Sean S. Yi / 이석원 (李碩遠)

Doing ML stuff for money.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@deredleritt3r @karpathy Had to learn this the hard way myself. The "please" is a must lmao.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@jconorgrogan @polynoamial Politicians are also supposed to be servants of the people lol.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@RealAdamMatthew @kchonyc Closed form solutions imply that the landscape is simple enough that we can just use an iterative approach and get comparable overall complexity. This becomes even more pronounced at larger scales. Think ordinary least squares.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@notnotstorm I'm a bit iffy on ruling out Black. Black had some nice trailing comma and vertical alignment that clashes with Ruff. But, yes, I ended up going with Ruff.

[Akshay 🚀 @akshay_pachaar]

ML researchers just built a new ensemble technique.

It even outperforms XGBoost, CatBoost, and LightGBM.

Here's a complete breakdown (explained visually):

[Wenhu Chen @WenhuChen]

I have very mixed feelings when reading these recent search + rl papers.
I understand the LLM folks are too young to read the 2020 RAG papers. But claims like "RAG lacks the flexibility for multi-turn, multi-query retrieval" and "We are the SoTA by achieving NQ = 41%, HotpotQA=38%" really baffle me.

1. Iterative RAG has been studied for quite a long time. There are plenty of great work done in this space.
2. The SoTA for NQ and HotpotQA are both around 70%!
There are some good curated RAG paper list in https://t.co/TK6T3oHiZj and https://t.co/PcpnUsHEly.

[@emilymbender.bsky.social @emilymbender]

With the OpenAI clownshow, there's been renewed media attention on the xrisk/"AI safety" nonsense. Personally, I've had a fresh wave of reporters asking me naive questions (+ some contacts from old hands who know how to handle ultra-rich man-children with god complexes). 🧵1/

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

Anyone using S3 for their pipelines will know what a big deal this is.

[PyTorch @PyTorch]

📣 Introducing CPP-based S3 IO DataPipes 📣

Utilizing Amazon S3, developers can quickly build DataPipe DAGs to access, transform, and manipulate data with shuffle, sharding, and batch features.

Check it out today: https://t.co/1vdw4RrWQR

#poweredbypytorch

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@synethetica @TheNFTCPA @tkkong @gdb

[Yanai Elazar @yanaiela]

Since you asked, here's my rant about adversarial examples papers in NLP:

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@ccanonne_ @jbhuang0604 Is the advice itself bad? I also thought the Zoom part was a funny joke.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@Sparksbet @emilymbender The NLP subreddit also specifies that they're not natural language processing for this reason lol.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@ShriramKMurthi @duolingo Very entertaining thread; thought this was a Humans of New York post for a second lol.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@arumshisky @sleepinyourhat @annargrs I personally think the time difference is the biggest issue. Attending EMNLP in person this year made me realize what I'm missing out on, even in Gather Town (usually in Seoul).

[Anna Rogers @annargrs]

Good news 🥳: https://t.co/axQAhVOnm7 will return for the 3rd iteration, this time co-located with @aclmeeting in Dublin! Teaser: invited insights by @tallinzen @barbara_plank!
This gives us about 6 months between 2nd & 3rd workshops. Time to save up any interesting failures😉
/1

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@LChoshen @nlphuji @IBMResearch I'm in the DR as well, if you're down I'd love to meet up.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@LChoshen @nlphuji @IBMResearch You still here?

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@jbhuang0604 Any chance you'll be releasing a recording or sth?

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@thegautamkamath Same deal for Korea I believe.

[Sean S. Yi / 이석원 (李碩遠) @seanswyi]

@hardmaru @TwitterEng Personally against the "downvote tweet" with a passion. I think one social media platform *ahem REDDIT* is enough for this.