Interview
Greg Brockman: OpenAI and AGI | Lex Fridman Podcast #17
- Aims to extract data to align systems with human values, potentially enabling safe AGI development within his lifetime or sooner.
- Projects that initial internet conditions like open academic mindsets and Wikipedia's ad-free model will shape digital progress and resource integrity for the next 40 years.
- Foresees AGI capabilities including reading all scientific literature to cure diseases, inventing technology for material abundance, and deploying biodegradable robots to convert ocean debris.
- Plans to scale language modeling techniques by factors of 10, 100, or 1,000 to observe emergent capabilities and out-of-distribution generalization.
- Anticipates that distinguishing human from AI content will become a losing battle, necessitating a shift toward reputation networks and physical-world identity verification.
- Established a new "reasoning team" to enable neural network reasoning, a long-term project targeting breakthroughs in mathematical logic or proving P equals NP that may not yield results in 2019.
- Expects significant scaling to produce qualitatively different behaviors and long-term planning capabilities not apparent in smaller models.
- Believes simulation of physical laws via massive training environments, such as Dota, may allow agents to transfer skills to the physical world.
- Acknowledges uncertainty regarding whether AGI is possible through current deep learning scaling alone, noting potential needs for new ideas or architectural tweaks regarding reasoning.
- Envisions a future where humans or AI systems may develop mutual affection, viewing this as a positive outcome of advanced general intelligence.
- Anticipates a competitive "run-up" to AGI where safety protocols risk compromise, committing to collaborative measures with other developers to prevent unsafe race conditions if not leading.
- Expects future regulation to be necessary, with the government's current primary role being measurement and literacy building rather than immediate strict rules.
- Plans to cap investor returns, ensuring that in a successful AGI scenario, the majority of created value legally belongs to the nonprofit to benefit everyone.
- Notes that while democratized compute utilities may not allow individual developers to compete in massive scale segments, impactful ideas can still be discovered without vast computational resources.
- Projects that the distinction between AI and human content will eventually become less critical than preventing deception by "smarter and smarter bots."