Tian and Karolina and team are at ICLR. Come say hi.

Tian Jin@tjin.bsky.social · last yr.

📣 The Journey Matters: Our #ICLR2025 paper shows how to pretrain sparse LLMs with half the size of dense LLMs while maintaining quality. We found that the average parameter count during sparse pre-training predicts quality, not final size. An MIT/Rice/Google/ISTA collab 🧵 1/N

So who’s read the 2027 article? What do you think?

Someone has suggested I check out bsky again. So I'm back looking around here. Notification list is kinda boring. So any good conversations going on? Perhaps about LLM/AI reasoning?

Anyone else have the worry that a lot of LLM research is .... just bad psychology?

Okay, so just a few* thoughts (*this got longer as I wrote 😅….long thread)-

Mary Beth Hunzaker@mbhunzaker.bsky.social · 2y ago

Having a lot of thoughts & feelings today as someone who’s worked both on FBs misinfo interventions (back when they were making investments) and Birdwatch (pre-Musk Community Notes) 💔. Debating if a thread is worth the inevitable headache 😓

I got to ski Revelstoke this winter break. Couple observations: the price of receiving 600 cm of snow by Jan 8 is that it is constantly snowing. Saw almost no sun the whole time and the peak was often in whiteout conditions (though North Bowl was always clear…). See image for more.

Multiple friends have likely lost their homes in Los Angeles. Can’t imagine how disorienting this would be. They had only minutes to flee and grab belongings.

OK. Practical question times. How are you adjusting your research given progress in reasoning style models? Also how are you adjusting the way you work?

Pretraining is not done. It's just that theorists haven't told the hackers how to do it better.

Nathan Lambert@natolambert.bsky.social · 2y ago

ILYA: "PRETRAINING IS DONE. WE ARE NOW IN THE POST TRAINING ERA."

Now that I've had a taste of X without post length limitations, I've got to say that it is quite annoying have to fit tweets into 256 characters here on bsky. On X, when they get to long, they go below the fold, and so you're still incentivized to make it short. Can't we have that here?

I've now read this paper carefully if anyone wants to discuss it.

Dan Roy@roydanroy.bsky.social · 2y ago

Link: arxiv.org/abs/2412.024...

I've never understood people who block people on Twitter, outside of cases of harassment. Such weak sauce.

Has anyone out there compared the GPU/memory specs for the new Macbook Pro M4 Pro/Max computers to the requirements to run inference in the more interesting open source LLM models, especially those aimed at reasoning? [1/2]

Has anyone read the new Shalev-Shwartz paper on PAC reasoning yet and want to discuss?

*AI for Mathematics and Theoretical Computer Science*, a workshop hosted jointly by the Simons Institute for the Theory of Computing and the Simons Laufer Mathematical Sciences Institute, will be held in Berkeley at April 7-11, 2025. simons.berkeley.edu/workshops/si...

Simons Institute for the Theory of Computing and SLMath Joint Workshop: AI for Mathematics and Theoretical Computer Science

This is an exciting time for mathematics, as new technologies for mathematical reasoning provide novel opportunities for mathematical research, communication, and discovery. Mathlib, a library of form...

simons.berkeley.edu

Huge community response to this. Much appreciated.

Dan Roy@roydanroy.bsky.social · 2y ago

My 9 yo has finished all the Harry Potters. Any suggestion on where to go from here? Maybe age appropriate sci fi?

My 9 yo has finished all the Harry Potters. Any suggestion on where to go from here? Maybe age appropriate sci fi?

Poem by Joseph Fasano

Congrats to Antonio Lijoi on being elected President of the International Society of Bayesian Analysis!

NeurIPS Test of Time Awards: Generative Adversarial Nets Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio Sequence to Sequence Learning with Neural Networks Ilya Sutskever, Oriol Vinyals, Quoc V. Le

As the semester is coming to an end, I've been able to steal some time to do research on my own. Such a buzz.

School’s out for summer. School’s out for ever!

“Next week we will derive [something awesome],” bragged Prof Roy. “This is our last class, Professor!”, said the brave student.

Late Sunday music.

Waterloo is facing a 75 million CAD deficit and so they are implementing a hiring freeze. As one of the world’s top universities, it is ridiculous that Ontario is not funding it fully. It has generated so much SV wealth, I suggest a SV billionaire solve this deficit and then some.

Hot take: Working another weekend this semester, I am left wondering: Why does the #NLProc publication process have so much artificial urgency baked into it? Requiring responses within 24h, emergency reviewers, everything urgent. Like, why? Slow science can also be a community goal.

Has anyone asked whether scaling laws are surprising or not? (Not only to meat puppets, but surprising in the real sense of "it could have been different".) Universal learning rates look like scaling laws in the non-extreme case, where the class isn't trivial or impossible. No compute there though.

Peter Loeb has passed away. Loeb made fundamental contributions to potential theory as well as the application of nonstandard models to mathematical analysis. [1/3] www.legacy.com/us/obituarie...

Peter Loeb Obituary (1937 - 2024) - Urbana, IL - The News-Gazette

View Peter Loeb's obituary, send flowers and sign the guestbook.

legacy.com

Hey... anyone know w.t.f. runs BlueSky? Does this sit on top of mastodon? Can we edit the software? How do we tweak our feeds? E.g., if I'm in the mood for ML or for politics or whatever?

Reviewing is dead. I just read some ICLR reviews. Absolutely junk. No insight. Written by humans, but could have been written by a chatbot or teenager who had read a few dozen reviews. Why do we think this is useful? Only 5% (if that!) of the community has worthwhile opinions.