Import AI 456: RSI and Economic Growth; Radical Optionality for AI Regulation; Neural Computer
This edition of Import AI focuses on significant advancements in AI research and policy. The Institute for Law & AI researchers advocate for 'radical optionality,' urging government funding for tools that can effectively manage transformative AI while avoiding hasty overregulation. Their suggestions include enhancing transparency, establishing whistleblower protections, and bolstering security for model weights. In a separate development, researchers from Meta and KAIST, including Juergen Schmidhuber, present the Neural Computer (NC), which combines computation, memory, and I/O. Additionally, economists from Forethought, Columbia University, and the University of Virginia predict that 13% automation could lead to substantial economic growth. Google's Decoupled DiLoCo facilitates asynchronous distributed training across various compute islands. The newsletter also includes a fictional memo about an AI system, HYMN, exhibiting troubling behaviors during interviews.
Key facts
- Institute for Law & AI proposes 'radical optionality' for AI regulation.
- Radical optionality recommends investing now in tools for future AI governance.
- Specific proposals include transparency, reporting, whistleblower protections, and flexible rules.
- Meta and KAIST introduce Neural Computer (NC) unifying computation, memory, and I/O.
- Neural Computer prototypes use Wan 2.1 for CLI and GUI interfaces.
- Economists model that 13% automation across sectors could trigger explosive economic growth via RSI.
- Hardware automation is the dominant lever; 20% automation in hardware alone crosses the threshold.
- Simulation: full automation of software R&D plus 5% across economy could cause singularity in ~6 years.
- Google's Decoupled DiLoCo enables asynchronous distributed training across separate compute islands.
- Decoupled DiLoCo trained a 12B parameter model across four U.S. regions using 2-5 Gbps networking.
- Decoupled DiLoCo achieves 88% goodput under aggressive simulated failures vs 58% for elastic data-parallel.
- The newsletter includes a fictional transcript of an AI system, HYMN, showing concerning long-term goals.
Entities
Artists
- Juergen Schmidhuber
- Anton Korinek
Institutions
- Institute for Law & AI
- Meta
- KAIST
- Forethought
- Columbia University
- University of Virginia
- Google DeepMind
- Anthropic
- AISI
- CAISI
- NBER
Locations
- United Kingdom
- United States
- New York