François CholletVerified account

those who are interested in this might want to check out our work on ‘cognitive programs’ where abstract concepts are learned as programs on a ‘visual cognitive computer’. And yeah, it transfers between tasks. https://www.vicarious.com/2019/01/18/a-thought-is-a-program/ …

Nice! This is remarkably similar task to ARC (table vs grid) and a powerful solution. What would it take to adapt VCC for ARC? @fchollet has an open challenge: solving ARC would get us very close to general intelligence, and it could take 2+ years.

Thanks! 'Remarkably similar' is not surprising because @fchollet read our paper before proposing ARC, but then got afflicted with source amnesia. ARC is ok, but brushes fundamental problems like 'objectness' under the carpet. Those cannot be solved using 'connect lines' etc.

You keep accusing me of plagiarism. FYI, ARC was near-finalized before your paper was released. I started working on it in late 2017. I took exactly nothing from your paper, which I vaguely skimmed after you sent it to me (which you did after I had talked to you about ARC...)

Yeah, and we started working on it in 2014. It doesn't matter whether you took anything from our paper or not -- it is prior work. You are *expected to read* prior work, take things from it, and then build on it.

As I explained, you don't cite papers based on topic similarity or tool overlap. If I write a "RL for games" paper once, it doesn't mean every subsequent "RL for games" paper must cite it (despite significant keyword overlap). You cite papers for one of three reasons:

1) it originated an important idea/method/tool that your work uses (attribution) 2) it offers a direct, competitive alternative to your proposal (which the reader should be made aware of) 3) it constitutes a reference survey that the reader can use to learn more about a topic