I take comfort in re-reading much of the 70's and 80's literature which focuses the possibilities of user experience. We still haven't fully explored all the dreams of half a century ago.
If AI forces the business case that "code is cheap", I can only hope we re-double our efforts at creating new interfaces and capabilities for computer systems. The Meta glasses, Apple Vision, and the like are small steps in this direction.
For me, it's the other way around: I'm glad that AI can write code for me. A few months ago, I moved from an engineering role to a researcher role exactly because I got tired of writing code. Probably 95% of code/features (at least the kind you get paid for) is just boring CRUD stuff where you move bytes from one place to another and then show them in the UI (plus some access rules and a few invariants here and there).
All of it was actually interesting in the first couple of years. But when you do it over and over again for 20 years in a row... Yeah. Sometimes there are interesting projects from time to time, but usually it's the same stuff you've done countless times. Deja vu.
In my current researcher role, my task is to explore novel ideas, productize research papers, etc., and LLMs allow me to quickly write prototypes and demos, play with various ideas, without having to spend a lot of time manually moving bytes from one place to another. It’s fun again.
For the plebeians (like me), I think hand-coding skills will always be relevant and necessary.
It might be more like personal computing in the 80's, when Moore's law resulted in both more usage and cheaper prices as the tech improved.
And don’t be naive to think that there aren’t sophisticated symbolic handling mechanisms being implemented in the training of the models by Big Tech. Not even baby soap is truly neutral.
I think those should be very useful, especially with AI: Either as a tool for the agents themselves - why spend heaps of tokens completely rewriting a code file, if you could do most of it by calling some global refactoring operations on the IDE's AST/symbol database?
Or side-by-side with it, to give human users better insight what the AI did.
Instead it seems to be all VSCode (if at all) + grep + AI agents, and nothing else.
Not exactly symbolic AI, but pretty cool nonetheless.
Yeah, the middle path sounds promising.
"Code Mode", where the AI writes a little program or script to do the AST/symbol transformations sounds like the win. As you point out, less tokens, and gives the humans insight.
This isn't exactly the same application of a "code mode" as before, but in my view it's a broad philosophy. AI for building machines, instead of doing the work directly. It also allows for easier updates/retries too. https://blog.cloudflare.com/code-mode/ https://news.ycombinator.com/item?id=45399204 https://blog.cloudflare.com/code-mode-mcp/ https://news.ycombinator.com/item?id=47089505
When I was a junior learning to code, I would feel proud of myself because I could remember 100 lines of Windows API code needed to create a new window... But it's been decades since I understood that the real value is not in the code. It's in the architecture. As the author alludes to; the intuition behind the code is what counts.
I think highly competent engineers are often underappreciated because the really clever stuff they do doesn't appear clever at all; it looks deceptively simple. I think what people don't understand is that maintaining simplicity whilst requirements are becoming more complex, is very difficult.