6 Jun 2024 | Edward Hughes * 1 Michael Dennis * 1 Jack Parker-Holder 1 Feryal Behbahani 1 Aditi Mavalankar 1 Yuge Shi 1 Tom Schaul 1 Tim Rocktäschel 1
The paper "Open-Endedness is Essential for Artificial Superhuman Intelligence" argues that open-endedness is a crucial property for achieving artificial superhuman intelligence (ASI). The authors define open-endedness as the ability of a system to produce increasingly novel and learnable artifacts over time, as observed by a human. They highlight that current foundation models, while making significant progress, lack the open-endedness necessary for ASI. The paper explores several research directions, including reinforcement learning, self-improvement, task generation, and evolutionary algorithms, to enhance the open-endedness of foundation models. It also discusses the safety implications of open-ended AI, emphasizing the need for responsible development to ensure that such systems can be safely and beneficially deployed. The authors conclude that open-ended foundation models have the potential to significantly advance society by enabling novel scientific and technological breakthroughs, enhancing human creativity, and expanding general knowledge across various fields.The paper "Open-Endedness is Essential for Artificial Superhuman Intelligence" argues that open-endedness is a crucial property for achieving artificial superhuman intelligence (ASI). The authors define open-endedness as the ability of a system to produce increasingly novel and learnable artifacts over time, as observed by a human. They highlight that current foundation models, while making significant progress, lack the open-endedness necessary for ASI. The paper explores several research directions, including reinforcement learning, self-improvement, task generation, and evolutionary algorithms, to enhance the open-endedness of foundation models. It also discusses the safety implications of open-ended AI, emphasizing the need for responsible development to ensure that such systems can be safely and beneficially deployed. The authors conclude that open-ended foundation models have the potential to significantly advance society by enabling novel scientific and technological breakthroughs, enhancing human creativity, and expanding general knowledge across various fields.