Marco
banner
mcognetta.bsky.social
Marco
@mcognetta.bsky.social
Language and keyboard stuff at Google + PhD student at Tokyo Institute of Technology.

I like computers and Korean and computers-and-Korean and high school CS education.

Georgia Tech → 연세대학교 → 東京工業大学.

https://theoreticallygoodwithcomputers.com/
Pinned
A lot of you followed me due to #NLP, but I like to post about #chess (especially computer chess), #programming (especially puzzles, code golf, etc), and machine learning.

And some less technical stuff like #Korean, #Esperanto, and #trains (mostly in Japan, just due to proximity).
January 26, 2026 at 4:26 AM
Reposted by Marco
Our journey at Sakana AI is just getting started.

We are looking for people to help us pioneer the next generation of AI—building from Japan to the world.

Join us: sakana.ai/careers
January 25, 2026 at 4:56 AM
Reposted by Marco
The #JuliaCon Global 2026 call for proposals is open! We have a host of exciting minisymposia and we look forward to your talks!

Find more information on juliacon.org/2026/cfp/ and submit your proposal until February 28th at pretalx.com/juliacon-202...

#julia @julialang.org
January 23, 2026 at 9:37 PM
Interesting developments in the #Julia ecosystem.
January 23, 2026 at 9:48 PM
Reposted by Marco
Remix timeline: mix your timelines with a custom distribution. Another awesome feature in the custom client. Exactly what I was looking for. Thanks @mariaa.bsky.social 😄👍
January 23, 2026 at 3:34 AM
Reposted by Marco
I can't track down who it was, but someone on here suggested making a "Remix" feed that pulls from all your pinned feeds while removing duplicates. I made something like that for Lea and it's my new favorite feed! Any other suggestions along these lines, please share.
January 22, 2026 at 7:21 PM
I'm always surprised when top_k doesn't work in numpy. It's such a weird omission from an otherwise amazing library.

There have been discussions about it for *years*, but it still hasn't landed.

Here's a cool write up from an undergrad that took it on as an intern project:
labs.quansight.org
January 22, 2026 at 5:59 AM
Do you like tokenization? Or better yet, do you hate tokenization and want to make it better?

Merge with us at the Tokenization Discord.
January 22, 2026 at 3:27 AM
This happens to me ~once a week and I always panic.

I'm printing this out to put on my wall.
oh shit! I tried to commit a file that should be ignored!

permalink: wizardzines.com/comics/oh-sh...
from our zine "Oh shit, git!": wizardzines.com/zines/oh-shi...
January 21, 2026 at 9:46 PM
Reposted by Marco
South Korea will release a new set of commemorative stamps featuring baby animals depicted with their lesser-known Korean names, in a move expected to attract collectors interested in linguistics as well as wildlife, the Ministry of Science and ICT said Wednesday.
www.koreaherald.com/article/1066...
Old Korean words come alive in new baby animal stamps
South Korea will release a new set of commemorative stamps featuring baby animals depicted with their lesser-known Korean names, in a move expected to attract c
www.koreaherald.com
January 21, 2026 at 6:56 AM
About a month before the attempted coup, I was on a KTX to Seoul and was standing in the area next to the VIP car.

A few minutes into the ride, Han Duck-soo walked out of the VIP car to take a phone call.

Anyway, he is now about to spend 20+ years in prison.
A South Korean court has found former Prime Minister Han Duck-soo guilty of insurrection and criminal conduct.
www.yna.co.kr/view/AKR2026...
[속보] 법원, 한덕수 '계엄 문건' 위증 혐의도 인정 | 연합뉴스
(
www.yna.co.kr
January 21, 2026 at 6:38 AM
Reposted by Marco
GSoC 2026 Mentoring org applications now open! 🚀

Inspire tomorrow’s developers with your open source project. Application period January 19 - February 3, 2026.

➡️ Read all the details in our GSoC Blog here: goo.gle/gsoc-2026-me...
January 19, 2026 at 8:30 PM
Huh, TIL `json.loads` is called that because the `s` stands for "string". I always just guessed between `json.load` and `json.loads` when I needed to do it on the fly.
January 19, 2026 at 6:04 PM
Reposted by Marco
Oh, the irony.
April 21, 2025 at 6:53 PM
Sakana has had some bangers in the past few months.
Introducing RePo: Language Models with Context Re-Positioning

Standard LLMs force a rigid linear structure on context, treating physical proximity as relevance. Cognitive Load Theory suggests this is inefficient—models waste capacity managing noise instead of reasoning.

arxiv.org/abs/2512.14391
January 19, 2026 at 1:05 AM
Oh this is a great list. A lot of the ML starter packs I know of are filled with the same names, but this one has a lot of new ones for me.
January 17, 2026 at 7:03 PM
Reposted by Marco
i updated this go.bsky.app/LFAZcGE
January 17, 2026 at 6:52 PM
I'm in awe of the jubeat* version naming scheme. Software companies could learn a thing or two.

*a popular Japanese rhythm arcade game; think musical electronic wack-a-mole

en.wikipedia.org/wiki/Jubeat?...
January 16, 2026 at 10:52 PM
Reposted by Marco
Meet TranslateGemma. 💎
​✅ Open weights (4B, 12B, 27B)
✅ 55 languages + 100s more in training data
✅ Multimodal capabilities (image text)
Blog: blog.google/innovation-a...
Paper: arxiv.org/pdf/2601.09012
Model: huggingface.co/collections/...
Cookbook: colab.research.google.com/github/googl...
TranslateGemma: A new suite of open translation models
TranslateGemma is a new family of open translation models built on Gemma 3.
blog.google
January 16, 2026 at 7:50 PM
Just for fun, I tested ChatGPT, Claude, and Gemini on this bug by asking them to review the original code (linked below).

ChatGPT, Gemini, and Claude Opus got it first try and Claude Sonnet got it after very slight prodding.

Pretty impressive, as they were given very little context.
January 16, 2026 at 7:36 PM
The tokenizer isn't even to blame here, but man it really feels like it should be.
January 16, 2026 at 6:31 PM
The @lichess.org yearly roundup had an interesting note in it: they now adjust their rating system to try to account for the empirical advantage of going first.

By analyzing their dataset, they found that white has a win rate of ~51.6, which corresponds to an ~11 point rating advantage.
January 16, 2026 at 8:44 AM
Gently reposting cause I think this book would be of interest to a lot of people here. First of all, Bopomofo is super cool and second this book is absolutely gorgeous.

And bopomofo is fun to say.
I recently learned about this beautiful book about Bopomofo, but it seems to have been a limited run and the designer is a bit hard to find (their self-listed site is gone).

Anyone know of a place that has a copy (or copies for sale)?
passport green covers taiwan’s hidden code in bopomofo tribute by memphis sun
www.designboom.com
January 15, 2026 at 7:35 AM
Reposted by Marco
We are hiring Members of Technical Staff (Research Engineers)!

Current LLM agents lack reliability, creating a gap between demos and production. We solve this by automating the complex workflow of debugging, evaluation, and iteration required to make agents robust. 👇
January 14, 2026 at 4:04 PM
TIL that Geoffrey Hinton's great-great-grandfather was George Boole (of Boolean fame).

Overall, a ridiculous family tree.
Geoffrey Hinton - Wikipedia
en.wikipedia.org
January 13, 2026 at 3:27 AM