer if they actually look at financial data or perform analysis on it. He said he does model iteration but uses data that is handed over to him with features already built. Now, it's important to understand that autoregressive models and attention mechanisms are computationally expensive and slow. If people genuinely want to save memory, there are numerous tricks available by changing a few lines of code using frameworks. I really don't see how these questions are relevant to actually doing ML and solving problems. Do they implement backpropagation themselves? No, they use popular frameworks. Honestly, I respect all the other interviews I've had with big and small companies, but not this one from G-Research. It was the most stupid interview I've ever had.