This looks incredible, looking forward to working with it! I could be wrong, but I think there’s a typo in this problem:

arcprize.org/play?task=27...
Specifically, the rectangle that moves to the right in example 2 loses its right border, and has too many orange squares as a result!

(Some of my Russian friends have said the word better translates as women’s legs, but Nabokov translates it as feet, so)

Happens a surprising amount in Pushkin as well—really put me off on my first go through of Oneigin!

I can’t speak for Chinese, but in classical Mayan there was a lot of variation in how logograms were written by different writers/cities! It was pretty common therefore to add phonograms for the first letter/syllable as decoration to the logogram. You could add more syllables, but didn’t have to!

One more aside on multimodals: your intuition that they being cross domain knowledge is very good, but they also will probably work if there’s no cross domain knowledge to impart (eg DNA strings and language)! en.m.wikipedia.org/wiki/Stein%2...

I don’t work on videos, so low confidence in this reply—but I do think we chunk and abstract them somehow! The way that time has a very subjective internal expression looks like some form of good compression algorithm.

Words tokenize very nicely, which is important for running good LMs. If you’re working with videos frame by frame, scaling in a transformer style (O(n^2)) is infeasible. So really, the answer is that people are trying, but there are still technical problems that need to be solved

Multimodal foundation models are very much supposed to be the next big thing! LLMs have access to very nicely formatted data, which is probably why they worked out first. Trying to incorporate video into your foundation model is probably a great idea—but our models still struggle with video, a lot.

I’ve been on team pareidolia for a while, but confabulation works as well!

Sure, but I’d actually like to see the question!

What’s the question??

Stephen Greenblatt seeing this post like

I mean. Denmark != Italy, but I’m also not opposed to it

If you don’t have home-made, store bought is fine

I dunno man. I’m pretty sure that Lego, like spaghetto or raviolo become spaghetti and ravioli, is supposed to be pluralized as legi.

It looks like he’s active on Threads :) www.threads.net/@yannlecun

A different direction that I remember seeing proposed was to use Morse homology on level sets to detect features, for example. Shared weights on patches aren’t the hardest idea, but it’s an important one, and he was the first person to write about it

They are very intuitive (especially now, when there’s lots of theory and explanations), and are also something that someone had to invent to really get computer vision off the ground. People were working on different ideas before CNNs were invented, it’s only in retrospect that they were inevitable.

For people not in the field: Yann is one of the biggest people in ML—and not just recently. For example, he’s the guy who invented convolutional neural nets, the most common tool for image recognition. Definitely one of the most cited people in the field.