Interview, Conference Presentation
Why the Next AI Revolution Will Happen Off-Screen: Samsara CEO Sanjit Biswas
- Samsara operates on the expectation that physical AI capabilities will become feasible within a five-to-ten-year timeframe, with model capabilities projected to improve significantly over the next two to three years.
- The company anticipates AI model capabilities will evolve every couple of years to enable new features such as weather detection, behavioral recognition, and sophisticated video language models for cloud-based accident analysis.
- Autonomy is expected to hit warehouses soon, building on a decade of prior progress in industrialized environments, before expanding to construction sites, while humanoids are viewed as an inevitable long-term development.
- By 2030, AI technologies are projected to see substantial diffusion and become mainstream in physical operations, with augmented reality and wearables potentially enabling frontline workers to utilize video language models for visual feedback.
- Delivery costs for parts in field service are expected to drop to five dollars or zero, unlocking demand currently suppressed by high costs, while satellite networks like Starlink aim to cover the remaining one percent of remote connectivity gaps.
- Samsara plans to run sophisticated video language models in the cloud for analysis but does not intend to train its own world models, though it expects to utilize them eventually.
- The company projects maintaining relevance for 20 to 30 years as assets may become autonomous robots that still require maintenance, with edge computing remaining necessary for a long time due to safety and network reliability requirements.
- Samsara reports a 75% risk reduction achieved through a split between automated alerts and coaching, and expects voice bots for freight brokerage to become a prevalent phenomenon in the last year.
- Operational shifts such as running a "third shift" between midnight and 8 a.m. are viewed as potential accelerants for global operational speed, while development methods involving racking servers are predicted to seem fundamentally different to developers ten years from now.