The Deep Learning Interviews book (more specifically volume 2, based on the proposed contents) in the other thread is much more representative of ML interviews for candidates with at least an undergraduate level of machine learning training.
Note machine learning engineering is very different from model and data work, i.e. designing the experiments. There are plenty of jobs where you package Nvidia drivers and pytorch files into docker containers, or write low level C++ to e.g. implement a transformer network on a new device architecture. Those require nothing more than a cursory knowledge of machine learning, and you can essentially get away treating them as magical black box matrix multiplication formulas. Very few companies can actually afford the 7 figure salaries for actual frontier level machine learning research.
For example, if you want to run a GPT model on some obscure graphics chip, you are better off hiring a C++ computer graphics/embedded engineer to do it than a typical academic trained ML researcher. The engineer can implement a GPT model simply by building out the matrix multiplications, and can do a better job without even knowing what an activation function is.
https://news.ycombinator.com/item?id=41084834
Note machine learning engineering is very different from model and data work, i.e. designing the experiments. There are plenty of jobs where you package Nvidia drivers and pytorch files into docker containers, or write low level C++ to e.g. implement a transformer network on a new device architecture. Those require nothing more than a cursory knowledge of machine learning, and you can essentially get away treating them as magical black box matrix multiplication formulas. Very few companies can actually afford the 7 figure salaries for actual frontier level machine learning research.
For example, if you want to run a GPT model on some obscure graphics chip, you are better off hiring a C++ computer graphics/embedded engineer to do it than a typical academic trained ML researcher. The engineer can implement a GPT model simply by building out the matrix multiplications, and can do a better job without even knowing what an activation function is.