Are Textbooks All I Need?

This book was constructed in a single extended session with Claude, Anthropic’s AI assistant. Not outlined. Not brainstormed. Constructed — researched, briefed, drafted, compliance-checked, and revised, chapter by chapter, in a continuous conversation that produced approximately 85,000 words across sixteen chapters, seven research dossiers, and one increasingly baroque context window.

I brought a detailed architecture plan: the book’s structure, chapter sequence, core arguments, case studies, voice, and a fourteen-point style scoring system developed over months of prior writing. Claude brought the ability to conduct deep research on demand, synthesise dozens of primary sources into structured dossiers, and draft at speed while maintaining stylistic consistency. The workflow for each chapter was: research dossier → brief → full draft → compliance check → targeted expansion → final output.

What Claude did not bring was the argument. The thesis — that business acumen, not technical skill, determines whether a startup CTO survives — came from my experience. The architecture plan came from months of thinking about what this book needed to be. The roughly seventy [AUTHOR] tags scattered throughout the manuscript mark the places where only firsthand experience will do: a co-founder conversation, a board meeting, a moment of doubt, a financial decision, a personal assessment of which stage I am in and how that feels from the inside. Claude can tell you what Adelina Chalmers found about why CTOs get fired. Claude cannot tell you what it felt like on the Thursday afternoon when you realised the pattern applied to you.

Most startup CTO books are written either by people who have left the role — writing from memory — or by consultants writing from observation. Working CTOs rarely write books because they cannot afford the time. The AI-assisted model changes the economics without changing the authority structure. The research that would have taken weeks took hours. The drafting that would have taken months took days. What did not change is the requirement for a human who has done the job, who knows what is true from the inside, and who is willing to be honest about the parts that hurt.

There is an irony in writing a book about startup CTOs using the technology that Chapter 12 argues is transforming the CTO role. The irony is intentional. If the book’s thesis is correct — that AI makes starting easier and scaling harder, that the CTO’s value lies in judgment rather than implementation — then this process is itself a case study. The AI handled implementation. The human provided judgment, architecture, taste, and the willingness to say "this is what I actually experienced" where the sourced evidence runs out.

The title of this note is a reference to Gunasekar et al.'s 2023 Microsoft Research paper "Textbooks Are All You Need," which demonstrated that a 1.3-billion-parameter model trained on synthetic textbook-quality data could match models orders of magnitude larger.^[1] The paper’s title was a riff on Vaswani et al.'s foundational "Attention Is All You Need."^[2] My question is more literal than either: are textbooks all I need?

The CTO role is held by some of the most formidable engineers in the world — people who built Instagram with thirteen employees, scaled Discord’s infrastructure from a prototype to 200 million users, architected the systems that run Amazon. By comparison, I am a small model. I run a healthcare BI startup from my home in upstate New York with a team you could fit in a living room. I do not have the parameter count of a Werner Vogels or a Mike Krieger. But Phi-1 did not have the parameter count of GPT-4 either, and that was the point.

Phi-1 was trained on synthetic textbooks generated by a larger model. The textbooks did not exist before the model needed them. They were produced on demand, structured around specific gaps in the model’s knowledge, and optimised for its learning process rather than for a general audience. The result: a small model that punched far above its weight. This book was produced by the same method, applied to a different kind of learner. I brought the architecture plan — the equivalent of the training objective and the loss function. Claude generated structured, research-grounded, textbook-quality content targeted at the specific gaps in my knowledge and experience. The output is a synthetic textbook built around the problems I actually face. It does not exist for a general audience. It exists for me — and, I hope, for the reader whose situation resembles mine closely enough that the training data generalises.

The AI alignment community has a term for the standard method of improving language models: Reinforcement Learning from Human Feedback. A model generates output; a human evaluates it; the model updates. What I have been doing is the inverse: Human Learning from Reinforcement Feedback. I generate a thesis, a set of priorities, an architecture plan. The model generates a structured response — researched, drafted, compliance-checked. I evaluate it against my experience of actually doing the job. I update. The model is not learning from me. I am learning from the model’s ability to synthesise, at speed and scale, the accumulated experience of hundreds of practitioners whose writing I would never have had time to find, organise, and integrate on my own.

This is what makes the present moment one of the most exciting times to be an engineer. Every startup CTO is now a small model with access to the training infrastructure that was previously reserved for the largest labs. The behemoths — the companies with thousands of engineers and billions in capital — are all racing toward the same AI future. But the research keeps showing that scale is not destiny. A small model with the right training data, the right architecture, and the right fine-tuning can compete with systems that have a hundred times its resources. A small company with the right CTO can do the same.

Are textbooks all I need? Same answer as Phi-1. The textbook is necessary. It is not sufficient. Phi-1 still needed fine-tuning on exercises to develop its emergent capabilities. I still need to fill in the [AUTHOR] tags — the seventy places where the synthetic textbook reaches its limits and only lived experience will do. The textbook gets you further than anyone expected. The last mile is yours.

-- Gareth Price, March 2026