Lightnews — Scholar-powered news

Reposted

Nicholas Lourie @ COLM @nicholaslourie.bsky.social · 5d

📄🔈✨ Deep learning is an empirical science, but we rely on basic empirical methods. What might a better foundation—a simple theory—for empirical work look like?

@kyunghyuncho.bsky.social, He He, and I move towards one in "Hyperparameter Loss Surfaces Are Simple Near their Optima" at #COLM2025!

🧵1/9

1 2 6

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · 7d

things initially get increasingly more complicated, but then there's a point after which they all start to become simplified.

arxiv.org/abs/2510.02300

i guess this can be thought of as learning a better-conditioned energy function.

Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models

We introduce Equilibrium Matching (EqM), a generative modeling framework built from an equilibrium dynamics perspective. EqM discards the non-equilibrium, time-conditional dynamics in traditional diff...

arxiv.org

8

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · 8d

i'll be at COLM on thursday (oct 9). DM me for a chat!

4

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · 8d

blog post: kyunghyuncho.me/mychartexplo...

MyChartExplorer: a vibe-coding journey toward accessing my own medical records – Kyunghyun Cho

kyunghyuncho.me

1

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · 8d

the past few weekends here and there spread through a few months have been fun for me. i vibe-coded and along the way learned various services that enable one to rapidly deploy a trial version of a web service.

an awsome era for software development!

1 2 7

Reposted

Mila - Institut québécois d'IA @mila-quebec.bsky.social · Sep 2

Exciting news! We're thrilled to announce the appointment of Professor Hugo Larochelle as Mila's new Scientific Director! A deep learning pioneer and former head of Google's AI lab in Montreal, Hugo's leadership will be pivotal in advancing AI for the benefit of all.

mila.quebec/en/news/hugo...

Hugo Larochelle becomes the new Scientific Director of Mila | Mila

Montreal (Quebec), September 2, 2025 – Mila, the Quebec Artificial Intelligence Institute, announces the appointment of Hugo Larochelle—Adjunct Professor at the Université de Montréal and former head ...

mila.quebec

2 7 37

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 28

amen

those with savior and superiority complex, obsessed with sci fi, blinded by dollar signs, devoid of empathy, and severely gullible.

Ben Recht @beenwrekt.bsky.social · Aug 28

I wrote about the tragic death of Adam Raine and the venal negligence of "AI Safety." www.argmin.net/p/the-banal-...

The Banal Evil of AI Safety

Chatbot companies are harmful and dishonest. How can we hold them accountable?

www.argmin.net

1 9

Reposted

FuriosaAI @furiosa.ai · Aug 14

Thank you to everyone in the AI community we met in NYC this summer and those who came to our happy hour at @nyu.edu’s Global AI Frontier Lab last week.

1 1 1

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 13

everyone says google will lose search biz to llm, but in reality, all llm co's are so busy trying to reproduce google search. irony

1 7

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 12

i meant each entry to a leaderboard is a hypothesis. in LM Arena, each entry is a chatbot product without any details, and i don’t believe we can consider it a scientific hypothesis.

2

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 12

but it's not a leaderboard in the context of a scientific experimental paradigm, since no scientific hypothesis needs to be stated anywhere to participate. who knows what these models are, how they were trained, etc.?

1 2

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 12

LMArena is a nice way to identify models in real time that are preferred by participants to this scheme. totally legit under this context.

...

1 1

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 12

burning bridges!

Burn GIF

ALT: Burn GIF

media.tenor.com

2

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 12

slide deck: drive.google.com/file/d/1rJEP...

Reality Checks.pdf

drive.google.com

1 3

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 12

recently gave a talk on <Reality Checks> at two venues, and discussed (and rambled) about how leaderboard chasing is awesome (and we want it to continue) but that this isn't easy because everyone (me! me! me!) wants to write more papers.

the link to the slide deck in the reply.

3 5 24

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 11

can’t really vibe code healthcare. a thoughtful and experience-rich talk by amazing Lavender Jiang at NYU Global AI Frontier Lab!

1

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 8

github: github.com/kyunghyuncho...

GitHub - kyunghyuncho/AdReader

Contribute to kyunghyuncho/AdReader development by creating an account on GitHub.

github.com

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 8

vibe coding continues with a MS Edge extension: the world's most expensive, slow and inaccurate ad blocker, featuring Gemini!

3 3

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Aug 7

the perfect response from Google cannot be simpler: just halve the price of everything.

please start with the Gemini Pro subscription. pretty please?

4

Reposted

Andrew Lampinen @lampinen.bsky.social · Aug 5

These biases can lead to dramatic downstream effects that cause unexpected conclusions from analyses. For example, RSA may identify two models computing the same, complex task as much less representationally-similar than either of them is to a model computing a much simpler task (right panel)!

RSA within and between different sets of models can give surprising results due to representation biases. This plot shows similarities within and between different models computing different types of features. Ideally the similarities would be highest in blocks on the diagonal (i.e. models computing the same features), and the blocks off the diagonal would show graded similarity corresponding to the functional overlap. However, that is not the case. (Left) When comparing a model trained to output both easy and hard features to ones that are trained on only one feature, the multi-task model appears very similar to the easy-task only model (cf. Hermann and Lampinen, 2020). In fact, the models trained only on the hard task do not even appear particularly similar to other models trained on the same exact task. (Right) When models are trained on multiple easy or multiple hard tasks, the models trained on only hard tasks appear less similar to other models trained on exactly the same tasks than they do to models trained on strictly easier tasks that use the same input units.

1 1 6

Reposted

FuriosaAI @furiosa.ai · Jul 29

🛬 Next week, we’re hosting our inaugural happy hour: Powered by Furiosa.

🍹 On August 8, we’re bringing together industry leaders and practitioners at the forefront of AI to discuss, share knowledge, and enjoy great food and drinks.

Powered by Furiosa | LinkedIn

Join us at our inaugural happy hour: Powered by Furiosa. This event will bring together industry practitioners and leaders at the forefront of AI inference. Connect with NYC researchers, engineers, a...

www.linkedin.com

1 1

Reposted

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · Jul 25

You know how RL is that whole big thing nowadays?

Present your work at the first-ever New York Reinforcement Learning Workshop (NYRL), co-organized by Amazon, Columbia Business School & NYU Tandon School of Engineering.

ny-rl.com!

1 4 26

Reposted

NYU Center for Data Science @nyudatascience.bsky.social · Jul 25

CDS professors have launched the Global AI Frontier Lab in Brooklyn with Korean partners like KAIST.

Profs @kyunghyuncho.bsky.social and @yann-lecun.bsky.social lead work on fundamental AI, responsible AI, and healthcare, and a biweekly seminar series.

nyudatascience.medium.com/building-bri...

Building Bridges Across the Pacific: How CDS Launched a New Model for International AI Research…

CDS has launched a new research lab in Brooklyn in partnership with Korean institutions, leading cross‑Pacific collaborations in AI.

nyudatascience.medium.com

1 1

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Jul 21

arxiv.org/abs/2104.01112 [2104.01112] NaturalProofs: Mathematical Theorem Proving in Natural Language

NaturalProofs: Mathematical Theorem Proving in Natural Language

Understanding and creating mathematics using natural mathematical language - the mixture of symbolic and natural language used by humans - is a challenging and important problem for driving progress i...

arxiv.org

kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · Jul 21

😂 i think this “challenging problem” may have been finally solved after five years.

===
Understanding and creating mathematics using natural mathematical language … used by humans is a challenging and important problem for driving progress in machine learning.
===

3 1 16