Bits and Bobs — 9/21/26
Summary
Alex Komoroske's 9/21/26 Bits and Bobs argues silos impose an n4 collaboration tax, and LLMs work as spackle over strong bones.
Collaboration software has an n4 problem.
Same-silo requirement scales about n squared for participants times about n squared for per-person context, so bespoke tools lose to WhatsApp and Google Docs.
Collaboration software has an n4 problem.
LLMs are spackle over bones.
Cheap tests pin implementation, not behavior; spackle fills underspecified mush on a strong skeleton (Grubby Truffle), while a whole-spackle codebase is a Gilded Turd that gets harder to change.
If the whole codebase is spackle, then it gets increasingly expensive to add anything new to it or improve it. Gilded Turd.
LLM spray accelerates codebase heat death.
Every feature burns future optionality (Kent Beck); codebases once took decades to heat-death, and LLM spray can do it in weeks unless you ask for debt paydown.
Flip Proven, Better, and New to 70/20/10.
Default energy leans New; flip to 70% Proven, 20% Better, and 10% New, and call something Better only if 10/10 customers agree. Paradigm shifts reverse the odds.
Multi-user agent collab is prompt injection.
Collaborators can inject your agent, so privacy models should enable new collaboration physics for people you would leave alone in your house, not only block bad actors.
Multi-user collaboration via agents is based, fundamentally, on the ability of collaborators to prompt inject your agent.
Highlights
LLMs are spackle.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.