We’ve stopped hardcoding rules inside AI models, but it’s still easy to trap ourselves in schemas and workarounds.
Rich Sutton’s The Bitter Lesson is one of my favorite essays on AI - the main insight from 70+ years of research is that general methods that scale with compute beat clever, human-designed knowledge.
I keep relearning it the hard way on my last few projects - Playground, Health, and more.
In the past, we used lots of specific health classifiers for each insight you wanted, surrounded by tons of scaffolding. Lately, the direction of travel is using large, multi-modal models to get better insights.
It’s even true for managing agent context and memory, where we went from designing complex graph databases, rigid schemas and rich taxonomies to a lot simpler structure with a fraction of complexity.
With Playground, we realized the same thing - we spent tremendous effort on the scaffolding and archetypes, and it all got thrown away when the new model landed.
We’ve seen this across the industry - including Boris Cherny, who created Claude Code and talks about scaffolding as temporary duct tape until the model catches up.
The lesson is bitter because the hand-built stuff is the part we’re proudest of.