Books That Shaped How I Think

• Why incumbents fail at disruptive technology — and it's not because they're bad at innovation. They're optimized for their best customers, which makes them blind to the low end.
• Directly maps to why large health systems struggle with AI adoption. They're not ignoring AI — they're rationally optimizing for existing workflows and customers.
• The 'jobs to be done' framing changed how I think about AI use case selection. What job is the user actually hiring this model to do?

Required reading for anyone wondering why enterprise AI POCs die in large organizations.

• Most decisions are bets under uncertainty, not equations with right answers. Good decisions can lead to bad outcomes. Bad decisions can lead to good outcomes.
• Directly applicable to AI product decisions — when do you have enough data to ship? When do you keep iterating? The answer is rarely 'when you're certain.'
• The 'resulting' trap: judging decision quality by outcomes rather than process. Enterprise AI teams fall into this constantly when a model failure gets blamed on the PM's go/no-go decision.

How I think about AI product decisions when I don't have enough data (which is always).

• The 'peacetime CEO vs. wartime CEO' mental model is one of the most useful in my toolkit. AI product leadership requires switching between these modes constantly.
• Nobody prepares you for the moments when everything is on fire. The value of this book is the honesty — Horowitz doesn't pretend there are clean frameworks for hard situations.
• Hiring and people decisions compound more than technical decisions. The AI team that wins is the one that gets the people equation right.

What I reread when things get hard at work.

• The most important question: is this a vitamin or a painkiller? Most enterprise AI is vitamins. The products that win are painkillers.
• Competition is for losers. Monopoly is the goal. I use this lens to evaluate whether an AI use case has real defensibility or just first-mover advantage that evaporates.
• Secrets: every great company is built on a secret others don't see yet. What's the secret in your AI product?

A useful counterweight to conventional enterprise thinking about AI adoption.

• The long view on AI trajectory. Useful not because the timelines are right but because it forces you to think about capability overhang and the pace of improvement.
• Reading this makes the hype cycles feel small. Calibrates your perspective on what current LLMs can and can't do.
• The control problem is real even at narrow AI scale. The Life Sciences & Life Sciences & Healthcare AI products I build have to answer 'who is responsible when this is wrong?' — and that question gets harder as capability increases.

For calibrating long-term thinking about where AI is going.

• The best book I've read for understanding cancer biology at a level that makes clinical AI actually make sense. Not a textbook — a biography of a disease.
• The history of clinical trial design in oncology is a masterclass in how hard it is to establish causation in biological systems. Directly relevant to AI validation methodology.
• Understanding the clinical context your AI operates in is non-negotiable. This book gave me the foundation to have real conversations with oncologists.

Essential reading for anyone building AI in oncology or clinical research.

• OKRs done right are a forcing function for clarity. Most teams skip the hard work of defining what 'good' looks like before building.
• The AI product metrics problem: teams default to model metrics (accuracy, F1) instead of outcome metrics (clinical decisions changed, time saved, errors prevented). OKRs force you upstream.
• Stretch goals have a specific role — they're not performance targets, they're direction-setters. The difference matters when you're building AI products where the ceiling is unknown.

How I think about AI product metrics beyond accuracy.

• Build-measure-learn is the right loop for AI POC design. Most enterprise teams skip the 'measure' step and go straight from 'build' to 'scale.'
• The pivot vs. persevere decision is the hardest call in AI product development. Most teams pivot too late because sunk cost bias is powerful.
• Validated learning, not output, is the unit of progress. Shipping 10 features doesn't matter if you didn't learn whether they moved the metric.

The framework I use to structure AI POCs and evaluation phases.

• Systems beat goals. Consistent weekly publishing compounds more than sporadic bursts of effort. I built my content operation around this principle.
• The 1% better every day compounding argument is real — I've seen it in model improvement curves, in team capability growth, in personal output quality.
• Identity-based habits: 'I am a person who writes' vs. 'I want to write more.' The framing shift is small but the behavioral difference is large.

How I think about building sustainable habits around content creation and continuous learning.

• System 1 vs. System 2 thinking is the most useful mental model for understanding AI adoption. Users default to System 1 — pattern matching, fast, low effort. If your AI output requires System 2 processing, adoption will be lower than you expect.
• Cognitive biases are features, not bugs, in human cognition. Designing AI that works with human biases rather than against them is a product design skill.
• The WYSIATI heuristic (What You See Is All There Is) explains why LLM hallucinations are so dangerous — models, like humans, construct plausible stories from available evidence without flagging what they don't know.

The most direct connection between cognitive psychology and AI product design.