The key idea: invest now for an uncertain future

Researchers with the Institute for Law & AI have proposed a strategy they call “radical optionality”, which suggests that governments should invest in tools and institutions now, even if those investments are not yet fully understood or needed. The goal is to ensure that governments can respond effectively to potential crises involving powerful AI.

Specific recommendations for AI regulation

• Information-gathering authorities: Transparency and reporting requirements, coupled with a mechanism for verifying this information.
• Whistleblower protections: Ensuring that employees at frontier labs can report risks to the government.
• Information-sharing within and between governments: Establishing effective channels for sharing sensitive AI-related information across different levels of governance.
• Flexible rules and definitions: Avoid premature regulation by setting broad goals (e.g., mitigating risk) and allowing companies to define specific methods to achieve these goals.
• Audits and evaluations: Developing government and third-party capacity to assess AI systems’ capabilities and safety.
• Security of model weights and algorithmic secrets: Investing in methods to secure the underlying data and algorithms that power AI models.
• Hiring and talent: Increasing funding for institutions like AISI (UK) and CAISI (US), which support technical expertise needed for these interventions.

The authors address some common counterarguments, such as the claim that their recommendations are not overly forceful or subject to abuse. They emphasize the need for robust information-sharing mechanisms and flexible regulatory frameworks to prepare for potential future AI developments.

Why this matters: setting the world up for success

The authors argue that investing in these areas now, even if they seem uncertain or costly compared to immediate benefits, is a prudent long-term strategy. They believe that while the initial costs are modest relative to potential future risks, failure to act could be catastrophic.

Read more about radical optionality.

A Schmidhuber Special: Neural Computers

Meta and KAIST have explored whether a neural network can perform tasks traditionally handled by conventional computers, such as executing commands or displaying graphical interfaces. This concept is known as a “Neural Computer” (NC).

• The big idea: A new machine form that could unify computation, memory, and input/output in a single learned runtime state.
• Two experiments: Demonstrations of using a powerful generative model to create simple NCs with both CLI and GUI interfaces. These prototypes show that the neural network can perform basic operations but are still very rudimentary.

The authors suggest that future developments in this area could lead to systems where all software is integrated into one large, unified neural network, potentially revolutionizing how we interact with technology.

Read the full paper on Neural Computers.

Recursive self-improvement and economic growth

Economists from Forethought, Columbia University, and the University of Virginia have developed models suggesting that recursive self-improvement (RSI) in AI could lead to an unprecedented economic boom. They identify two key channels through which this might occur: technological feedback loops within the innovation network, and economic feedback loops driven by automation.

Read more about recursive self-improvement.