@TheSpeculator0 Landlady won't stand it, I'm afraid. Also much easier to plan a trip to Asia for the 3-4 weeks of the worst heat 🤷♂️
English
Indraneil Paul
169 posts
Indraneil Paul
@androneil54
@ELLISForEurope PhD Candidate @UKPLab @TUDarmstadt ex. Research Intern @awscloud Working on RL + Code-Gen + Function Calling
Berlin, Germany Katılım Mayıs 2018
286 Takip Edilen147 Takipçiler
Places where it's warm for a long time all have AC units. Countries like Germany etc. have lower AC adoption because a lot of people are completely fine being slightly warm in their home for like 3 weeks a year.
Especially because the homes are generally much better insolated than in the US.
Antonio García Martínez (agm.eth)@antoniogm
This is counter-intuitively the correct take. All the Ameribros with their hur-dur Europoor takes are so provincial they’ve never been to Southern Europe or any developing country where there’s an AC compressor bolted to every balcony. It’s not about money, it’s a pure cultural artifact that stymies AC adoption, which is why it’s so fascinating.
English
@TheSpeculator0 Yes, and this is just the end of May. Clearly not ideal to have to stay at work till 2000 as the offices are one of the few places that have bucked the Euro aversion for AC!
English
@androneil54 which means in a normal insolated home it's like 25. Do you really need AC for that as an able bodied man
English
@Moebius1974 @TheSpeculator0 I live on the top floor of my (rather old) building and plenty warm, Thanks!
English
@androneil54 @TheSpeculator0 If we're lucky, yes we get three weeks of great summer weather. But it has been a couple of years ...
Tell global warming to hurry up will you?
English
Neat extension to the Physics of LM/Quark style quality prefix tags by prefixing NL feedback (I suspect this is common in frontier labs).
Also backs up my experience on long-context code training, where keeping low-quality or forked repos with metadata beats filtering them out.
Prithviraj (Raj) Ammanabrolu@rajammanabrolu
Ever wished we had fewer X-training hyphenates? Pre, mid, post etc. Why not just Training? Trying to bridge the divides (and get all our friends into one team again), we intro *Introspective X Training*, an offline RL inspired method that scales effectively across any LLM stage by annotating your data with a thinking reward generated language critique! Up to 2.8x FLOP efficiency + 5-10 point score gains (esp with math and code) at any stage from scratch to 24T tokens on 8b (active) sized models!! We burned much compute ablating so you wouldn't have to Moral of the story is‼️don't throw out any data via filtering, just feedback condition it‼️ You can spend FLOPs up front on inference to *classify* data quality and then train so that tokens aren't all treated equally based on the feedback starting early in training itself. Right now they're really only separated out much later during mid/post training This improves overall compute efficiency and gives us benchmark perf not possible with just baseline methods! Paper here: arxiv.org/abs/2605.20285 Thanks to @BrandoCui and @GXiming for leading this w/ @__SyedaAkter @davidjesusacu @hyunw_kim @jaehunjung_com Yuxiao Qu @shrimai_ @YejinChoinka
English
An absolute lifesaver for anyone who's lived in Berlin. The future of the internet as a source of info may be bleak if the ladder gets pulled on anyone trying to replicate something like this today.
All About Berlin@aboutberlin
AI is killing All About Berlin. When you Google something, you used to get a link to my website, but now you get an AI-generated answer trained on my work. This has a devastating impact on traffic.
English
@tatsu_hashimoto Interesting result backing my hunch that a lot of pre-training data is too sanitized!
I wonder if you guys have tried the "Physics of LM" approach of marking highly curated and unfiltered data with dedicated, distinguishing tags to improve robustness without succumbing to noise.
English
Some new results I found surprising that I’m tweeting for Chris (who isnt on here). With enough compute, the best data filter for LMs (on DCLM) might be no filter. Why? Large models can tolerate a surprising amount of nominally 'low quality' data, and can sometimes even benefit.
English
Indraneil Paul retweetledi
For the past 12 years, cuDNN has been completely closed sourced (besides the .h files), until this week! OVER 20 MoE kernels & NSA sparse attention kernels from cuDNN has been open sourced! Great work to @manicely6005 & the rest of the team on seeing that parts of NVIDIA are moving towards open kernels! open source kernels drive innovation! (1/3) 🧵
English
@mgill25 The new EU defamation law has been badly misused in the country. At this point, I don't trust any public ratings in Germany.
English
Germany can never progress because you get sued the shit out of you for a 1-star review on Google Maps
English
@YouJiacheng @nvidia Even the Nemotron Code datasets were just fake open-sourced on HF and gated 🤦♂️
They won't accept any access requests.
English
everything is good, except the license.
@nvidia come'on, you sell GPUs, free data help your business. CC-BY-NC is ridiculous.
Shizhe Diao@shizhediao
Time to upgrade your pretraining dataset. Instead of FineWeb-EDU / DCLM / X, try ClimbMix-400B. 📄 Paper: arxiv.org/pdf/2504.13161 📦 Data: huggingface.co/datasets/nvidi… CLIMBMix uses clustering-based iterative data mixture to improve pretraining efficiency and data quality. Would love to see the community experiment with it and push it further 🚀
English
Indraneil Paul retweetledi
Indraneil Paul retweetledi
Releasing the official SkyRL + Harbor integration: a standardized way to train terminal-use agents with RL.
From the creators of Terminal-Bench, Harbor is a widely adopted framework for evaluating terminal-use agents on any task expressible as a Dockerfile + instruction + test script.
This integration extends it: the same tasks you evaluate on, you can now RL-train on.
Blog: novasky-ai.notion.site/skyrl-harbor
🧵
English
Indraneil Paul retweetledi
Highly underrated aspect of Japan is that on street parking is just straight up banned everywhere.
Jonathan Berk@berkie1
Under Mayor Anne Hidalgo, Paris has created nearly 300 “rues aux écoles in an effort to; improve air quality, reduce crashes, and give kids more safe spaces in their neighborhoods to walk, bike, play and just be kids. 🇫🇷
English
Indraneil Paul retweetledi
In my recent blog post, I argue that "vision" is only well-defined as part of perception-action loops, and that the conventional view of computer vision - mapping imagery to intermediate representations (3D, flow, segmentation...) is about to go away.
vincentsitzmann.com/blog/bitter_le…
English
Reads like a neat way to exploit the fact that in high dimensions almost every thing is nearly orthogonal to everything else.
Also by virtue of being in spherical space, better placed to take advantage of any dimensional collapse present in the embedder: arxiv.org/abs/2110.09348
Jina AI@JinaAI_
Convert your embeddings to spherical coordinates before compression - this trick cuts embedding storage from 240 GB to 160 GB, and 25% better than the best lossless baseline. Reconstruction is near-lossless as the error stays below float32 machine epsilon - so retrieval quality is preserved perfectly. Works across text, image, and multi-vector embeddings. No training, no codebooks.
English
To make a bit of an excuse for Microsoft: the world is just waking up to the fact that coding agents are general agents.
It’s bitter lesson adjacent: Writing and executing code will likely outperform years of handcrafting vertical-specific agents with expert knowledge.
Actually it might exactly map in bitter lesson: Program synthesis is a form of scalable search.
English
@amasad @GavinSBaker Fully on point! This lesson keeps being re-learnt (like the advent of VLAs in robotics).
English
Claude Cowork is what Copilot should have been.
Evidently built in 10 days with Claude Code while Microsoft has been working on Copilot for years.
English
@orionweller @huggingface I can't believe I'm saying this but I'm going to have to check out Modelscope 😅
(ms-swift is legit nice tho!)
English
@androneil54 Sadly I have never gotten an answer from @huggingface. It seems perhaps they are trying to juice revenue numbers by squeezing it out of us?
I didn’t want to believe that but the lack of communication despite repeatedly reaching out leaves little other explanations :(
English
Got blindsided by this as well! It keeps creeping down without announcement irregularly. Last I saw it was 3.8TB and today I see it is 1TB.
Only realized this was the culprit when I couldn't even upload some of my public model checkpoints 🤷♂️
Orion Weller@orionweller
Anyone else have this? @huggingface has been shrinking my storage at the end of each month and charging me $100+ I then delete data to meet the limit but the next month it lowers/charges me again. When I emailed they didn’t give a straight answer on the storage lowering.. @Thom_Wolf @ClementDelangue would be nice to get an email just telling us straight up the max we're allowed and the period we have to get to that number. Right now it seems like you’re shrinking our storage at the last minute to “catch” and charge us
English
@giffmana Loved their enroot and pyxis additions to the ecosystem tho. Whether it is entend and extinguish, time will tell 🤷♂️
English
Yo what the heck?! Wasn't on my bingo card at all.
Not sure what to make of it. I haven't been fan of slurm, and am really on the fence of whether this means improvements are coming, or the lock-in is slowly starting...
Time will tell i guess!
English