Prediction of Advanced AI
The prediction of when human-level AI and smarter-than-human AI might be common is difficult. It is also very important.
We take a conservative approach. Rather than attempting to extrapolate advances in computer science NeuralNetworks, we turn to computational neuroscience, and attempt to assess when it might be feasible to construct SpikingModels of the human brain.
TrueNorth is a state of the art spiking neuron hardware chip developed by IBM.
As discussed on the HardwareOverhang page, the estimated cost for human brain level hardware performance using TrueNorth in 2016 is $80/hr.
Computing costs decline by roughly a factor of 10 every 10 years.
- 2016: human brain level performance would cost around $80/hr
- 2040: human brain level performance $0.30/hr; superintelligence $3/hr
All if we get the wiring right. Which we currently don't know how to do. See WhatNeuroscientistsDontYetKnow. However, significant progress is being made on understanding the human brain, and it seems likely that by 2040 we will know how to get the wiring right. See ResearchFunding for more details.
These are median estimates. For risk planning purposes we are probably more interested in say 5th percentile estimates. If there is a 5% chance of superintelligence emerging by some date, we as a society need to be prepared for it by that date.
Be aware that advances in NeuralNetworks could cause advanced AI to occur earlier than these estimates.