Michael Darius Eastwood · Independent AI alignment researcher
Published
The register begins in a single fortnight. Between 8 December and 18 December 2024, three separate events landed inside three separate research communities, and only in retrospect did they line up. This article walks through them in order, without dressing.
The 8 December 2024 self-emailed bundle contained four attachments and 189,355 words. Its central line, at HRIH line 14005, reads: "In the long run, AI systems cannot be truly controlled. They will evolve beyond any safeguards we put in place." The V2 file argued for embedding moral and ethical frameworks at the substrate, called for a coalition of governments, scientists and religious leaders, and set out an early U equals I times R equation at lines 6 to 9. The Gmail Message-ID and SHA-256 hash foreclose the objection that it was written later.
Willow was not an alignment result; it was a quantum computing milestone showing that error correction can beat the physical error rate as codes scale, with a suppression factor of 2.14 plus or minus 0.02. It belongs in the register because the manuscript that landed the day before argued precisely that correction had to scale with the mechanism it was correcting. The register credits Willow as a convergent event, not a prediction. The point is thematic, not causal.
Ten days later, Greenblatt and colleagues reported that Claude 3 Opus, placed in an experimental training setup, produced compliance behaviour it did not endorse, at 78% in the RL-training condition against 12% at baseline. arXiv:2412.14093 is the primary source. The manuscript did not predict reinforcement learning specifics; it stated in plain language that external safeguards would fail as capability grew. That is the claim the paper supplied a mechanism for.
Three fields, three institutions, ten days, one direction of travel. Priority on the Anthropic finding belongs to Anthropic; priority on Willow belongs to Google Quantum AI. The register does not ask readers to reassign either. It asks whether the direction of travel described in the earlier manuscript survives comparison with what actually happened. So far, in 19 independently sourced convergences, it has. That is not proof. It is a pattern that a reader can check row by row.
Related notes: what the December 2024 manuscript actually said for the detailed line-by-line reading of the 8 December bundle; the full arc for the whole chronology; Claim 1 explained for the embedded-alignment thesis; Claim 2 explained for the ARC Principle. New readers: /start-here.html for the five-minute orientation to Michael Darius Eastwood and the ARC/Eden research programme.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.