아둔 코딩계 @AdunDebugging

CS 박사과정 / 컴공이지만 코딩보다 수학을 많이 합니다 / 프사는 100x100님 / 이 계정은 모든 팔로잉의 리트윗을 끄고 있어요 Joined October 2021

Tweets

5K
Followers

86
Following

66
Likes

602

아둔 코딩계 @AdunDebugging

2 hours ago

뉴립스도 자체적으로 리뷰 LLM에 그냥 넣는 멍청이들 잡는 프롬프트 넣어놨네

0 0 0 33 0

View Details

Machine Learning (ML) Papers @Memoirs

3 days ago

Score Based Error Correcting Code Decoder Alon Helvits, Eliya Nachmani arxiv.org/abs/2605.28358 [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙸𝚃] 💬Accepted to ICML 2026

0 4 10 677 5

View Details

Hi! We answered Jack and Christian in detail on Zulip, but TL;DR: - We agree, and clearly state, that the code is not perfect: each book has a completion percentage, and each statement has quality scores. Some books have a low completion percentage, and some statements are flagged as low quality. Most of Geometry of Manifolds is flagged as bad. We released all our outputs, including the weaker ones, on purpose; now our goal is to keep raising the code quality over the coming months. - There is a separate issue of whether the eval harness used to score the code is trustworthy. In most cases, we find that it is. When it does fail, it is often because it spots the error but judges it to be less severe than it is (this was the case with the error highlighted by Jack). Overall, our goal is to explore both the potential and the limitations of autoformalization at scale with the current generation of models, and community feedback is very helpful - we will build upon it to improve the project.

1 3 22 4K 2

View Details

Sujeet Bhalerao @sujeetbhlr

6 days ago

Here's a "theorem" in the complex variables section proved by sorry: "If g is complex differentiable at every point with Re z \geq 0, then g is complex differentiable everywhere." This is then used in the proof of another theorem in Lecture21.lean. All of this is marked as a "pass".

Charles Arnal @arnal_charles

7 days ago

Our team at @AIatMeta is excited to announce ATLAS: one of the largest automated formalization efforts to date. ATLAS contains Lean 4 formalizations of both statements and proofs from 25+ mathematics textbooks, spanning dozens of domains, for a total of 500k lines of code. We

29 88 423 361K 276

14 37 323 48K 74

View Details

아둔 코딩계 @AdunDebugging

3 weeks ago

@cfomega 제 느낌으로는 어쨌든 AI 논문은 실험이 동반되는 경우가 많아서 그렇게 심하지는 않은데, 이론 논문이 좀 심해요. 그리고 말만 보면 꽤 그럴듯한 경우가 종종 있는데 정작 보면 말이 이상하거나 보통 글 쓰는 식이 아니거나 (논문 전체가 bullet point라던지)

1 0 1 54 0

View Details

아둔 코딩계 @AdunDebugging

3 weeks ago

논문을 쓰는게 맞는지 의심스러운 사람들이 아카이브 욕해서 뭐 어쩌자는거지 싶음

2 0 0 76 0

View Details

아둔 코딩계 @AdunDebugging

3 weeks ago

@cfomega x.com/i/status/20550…

Thomas G. Dietterich @tdietterich

3 weeks ago

Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/

140 918 6K 1.1M 1K

1 0 0 30 0

View Details

아둔 코딩계 @AdunDebugging

3 weeks ago

@cfomega 아 그건 현재 기준이고 아카이브 담당하는... 교수님? 이 어젠가 트윗을 올렸어요

1 0 0 19 0

View Details

아둔 코딩계 @AdunDebugging

3 weeks ago

@cfomega 제가 이해가 되지 않던 것은 내가 그냥 조금만 들어간 논문에서 1저자 같은 사람이 실수한 것으로 같이 처벌받아야하냐 논문이 억셉될 것을 요구하는건 게이트키핑이다 내 논문을 어떻게 다 읽냐 가 있었습니다

1 0 0 28 0

View Details

아둔 코딩계 @AdunDebugging

3 weeks ago

@cfomega 제가 이해가 가던 것들로는 citation 실수는 이전에도 있었다 (실제로 통계적으로 한 10% 정도는 있었다는 것 같아요 2020년 전에도) 그래서 얀 르쿤 같은 빅네임이 규정 어기면 밴할거냐 안 할 거 아니지 않냐 (모르는 일이라고 보지만...) 정도가 있었고요