robertskmiles
patreon
New(ish) Video: Intro to AI Safety, Remastered
🕑 Added 2021-06-22 15:36:58 +0000 UTCI spent the last 3 weeks or so in the USA, doing various things which have given me a lot of video fuel but no actual videos, so here's a remaster of an existing second channel video:
A while back I gave a talk at "AI and Politics" in London, which I think was a decent introduction to AI Safety. I published a quick version of it my second channel, which probably a lot of you have seen, being patrons, but most people didn't. So I've remastered that video, with better editing, better graphics, more information, etc, and put it on the main channel. This is that. Maybe you never saw the original, maybe it's been a while since then, either way, enjoy!
https://youtu.be/pYXy-A4siMw
Comments
Totes. I have artificial neural networks playing Prisoners dilemma in a virtual space.
The second channel is here: https://www.youtube.com/channel/UC4qH2AHly_RSRze1bUqSSNw And yeah, the definition isn't inherently adversarial. It means being good at choosing actions in order to get what you want. An obvious and neat way to determine the more intelligent agent is if the two agents are in a shared, symmetrical environment, and have opposing goals. But in settings that aren't zero sum, intelligence can often mean cooperating, if cooperating tends to result in better outcomes by the agent's lights.
Interesting - and disconcerting- that one of the foundational definitions of intelligence is adversarial. “The more intelligent agent gets what it wants”. Where’s the second channel, Rob?
