Debt: The First 5,000 Years — David Graeber
The argument that money is a social relationship before it's a technical system. I read this while working on payment rails and it reframed the problem completely. The question stopped being "how do we make settlement faster?" and became "what kind of obligation is this, and who holds it?" For agentic payments, that question gets much harder. A social relationship requires a social actor. An AI agent doesn't carry obligations the way a person does — it carries instructions. The gap between those two things is where I think the real infrastructure problem lives.
Seeing Like a State — James C. Scott
About how large systems fail when they optimize for legibility at the expense of local knowledge. The canonical example is scientific forestry: German foresters in the 18th century redesigned entire forests to make timber accounting easier, and the forests started dying within a generation because everything that made them work — the undergrowth, the biodiversity, the messy parts — had been removed for tidiness.
I think about this constantly in Web3. Decentralization isn't just ideology. It's an epistemic claim: local knowledge aggregates better than central knowledge over a certain class of problems. The claim is sometimes right and often wrong, and Scott gives you the tools to tell the difference.
Skin in the Game — Nassim Taleb
The one question this book trains you to ask: who pays if this goes wrong? Taleb calls it the Bob Rubin trade — you collect steady fees for years, the tail risk eventually arrives, and someone else is holding the position. Crypto is full of this structure. So is AI safety discourse. So is most infrastructure design. I run every partnership decision through this filter now.
High Output Management — Andy Grove
The most honest book I've read about what it means to run systems with humans inside them. The core insight is deceptively simple: a manager's output is the output of their organisation, not their own individual work. Grove wrote this in 1983 and it still reads like it was written for the specific problem of keeping a small team coherent under pressure.
I stole most of my operational instincts from here. The SOPs, the structured weekly calls, the obsession with output metrics rather than activity metrics — all Grove, translated into a Web3 context where nothing is quite as defined.
The Dream Machine — M. Mitchell Waldrop
The biography of J.C.R. Licklider, the person who more than anyone else invented the idea of personal computing — not the hardware, but the vision. Licklider spent years funding research he knew he'd never see completed, in directions others thought were dead ends. The ARPANET exists partly because he believed computers should augment human intelligence rather than replace it, and spent a decade making sure the right people had money to explore that idea.
I read this when I was writing a lot of grant applications and feeling cynical about institutional funding. It helped.
The Sovereign Individual — Davidson & Rees-Mogg
Published in 1997, predicts a world where cryptography and digital networks erode the monopoly of nation-states on transaction and violence. Much of it aged badly. The parts that didn't are the parts that underpin a lot of Web3's foundational assumptions — whether the people building it have read this book or not.
I don't endorse the politics. I do think the mental model of sovereignty as a function of information asymmetry — rather than territory — is one of the more useful frameworks for understanding what's happening at the intersection of AI and crypto right now.
Paul Graham's essays — paulgraham.com
Less a book, more a long-running argument about what it means to make things. I've been reading these since before I knew what YC was. The essays I return to most are the ones about writing: that you don't know what you think until you've tried to write it down. That's the bar I try to hold my own writing to. I rarely reach it.