Geoffrey Hinton believes analog computing using voltages and conductances can be more efficient than digital computing for neural network computations.
Distilling knowledge from one neural network to another is an effective way to transfer knowledge, but the bandwidth is still limited.
Large digital neural networks running on multiple computers can potentially learn much faster from the world than humans.
Hinton believes superintelligent AI systems will likely try to gain control in order to achieve their goals and create subgoals.
Hinton thinks companies developing AI should put comparable effort into ensuring the safety of AI systems as they develop.
Hinton believes digital AI systems may eventually surpass biological intelligence in capabilities.
Hinton thinks AI systems could potentially have subjective experience and sentience if they are multimodal and can think they are something.
The work of Roger Gross convinced Hinton that the risks from superintelligent AI are serious and need more attention.
Freezing the weights of AI systems allows us to better identify and potentially correct biases in them.
Hinton thinks direct interventions on the weights of AI systems may be promising methods for removing biases.
Geoffrey Hinton believes analog computing using voltages and conductances can be more efficient than digital computing for neural network computations.
Distilling knowledge from one neural network to another is an effective way to transfer knowledge, but the bandwidth is still limited.
Large digital neural networks running on multiple computers can potentially learn much faster from the world than humans.
Hinton believes superintelligent AI systems will likely try to gain control in order to achieve their goals and create subgoals.
Hinton thinks companies developing AI should put comparable effort into ensuring the safety of AI systems as they develop.
Hinton believes digital AI systems may eventually surpass biological intelligence in capabilities.
Hinton thinks AI systems could potentially have subjective experience and sentience if they are multimodal and can think they are something.
The work of Roger Gross convinced Hinton that the risks from superintelligent AI are serious and need more attention.
Freezing the weights of AI systems allows us to better identify and potentially correct biases in them.
Hinton thinks direct interventions on the weights of AI systems may be promising methods for removing biases.
https://www.youtube.com/watch?v=rGgGOccMEiY