CS 박사과정 / 컴공이지만 코딩보다 수학을 많이 합니다 / 프사는 100x100님 / 이 계정은 모든 팔로잉의 리트윗을 끄고 있어요
Joined October 2021
- Tweets4.8K
- Following67
- Followers117
- Likes630
아둔 코딩계 retweeted
How do we understand neural nets with so many interacting params that we can’t describe what they do in words?
Our take: objective functions and regularizers offer a concise description of "what models learn"
Can we *empirically* study implicit regularization in DL?🧵
아둔 코딩계 retweeted
The key idea is gradient matching: if learned weights minimize loss + a regularizer, their gradients must cancel. (see gif in opening tweet)
Holding the weights fixed, we fit the parameters of a candidate regularizer to balance the residual loss gradient.
아둔 코딩계 retweeted
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
아둔 코딩계 retweeted
The most important sentence of the GitHub repository:
"On average, each result used three hours of ChatGPT Pro thinking compute with that model."
The important result on that list I will publish with my collaborator on arxiv tomorrow was found in a session where Astra thought 6 hours in total and probably used my local context. The main problem solving step was done in 2.5 hours.
If results on this level are possible with one-shot solutions in three hours, that is insanely impressive.
아둔 코딩계 retweeted
Erdős reciprocal sum??? Special cases of this one contributed to at least 3 Fields medals and 1 Abel prize...
github.com/openai/math/blob/…
아둔 코딩계 retweeted
OpenAI: determined the exact crossing number of complete graphs and complete bipartite graphs
github.com/openai/math/blob/…
github.com/openai/math/blob/…
아둔 코딩계 retweeted
OpenAI: The Euclidean plane cannot be colored with five color github.com/openai/math/blob/…
아둔 코딩계 retweeted
OpenAI: A counterexample to Sidorenko's conjecture
github.com/openai/math/blob/…
아둔 코딩계 retweeted
This year at @iclr_conf, we've policies where reviewers would be required to (a) disclose AI tools used for assisting in writing peer reviews, and (b) report original self-written assessment and any LLM interactions.
아둔 코딩계 retweeted
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
아둔 코딩계 retweeted
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
아둔 코딩계 retweeted
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.
아둔 코딩계 retweeted
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".
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 are also releasing a flexible formalization harness and a companion paper.
External contributions are welcome!
Joint work spearheaded by our amazing PhD student Ahmad Rammal (@Ahmad3Rammal), together with Niket Patel (@niketnpatel ), Fabian Gloeckle (@FabianGloeckle), Amaury Hayat (@Amaury_Hayat), Remi Munos (@MunosRemi), Julia Kempe (@KempeLab), Vivien Cabannes, and myself from @AIatMeta, @NYUDataScience , and Ecole des Ponts. This is an ongoing effort; more details in the thread below.
(1/9)
아둔 코딩계 retweeted
Really cool
When they trained GPT3 they had loss spikes because they scraped from a subreddit of microwave noises
That training batch was literally text like "mmmmmmmmmmmmmmm"
아둔 코딩계 retweeted
there will soon be no point in playing Slay the Spire 2, as superintelligent AI will be able to play it better than you, and will have more fun doing so than you ever could have
아둔 코딩계 retweeted
Deep Learning as a Convex Paradigm of Computation: Minimizing Circuit Size with ResNets ift.tt/2glZjtV
아둔 코딩계 retweeted
ICML journal editors have added hidden prompt injections to every paper sent to reviewers, to detect when reviewers are using AI
It secretly tells the AI to use 2 specific phrases in the review
A reviewer found it & was about to desk reject, assuming the author did it
아둔 코딩계 retweeted
The Geometry of Algorithmic Stability: A Hodge Theoretic View on Structural vs. Statistical Insta...
Karen Sargsyan.
Action editor: Alberto Bietti.
openreview.net/forum?id=rFqs…
#robustness #stability #instability
아둔 코딩계 retweeted
TIME copied this iconic photo of the workers who built our cities to crown the tech billionaires who are using AI to decimate and destroy the working class.
2025 was the year when artificial intelligence’s full potential roared into view, and when it became clear that there will be no turning back.
For delivering the age of thinking machines, for wowing and worrying humanity, for transforming the present and transcending the possible, the Architects of AI are TIME’s 2025 Person of the Year. time.com/7339685/person-of-t…
