Dataflow matrix machines (by Anhinga anhinga)

"WorldCoder, a Model-Based LLM Agent: BuildingWorld Models by Writing Code and Interacting with the Environment", arxiv.org/abs/2402.12275

Not a widely known paper (the authors don't promote it), but pretty spectacular (a friend of mine said, "Is it AGI already?").

I think I mostly understand how this works and I made some notes yesterday.

A meta-note here: GPT-4-level models mostly understand what they are doing, but are unreliable; so the question is, can one organize a process which reliably produces needed results based on that. There are plenty of papers trying to push in this direction, but this one is very elegant, and the results are quite good.

******

www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction - very elegant and simple


******

May 9, 2024 update: Since this is access-list-only at the moment (although this post is likely to become public eventually), it's a good place for my notes on switching to Twitter "X Premium" experience (in comments).

May 13: let's move this post to being public.

I updated my Mastodon for the first time since I created it in November: dmm.dreamwidth.org/65045.html

It turns out I follow this cool account: arXiv Highlights AI/ML sigmoid.social/@arxiv@creative.ai

Twitter seems to gradually becoming somewhat less effective as a source of information; I should probably read Mastodon more...

Sigmoid Social is a Mastodon instance for people researching, working on, or just interested in AI. It seems that a number of AI people are creating accounts there. I followed the example set by Ken Stanley and created one for myself too:

sigmoid.social/@DataflowMatrixMachines

It has a different optional verification setup: one can verify that one controls a Web page referenced from one's Mastodon profile if one feels like it. For example, I used that verification mechanism to put a green checkmark on the link to the GitHub mirror of my resume page.

It's often quite annoying that Twitter interferes with reading it anonymously, without a login. This is one of the better workarounds:

nitter.net/

Lexica.art lets one to search for AI-generated art created by other people (no account required) and to generate one's own AI art (a free account is required):

lexica.art/

Mage.space lets one to generate AI art (no account required, but it is possible to create one):

www.mage.space/

“There are so many people doing machine learning research that it’s best to work on things that you’re the most likely person to make progress on.”

:-) This is the advice I seem to follow, mostly :-)

***

I had a dormant twitter account for quite some years: twitter.com/ComputingByArts

Nevertheless, somehow Twitter figured out that it should notify me of interesting things in machine learning ( twitter.com/hardmaru is my favorite).

I recently started to use it a bit more, retweeting particularly interesting things occasionally. The other retweet is about the phenomenon of "shortcut learning" arxiv.org/abs/2004.07780, which seems to be responsible for much of under-performance of both AI and human-based systems (and, at the same time, is almost the essence of life).

By the way, this famous collection of anecdotes is mostly examples of shortcut learning: arxiv.org/abs/1803.03453 "The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities".