[rpedela]

I have definitely run into memory bugs in managed languages that someone without my experience would have no clue how to debug since the managed language hides that information. Having said that, learning to be "careful" is much more than just avoiding memory bugs. It is about learning to avoid mistakes by really thinking about the code you are writing so that you don't end up screwing yourself or the team later. That can be applied to memory management, class design, API design, complex algorithm development, etc. There are many ways to learn to be careful but manual memory management is one of the best because the consequences are so severe, in terms of debug time, if you aren't careful.

Big O is useful but I think its usefulness is overblown. There are cases where understanding how the kernel allocates memory or how CPU caching works or how networking works, you can develop an algorithm that is inefficient in terms of Big O but is still performant. Sometimes you can even beat the best theoretical algorithm. The reason is that k is extremely important. If I can make k extremely small using my knowledge of the computer and n is within reason, then Big O often doesn't matter. So I don't have to waste time implementing fancy algorithms as a result. But I also have a better feel when Big O does matter. If I can't make k small or n is a huge number or both, then I spend the time on algorithm optimization and then understanding Big O is helpful.

I will give you a real example with GPUs. Let's say you want to do some GPGPU and you have data that is too big to fit in RAM. How do you process this data efficiently? Knowing how the GPU works is very important to solving this problem. You could spend days, weeks, years optimizing the crap out of Big O in your code, but it won't really move the needle in a lot of cases because that isn't the bottleneck. The primary bottleneck in this case is the PCIe bus and a high-level understanding of how it works is needed to keep it full of data as the GPU is processing it. Once that is solved, the next bottleneck will likely be the data format. The GPU is most efficient when each data sample is independent which is related to how the GPU cores work, caching works, etc. So putting the data in GPU-friendly format (not always possible) will make everything go faster even before worrying about Big O.

What C does is it forces you to learn how the computer really works because it is hard to be productive in C without that knowledge. And that knowledge is largely transferable.

I certainly agree with learning and using Python or JS when it makes sense because they are very productive languages. But from an education perspective, people who only learn JS are less likely to be able to solve the really hard problems because they lack that foundation of how the computer works. And there are plenty jobs where that doesn't matter too much but you should probably have at least one person around that really understands how the computer works for the times when it does.