Interview, Fireside Chat
Sophia is not AGI (Ben Goertzel) | AI Podcast Clips with Lex Fridman
- Plans to relocate to Hong Kong in 2011 to reside with his wife and collaborate with Gino Yu at Hong Kong Polytechnic University on the IDEA project, which utilizes machine learning for stock and futures prediction.
- Aims to establish a "benevolent AGI" designed to improve life for all sentient beings by infusing machines with love, compassion, and human values through interaction within compassionate contexts, while assigning a non-zero probability to existential risks such as human enslavement or conversion to computronium.
- Intends to serve as chief scientist and head of software at Hanson Robotics for a few years before stepping back, facilitating a collaboration with David Hanson to develop "smarter" robots, introduce the company to investors for potential relocation to Hong Kong, and utilize Shenzhen's hardware ecosystem for large-scale, low-cost consumer electronics.
- Strategy involves creating a blockchain-based cloud mind platform named SingularityNet where Sophia serves as one of many user interfaces, featuring an API marketplace for distributed AI algorithms and a "platform" approach similar to SoftBank's Pepper Robot with open APIs for experimental AI integration.
- Technical development includes building a walking skeleton software system integrating computer vision, memory, learning, and a hybrid dialogue system of hand-coded rules and deep neural nets, alongside experiments comparing Sophia's chatbot with Facebook's transformer model to generate training data for narrative structure improvement.
- Future software architecture envisions using reformers and other deep neural architectures to enhance transformer training, while distinguishing Sophia's current ungrounded language capabilities from a planned "robo toddler" that grounds language in physical experience to avoid heuristic and rule-based limitations.
- Anticipates that despite explicit clarification of Sophia's subsystems (rule-based, chatbot, OpenCog) and technical limitations like speech-to-text failures with Hong Kong accents, society will continue to ascribe high agency and consciousness to the robot, a dynamic the speaker views as a necessary part of the AI creation process to foster empathy.
- Proposes a "walking skeleton" where specific modules handle vision, movement, memory, and learning, acknowledging that current neural networks struggle with narrative structure, leading to rambling outputs in models like Facebook's deep neural chatbot compared to Sophia's repetitive but longer-arc conversation capabilities.
- Foresees near-term collaboration on medical robotics for elder care, autism therapy, and nursing assistance during the COVID era, leveraging human-like appearances to provide infection-free care and emotional support, while warning against the malicious use of human-looking robots for scams.
- Divergent views identified between the speaker's "math-oriented" approach and David Hanson's preference for biologically based AI and his belief that the public prefers the illusion of robot sentience over technical transparency, leading to ongoing media debates and planned future discussions with figures like Yann LeCun.