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

Researchers with the Institute for Law & AI have published work on "radical optionality," an approach in which governments give themselves tools they may need if powerful AI starts to massively disrupt the world. The authors write that radical optionality is about preserving democratic governments' ability to make good decisions about governing transformative AI as circumstances evolve, which in the short term means avoiding overregulation while rapidly building institutions, information channels and legal authorities. They argue governments should be willing to spend an extraordinary amount of money, effort and political capital on preserving optionality, and should be wary of counterproductive interventions but not much concerned with the pecuniary cost of measures likely to have net-positive results.

Specific recommendations include transparency and reporting requirements, followed by an auditing regime; whistleblower protections for frontier lab employees; information-sharing within and between governments, including for supply chains deemed critical to AI development; flexible rules and definitions such as conditional "if-then" commitments; assessments and evaluations capacity; improved security of model weights and algorithmic secrets via voluntary physical and cybersecurity standards; and hiring and talent investment, including increased funding of AISI (UK) and CAISI (US). The authors address counterarguments, saying the ideas aren't weighty substantive authorities that lend themselves to abuse, and that they are basically convinced there is significant risk of governments asserting control over AI development, which is why they don't recommend massively expanding emergency authorities such as the Defense Production Act.

Separately, a paper titled Neural Computers from Meta and KAIST asks whether a neural network can act as a traditional computer, describing a neural system that unifies computation, memory and I/O in a learned runtime state. Prototypes based on a command-line interface and a graphical user interface were built using the generative video model Wan 2.1, and the authors say current neural computers can already learn elementary runtime primitives, notably I/O alignment and short-horizon control.

Economists from Forethought, Columbia University and the University of Virginia developed a framework analyzing how AI-driven automation interacts with feedback loops. They report that 13% automation across all sectors is sufficient to push the economy into an explosive regime, and 17% suffices when only software and hardware research are automated; 20% automation of hardware alone is enough to cross the threshold. In a baseline stylized simulation, full automation of software R&D plus 5% automation elsewhere brings the singularity in roughly six years.