What I've Learnt Since December

The question

What have I learnt since December?

It took me longer to answer than it should have.

Why it's hard to answer

For anyone outside the tech world: December was the inflection point where AI got so good at coding that developers stopped typing code (or even correcting it manually) and moved over to accepting and inspecting diffs (if they even do that). If the code had a bug, the natural next step was to paste the bug into your coding agent and expect it to magic-fix everything. And it almost always does. Don't get me wrong, it's incredible. But somewhere along the way it kept getting better, and I started off-loading more and more of my thinking to it.

Designing a table schema? Take it. Whatever you think is right, and efficient. I was pinning a lot of hope on the word efficient to produce the right schema.

This blog post is an explanation of how quickly things have changed, and a personal note on my plight in this AI-driven everything: you have a problem, let me build an app for it. And most of those apps are bad. There are so many shitty applications on the internet right now, and I honestly don't know why. Competitive exams deduct marks for wrong answers; nobody deducts marks for shitty codebases.

The list

A few things stuck. Most came from people, not prompts.

From co-workers:

• StepFunctions for testing parts of a pipeline and not re-running the whole thing.
• Make pre-commit to check your ruff/lint/mypy checks before letting GitHub Actions do its thing. (GitHub Actions can quickly compound in cost over time and this has been a good alternative for me.)

On my own:

• I learnt about uv by watching a full 27-minute YouTube video instead of asking the AI.
• Deploying a Lambda function on AWS, though I'd have to refresh my memory.

What I noticed

Something that bugs me: the things AI teaches me don't stick, most of the time. Not as much as hearing them in a conversation with human beings, or reading someone's blog to get a richer understanding of things.

I think the reason is friction. A conversation makes me hold an idea long enough to respond to it. A blog post makes me sit with the author's framing for a few minutes. An AI answer arrives pre-chewed, complete, and instantly forgettable. I never had to do anything to earn it, so there's nothing for the memory to hang on.

There have been good times too. Skill exploration has helped me stay on top of new developments, and hearing how people use AI has been fascinating, and daunting at times. I follow a few people on the internet who constantly teach me new things and introduce me to new tools that make my workflow better, both in my understanding and in how fast I can deliver.

An aside on the algorithm

I do have a love-hate relationship with Twitter right now. The algorithm has been good at surfacing relevant topics for me, but most of the people posting under those topics seem to be using the same AI prompt, and you can tell from a mile away: excessive em-dashes, "[X] killed [Y]" framings, every point posed as a question-and-answer, too much use of the phrase "the catch is?". All of these are going on my block list on X.

What I'm doing about it

This isn't just a complaint. I've started changing how I work.

I used to be a pen-and-notebook kind of note taker. I switched to Apple Notes and Bear, partly because paper was hard to feed into the AI and the apps fit neatly into the workflow. Clean, efficient. And I never go back and read those notes.

Wispr Flow made it worse and better at the same time. Being able to blurt every half-formed idea straight into an app is incredible. But when I wrote in a notebook, I needed a level of conviction and clarity before the pen touched paper. Now I put down whatever comes out, on the assumption that the AI will clean it up later. It usually can. The problem is that I never revisit the thought: blurt, done, next. Revisiting was where the learning actually happened.

So the active steps are boring ones: reading blogs in full instead of skimming, and actually going back over the notes I take. Nothing clever. Just adding back the friction I optimized away.

The landing

I should be learning things every week. That's my point. I should not struggle to think about what I learnt. This should be ingrained and part of my routine.

The goal isn't to learn faster. It's to learn through channels where it sticks.