The inconvenient discovery.
I did not start this book as a fan. I came looking for the limits of a strategy that seemed to be winning everywhere. Christensen supplied the uncomfortable question: could the habits that made NVIDIA exceptional also obstruct its next transformation?
The inconvenient discovery was Jensen Huang himself. The internal rebel was already running the company. That did not end the questions. It changed what I wanted from them.
What began to affect me was not only what NVIDIA built, but how its example changed what I thought I could attempt. My personal analogy is Cobain, Sid Vicious and Keith Flint: not always my favourite musicians, but people who changed my sense of how a person could think and act. Influence on the how, not a request to copy the what.
The question that left the page.
I am not a coder or an AI engineer. I came from banking, compliance, digital identity and transactions. I had founded Arival Bank. An instruction reaching a system and an institution having authority to act on it were never the same thing.
Chapter 10 asks how one agent knows that another is authorised to act for a company. A narrower version became AuthorityClaw: a model reads invoices; a separate gate checks the live chain of authority, suppliers, bank changes and limits. The payment rail is simulated. Revoke the delegation above the agent, and the next ordinary invoice is rejected.
The agent cannot authorise itself.
AuthorityClaw was selected for First Place in NVIDIA’s Claw Agent Challenge: London. The award recognises the experiment, not the book’s forecasts. No real money moved. Project record ↗
The Coda returned through logic.
Professor Ross King appears in the book’s Coda. His influence here came through our conversations about formal logical approaches to AI — not through DNA computing.
After the challenge, a first-order model checked with Z3 exposed edge cases in the original implementation. They were replayed, fixed and retained as regression tests. The published checks have stated assumptions and bounds; they are not a proof of every line of code. Checks and limits ↗
The prize was not permission to stop checking. Chapter 15’s counterargument had found a job in the code.
A participant the book had missed.
The possibility that interests me now is not “non-engineers replacing engineers”. It is people with hard-earned knowledge of consequential problems gaining a more direct route from that knowledge to an executable test.
Not people with no expertise. People whose expertise previously had to stop at a document, a proposal or a request to somebody else’s engineering team. My case does not establish a market size. It gives me a reason to investigate.
The book needs the same discipline. Its original text stays intact; dated notes distinguish corrections from subsequent developments. A company has not failed to answer my question merely because it did not use my diagram.
I started by asking what NVIDIA might make possible next. Then I had to try something myself.