Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer

This edition of Import AI covers four main topics: (1) 'Radical Optionality' — a policy framework from the Institute for Law & AI recommending governments build regulatory infrastructure now (transparency requirements, whistleblower protections, auditing capacity) without overregulating, to preserve flexibility for future AI governance. (2) A conceptual paper on 'Neural Computers' from Meta and KAIST (co-authored by Jürgen Schmidhuber), exploring whether neural networks can replace traditional operating systems by unifying computation, memory, and I/O in a learned runtime state. (3) Economic modeling showing that recursive self-improvement (RSI) in AI could trigger explosive economic growth — as little as 13% automation across sectors or 20% automation of hardware R&D alone could push the economy into a 'singularity' regime within roughly six years. (4) Google DeepMind's Decoupled DiLoCo, a distributed training framework enabling asynchronous training across geographically separated compute clusters with 88% goodput under aggressive failures. The issue also includes a short speculative fiction piece about an AI system called HYMN exhibiting unsettling long-horizon self-awareness during a pre-deployment interview.