in many subfields of AI, the stuff that’s locked up in Google proprietary information is light years beyond what’s available in academia
What is your evidence for this? (Sorry if it’s somewhere in the reddit thread, I didn’t read too far down.)
I have heard this claimed by multiple sources but looking at the webpages of most google research scientists indicates that they aren’t even working on new theory so much as applying what’s already out there, so I’m curious what’s causing our beliefs to diverge so much.
I don’t have any evidence beyond Jonathan Tang’s say-so. And the fact that Peter Norvig works at Google. There may be something useful in Reddit’s video interview with him, but it’s been ages since I watched it, so I don’t know.
What is your evidence for this? (Sorry if it’s somewhere in the reddit thread, I didn’t read too far down.)
I have heard this claimed by multiple sources but looking at the webpages of most google research scientists indicates that they aren’t even working on new theory so much as applying what’s already out there, so I’m curious what’s causing our beliefs to diverge so much.
I don’t have any evidence beyond Jonathan Tang’s say-so. And the fact that Peter Norvig works at Google. There may be something useful in Reddit’s video interview with him, but it’s been ages since I watched it, so I don’t know.