Google researchers and Google DeepMind declare About the creation of the DeepMind Institute – a new platform designed to enhance discussion on issues related to strong artificial intelligence (AGI).
Image source: institute.deepmind.com
The organization’s directors include DeepMind co-founder Shane Legg, Google executive James Manyika and Google DeepMind chairman Demis Hassabis; Mr. Legg takes over as editor-in-chief. The institute aims to identify differences in the positions of Google, Google DeepMind and the global community on AGI issues. “Opinions are not always unanimous and may change as new data and information becomes available in this rapidly evolving field.”the organization said in a statement.
The first set of four papers covers a wide range of issues: economic policy for managing the likely impact of general AI; maintaining the ability to trace model reasoning in human-readable form; principles of human flourishing; and methods for evaluating advanced AI models.
Security researchers at DeepMind note in a paper that shrinking the transparency window—the ability to see and inspect a model’s step-by-step reasoning—is not inevitable. With the new architecture, monitoring the most powerful models becomes more difficult, and the authors encourage developers and regulators to openly address security trade-offs. For example, limit “opaque sequential depth”that is, the amount of sequential computations a model performs without forming an understandable reasoning scheme; or requiring developers to prove that less transparent systems can still be audited.
In another paper, Mr. Hassabis suggested creating a standardized body for advanced artificial intelligence systems to evaluate models, with developers voluntarily submitting their systems 30 days before release. Once the evaluation system is proven effective, pretesting may become a requirement for advanced model deployment in the United States. First, the evaluation criteria will be developed jointly with the Artificial Intelligence Laboratory, and over time the agency will introduce its own independent and closed testing – which will not give the laboratory the opportunity to adjust model responses based on known test parameters. If the situation becomes serious, the mechanism can be tightened to the point of slowing down the development of advanced artificial intelligence systems.
If you find an error, select it with your mouse and press CTRL+ENTER.
