Wensen
Vincent
Wu

Healthcare, aviation, finance, government — the industries where software matters most are decades behind consumer tech. It's not just bureaucracy. Safety certification takes years, so vendors ship slowly. Procurement favors incumbents, so users can't choose alternatives. Risk aversion punishes new approaches more than it punishes bad ones. Each of these is rational individually. Together they create a system where the places where bad software is most dangerous are exactly the places most likely to have it.

Most AI discourse is about chatbots and content generation. The harder question is what happens when models are reliable enough to make decisions in environments where mistakes kill people — air traffic control, surgical systems, nuclear monitoring. The regulatory frameworks for this don't exist yet. The engineering standards don't exist yet. Whoever figures this out first will define how an entire generation of critical infrastructure works.

React was a Facebook internal tool. Rails was extracted from Basecamp. Most of the open source tools I've built that people actually use started as weekend experiments with no business plan. There's something about the absence of market pressure that produces better tools — maybe because the builder is the user, and there's no incentive to compromise on taste.

Small teams with strong taste make beautiful products. But at some point, every company crosses a threshold where decisions get made by committees, design reviews become political, and the product starts to feel like it was built by consensus. Apple seems to have delayed this longer than most. Teenage Engineering has avoided it entirely, maybe by staying small. Is taste fundamentally a small-team property, or is there a way to preserve it at scale?

Every note app, wiki, and knowledge base assumes you'll organize things yourself. But the real problem isn't storage — it's retrieval. How do you query a lifetime of context? Embedding search gets you similarity, not reasoning. RAG gives you fragments, not understanding. The interesting question is whether there's a system that can hold everything you've ever read, written, or thought — and surface the right piece at the right moment without you asking. That's not search. That's something closer to memory.

Rams had ten principles for good design, mostly about physical products. But software has properties that physical objects don't — it can update, it has state, it exists in time. The design discourse around software is still mostly about surfaces: typography, spacing, color. Almost nobody talks about the design of systems themselves — the APIs, the data models, the way errors propagate. I think there's a missing discipline here.