All posts

A pattern I keep running into across very different parts of learning: every real shortcut turns out to just move the difficulty somewhere less visible, not remove it.

Being able to retrieve an answer and understanding it are different capacities, and it's never been easier to confuse them for each other.

Backpropagation is simple in theory and unforgiving in practice. Here's what actually went wrong when I built one with nothing but numpy, and how gradient checking saved me.

The advice is simple to state and uncomfortable to actually follow, for reasons that have less to do with time and more to do with ego.

The part of the job that's shrinking isn't the part most people worry about. It's not thinking that's getting outsourced — it's typing.

Not reading advice in general — specifically what to do with a book full of exercises and proofs that's too dense to read the way you'd read anything else.

Pipelines rarely fail loudly. They fail by quietly producing plausible-looking wrong answers, which is a much harder problem than a crash.

We talk about time management, but time was never really the constraint. What actually runs out, faster and less predictably, is attention.

Motivation runs out. What's replaced it is a small set of habits that make studying a hard subject less dependent on how I happen to feel that day.

Stripped of the vector calculus notation, gradient descent is just a rule for guessing better. Here's the mechanical version I wish someone had shown me first.