kyuxu @xuukyu

cs, ai, emergence, evolutionary optimization, open-endedness github.com/kevin ⯎ ⯎ ⯎ ⯎ ⯎ ⯎ Joined January 2025

Tweets

378
Followers

107
Following

307
Likes

2K

kyuxu @xuukyu

3 weeks ago

Happy to say my first paper from my first semester working on research was accepted for presentation at ALIFE later this year in Kyoto. Hopefully more ambitious projects coming soon 😺

0 0 4 81 0

kyuxu @xuukyu

a month ago

fair enough tbh if true. benchmark farming is like 50% for hype in the tech space and it would be bad if the most mainstream-used model had negative side effects from that

Sam Altman @sama

a month ago

fair enough tbh if true. benchmark farming is like 50% for hype in the tech space and it would be bad if the most mainstream-used model had negative side effects from that

1K 805 11K 1.2M 1K

0 0 1 158 0

a certain loud portion of the internet is so unfathomably against “ai” as a concept but as a social movement that was never rooted in legitimate argument very difficult to tell what the actual average person knows or thinks

0 0 0 55 0

kyuxu @xuukyu

2 months ago

results shape beliefs, but the converse is also true

0 0 0 60 0

kyuxu @xuukyu

2 months ago

semi successfully vibecoded after lots of pain but now i realize i probably have no use for this as long as i'm still using jupyter notebooks like a normie. so i have to be a nerd and do everything in a python shell from now on

Quentin @qtnx_

6 months ago

9 0 35 4K 4

0 0 6 318 0

Download Image

kyuxu @xuukyu

2 months ago

@_ueaj evolutionary techniques are inherently really parallelizable, but still turns out to be less sample efficient than gradient/hill-climbing methods. but even a simple evolutionary process that doesn't intelligently balance exploration/exploitation based on task structure (i.e. more…

2 1 5 321 1

ueaj @_ueaj

2 months ago

New article! It's about a proposal for a new learning algorithm inspired by evolutionary algorithms and string theory. It's not something I will actually pursue b/c I'm lazy and don't care to learn kernel writing but it's cool. You can just make all the various axis in a GPU…

4 8 164 16K 132

Download Image

kyuxu @xuukyu

2 months ago

i remember LLM API costs being so expensive just a year ago, for far less capable models. so much scaffolding and tons of API calls needed to be able to make 4o useful for mass document processing

1 0 3 80 0

kyuxu @xuukyu

2 months ago

seeing if codex can port this julia unicode plotting library to python for me :)

0 0 1 74 0

Download Image

William Gilpin @wgilpin0

2 months ago

A UT undergrad wrote a very interesting paper that uses neural cellular automata to solve some of the Abstraction and Reasoning Corpus (ARC) AGI benchmark problems. @TexasScience arxiv.org/abs/2506.15746

8 54 447 26K 277

Download Image

kyuxu @xuukyu

3 months ago

Neural and Evolutionary Computing Papers @TAL

3 months ago

0 0 2 804 0

0 1 4 367 1

Mithil Vakde @evilmathkid

3 months ago

How far can vanilla NCAs go on ARC-AGI? Farther than you'd expect. I benchmarked fixed-step, unmodified NCAs on all public tasks from ARC-AGI-1 and ARC-AGI-2 They hold up remarkably well. Its also dirt cheap. Results + visuals: