Computable Artificial General Intelligence
I recommend a panel for anyone interested in artificial general Intelligence (AGI). David Wood has chosen three panelists who have deep insights into the possible regulation of AGI. Dan Faggella was one of the panelists, and he was as informative and eloquent as ever. David Wood chose two other panelists for this session of London Futurists who had very different opinions about how to restrain AGI.
A general artificial intelligence (AGI) is an agent which requires less information to make a prediction than any other. One could argue that AIXI, the general reinforcement-learning agent, not only met the definition but was also the only mathematical formalism which did so. AIXI had a significant outcome, but was not computable. Its performance was also subjective. This paper proposes a formalism for AGI that overcomes both of these problems. This paper provides formal proof of the performance, as well as a simple implementation. It also includes experimental results to support this claim.
Source:
https://arxiv.org/abs/2205.10513v2
Leave a Reply