📄🔈✨ 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

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.

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

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!

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...

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

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.

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

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.

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

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

the perfect response from Google cannot be simpler: just halve the price of everything. please start with the Gemini Pro subscription. pretty please?

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)!

🛬 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.

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!

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...

😂 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. ===

thanks all for the interest! i have put the postdoc in touch with many amazing folks.

japan’s let-them-eat-cake moment: which japanese politician will say “let them eat cake”?

ok truly done with my two weeks of vibe-coding (catching up with the trend): the final day ended with local inference using gemma3 via ollama for the medical advisor app for Mac OS. it's slow (bc my laptop) but it works to some extent. local models, local inference and privacy will win!

EPIC never sent me any offer, not to mention a zuck-level one, and thus, out of spite, i've tried to build a Mac native app of my personal medical advisor. still waiting for a call with a zuck-level offer from EPIC 🤑

some people are way too comfortable with unfalsifiable assumptions over empirical results or simulation based results. sad

yet another day of vibe-coding: people use LLM as their bespoke medical advisor. what if this advisor can be super-charged with your own medical records? download your medical records from Epic App, ingest them into db, and have Gemini answer your question based on your own medical records for you!

to cap the weekend of vibe coding, i asked Gemini to create a web app version, and it did it perfectly in one go.

vibe-coding day 2: i decided to rewrite the app from yesterday using a more native app platform. never used xcode nor swift before ... and voila. two hours and here's a full-blown app. am i now a swift developer?

since vibe coding is all the rage, i vibe coded an absolutely useless app today with Gemini. `gemini-vision` lets you capture a screen and query Gemini to answer your question about the captured screen. vibe coding is fun, but i infinitely prefer bbq over vibe coding

some educational systems are exceptionally good at producing cheaters.

Pearson's RNGs: www.argmin.net/p/sounds-lik... That said, I could make the case that computers today wouldn't look the way they do without the Statistics revolution of the 1930s.

"But I’m particularly fascinated by the weird obsession with statistics" amen

really, apple? this is the most important thing to say about this gigantic update?

i was asked by a few (inc. @yuanqingwang.bsky.social) what i meant by this earlier tweet, and since i'm pretty busy, i decided to write a blog post to answer the question 🤦

can someone please QAT gemma-3n?

i have the perfect solution to the nuclear power plant debate: build permanent high-grade nuclear waste facilities in Aspen and Park City. both have low pop density, solid foundation, good road access and are relatively remote. you’re welcome

wow … we could’ve collectively saved so much time of reviewers only if this paper from 2005 was more widely known.

i was finally convinced by sungmin and sangheum to work on unlearning. the first we did (or i learned) together with sungjin and dason was to learn how people evaluate unlearning, and horror 😱 here is our short writeup and report on what we found and what we propose as a remedy.

the new administration in Korea has managed to bring onboard Jungwoo as the secretary of AI and future technologies to directly report to and advise the President. some real AI researcher, developer and director! a great job!

finally, wind is changing its direction: causal inference becomes easier if we give up on designing a new estimation algorithm ourselves (i don't think we've evolved to do so ourselves well.) let learning find one for you!

“every bar is local” - Henri Dwyer (2025)

ai for drug discovery looks a lot like machine translation research/development during the cold war.

is consulting as a job finally at its end? 🥳 for some uninteresting reasons, i thought it would be interesting to find all AI-related courses at NYU and create a report on AI education at NYU, but i'm lazy. (1/3)

CDS PhD graduates @activelearner.bsky.social & Omar Mahmood, with CDS Prof @kyunghyuncho.bsky.social, former CDS co-Director @richbonneau.bsky.social, and others developed a machine learning system that achieved up to 100x improvements in antibody binding. nyudatascience.medium.com/making-drugs...

you can say the authors of <BLEND: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages> were dedicated by looking at Table 1.

society has decided we need to invest in long-term science for the benefit of all not just shareholders or rich patrons, and universities have agreed to step up beyond education to provide venues for conducting science on the society’s behalf via federal funding based on tax.

a thought provoking and also much better grounded interview of Mina Lee at U Chicago about AI and its impact (in korean): www.newstheai.com/news/article...

BREAKING: single-issue voters continue to be punched in their faces over and over again.

it's been more than a decade since KD was proposed, and i've been using it all along .. but why does it work? too many speculations but no simple explanation. Sungmin and i decided to see if we can come up with the simplest working description of KD in this work. (1/4)

don't try to make a consensus about the future. future is uncertain.

had an interview on AI in korean.

together with Korea's Ministry of Science and ICT, NYU recently opened the Global AI Frontier Lab to facilitate international collaboration on AI. as part of this newly established lab, we've been hosting a seminar series, and i share our experiences running this series as a blog post.

Gemma-3 is very humble and never recommends itself