I want to be a Journey Programmer Again

Relevant: LLMunix - A Pure Markdown Operating System - https://news.ycombinator.com/item?id=44279456 - Jun, 2025 (1 comment)

I get your point, I think the difficult thing is that these tools are not delineated: preground-ink does not have the capability to write your stories for you, but with llms we constantly have to reasses which parts of the thing we are building merrit our attention.

If llms get better, will you have to decide whether you actually care about writing decision trees, or if instead you just want to, more generally, curate procedural interactions (or something)?

My point is: if these next few years every project becomes an exercise in soul-searching for which parts of my work actually interest me, it is maybe less work not to use these tools, or alternatively, find something fullfilling that doesn’t involve making something.

> hollow destination.

I can say that in the last 2 years chatgpt/claude have added more code to my projects than me, and I am programming for 25 years (counting the rejected tokens as well).

When I use copilot/cursor it is so violent, it interrupts my thoughts, it makes me a computer that evaluates its code instead of thinking about how my code is going to interact with the rest of the system, how it evolves and how it is going to fail and so on.

Accept/Reject/Accept/Reject.. and in the end of the day, I look back, and there is nothing.

One day, it lagged a bit, and code did not come out, and I swear I didn't know what to type, as if it was not my code. On the next day I took time off work to just code without it. During that time I used it to write a st7796s spi driver and it did an amazing job, I just gave it 300 pages docs, and told it what api to make and it made amazing driver, I read it, and I used it, saved me half a day of work easily.

Life is what overcomes itself, as the poet said, I am not sure "destination programmers" exist. Or even if they do, I don't know what their "destination" means. If you want to get better, reflect on what you do and how you do it, and you will get better.

I wrote https://punkx.org/jackdoe/misery.html recently out of frustration, maybe you will resonate with it.

PS: there is no way we will be able to read llm's code in near future, it will easily generate millions of lines for you per day, so we will need to find am interface to debug it, a bit like Geordi from Star Trek. LLMs will be our lens into complexity.

Students of ancient languages fall into one of two camps: those who use translations for 'assistance' and those who don't. Classroom experiences have shown me that the two groups of students learn vastly different skills.

The group who struggle through texts by themselves with relying on any shortcuts -- they just sit with the text -- probably won't become top-shelf philologists, but when you give them a sentence they haven't seen before from an author they've read, the chances are very good that they'll be able to make sense of it without assistance. These students learn, in other words, how to read ancient languages.

The group who rely on translations learn to do precisely that: rely on a translation. If you give them a text by an author they've 'read' before and deny them use of side-by-side translation, they almost never had any clue how to proceed, even at the level of rudimentary parsing. Is that word the second-person-singular aorist imperative middle or is it the aorist infinitive active? They probably won't even know how to identify the difference -- or that there is one.

Our brains are built for energy conservation. They do what, and only what, we ask of them. Learning languages is hard. Reading a translation is easy. Given the choice betweem the harder skill and the easier, he brain will always learn the easier. The only way to learn the harder one is to remove the option: sit with the text; struggle.

So far I've been able to avoid LLMs and AI. I've written in other comments on HN about this. I don't want to talk to an anthropmorphic chat UI, which I call "meeting-based programming." I want to work with code. I want to become a more skillful SWE and better at working with programming languages, software, and systems. LLMs won't help me do this. All the time they save me -- all the time they steal from reading code, thinking about it, and consulting documentation -- is time they've stolen from the work I actually want to do. They'll make me worse at what I do and deprive me of the joy I find in it.

I've argued with teammates about this. They don't want to do the boring stuff. They say AI will do it for them. To me that's a Faustian bargain. Every time someone hands off the boring stuff to the machine, I'd wager they're weakening and giving up the parts of themselves that they'll need to call upon when they find something 'interesting' to work on (edit: and I'd wager that what they consider interesting will be debased over time as well, as programming effort itself becomes foreign and a less common practice.)

LLMs will usher in a programming future that looks nothing like today’s programming.

Today, we are shoveling the old way into LLMs.

In the future, programming will be optimized for LLMs and not humans.

Do you understand the assembly language that the compiler writes today? Do you inspect it? Do you Analyse it and not trust it? No, you ignore it.

That’s the future.

Languages written purely for LLMs have not yet been invented but they’re coming for sure.