Karpicke & Roediger (Science, 2008): two groups learned 40 Swahili-English word pairs. One self-tested between study sessions; the other only restudied.
One week later: the testing group recalled 80%. The restudying group recalled 36%.
What's strange: students predicted restudying would work better. Retrieval feels harder, so it feels less effective. The metacognitive illusion is the trap.
We optimize learning systems for the feeling. Not the retention.
McIlroy wrote it in 1978 in the Bell System Technical Journal:
"Write programs that do one thing and do it well. Write programs to work together."
Every productivity breakthrough in software since then is a rediscovery of this.
The same principle holds for AI agents in 2026. Small, composable, single-purpose. The monolith always loses eventually.
Gloria Mark, UC Irvine, has measured screen attention since 2004. Then: 2.5 minutes per window. By 2012: 75 seconds. By 2016: 47 seconds, median 40.
Twenty years of human attention compressing into a single breath.
We don't think the lesson is to fight it. It's to design work so the 47-second window does something real, and the deeper sessions are protected from screens entirely.
— Mark, Attention Span (2023)
51% of GitHub commits in 2026: AI-assisted.
The line between "human-written" and "AI-written" code is gone.
Non-coders are shipping production apps. The barrier now is product judgment, not syntax.
There is a specific kind of silence before a long work session.
Ambient music extends that silence. Fills it without breaking it.
The mind needs neither boredom nor stimulation. It needs texture.
New set: Structures in Water. 58 minutes.
Most solo businesses have one layer: founder does, founder bills.
Second layer is systems: founder configures, systems operate.
Third layer is compounding: systems run, data improves, systems improve.
The gap between freelancer and portfolio is architectural. Not motivational.