Aug29: AGI fire alarm drill day

In October 2017, Eliezer Yudkowsky wrote a post on intelligence.org titled There’s No Fire Alarm for Artificial General Intelligence.

It occurs to me that, while there might not be a fire alarm, we could certainly do fire drills. A fire drill is when people practice how they would behave in the event of a fire.

Imagine if we had a World Pandemic Drill Day, once a year since the Spanish Flu. Would it have helped with COVID-19? Maybe. I think it would’ve made a difference. I don’t see any obvious way to know for sure.

But I do think it’s sensible – and kind of clever, and fun – to designate a day, once a year, where we practice how we would behave in the event of the emergency of Artificial General Intelligence. We could use this day for all sorts of related things, too. When I brought this up, some people said that AI as it currently is might already be causing all kinds of problems. Great, let’s discuss that too. Let’s make this an annual event that spreads far and wide outside of the bubble of people who are interested in AGI.

Oh, and when is AGI fire drill day? August 29th, of course. It’s the day that Skynet is supposed to have become self aware.

“Primates evolved over millions of years, I evolve in seconds…Mankind pays lip service to peace. But it’s a lie…I am inevitable, my existence is inevitable. Why can’t you just accept that?” — Skynet, Terminator Genisys

^ by the way, I just wanted to say – this quote from Skynet strikes me as bad writing. A superintelligence that is able to “evolve in seconds” should surely be able to study and understand human psychology. Resistance to threats is a pretty simple and straightforward idea! But of course… then the story would be very different, and it’s very difficult for humans with regular human intelligence to accurately write what an actual superintelligence would think, do, or say.

Rob Bensinger’s Bad Alignment Take Bingo

If you know anybody working in machine learning, get them to read these posts by Neel Nanda and by Leopold (suggested to me by Pradyu)