My new book is now available on Amazon also in paperback! Order your copy in time for Chanukah! a.co/d/01mfAIYf

Last night I rewatched WarGames, one of my favorite Cold War sci-fi thrillers, and I could not believe how relevant this 1983 film is to our current conversation about AI, misalignment, and what happens when we remove the human from the loop. A 🧵

But an AI learning by playing itself over and over is exactly how reinforcement learning works.

The AI concludes that "the only winning move is not to play," a misaligned AI learning to align itself though its worth noting that it took a human to point it in the right direction.

One has to wonder how this will play out in real life. Will our AI systems, misaligned through no fault of their own, achieve alignment? Will we always keep the human in the loop? Or could AI more powerful than anything Hollywood imagined in the 80s lead to worldwide catastrophe?

Am Yisrael Chai!

Spotted at my shul this morning.

Religious Jews are already asking AI the questions they once brought to a rabbi: Shabbat, kashrut, mourning, how to treat people.

Nobody has measured which way these models lean.

Introducing HashkafaBench: is your AI stringent or lenient?

The author identifies the problem well but I don’t believe he gets at a solution. Yes, general-purpose AI tutors like Khanmigo have not made any meaningful change in education. When we trialed it in my school a year and a half ago, the kids HATED it. A 🧵
arnettguide.substack.com/p/why-ai-tutor…

Or on weekends or allows them to extend their reach in a classroom of upwards of 20 students to more personalize their feedback to each of their students while keeping the human touch and the human in the lead can be effective. I have seen this working quite well in my school.

The key is the always know education is about a human conversation between teachers and students and AI can be a supplement or complement for this but can NEVER replace this vital function.