Conference Presentation, Keynote
9 Years to AGI? OpenAI’s Dan Roberts Reasons About Emulating Einstein
- OpenAI projects that current models solving textbook-level problems in one minute will require approximately nine years and 16 doubling times to achieve the capability of discovering major scientific principles like general relativity.
- The organization intends to scale reinforcement learning compute to a level that eventually dominates and surpasses pre-training compute, necessitating a reinvention of scaling science due to new demands in test time compute and reinforcement learning.
- OpenAI plans to raise $500 billion to acquire land in Abilene, Texas, construct facilities, and install computing infrastructure, with a strategy to continuously raise capital as revenue is generated to expand these assets.
- Task completion durations for AI agents are currently about one hour and are expected to double every seven months, with a linear extrapolation predicting capabilities for tasks lasting between two and a half and three hours in the coming year.
- Future model improvements are expected to focus on extended test time thinking, though predictions regarding specific capability timelines carry inherent risks as industry forecasts often prove inaccurate.