Writing.

Philip Anderson's 1972 paper is the most important paper on emergence in 20th-century physics. Its central claim — that complex systems cannot be understood by reducing them to their components — applies with full force to the AI systems we are building right now.

Nuclear chain reactions don't scale gradually — they have a critical point. Add one neutron too many and you go from subcritical to supercritical. Agent systems have the same structure. Most teams discover this in production.

No ant understands architecture. The colony does. This is not a metaphor for multi-agent AI — it's a precise description of a phenomenon we are actively building and don't fully understand.

After four years building engineering teams at an AI-first company, I have a clear answer to the question I wish someone had given me in 2021: here's exactly how I'd structure the first 15 people.

In quantum mechanics, observation collapses the wave function. In LLMs, evaluation changes the behavior being evaluated. This isn't philosophy — it's an engineering problem with concrete production consequences.

In five years, the recruiting technology stack has gone from system of record to intelligent pipeline. Here's the full arc — what changed, what it meant architecturally, and where it's going next.

Physics has known for 60 years that symmetric equations can produce asymmetric outcomes. LLM bias isn't a data problem or an alignment problem — it's a symmetry breaking problem. Here's why that distinction matters.

Grokking isn't a curiosity — it's evidence that neural network learning is phase-transition-shaped. Here's what the physics tells us, and why it should change how we think about training, evaluation, and trust.