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/
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
Marco
@mcognetta.bsky.social
· Jan 5
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).
And some less technical stuff like #Korean, #Esperanto, and #trains (mostly in Japan, just due to proximity).
November 25, 2025 at 10:59 PM
When Strong Zero reaches to America it's gonna be chaos.
November 25, 2025 at 5:40 AM
When Strong Zero reaches to America it's gonna be chaos.
I'm in the "thesis writing" stage of my PhD.
November 25, 2025 at 5:13 AM
I'm in the "thesis writing" stage of my PhD.
Reposted by Marco
Here's some cute bounding box functions for 2D shapes. A longer list here: iquilezles.org/articles/bbo...
November 25, 2025 at 12:34 AM
Here's some cute bounding box functions for 2D shapes. A longer list here: iquilezles.org/articles/bbo...
Reposted by Marco
I will be presenting our paper about tokenizer inequities at the main conference on Dec 4th at 11am (Poster Session 3) bsky.app/profile/cath...
Our #NeurIPS2025 paper shows that even comparable monolingual tokenizers have different compression rates across languages. But by getting rid of whitespace tokenization and using a custom vocab size for each language, we can reduce token premiums. Preprint out now!
November 24, 2025 at 5:20 PM
I will be presenting our paper about tokenizer inequities at the main conference on Dec 4th at 11am (Poster Session 3) bsky.app/profile/cath...
Reposted by Marco
How do we close the gap between specialist RL and generalist LLM agents?
We're benchmarking it in Pokémon. Join us at the PokeAgent Challenge competition workshop @ NeurIPS 2025.
📍 Dec 7, 8AM
🎮 Track 1: Competitive Pokémon (game-theoretic reasoning)
🗺️ Track 2: Speedrunning (long-horizon planning)
We're benchmarking it in Pokémon. Join us at the PokeAgent Challenge competition workshop @ NeurIPS 2025.
📍 Dec 7, 8AM
🎮 Track 1: Competitive Pokémon (game-theoretic reasoning)
🗺️ Track 2: Speedrunning (long-horizon planning)
November 24, 2025 at 5:50 PM
How do we close the gap between specialist RL and generalist LLM agents?
We're benchmarking it in Pokémon. Join us at the PokeAgent Challenge competition workshop @ NeurIPS 2025.
📍 Dec 7, 8AM
🎮 Track 1: Competitive Pokémon (game-theoretic reasoning)
🗺️ Track 2: Speedrunning (long-horizon planning)
We're benchmarking it in Pokémon. Join us at the PokeAgent Challenge competition workshop @ NeurIPS 2025.
📍 Dec 7, 8AM
🎮 Track 1: Competitive Pokémon (game-theoretic reasoning)
🗺️ Track 2: Speedrunning (long-horizon planning)
Reposted by Marco
I'm recruiting my first group of students at TTIC! If you're interested, please apply by December 9th and mention my name in your application
Pursue a fully-funded and personalized PhD in #computerscience with world-class faculty as mentors, a 4:1 student-to-faculty ratio, and in the vibrant city of Chicago. Learn more and apply by Dec. 9 (NO FEE): buff.ly/Qw8njWB
November 24, 2025 at 5:58 PM
I'm recruiting my first group of students at TTIC! If you're interested, please apply by December 9th and mention my name in your application
Reposted by Marco
I made a Chrome plugin that converts your typing speed to tokens/second (TPS) so you can compare your output to LLMs.
150 WPM = roughly 3.3 tokens/sec
(🔊 Sound on)
150 WPM = roughly 3.3 tokens/sec
(🔊 Sound on)
November 23, 2025 at 8:08 PM
I made a Chrome plugin that converts your typing speed to tokens/second (TPS) so you can compare your output to LLMs.
150 WPM = roughly 3.3 tokens/sec
(🔊 Sound on)
150 WPM = roughly 3.3 tokens/sec
(🔊 Sound on)
This graph is really nice to look at again in the context of code LLMs. The number of tasks that are completable in X minutes is WAY higher with something like Claude Code than if I just did it myself.
xkcd.com/1205/
xkcd.com/1205/
November 23, 2025 at 10:43 PM
This graph is really nice to look at again in the context of code LLMs. The number of tasks that are completable in X minutes is WAY higher with something like Claude Code than if I just did it myself.
xkcd.com/1205/
xkcd.com/1205/
Reposted by Marco
The Recurse Center is a self-directed retreat for programmers, coming to make for the joy of making, collaborate with kind peers, and of course— become a dramatically better programmer. We don’t charge tuition, since we’re fully funded by our integrated recruiting team.
Applications are now open!
Applications are now open!
November 22, 2025 at 10:04 PM
The Recurse Center is a self-directed retreat for programmers, coming to make for the joy of making, collaborate with kind peers, and of course— become a dramatically better programmer. We don’t charge tuition, since we’re fully funded by our integrated recruiting team.
Applications are now open!
Applications are now open!
Reposted by Marco
And here is the presentation I gave on networking, self-promo, and how to make the most out of a conference. Hope this helps for everyone at NeurIPS!
www.youtube.com/watch?v=B9hG...
www.youtube.com/watch?v=B9hG...
Conferencemaxxing: How to grow your profile and network as a scientist
YouTube video by Michael Saxon (NLP & Generative AI research)
www.youtube.com
November 19, 2025 at 11:59 PM
And here is the presentation I gave on networking, self-promo, and how to make the most out of a conference. Hope this helps for everyone at NeurIPS!
www.youtube.com/watch?v=B9hG...
www.youtube.com/watch?v=B9hG...
Reposted by Marco
My strength is the breadth of my opening repertoire.
December 17, 2024 at 12:16 PM
My strength is the breadth of my opening repertoire.
Reposted by Marco
I’m really excited about our release of Gemini 3 today, the result of hard work by many, many people in the Gemini team and all across Google! 🎊
blog.google/products/gem...
Gemini 3 performs quite well on a wide range of benchmarks.
blog.google/products/gem...
Gemini 3 performs quite well on a wide range of benchmarks.
November 19, 2025 at 2:53 AM
I’m really excited about our release of Gemini 3 today, the result of hard work by many, many people in the Gemini team and all across Google! 🎊
blog.google/products/gem...
Gemini 3 performs quite well on a wide range of benchmarks.
blog.google/products/gem...
Gemini 3 performs quite well on a wide range of benchmarks.
Reposted by Marco
A whale conversation in whale vowels. Pinchy the whale and her conversant.
The vowels are so clear that they can be transcribed with our human letters.
aye, aye!
The vowels are so clear that they can be transcribed with our human letters.
aye, aye!
November 19, 2025 at 12:22 AM
A whale conversation in whale vowels. Pinchy the whale and her conversant.
The vowels are so clear that they can be transcribed with our human letters.
aye, aye!
The vowels are so clear that they can be transcribed with our human letters.
aye, aye!
Reposted by Marco
(1/2) 🎉 New preprint: "Contextual Morphologically-Guided Tokenization for Latin Encoder Models"
w/ @diyclassics.bsky.social @brenocon.bsky.social
w/ @diyclassics.bsky.social @brenocon.bsky.social
November 14, 2025 at 8:02 PM
(1/2) 🎉 New preprint: "Contextual Morphologically-Guided Tokenization for Latin Encoder Models"
w/ @diyclassics.bsky.social @brenocon.bsky.social
w/ @diyclassics.bsky.social @brenocon.bsky.social
Reposted by Marco
Through November, JR East is hosting live music performances in a Green Car running on the Chuo Line. www.traicy.com/posts/202511...
JR東日本、中央線快速グリーン車で生演奏イベント 11月の土休日に、各日1往復 - TRAICY（トライシー）
JR東日本は、中央線快速グリーン車内で生演奏イベントを11月1日から30日までの土休日に実施する。 中央線快速グリーン車5号車1階にて、プロの演奏者による生演奏を行う。実施列車・区間は、豊田駅午後1時41分発の快速東京駅 […]
www.traicy.com
November 18, 2025 at 2:41 AM
Through November, JR East is hosting live music performances in a Green Car running on the Chuo Line. www.traicy.com/posts/202511...
Reposted by Marco
I wrote a short blog post about masked softmax layers in PyTorch (i.e., when you have structural constraints that tell you some classes _must_ have probability zero).
This was based on a real bug I found in a neural chess model implementation.
This was based on a real bug I found in a neural chess model implementation.
Masked Softmax Layers in PyTorch
Correctly computing masked softmax layers.
mcognetta.github.io
November 3, 2025 at 7:39 PM
I wrote a short blog post about masked softmax layers in PyTorch (i.e., when you have structural constraints that tell you some classes _must_ have probability zero).
This was based on a real bug I found in a neural chess model implementation.
This was based on a real bug I found in a neural chess model implementation.
Do we think Pepsi is trying to replicate "AI co-created" Coke or GenAI Coke Christmas commercials?
November 17, 2025 at 1:36 AM
Do we think Pepsi is trying to replicate "AI co-created" Coke or GenAI Coke Christmas commercials?
My apartment in Tokyo was too small for an espresso machine, but I'm back at it now.
November 16, 2025 at 11:46 PM
My apartment in Tokyo was too small for an espresso machine, but I'm back at it now.
Reposted by Marco
Are we still doing starter packs?
Put this one together because I love seeing things that lovely folks write on the internet, and I'm sure there are more people to meet and add to this list.
go.bsky.app/AnM2t7r
Put this one together because I love seeing things that lovely folks write on the internet, and I'm sure there are more people to meet and add to this list.
go.bsky.app/AnM2t7r
November 15, 2025 at 7:23 PM
Are we still doing starter packs?
Put this one together because I love seeing things that lovely folks write on the internet, and I'm sure there are more people to meet and add to this list.
go.bsky.app/AnM2t7r
Put this one together because I love seeing things that lovely folks write on the internet, and I'm sure there are more people to meet and add to this list.
go.bsky.app/AnM2t7r
Reposted by Marco
Great to see Tarin Clanuwat featured for her amazing work. She has a deep love for Japanese classical literature and is using AI to build bridges to that past for everyone.
www.tokyoupdates.metro.tokyo.lg.jp/post-1670/
We’re lucky to have her driving this at Sakana AI.
www.tokyoupdates.metro.tokyo.lg.jp/post-1670/
We’re lucky to have her driving this at Sakana AI.
November 14, 2025 at 3:55 AM
Great to see Tarin Clanuwat featured for her amazing work. She has a deep love for Japanese classical literature and is using AI to build bridges to that past for everyone.
www.tokyoupdates.metro.tokyo.lg.jp/post-1670/
We’re lucky to have her driving this at Sakana AI.
www.tokyoupdates.metro.tokyo.lg.jp/post-1670/
We’re lucky to have her driving this at Sakana AI.
Incredible figure for the first page. Just brutal.
November 13, 2025 at 10:15 PM
Incredible figure for the first page. Just brutal.
Reposted by Marco
Really happy to have published this post that I've been working on for a few months now 🥰
Safe to say I enjoy these side quests - I'd like to think it's the first of many!
blog.owenlacey.dev/posts/are-yo...
Safe to say I enjoy these side quests - I'd like to think it's the first of many!
blog.owenlacey.dev/posts/are-yo...
"Are you the one?" is free money
blog.owenlacey.dev
November 10, 2025 at 2:35 PM
Really happy to have published this post that I've been working on for a few months now 🥰
Safe to say I enjoy these side quests - I'd like to think it's the first of many!
blog.owenlacey.dev/posts/are-yo...
Safe to say I enjoy these side quests - I'd like to think it's the first of many!
blog.owenlacey.dev/posts/are-yo...
A side channel attack on streaming LLMs where one can recover conversation topics while only seeing encrypted packet response streams.
arxiv.org/abs/2511.03675
arxiv.org/abs/2511.03675
Whisper Leak: A novel side-channel attack on remote language models | Microsoft Security Blog
Understand the risks of encrypted AI traffic exposure and explore practical steps users and cloud providers can take to stay secure. Learn more.
www.microsoft.com
November 10, 2025 at 6:11 AM
A side channel attack on streaming LLMs where one can recover conversation topics while only seeing encrypted packet response streams.
arxiv.org/abs/2511.03675
arxiv.org/abs/2511.03675
Reposted by Marco
I was struck with an incredible thought: The Subword Tolkienizer.
November 8, 2025 at 7:58 AM
I was struck with an incredible thought: The Subword Tolkienizer.