EFTA02488879.pdf

From: Joscha Bach < Sent: Wednesday, August 26, 2015 7:45 PM To: Jeffrey Epstein Subject: Decision making Re motivation: Have you seen the recent movie "Ex Machina"= I liked it; one of the few Al movies that have not been dumbed down. =he main character is a beautiful female looking Al, clearly =ntelligent, and able to manipulate humans to an arbitrary degree. What =akes her inhuman is that she is probably motivated by a single =rinciple, like option maximization. That would make her an inscrutable =sychopath. Option maximization would entail energy, physical integrity, =erhaps reproduction, certainly learning, but you will have an agent =hat you won't like to share a prisoner's dilemma with. = How do you approach decision making? I have recently learned that people =hat subscribe for Cryonics (freezing one's head or body in the =ope to be revived when future technologies make it possible) assign a =ower probability to that it works than the general population. But =hereas "normal" people tend to make binary models about =he world: something is "probably not going to work, so let us =ot bother", many of the Cryonics folks will argue that paying =500 a month for a 1% chance of immortality seems like a bargain. This seems to generalize: in principle, we should perform a Bayesian =pproximation for all our major decisions, attach probability =istributions to everything in the space of possible beliefs, and be =ble to outperform the vast majority of folks that relies on narratives =i.e. binary yes/no decisions about the facts in the world). Gigerenzer, =ahnemann and many others have shown that human brains are terrible of =etting this intuitively right, to the point where an absence of =ine-grained domain knowledge often leads to better management decisions =tc. The divide between probabilistic models vs. narrative models is =eflected to some degree in the conflict between probabilistic and logic =ased Al. In practice, we will probably need to combine both, but I =onder if there is an intrinsic limit to probabilistic descriptions in a =ighly complex world, where we cannot observe baseline probabilities =nyway. How do you decide? Do you Solomonoff-induce and Bayes the hell out of =he stock market, do you reason, do you soak up data and let your =ntuitions guide you, or is most of the important stuff depending on =ommunication and negotiation? Is there are general approach, or how =uch should theories of decision making be dependent on the domain? Cheers, Joscha=?xml version=.0" encoding=TF-8"?> <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd"> <plist version=.0"> <dict> <key>conversation-idgkey> <integer>106419</integer> <key>date-last-viewed</key> <integer>0</integer> <key>date-received</key> <integer>1440618272</integer> <key>flags</key> <integer>8590195717</integer> <key>gmail-label-ids</key> <array> <integer>6</integer> <integer>2</integer> </array> EFTA_R1_01609123 EFTA02488879  <key>remote-ick/key> <string>537356</string> </dict> </plist> 2 EFTA_R1_01609124 EFTA02488880