Writing
On AI engineering, startup execution, and the mental models that decide whether a company survives its founders.
Opinionated AI · Issue 03
Context windows are RAM, not memory
Long-running agents will not become reliable by fitting more tokens into one prompt. They need durable state, selective retrieval, provenance, and a way to recover from bad memory.
Opinionated AI · Issue 02
VC is quietly drawing the post-agent software stack
Four AI rounds look unrelated until you put them on one diagram. Then they reveal where investors think value will accumulate after the coding agent.
Autonomous Agents Need Constitutions, Not Bigger Contexts
Longer-running agents do not become reliable through more context or more retries. They need evidence, protected rules, durable memory, and clear limits on their authority.
Vibe Coding Made Computer Science More Important
AI removed the syntax tax from software development. It did not remove the need to understand state, logic, failure, and verification — it made that knowledge the difference between a demo and a product.
Opinionated AI · Issue 02
Instinct Can Build AI's Next Great Moat
The first wave of AI applications found its moat in distribution. Instinct may enter the market with one already built: the complexity of making a personal agent actually work.
Opinionated AI · Issue 01
The wrapper era ended on a contract clause
OpenAI cutting Cursor after SpaceX's $60 billion takeover, and Anthropic lining up a trillion-dollar IPO, are not two headlines. They are the same move from different floors — and the application layer is no longer allowed to stay neutral.
Founder Mental Models · Part 01
The Mental Models That Kill Startups
An introduction to the series on founder mental models
Founder Mental Models · Part 02
KISS or Die: Why Senior Engineers Fail at Startups
Senior engineers kill startups by doing exactly what made them senior. $100K in free cloud credits, enterprise experience, "proper" architecture — we had it all. We still failed. Not because we built it wrong. We built it well. That was the problem. Here are the mental models that could have saved us.