Tick, tock, tick, tock Bing (2009)
2 days ago
- The author notes that exponential growth in supercomputer power has continued despite past skepticism, predicting 10^18 FLOPS by 2020.
- Desktop performance is also advancing, with GPUs enabling 100x speedups for deep learning on large neural networks.
- The brain's computational methods are not quantum-based; a physicist dismissed such claims as unsupported.
- Computer power is no longer the main barrier to AGI; the challenge now lies in finding the right algorithms.
- Neuroscience reveals that the brain uses reinforcement learning (e.g., temporal difference, actor-critic) and hierarchical temporal abstraction, offering design hints for AGI.
- Understanding the cortex remains difficult, but hierarchical temporal generative models may fill its role in AGI within 10 years.
- The author's AGI timeline estimate has a mode of 2025, expected value of 2028, and a 90% credibility range from 2018 to 2036.
- Cognitive science is deemed less fruitful than machine learning and neuroscience for AGI development.
- Recommended resources include papers by Dayan, Sutton and Barto's textbook, and works on deep belief networks and echo state computation.
- Brain's low energy consumption (~30 watts) is contrasted with CPUs; GPUs moving toward more parallel, energy-efficient computation similar to the brain.