Reading papers before cognitive surrender

I really enjoyed The Pragmatic Engineer's conversation with Casey Muratori. We're not all in the same position as him, but it's refreshing to hear someone take a different stance in this AI era and still make a living from it.

I don't need to agree with every position to find value in that. Sometimes a conversation is useful because it makes me curious about something I've stopped paying attention to.

This one did exactly that.

Looking beneath the abstractions

I haven't touched assembly since college.

Most of my daily work happens several layers above that. Frameworks, libraries, containers, APIs: useful abstractions that let me build things without thinking about every instruction the machine executes.

But not having to think about something isn't the same as having nothing to learn from it.

Understanding how things work at different levels opens doors. A slow application looks different when you can ask questions about memory access, data layout, or the work a library is actually doing. You don't necessarily need to write assembly to benefit from knowing a little more about what happens beneath your code.

The conversation made me want to understand more about CPUs, GPUs, and TPUs. Not to become an expert in all three, or to replace every abstraction with a hand-written implementation. Just to have a better mental model of the machines doing the work.

The book recommendation that wasn't a book

Of everything that stood out, my favorite was Muratori's answer to the book recommendation question: read papers.

I've only read a few papers in my career. Most of my learning has come through other formats, and it's easy to keep reaching for the familiar ones.

A paper asks something different of me. What problem were the authors trying to solve? What did they assume? What did they test? Where does their evidence stop?

Those questions take time. I might need to look up a concept, revisit some mathematics, or read the same paragraph several times. That isn't necessarily a sign that the reading is going badly. It may be where the learning starts.

Delaying cognitive surrender

With AI everywhere, getting an explanation is easier than ever. That's useful. The risk, for me, is confusing an explanation I can follow with an idea I actually understand.

I can read a fluent summary and feel that I've learned something. Whether I could explain the idea without it, question its assumptions, or apply it somewhere else is another matter.

That's what I mean by cognitive surrender: letting the availability of an answer replace the effort of understanding it.

Reading papers isn't a guarantee against that. I can skim a paper just as superficially as anything else. But setting aside time to work through one, notice what I don't understand, and stay with those gaps seems like a good way to resist the habit.

I don't see this as a reason to stop using AI. I want to keep making room for the kind of learning where I have to do the thinking myself.

First on my list

I'd start with papers from before the GPT era. First on my list: Attention Is All You Need, published in 2017.

There's a certain irony in choosing a foundational Transformer paper as a way to resist outsourcing my understanding to AI. I like that.

My goal isn't to finish a stack of papers or pretend I've understood every equation. It's to come away able to explain one idea in my own words, including what I still don't understand.

That feels like a worthwhile place to start. 🙂