BusyBrain⌨

Happy Friday everyone! I just posted what I think is an important blog post on my website. It is a critique of meta-meta-analyses: meta-analyses of meta-analyses. Link: matthewbjane.github.io/blog-posts/b... #stats #metascience

1/🧵ICLR 2025 Spotlight Research on LM & Memorization! Language models (LMs) often "memorize" data, leading to privacy risks. This paper explores ways to reduce that! Paper: arxiv.org/pdf/2410.02159 Code: github.com/msakarvadia/... Blog: mansisak.com/memorization/

Microtubule regulation drives an asymmetry in the regeneration of sensory neurons, with specific proteins controlling growth. buff.ly/4ijksHC

1/13 New Paper!! We try to understand why some LMs self-improve their reasoning while others hit a wall. The key? Cognitive behaviors! Read our paper on how the right cognitive behaviors can make all the difference in a model's ability to improve with RL! 🧵

Reading and Writing Google Sheets in DuckDB duckdb.org/2025/02/26/g...

www.nature.com/articles/s41... awesome new work out today! From the Lee lab in the intramural research program at NIMH!

Our online book on systems principles of LLM scaling is live at jax-ml.github.io/scaling-book/ We hope that it helps you make the most of your computing resources. Enjoy!

Excited to share 𝐈𝐧𝐟𝐀𝐥𝐢𝐠𝐧! Alignment optimization objective implicitly assumes 𝘴𝘢𝘮𝘱𝘭𝘪𝘯𝘨 from the resulting aligned model. But we are increasingly using different and sometimes sophisticated inference-time compute algorithms. How to resolve this discrepancy?🧵

what kind news is this? lol

On monday in our reading group we discuss "Flow Matching with General Discrete Paths: A Kinetic-Optimal Perspective" arxiv.org/abs/2412.03487 With Neta Shaul. Join on zoom on Monday at 9am PT / 12pm ET / 6pm CET: portal.valencelabs.com/logg

🚨 Researchers uncover 4.5M fake stars on GitHub 🌟, often boosting malware disguised as pirated software & crypto bots. Fake stars surge in 2024, posing major risks to open-source trust & security. #CyberSecurity #GitHub #OpenSource #SupplyChainSecurity arxiv.org/abs/2412.13459

That exhilarating feeling that *everything is possible* when you open an editor to code, it hopefully never goes away.

Binary code is pervasive, and binary analysis is a key task in reverse engineering, malware classification, and vulnerability discovery. So, they created Assemblage - the dataset of source-to-binary projects compiled from GitHub. Assemblage - A dataset of binary executable corpuses

this morning walk, an ideas stuck me: can you play chess on Rubik's Cube (does not have to be 3x3 one)? not just chess with 6 sides, but normal chess board abstracted away to Rubik's Cube representation and operation

Half of Twitter right now is people getting mad at some random lady that got a literature PhD. Seems a bit crazy to get so mad about, but I do agree woke academia has become silly and we need to go back to when it was about real solid research, like measuring skull sizes to determine personalities

The amazing, new Qwen2.5-Coder 32B model can now write SQL for any @hf.co dataset ✨

Good news everyone! A new version of graph-tool is just out! @graph-tool.skewed.de graph-tool.skewed.de Graph-tool is a comprehensive and efficient Python library to work with networks, including structural, dynamical, and statistical algorithms, as well as visualization. 1/N #networkscience

An aspect of flow matching which I find a bit interesting is that it is covariant under affine changes of coordinate (c.f. optimal transport, which need not be). This allows for a few nice WLOGs, which I imagine have more applications than I realise.

If you think the out (site) group isn't enjoying thinking like your ingroup, I've lost respect for you. Sorry.

More formal verification, this time from the engineers at Cloudflare using a lesser-known verification stack: Cloudflare uses racket & rosette, a solver-aided programming system to, ensure the correctness of their DNS query engine configuration blog.cloudflare.com/topaz-policy...

Some recent discussions made me write up a short read on how I think about doing computer vision research when there's clear potential for abuse. Alternative title: why I decided to stop working on tracking. Curious about other's thoughts on this. lb.eyer.be/s/cv-ethics....

Done, Five million. Compressed to 447 mb. Getting more on the main set, it's already at 7+ million. There is a surprisingly good amount of quality data. There will be more curation/variations but won't post more about it unless I manage a billion. huggingface.co/datasets/Ror...

Very useful looking Python package by BASF on evaluating MLIPs! github.com/basf/mlipx

today feels like a good day to pick this book up again. be back in a while 👋 also shoutout to @welltypedwit.ch for always teaching me new stuff

Qi meeting notes https://github.com/drym-org/qi/wiki/Qi-Meeting-Nov-29-2024 -producers, transformers, and consumers -- stream abstractions for deforesting any list-oriented operation -encoding two runtimes - naive and optimized - in "deep" macros -Michael is teaching an exciting new course on DSL…

A few days ago Lawrence Hollom posted a preprint that solves a problem about posets, informally known as the fishbone conjecture, which I am told was extremely central in the area. 🧵 arxiv.org/abs/2411.16844

I am running my statistical rethinking course again starting in January. But this time just for Leipzig locals, so I can work with a smaller group this year and track individual progress better. I think that will help me tune the material, espeically homework, better. Materials still open however.

Every time I'm having a hard time to motivate myself to work on things I just grab my laptop and move to a different room. It kinda resets my brain 'cause the environment has changed. I heard people recommend to rearrange your furniture from time to time for the same reason.

Dear algorithm, I would like to view: 70% new ML and LLM research and cool results. 10% funny videos with cute animals. 5% sports (but no spoilers if I'm planning to watch the reply later). 5% travel and life hacks. 5% general tech. 5% random. Regards,

When can a sum of reciprocals of natural numbers sum to a rational number? There are many unsolved problems in this area, by Erdős and others. With Vjeko Kovac, we have been able to resolve some open questions and make progress on others: terrytao.wordpress.com/2024/11/27/o...

wow, nobody steal this promo link. BSky high trust level looking good

Hey #EconSky! I’m a postdoc in psychology and economic theory @ Harvard My #EconJMP studies how shared context supports efficient communication and cooperation 🧵⬇️

An AI researcher that wants to stop big tech owning everything was permabanned here for releasing a dataset of 2M posts. A librarian received death threats for a 1M post dataset. The EU funded the creation of this dataset of 235M posts months ago, and… nothing? zenodo.org/records/1108...

This Thanksgiving, I'm grateful for peace, health, and happiness for my family and me—what are you thankful for?

Thank you @bsky.app team for correcting the mistake. Glad to be back!

"You just ~20 textbooks away from your cracked friends?" what if I told you that what matter is NOW! what you can figure out at real time challenge the first time you face it? when you notice "huh, I never thought about this before" that's what matter the most. Tune in, control yourself

hearing a lot of techno-ableism issue lately here. I just found one book on SERP with that word. I guess this is what I gonna read this weekend. www.google.com/books/editio...

hahahahhahaah holy cow, paligemma is good =)

Bacteriorhodopsin is a membrane-bound light-driven proton pump. Using neural networks, it's possible to redesign it so that it's soluble but retains its photocycle! @ivangushchin.bsky.social www.biorxiv.org/content/10.1...

FYI, here's the entire code to create a dataset of every single bsky message in real time: ``` from atproto import * def f(m): print(m.header, parse_subscribe_repos_message()) FirehoseSubscribeReposClient().start(f) ```

Yeah, That's what they want – for us to give up on this platform. That's a huge loss for everyone. This is the kind of open web we've always dreamed of. We shouldn't bother with a bunch of angry people who don't understand open protocols tell you what to do

people on this platform will take your words out of context, twist, not mention your correction, because they just want to hate on what you work on, and insult you comfortably. I'll keep posting here about my work but will not be interacting with anyone who wants to bash on my company.

One of my favourite data discoveries this year: Google's mind-blowing ARCO-ERA5 dataset: hourly data for ~300 climate variables, available globally from 1940! 🤯 Loadable with a single line of Python code from a single cloud-friendly Zarr file! Below: a month of wind waves + swell: 🌊

Jeovane Honorio Alves, Radu State, Cinthia Obladen de Almendra Freitas, Jean Paul Barddal LongKey: Keyphrase Extraction for Long Documents https://arxiv.org/abs/2411.17863

framing atproto scraping as an issue of “consent” is really offputting because it feels like an attempt to pull norms around bodily autonomy into a discussion about data use. but someone’s posts are not in fact an extension of their body and those norms are not appropriate.

the people hating on Hugging Face are destroying any chance society has of a future where technology is created for the public good. People are doing crazy shit with your data, in private, for years. This is the first time we're starting to get these tools in the hands of good people, you & me