Thennal D K
@thennal.bsky.social
cs undergrad student, nlp researcher, fan of fungi
any/all, english/മലയാളം/日本語
https://thennal10.github.io
any/all, english/മലയാളം/日本語
https://thennal10.github.io
Reposted by Thennal D K
What do you do after you’re done jumping the shark?
Whatever it is, Nature Careers is all in.
Whatever it is, Nature Careers is all in.
November 11, 2025 at 10:41 PM
What do you do after you’re done jumping the shark?
Whatever it is, Nature Careers is all in.
Whatever it is, Nature Careers is all in.
Stop using Word Error Rate! The lovely @sthirkal.bsky.social and I made a poster for my recent NAACL 2025 Findings paper, highlighting the issues in multilingual ASR evaluation and proposing viable alternatives.
April 8, 2025 at 2:35 PM
Stop using Word Error Rate! The lovely @sthirkal.bsky.social and I made a poster for my recent NAACL 2025 Findings paper, highlighting the issues in multilingual ASR evaluation and proposing viable alternatives.
Reposted by Thennal D K
Modern LLMs "speak" hundreds of languages... but do they really?
Multilinguality claims are often based on downstream tasks like QA & MT, while *formal* linguistic competence remains hard to gauge in lots of languages
Meet MultiBLiMP!
(joint work w/ @jumelet.bsky.social & @weissweiler.bsky.social)
Multilinguality claims are often based on downstream tasks like QA & MT, while *formal* linguistic competence remains hard to gauge in lots of languages
Meet MultiBLiMP!
(joint work w/ @jumelet.bsky.social & @weissweiler.bsky.social)
✨New paper ✨
Introducing 🌍MultiBLiMP 1.0: A Massively Multilingual Benchmark of Minimal Pairs for Subject-Verb Agreement, covering 101 languages!
We present over 125,000 minimal pairs and evaluate 17 LLMs, finding that support is still lacking for many languages.
🧵⬇️
Introducing 🌍MultiBLiMP 1.0: A Massively Multilingual Benchmark of Minimal Pairs for Subject-Verb Agreement, covering 101 languages!
We present over 125,000 minimal pairs and evaluate 17 LLMs, finding that support is still lacking for many languages.
🧵⬇️
April 8, 2025 at 12:27 PM
Modern LLMs "speak" hundreds of languages... but do they really?
Multilinguality claims are often based on downstream tasks like QA & MT, while *formal* linguistic competence remains hard to gauge in lots of languages
Meet MultiBLiMP!
(joint work w/ @jumelet.bsky.social & @weissweiler.bsky.social)
Multilinguality claims are often based on downstream tasks like QA & MT, while *formal* linguistic competence remains hard to gauge in lots of languages
Meet MultiBLiMP!
(joint work w/ @jumelet.bsky.social & @weissweiler.bsky.social)
Reposted by Thennal D K
International students will stop coming to American universities if their visas are going to be at risk. This will make our intellectual community poorer and also make tuition more expensive for domestic students.
UPDATED: At least 83 students -- at campuses for University of California, California State University and Stanford -- have had their visas revoked as of Monday evening.
LATEST: At least 45 student visas across the state have been revoked by the Trump administration, California universities report, as numbers grow. A lawsuit has been filed in a Los Angeles federal court against DHS and Kristi Noem www.latimes.com/california/s...
April 8, 2025 at 1:52 AM
International students will stop coming to American universities if their visas are going to be at risk. This will make our intellectual community poorer and also make tuition more expensive for domestic students.
Reposted by Thennal D K
FBI Uncovers Al-Qaeda Plot To Just Sit Back And Enjoy Collapse Of United States
FBI Uncovers Al-Qaeda Plot To Just Sit Back And Enjoy Collapse Of United States
WASHINGTON—Putting the nation on alert against what it has described as a “highly credible terrorist threat,” the FBI announced today that it has uncovered a plot by members of al-Qaeda to sit back an...
theonion.com
February 5, 2025 at 6:31 PM
FBI Uncovers Al-Qaeda Plot To Just Sit Back And Enjoy Collapse Of United States
Reposted by Thennal D K
Most of my colleagues are shocked when I bring up these comments to them. Being in academia doesn't mean we are protected, or even simply ignored by this. We and our institutions are being actively attacked by policies like funding freezes, and I don't see why it wouldn't get worse than that.
Reminder that J.D. Vance called for the MAGA movement to "aggressively attack the universities" www.peoplefor.org/rightwingwat...
February 1, 2025 at 7:12 PM
Most of my colleagues are shocked when I bring up these comments to them. Being in academia doesn't mean we are protected, or even simply ignored by this. We and our institutions are being actively attacked by policies like funding freezes, and I don't see why it wouldn't get worse than that.
Reposted by Thennal D K
We need a name for the 'paradox' where frontier LLMs are useful in domain x if your expertise in domain x go much deeper than theirs and harmful otherwise
February 1, 2025 at 5:58 PM
We need a name for the 'paradox' where frontier LLMs are useful in domain x if your expertise in domain x go much deeper than theirs and harmful otherwise
Reposted by Thennal D K
My university (Northeastern) got rid of their page on Diversity, Equity, and Inclusion, and redirected the URL to a page on "belonging". It's not clear to me how a private university having a page on their values violates the law?
web.archive.org/web/20250124...
web.archive.org/web/20250124...
Office of Belonging – Belonging at Northeastern
web.archive.org
January 31, 2025 at 9:15 PM
My university (Northeastern) got rid of their page on Diversity, Equity, and Inclusion, and redirected the URL to a page on "belonging". It's not clear to me how a private university having a page on their values violates the law?
web.archive.org/web/20250124...
web.archive.org/web/20250124...
Reposted by Thennal D K
We are launching HALoGEN💡, a way to systematically study *when* and *why* LLMs still hallucinate.
New work w/ Shrusti Ghela*, David Wadden, and Yejin Choi 💫
📝 Paper: arxiv.org/abs/2501.08292
🚀 Code/Data: github.com/AbhilashaRav...
🌐 Website: halogen-hallucinations.github.io 🧵 [1/n]
New work w/ Shrusti Ghela*, David Wadden, and Yejin Choi 💫
📝 Paper: arxiv.org/abs/2501.08292
🚀 Code/Data: github.com/AbhilashaRav...
🌐 Website: halogen-hallucinations.github.io 🧵 [1/n]
January 31, 2025 at 6:27 PM
We are launching HALoGEN💡, a way to systematically study *when* and *why* LLMs still hallucinate.
New work w/ Shrusti Ghela*, David Wadden, and Yejin Choi 💫
📝 Paper: arxiv.org/abs/2501.08292
🚀 Code/Data: github.com/AbhilashaRav...
🌐 Website: halogen-hallucinations.github.io 🧵 [1/n]
New work w/ Shrusti Ghela*, David Wadden, and Yejin Choi 💫
📝 Paper: arxiv.org/abs/2501.08292
🚀 Code/Data: github.com/AbhilashaRav...
🌐 Website: halogen-hallucinations.github.io 🧵 [1/n]
Reposted by Thennal D K
This is what the government did with 120K+ Japanese Americans in 1942.
I know. I was there in those camps.
I know. I was there in those camps.
January 31, 2025 at 7:11 PM
This is what the government did with 120K+ Japanese Americans in 1942.
I know. I was there in those camps.
I know. I was there in those camps.
Happy to share that this paper has now been accepted at NAACL 2025!
I believe the natural language processing as a field has severe issue with multilingual evaluation. This paper focuses specifically on automatic speech recognition evaluation, and advocates for an a more inclusive, more accurate evaluation scheme than the one currently employed by e.g. OpenAI.
Advocating Character Error Rate for Multilingual ASR Evaluation
arxiv.org/abs/2410.07400
arxiv.org/abs/2410.07400
January 23, 2025 at 4:27 PM
Happy to share that this paper has now been accepted at NAACL 2025!
Reposted by Thennal D K
I'll get straight to the point.
We trained 2 new models. Like BERT, but modern. ModernBERT.
Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff.
It's much faster, more accurate, longer context, and more useful. 🧵
We trained 2 new models. Like BERT, but modern. ModernBERT.
Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff.
It's much faster, more accurate, longer context, and more useful. 🧵
December 19, 2024 at 4:45 PM
I'll get straight to the point.
We trained 2 new models. Like BERT, but modern. ModernBERT.
Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff.
It's much faster, more accurate, longer context, and more useful. 🧵
We trained 2 new models. Like BERT, but modern. ModernBERT.
Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff.
It's much faster, more accurate, longer context, and more useful. 🧵
Reposted by Thennal D K
Here's why "alignment research" when it comes to LLMs is a big mess, as I see it.
Claude is not a real guy. Claude is a character in the stories that an LLM has been programmed to write. Just to give it a distinct name, let's call the LLM "the Shoggoth".
Claude is not a real guy. Claude is a character in the stories that an LLM has been programmed to write. Just to give it a distinct name, let's call the LLM "the Shoggoth".
December 19, 2024 at 11:15 PM
Here's why "alignment research" when it comes to LLMs is a big mess, as I see it.
Claude is not a real guy. Claude is a character in the stories that an LLM has been programmed to write. Just to give it a distinct name, let's call the LLM "the Shoggoth".
Claude is not a real guy. Claude is a character in the stories that an LLM has been programmed to write. Just to give it a distinct name, let's call the LLM "the Shoggoth".
Reposted by Thennal D K
I saw recently that this had been cited in some lecture notes, so I decided to try to preserve it for posterity (since the original tweet had been deleted by me in the interim).
December 27, 2024 at 9:06 PM
I saw recently that this had been cited in some lecture notes, so I decided to try to preserve it for posterity (since the original tweet had been deleted by me in the interim).
Reposted by Thennal D K
At risk of repeating myself: The Luddites weren’t against technology. They were against getting put out of work by a technology that did a version of their job faster but worse, in the service of increasing profits for their bosses.
December 26, 2024 at 7:13 PM
At risk of repeating myself: The Luddites weren’t against technology. They were against getting put out of work by a technology that did a version of their job faster but worse, in the service of increasing profits for their bosses.
Reposted by Thennal D K
Is artificial intelligence supercharging the problem of political misinformation? In a word: no. New from @randomwalker.bsky.social & @sayash.bsky.social at @knightcolumbia.org. knightcolumbia.org/blog/we-look...
We Looked at 78 Election Deepfakes. Political Misinformation Is Not an AI Problem.
knightcolumbia.org
December 13, 2024 at 8:25 PM
Is artificial intelligence supercharging the problem of political misinformation? In a word: no. New from @randomwalker.bsky.social & @sayash.bsky.social at @knightcolumbia.org. knightcolumbia.org/blog/we-look...
Reposted by Thennal D K
How do disparities in healthcare access affect ML models? 💰📉🧐 We found that low access to care -> worse EHR data quality -> worse ML performance in a dataset of 134k patients. Work with Anna Zink (on the faculty job market rn!) + Hongzhou Luan, presented at #ML4H2024
December 20, 2024 at 1:04 AM
How do disparities in healthcare access affect ML models? 💰📉🧐 We found that low access to care -> worse EHR data quality -> worse ML performance in a dataset of 134k patients. Work with Anna Zink (on the faculty job market rn!) + Hongzhou Luan, presented at #ML4H2024
Reposted by Thennal D K
I’m a Bourdieusian, when I give a movie 7/10 it means it goes here
December 25, 2024 at 8:10 PM
I’m a Bourdieusian, when I give a movie 7/10 it means it goes here
Reposted by Thennal D K
I’m dreaming of a slop Xmas…
Merry Slopmas!
AI-generated Christmas classics that dwell in the uncanny valley are giving listeners the creeps.
www.404media.co
December 25, 2024 at 8:18 PM
I’m dreaming of a slop Xmas…
Reposted by Thennal D K
As data centers cause America to run out of power, Arizona is emblematic of the sacrifices ordinary people face to fuel the boom.
In the shadows of Arizona’s data center boom, thousands live without power
As data centers drain America’s power grids, a fierce battle is being waged for electricity. On Navajo Nation land, many are on the losing end.
www.washingtonpost.com
December 23, 2024 at 3:10 PM
As data centers cause America to run out of power, Arizona is emblematic of the sacrifices ordinary people face to fuel the boom.
honey you didnt close the fridge door again im adjusting my priors on the divorce
My pet peeve is shallow invocations of machine leaning principles to justify faulty human heuristics
My pet peeve is shallow invocations of neuroscience to justify a heuristic ML algorithm
December 16, 2024 at 6:27 PM
honey you didnt close the fridge door again im adjusting my priors on the divorce
extremely funny sequence of events that's emblematic of issues in the field in general. and ofc new work will ignore both these datasets and just use the original MMLU anyway
Very cool work! 👏🚀 Unfortunately, errors in the original dataset will propagate to all new languages 😕
We investigated the issue of existing errors in the original MMLU in
arxiv.org/abs/2406.04127
@aryopg.bsky.social @neuralnoise.com
We investigated the issue of existing errors in the original MMLU in
arxiv.org/abs/2406.04127
@aryopg.bsky.social @neuralnoise.com
Is MMLU Western-centric? 🤔
As part of a massive cross-institutional collaboration:
🗽Find MMLU is heavily overfit to western culture
🔍 Professional annotation of cultural sensitivity data
🌍 Release improved Global-MMLU 42 languages
📜 Paper: arxiv.org/pdf/2412.03304
📂 Data: hf.co/datasets/Coh...
As part of a massive cross-institutional collaboration:
🗽Find MMLU is heavily overfit to western culture
🔍 Professional annotation of cultural sensitivity data
🌍 Release improved Global-MMLU 42 languages
📜 Paper: arxiv.org/pdf/2412.03304
📂 Data: hf.co/datasets/Coh...
December 6, 2024 at 4:12 PM
extremely funny sequence of events that's emblematic of issues in the field in general. and ofc new work will ignore both these datasets and just use the original MMLU anyway
Reposted by Thennal D K
Here are some nice mushrooms
December 2, 2024 at 1:52 PM
Here are some nice mushrooms