Tick, tock, tick, tock Bing (2009)

Shane Legg surveys computing and neuroscience to argue that raw compute is unlikely to block the arrival of artificial general intelligence (AGI). He plots historical LINPACK performance and expects roughly 10^18 FLOPS by 2020 and around 10^20 FLOPS before 2030, noting desktop GPUs already reach ~10^13 FLOPS and yield large speedups for deep learning. Quantum-brain hypotheses are dismissed as unsupported. With computation exploding, the bottleneck shifts to algorithms: we need the right architectures and learning methods rather than more cycles.

Legg reports that contemporary neuroscience is yielding concrete, AGI-relevant hints, particularly about reinforcement learning: the brain appears to use temporal-difference-like mechanisms, mixing model-based and model-free control, with components like pseudo-rewards, inverse-reward inference, uncertainty modeling and mechanisms that promote conceptual knowledge. Cortex function remains much less understood and observing it in action is limited, so progress on cortical algorithms may lag; instead, advances in deep belief networks, hierarchical temporal models, liquid computing and slow feature analysis could stand in. He gives a modal AGI date of 2025, an expected date of 2028, and a 90% credibility interval of 2018-2036, acknowledging some expert disagreement but attributing it to differences in background and focus.