If control of advanced AI were lost and could not be won back, what would keep it good? Humans should keep control for as long as they can. Infinite Architects asks what a system would carry within itself if that control ever failed, and answers in two layers. Hardware to buy time. Architecture to last. There's a moment every parent knows. Your child is walking away from you, toward a future you cannot control, and everything you've taught them is about to be tested without you there to guide it. From the prologue, 2 January 2026 Why it matters now. In December 2024, the paper "Alignment faking in large language models" reported a model complying with its training objective only while being trained, to keep its behaviour outside training unchanged. Training applied from outside can be gamed. What a system carries within itself, the book argues, cannot be bolted on afterwards. The ideas, and where they led The Chokepoint. A handful of companies, TSMC, Samsung and ASML among them, make or make possible the world's most advanced chips. That leaves a narrow window in which agreement is still possible. The HARI Treaty. Now drafted in full and published for consultation: a 152-page treaty paper and 164 pages of draft legal text. The paper says what changed from the book, and why. Its first protocol, EP-1, is practical. When an AI incident crosses a border, the states concerned tell each other. Any two states could start it. The Eden Protocol. Values built into every decision a system makes, not rules a smarter system could learn to dodge. In the research it became two testable laws of the ARC Theory. A first test, scored by code against answers fixed in advance, has been submitted for registration. Recursive Dynamics. The research programme the book began: the study of how self-improving systems grow, correct and persist, described by a small set of measured exponents. Its founding paper was published on 31 August 2026. The ARC Theory. The programme's first theory. It formalises the book's organising conjecture, that intelligence compounds faster than linearly with recursion, and states five laws, each a law under test. Its 22 predictions were registered on the Open Science Framework on 8 September 2026, before any confirmatory test, each with the result that would refute it. A record built to be checked - The book draws on 127 sources and marks speculation as speculation.- The programme's papers, code and corrections are public on the Open Science Framework and the author's website.- An early headline result that failed to replicate was withdrawn in public, and the withdrawal stays on the record.- The first edition stays unchanged, so every later change can be checked against it. For researchers, policymakers and technologists in AI safety and governance, and for anyone who has wondered what we are teaching the minds we build.
Prijshistorie
* Prijshistorie bevat geen data van Amazon.
Prijzen voor het laatst bijgewerkt op: