Interview
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
- AI methods are expected to follow oscillating trends where specific approaches become fashionable, fade, and resurface later, yet the field's high watermarks will continue to rise despite inevitable cycles of overhype and crashes.
- The community anticipates a shift from chasing optimal solutions for intractable problems to developing formal approximate solution concepts and regret bounds, while also aiming for engineering advances to evolve into science that predicts system principles before implementation.
- Future algorithms are predicted to automatically construct sophisticated abstractions useful for reasoning, moving the focus of competence toward engineering objective functions rather than designing algorithms manually.
- Researchers will need to be educated on the paradigm of optimization within hypothesis classes rather than explicit step-by-step programming, with a hope that future learning systems will formalize these principles.
- Longer research horizons are considered necessary for students to spend years on difficult problems, contrasting with current short-term publication incentives that risk driving away such talent.
- The publication landscape is expected to see paper volumes vastly exceeding movie releases, necessitating a transition to public, curated reviews to replace or supplement traditional journal models.
- Risks include investor disappointment if grand AI capabilities are not delivered, potentially crashing the field, alongside structural pressures that may deter deep research due to the demand for rapid paper output.