I have a slightly different theory for AI fatigue. Programming essentially comprises three aspects: problem-solving, coding up the solution, and seeing the solution work. We used to have a balance between them that AI agents are disrupting.
Consider that each aspect has different psychological impacts. The first aspect is the hardest, more intense one.
The second one, translating the ideas in your head into code -- where "flow state" typically comes into play -- is mostly a mechanical, somewhat repetitive task, and hence actually meditative. It is relaxing, almost "cozy", in a counter-intuitive way.
The third part, seeing the code work at the end, is a nice little dopamine reward.
Originally we used to shift between these aspects at a much more balanced cadence: we typically spent a lot of time on the stressful problem-solving part, and even longer on the meditative coding stage. The dopamine hits were spread out across longer stretches of time. (Note that this pattern is fractal, because not only is "high-level algorithm design / architecture -> coding" one cycle, the coding stage itself has similar smaller-scale sub-cycles, e.g. "encountering syntax errors / implementation bugs -> fixing them.")
With LLMs this balance has been disrupted. Now that agents are doing most of the coding, we are spending almost all our time in the stressful problem-solving mode, and then directly jumping to the dopamine hits when the agent delivers code that mostly just works. If you're not YOLOing it and do code reviews, those are also intense, as TFA points out. We get almost none of the meditative coding activity in between.
So it's not surprising that oscillating constantly between two intense modes is going to be much more draining. Vibe-coding may actually be the coding equivalent of doom-scrolling.
I think a lot of recent talk about missing the old way of coding is rooted in the loss of that meditative aspect.
Personally I've noticed I prefer a pair-programming approach with AI, which still lets me spend a lot of my time hands-on coding. It may be because this maintains some of that balance and is more sustainable in the long run. However the allure of productivity from wrangling multiple agents is also very, very seductive...
Consider that each aspect has different psychological impacts. The first aspect is the hardest, more intense one.
The second one, translating the ideas in your head into code -- where "flow state" typically comes into play -- is mostly a mechanical, somewhat repetitive task, and hence actually meditative. It is relaxing, almost "cozy", in a counter-intuitive way.
The third part, seeing the code work at the end, is a nice little dopamine reward.
Originally we used to shift between these aspects at a much more balanced cadence: we typically spent a lot of time on the stressful problem-solving part, and even longer on the meditative coding stage. The dopamine hits were spread out across longer stretches of time. (Note that this pattern is fractal, because not only is "high-level algorithm design / architecture -> coding" one cycle, the coding stage itself has similar smaller-scale sub-cycles, e.g. "encountering syntax errors / implementation bugs -> fixing them.")
With LLMs this balance has been disrupted. Now that agents are doing most of the coding, we are spending almost all our time in the stressful problem-solving mode, and then directly jumping to the dopamine hits when the agent delivers code that mostly just works. If you're not YOLOing it and do code reviews, those are also intense, as TFA points out. We get almost none of the meditative coding activity in between.
So it's not surprising that oscillating constantly between two intense modes is going to be much more draining. Vibe-coding may actually be the coding equivalent of doom-scrolling.
I think a lot of recent talk about missing the old way of coding is rooted in the loss of that meditative aspect.
Personally I've noticed I prefer a pair-programming approach with AI, which still lets me spend a lot of my time hands-on coding. It may be because this maintains some of that balance and is more sustainable in the long run. However the allure of productivity from wrangling multiple agents is also very, very seductive...