The Human Energy required for AI Coding
Building things with AI takes a lot of human energy.
https://steve-yegge.medium.com/the-ai-vampire-eda6e4f07163
Beautifully put, and very true.
I feel like I've been burning the candle at both ends the last few months, building things with AI and trying to learn a new set of skills to operate in the post-November 2025 (post useful AI models like Opus 4.5/GPT 5.4) world. It's intoxicating and exciting, and very valuable, but it exposes the limits of my attention and stamina.
I'm trying my best to stay focussed, calm and deliver at a higher but still maintainable pace with all these new tools.
Throughout my career as a developer I had to pivot and build skills for new platforms as technology evolved and the products I built adapted to where customers were running it:
1990s - Desktop/CD Mac: HyperCard/C
2000s - Desktop/DVD Windows: Visual Basic/C++/Custom DB
2005+s - Web: C#/Flash/SQL Server
2010s - Web/Cloud: C#/Angular/SQL Server
2015s - Web/Cloud: Java/Angular/MongoDB/Cloud/Python
2020s - Web/Cloud/Mobile/Serverless: Java/Angular/MongoDB/Swift/Java/Cloud/Python/Serverless/TypeScript
2026: - Web/Mobile/Anything - AI -> any language+stack
Each time I had to change significantly involved a period of hard study and intense practice to become proficient. As time went on the number of technologies I worked on expanded, and my depth became less as my breadth increased.
I think with any language, stack and platform there are nuances to how to build reliable, repeatable high quality software.
As the needs of people using the software change, what is valuable and necessary changes too.
In this new AI world we also need to apply experience and nuance to what we build. We may have the capacity to build 100x the code, but we should have the discernment to know what is useful, what will work well, what people want and what we can sustain.
One of my favourite things to do as a developer is to delete 'dead' (no longer used) code. It is an important maintenance activity to reduce the code being compiled, prevent unnecessary work in the future and keep the product lean and working efficiently.
Maybe the skills we need now are less about the language with which to build what we build, but the structure, quality, value and reliability to demand and validate from the AI.
The tools (Codex, Claude Code) are impressive, and if used right can produce incredible, high quality software very fast.
We need to have a balance and choose a sustainable pace despite our increased capacity to build more and faster. Importantly, we need also to know what not to build.
Just because we can, doesn't mean we should.
