Enyan Zhang 张恩言
@EnyanZhang
Followers
57
Following
1K
Media
7
Statuses
46
PhD-ing in CS @Yale, prev. undergrad @BrownUniversity. He/him
Joined October 2023
It’s often hard to predict/control what neural networks learn during training—can we encourage them to adopt the solutions we want? In new work with @brown_nlp and @michael_lepori, we introduce subtask induction, a method to bias any model toward particular algorithms. (1/12).
1
1
13
I love how gptzero gives "your most human sentences", so in a future world of ASI I can brag "the most human I've ever been is. ".
0
0
0
RT @Aflah02101: THOUSANDS OF PAPERS ARE NO LONGER REPRODUCIBLE (they never were perfectly reproducible but any kind of reproducibilty is no….
0
1
0
RT @willccbb: removed the sentence from the readme that said it's fine to use pip instead of uv if you really want to.
0
8
0
RT @RTomMcCoy: 🤖🧠New paper!🧠🤖. A common assumption: In neural networks, "inductive bias" = "model architecture". But in fact initial weight….
0
11
0
Maybe now is the right time to say, I’ve always felt weird about scaling law: loglinear price/perfmance for something you pay for? That sounds terrible!.
0
0
1
how come the model class in hf is "Pre*T*rainedModel" but the config class is "Pre*t*rainedConfig"??? and how come i didn't realize this until now??.
1
0
1
tried out Cline: The same model that got 18th place in IOI yesterday can't code independently yet, it seems.
0
0
0
RT @random_walker: One of the biggest areas of AI hype right now is the notion that it will hyperaccelerate scientific progress. I understa….
0
118
0
Deepseek R1 is the only model (o1, o3, gemini 2 flash thinking incl.) I've used so far that proves things reliably on problems I care about. It's actually surprising --- a lot of the problems I ask are supposed to be quite common!.
0
0
1
RT @patelmaitreya: Hey @huggingface, can we get a lightweight library over transformers & diffusers that extracts the bare minimum code for….
0
2
0
RT @_KarenHao: As someone who has reported on AI for 7 years and covered China tech as well, I think the biggest lesson to be drawn from De….
0
3K
0
I feel this is the wrong takeaway: it's surprising that huge leaps in math/science, often used as criteria for human intelligence, did not correlate to user perception of model intelligence. Either it's an UX problem, or these benchmarks don't accurately reflect latent capability.
🚨 New piece in @TIME: AI progress hasn't stalled — it's just become invisible to most people. 🚨. I used to think that AI slowed down a lot in 2024, but I now think I was wrong. Instead, there's a widening gap between AI's public face and its true capabilities. 🧵
0
0
1
RT @rohan99pandey: AI is moving fast, and I love what we're doing as a community. But, I want to talk and share how the constant push and….
0
3
0
Maybe the takeaway from this year's Nobel prize is computer science is "science" after all.
0
0
2
RT @BrownUniversity: There is no denying it — it's officially Commencement season! 🐻🎓 Congratulations, Class of 2024. #Brown2024 #BrownU. 📸….
0
8
0
RT @roydanroy: The original post by Max, as well as the reactions by Yann, etc., are not even wrong. As a computer scientist, I'll get us….
0
10
0
RT @amyxlu: us: writes profusely on the internet about evaluating LLM retrieval of niche content in long contexts .us: trains LLM on said i….
0
20
0