21 days ago
- The computational cost of human visual perception is immense: simulating the visual input of just one person's eyes in real-time would require processing about 1.9 million tokens per second.
- Current AI inference capacity is vastly insufficient for widespread humanoid robotics: Google's entire fleet can only handle the visual feed of about 650 people, and scaling to one humanoid per human would require millions of times more compute than today.
- Historical predictions about compute ceilings (e.g., '640K is enough') consistently fail because they underestimate human imagination and new use cases, but AI may break this cycle by generating its own demand for compute through self-improving intelligence.
- Text-based AI models are a distillation of human knowledge with inherent limits; surpassing human understanding requires AI with direct sensory experience, leading to unprecedented compute demands for multimodal and robotics applications.
- Projections suggest AI inference demand for humanoid senses could match supply by the 2030s, but this assumes constant growth rates and ignores that AI itself may accelerate demand by inventing new compute-intensive applications.