One can argue that today’s coder needs to understand how code-generating LLMs work. Which means learning a bit of linear algebra, calculus, and probability/stats.
I don’t think that’s the case really; if you go in the HPC/scientific computing direction and learn a bunch about linear algebra, how much does it teach you about LLMs? I did some linear algebra stuff… but LLMs just look like oddly shaped GEMVs.
I agree that most coders would be well served to learn some linear algebra and probability theory as those are ubiquitous. Calculus and stats are also useful depending on your specialty. LLMs don't really change any of this.
"A bit of linear algebra, calculus and probability/stats" would make you better equipped at leveraging LLM? Or do you mean that having "a bit" of those you would be able to implement/train you own custom LLM to write code for you?
I just meant that if you want to understand the tools you use as a SWE you will need to learn the relevant math to understand LLMs. Not sure if that would help with “leveraging”. Similar to any practical benefits of learning something like lambda calculus.
I know a few people that are already using LLMs in their own professional fields. I can assure you that none of them needed to learn (or recall) "the relevant math".
Just like the majority of MS Word users do not need to know that the paragraphs are internally represented as XML encapsulated in <x:p> tags.
You seem to conflate the idea of "building models from scratch" together with "using a model created by others to do meaningful work.
We are discussing "tools" here, you do not need to know much about physics or metallurgy or woodworking to use a hammer. And hammer users will outnumber hammer makers by various orders of magnitude.