Michael Darius Eastwood · Independent AI alignment researcher
Published
April 2026 lands two convergence entries at once. Row 16 of the register is a venture funding round; row 18 is a peer-reviewed mathematics-and-cognition paper. Different institutions, different fields, same month, same direction of travel as the December 2024 thesis. This article explains both, in the plain language the register uses.
A four-month-old London company raised half a billion dollars at a four billion dollar valuation to build the exact thing the December 2024 manuscript named as the failure class: closed-loop recursive self-improvement. The register credits this as a convergent naming, and only as a convergent naming. It is not a claim that the founders knew of the manuscript. It is not a claim that they named their company on the strength of anything Eastwood wrote. It is the observation that the name at the centre of a hostile funding round is the name at the centre of an eighteen-month-old private thesis. The programme's rule on causal attribution is absolute: no such claim is made, and the truth gate blocks it if slipped in.
Because it is the money-following-thesis event. The manuscript argued in December 2024 that recursive self-improvement was where control failure would concentrate. Sixteen months later, the venture capital market priced that concentration at four billion dollars pre-money for a company with no published safety architecture. That is not proof of the manuscript's argument. It is evidence that the argument's target is now a funded market, which changes the incentives around every other row in the register.
The PNAS Nexus paper argues, formally, that alignment is undecidable in the classical sense and that the most tractable contingent response is cognitive diversity across the aligned agents. The December 2024 manuscript, sixteen months earlier, argued informally that control from outside would fail and that the plausible response was a distributed coalition of governments, scientists and religious leaders across whose diverse commitments the failure modes would not correlate. The mathematical content is not the same. The direction of travel is.
The two rows share nothing except the month, and neither cites the other. That is the pattern the register measures: independent starting points, external dates, and a direction of travel that runs the same way. Priority on the PNAS Nexus paper belongs to the authors named on it. Priority on Recursive Superintelligence's funding round belongs to the company and its investors. What the register claims is not priority; it is that a hostile reader can walk row by row through 19 such events, check the sources, and conclude that the direction of travel is a fact about the world rather than an artifact of confirmation bias.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.