François CholletVerified account

I am irrationally excited about this. @fchollet has asked a set of fundamentally interesting questions with ARC and I'll be playing for fun! It's like Sudoku for AI addicts ;) (n.b. I personally got many examples wrong - huzzah, a dataset where you question your own cognition )https://twitter.com/benhamner/status/1228070207400202240 …

How is this asking fundamentally different questions from our prior work on this? Of course it is cool to have kaggle competitions on it, but I am asking about how the scientific setting is fundamentally different. https://www.vicarious.com/2019/01/18/a-thought-is-a-program/ …

To be clear, concept/program discovery from 2D input & output grids has been around since the 1930s in the context of IQ testing, and since the 60s in the context of CS/AI. It's a classic format in inductive programming, an entire (tiny) subfield of AI (see eg. the Karel dataset)

I don't claim to have invented it. There's a long intellectual history behind ARC, which is retraced in the ARC paper. Imagine publishing a "RL for games" paper in 2016, then claiming credit for every subsequent paper doing RL for games. You have more productive things to do

What puzzles me though, is that you simultaneously claim credit for ARC and say that you profoundly disagree with every idea that ARC is built upon (in particular Core Knowledge). It's not coherent. It can't be both at the same time.