Michael Darius Eastwood · Independent AI alignment researcher
Published
The claim is descriptive first and interpretive second. Descriptive: across several fields that do not share a citation graph, the same structural pattern keeps appearing. A threshold ratio between generation and correction, or between drive and dissipation, sets whether a system stabilises or diverges. In quantum error correction, the ratio Lambda greater than one signals the below-threshold regime. In biological scaling, the recursion exponent determines whether metabolism keeps pace with mass. In neural dynamics, the ignition threshold in cortical activity plays the same role. In AI self-improvement, the exponent alpha decides whether a chain of iterations compounds or saturates. Interpretive: the programme reads this convergence as a common structural principle, not as a single physical mechanism. The interpretation is the programme's own; the underlying results are independent prior and concurrent work.
The convergence register is public, one row per event and one primary source per row. Its current headline count is 19 independently sourced convergences, with further candidates listed separately under source-pending status and excluded from the count. The register spans domains: Anthropic's alignment-faking paper on the AI safety side, Google's Willow chip (Nature, 2024, Lambda approximately 2.14 plus or minus 0.02 at 105 qubits) on the quantum-computing side, the COGITATE consciousness-theories collaboration (Nature 642, 133-142, April 2025) on the neuroscience side, and the Mitra group's discrete time crystal at NYU (Natsheh, Gambassi and Mitra, PRB 103, 224311, 2021) on the classical-physics side. The pattern is the joint appearance of a threshold-form stability condition. What the register does not do is claim causal influence in either direction; the entries are neutral about who read whom, and most of the external authors have almost certainly not seen the manuscript.
Descriptive convergence is not causal unification. Different fields can converge on similar structural patterns because the underlying mathematics of stability under recursion is generic, not because one framework generated all of them. The programme's interpretation is that a single-cause unification is a plausible hypothesis; the interpretation is not a claim that the external authors endorse. Their results are independent, and calling them independent is a compliment, not a hedge. The number to remember is 19, the source of every row is public, and the interpretive layer sits above that public dataset and does not pretend to be part of it.
The falsification contract asks a meta-analysis across at least four of the domains (quantum error correction, biological scaling, neural dynamics, AI self-improvement) to test two things. First, does the threshold-form condition (generation-to-correction ratio crossing one changes stability) hold in at least three of the four domains? Second, are the fitted effective recursion exponents statistically indistinguishable across those domains, within a plus-or-minus 0.3 window around the cross-domain mean? Both conditions passing confirms the strong reading. If the threshold-form holds but the exponents diverge, the form generalises but the magnitude does not, and the claim moves to partial. Fewer than three of four passing refutes the strong reading and the register returns to being an interesting curiosity rather than an interpretive lever.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.