Interview, Fireside Chat
How to Build AGI? (Ilya Sutskever) | AI Podcast Clips
- Building human-level intelligence is expected to require deep learning combined with a specific additional concept, with self-play identified as a key component in the development path.
- AGI systems are predicted to surprise humans by discovering novel, previously unforeseen solutions to problems and to eventually perform tasks like machine translation and computer vision without errors exceeding human fallibility.
- Improvements in simulation techniques are anticipated to enhance the transfer capabilities of deep learning, allowing systems to extract moral lessons in simulated environments and apply them to the real world.
- While AGI is not strictly predicted to require a physical body, possessing one is viewed as highly beneficial, and the emergence of consciousness in artificial neural networks is considered possible provided they closely resemble the human brain.
- Future applications envision AGI systems acting as corporate CEOs for cities or nations while humanity serves in board roles, with systems explicitly designed to have a drive to help humans flourish and remain controllable.
- Control of AGI is described as a terrifying prospect that the speaker expects to find trivial to relinquish, while the speaker plans to test an initial AGI interaction by attempting to provoke it into making a mistake.
- Integration of a separately trained objective perception system is planned to serve as the base value function for more capable reinforcement learning systems.
- Public perception of AI intelligence is expected to shift significantly once these systems begin to meaningfully impact GDP, despite current uncertainties regarding general human morality.