Best of Twitter - August 22-Sep 11, 2022 - AI supplement
.@GothamChess youtube.com/watch?v=8dT6CR… says that AlphaZero really loves sacrificing pieces for speed & initiative. "AZ loves suppressing movement." Web-based stockfish thinks AZ does dumb moves. Why? Because SF evaluates piece adv. If you zoom out, "dumb moves" are actually smart.
Short written assignments are FINISHED pedagogically. Educators need to wake up to this reality. There are alternatives: in-class blue books, oral exams, longer research papers (for now). But AI forces us to rethink our pedagogy.
This took me 10 seconds in GPT-3.
DALLE-2 was paywall-released recently by an extremely well-funded company.
Just yesterday, a group of independent researchers released their own model (Stable Diffusion) that you can use in a few lines of code for FREE.
The speed of the ML research community is insane 🤯
"You are GPT‑3, and you can't do math": Prompting GPT‑3 via zero-shot instruction to answer calculation/math questions by consulting a Python REPL.
"AI And The Limits Of Language" noemamag.com/ai-and-the-lim… good article on a big open question in my mind - how much can an AI learn from internet text alone? what if added a lot of images/videos from the internet? do we have to reach all the way to embodied agents?
Transformers are Sample Efficient World Models
“With the equiv. of 2 hours of gameplay…our approach sets a new SOTA for methods without lookahead search, and even surpasses MuZero.”
Love the simplicity of this approach!
pdf arxiv.org/abs/2209.00588
code github.com/eloialonso/iris