Aleksandr Drozd (@bkbrd) header
Aleksandr Drozd (@bkbrd) profile picture

Building stuff. Writing about everything. AI and HPC among other things. Based in Japan for past 17y. Raising two mixed-race kids. にゃ ^_^

Tokyo-to, Japan
Joined August 2009
Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)Photo by Aleksandr Drozd (@bkbrd)
Aleksandr Drozd retweeted
📢 #NLProc I'm hiring! 2x PhD students (apply by Nov 22) & postdoc (Nov 15). These positions are part of a large project on the problem of attribution for LLM outputs to the original data sources, on which LLMs were trained. Job links on my website (in the bio) /1
3
48
7
173
31,524
Aleksandr Drozd retweeted
Great news! Based on requests, the submission deadline for the LLMxHPC workshop (@ IEEE #Cluster) has been extended to June 29, 2024. Submit your novel research on using #HPC to advance #LLMs and using LLMs to improve the design/operation of HPC systems. llmhpc.github.io/2024/
1
2
177
Aleksandr Drozd retweeted
arxiv.org/abs/2404.08676 🚨Introducing ALERT🚨 , a large-scale safety benchmark of > 45k instructions categorized using our new taxonomy. ✨Using prior work such as Bo Decoding Trust and our Aurora-m huggingface.co/datasets/auro… and @jackclarkSF's awesome Anthropic-HH. 🙏
2
17
1
53
13,305
#hpcmyths 6: We Will Soon Run at Zettascale! Read our discussion in @ProfMatsuoka et al. arxiv.org/abs/2301.02432 and discuss here at the #HPC twitter! Will we reach zettaflop/s performance or will fp64 lose relevance before?
38%zettaflop in fp64 it is
39%fp64 will not be relevant
23%fp64 is irrelevant now
69 votes • Final results
1
3
6
2,385
Aleksandr Drozd retweeted
The 'Myths' paper is out: Satoshi Matsuoka, Jens Domke, Mohamed Wahib, Aleksandr Drozd, Torsten Hoefler. "Myths and Legends in High-Performance Computing" arxiv.org/abs/2301.02432" undergoing peer review. Hopefully the community will be engaged in yes/no discussions on Twitter!
17
107
22
359
123,154
Aleksandr Drozd retweeted
📢 #NLPaperAlert: Outliers Dimensions that Disrupt Transformers Are Driven by Frequency with @gpuccetti92 @bkbrd Felice Dell'Orletta TLDR: remember how BERT over-relies on just 48 magic params? They're related to token frequency in training! /1 arxiv.org/abs/2205.11380
4
5
39
Aleksandr Drozd retweeted
.@aclmeeting will soon be over, but the fun continues. The Third Workshop on Insights From Negative Results is happening tomorrow! Here's what's on the menu: insights-workshop.github.io/… /1
1
5
3
34
Aleksandr Drozd retweeted
[#CFP Extended] 4th Workshop on Benchmarking in the Data Center: Expanding to the Cloud @ PPoPP 2022 Submit your novel/early results on benchmarking, performance optimization, et al. by Feb. 14, 2022. Check out full topics list: parallel.computer/ #HPC #Cloud #datacenter
3
1
Aleksandr Drozd retweeted
📢 The 2nd CFP for insights-workshop.github.io is out! About one month left to submit a short paper about NLP things that don't work / reproduce / generalize. Deadline for submissions to be reviewed directly by the workshop: Feb 28 Deadline for papers with ARR reviews: Mar 21 /1
3
13
2
45
Aleksandr Drozd retweeted
#NLPaperAlert: Generalization in NLI: Ways (Not) To Go Beyond Simple Heuristics With @prajjwal_1 @bkbrd, accepted by insights-workshop.github.io arxiv.org/abs/2110.01518 A proud example of a paper genre I'd call "research realism": a mix of negative & (cautiously) positive results /1
3
7
1
33
The 3rd R-CCS Symposium is online starting now, and I will give a keynote talk starting from 8:25 AM for 40 min or so. The symposium is online and free to attend, and we really have some exciting sets of talks on the cutting edge HPC on #Fugaku. r-ccs.riken.jp/R-CCS-Symposi…
1
3
1
11
Aleksandr Drozd retweeted
Congratulations to @riken_en & @fujitsuhpc for again taking the top position for HPL, HPCG and HPC-AI - the triple crown of supercomputing! #HPC #AI #SC20
1
12
2
23
Aleksandr Drozd retweeted
The final version of our BERTology primer is out (to appear in TACL), updated from 40 to 150 papers between March and October. Check it out quick, before it gets outdated again next week at #EMNLP2020! The multilingual section is now only available in v1. arxiv.org/abs/2002.12327
2
39
3
177
We just announced the Fugaku config at a press briefing. The whole system is 158,976 A64FX nodes, and running at 2.2Ghz peak performances for 64bitFP/32bitFP/16bitFP/8bitINT are 537Peta/1.07Exa/2.15Exa/4.30Exa (FL)ops respectively, as well as the total memory BW 163 PetaByte/s.
12
92
26
263
Aleksandr Drozd retweeted
New #NLProc blog post: crusade against reject-if-not-SOTA conference reviews! hackingsemantics.xyz/2020/re… TLDR: SOTA performance is just one way in which a system could be interesting. Conference guidelines should clearly include others, such as efficiency and generalizability. /1
8
31
8
92
Aleksandr Drozd retweeted
Something is wrong with my news feed today. - Koalas are burning alive as you are reading this - New Samsung phones are coming, and are obviously the best use of your money Dear Aussies, any donate links recommendations? The koalas seem to be low on PR staff. @ivrik @eltimster
2
1
4
Aleksandr Drozd retweeted
The code for hosting papers on the excellent @aclanthology is open source and free to use: github.com/acl-org/acl-antho… Does anyone want to do a global search and replace of ACL to ACM to show @TheOfficialACM how they could easily host all their papers open access for virtually free?
1
31
2
148
Free and open-source textbooks on calculus, linear algebra, differential equations, real analysis, abstract algebra, combinatorics, and introduction to proof -- a partial list, helpfully compiled and curated by Dana Ernst @danaernst. danaernst.com/resources/free…
12
357
11
976
Aleksandr Drozd retweeted
New tech report with Junghyun Min and @TalLinzen: "BERTs of a feather do not generalize together" Across 100 re-runs, BERT fine-tuned on MNLI has a consistent score on MNLI but extreme variation in syntactic generalization (measured w/ HANS). Link: tommccoy1.github.io/BERTs_of… 1/7
1
17
6
86
Aleksandr Drozd retweeted
Applications still open this week for Postdoc positions at Cross Labs, for research in AI, Neuroscience and Artificial Life! Informations: crosslabs.org/jobs Deadline: Friday Oct 31, 2019 at midnight (AoE) crosslabs.org/jobs