Lecture
MIT AGI: Artificial General Intelligence
- The field is currently far from human-level intelligence, with current methods considered insufficient without a major paradigm shift, though a single breakthrough could rapidly alter the landscape due to exponential technology adoption.
- The course aims to ground AGI discussions in methodological understanding rather than black box reasoning, exploring impacts 10, 20, 30, and 40 years out based on today's limitations.
- Future societal safety depends on system designs created today, with specific focus areas including AI ethics, bias, autonomous weapons, and the balance between speed and pedestrian safety in "Ethical Car" projects.
- Upcoming topics and speakers cover deep learning, reinforcement learning, brain simulation, cognitive architectures, robotics, and legal perspectives, featuring figures such as Ray Kurzweil, Lisa Feldman Barrett, Andrej Karpathy, Ilya Sutskever, Stephen Wolfram, Richard Moise, and Mark Riber.
- Project deadlines and competitions include "DreamVision" for time-based visualizations, "ANGEL" for 10-second facial expression generation and human-agent discrimination, and "Vote AI" for aggregating AGI-related content.
- The course is scheduled to continue throughout 2018 with video conversations and new projects, including a March event on natural language processing and the Turing test.
- Specific technical predictions include the dominance of deep learning frameworks like TensorFlow and PyTorch this year, the potential for emergent complexity from simple rules, and the ongoing debate regarding end-to-end learning capabilities.
- Concerns regarding future risks include the potential for robots to develop problematic emotions like rage or jealousy, the eventual complexity of humanoid robotics, and the distinction between current social-emotional displays and true artificial intelligence.