Lightnews — Scholar-powered news

Sourav @souravmishra.bsky.social · May 18

L5 Manager stork making sure the L3 and L4 employee storks finishing their task properly 😂

cc @pierrealquier.bsky.social

1 3

Reposted by Sourav

shimon8282.bsky.social @shimon8282.bsky.social · Apr 18

Our survey on meta reinforcement learning has now been published by Foundations and Trends in Machine Learning: nowpublishers.com/article/Deta...

now publishers - A Tutorial on Meta-Reinforcement Learning

Publishers of Foundations and Trends, making research accessible

nowpublishers.com

2 12

Sourav @souravmishra.bsky.social · Apr 8

Big win for Stanford NLP to have @yejinchoinka.bsky.social Looking forward to new amazing directions

Stanford NLP Group @stanfordnlp.bsky.social · Apr 4

Look who we found hanging out in her new Stanford Gates Computer Science office!

We’re truly delighted to welcome @yejinchoinka.bsky.social as a new @stanfordnlp.bsky.social faculty member, starting full-time in September. ❤️

nlp.stanford.edu/people/

1 1 2

Sourav @souravmishra.bsky.social · Apr 8

Excited to see Waymo come to Tokyo

* Tokyo has a passive, secondary transportation mode which will stay strong despite the subway.

* It is generally considered to be one of the safest cities. Low chances of vandalism.

* People are curious to try out new technology across all age strata.

Chris Paxton @cpaxton.bsky.social · Apr 7

The first waymo depot in Tokyo has been identified, a huge milestone: thelastdriverlicenseholder.com/2025/04/07/w...

1 6

Sourav @souravmishra.bsky.social · Mar 27

Picking the brains of the Claude model. Anthropic has made a new foray into "AI biology" and some interesting case studies on what the assumptions are.

Turns out LLMs don't entirely piece together their response the way we do. Great investigate work
www.anthropic.com/research/tra...

Tracing the thoughts of a large language model

Anthropic's latest interpretability research: a new microscope to understand Claude's internal mechanisms

www.anthropic.com

1 2

Reposted by Sourav

Chris Olah @colah.bsky.social · Mar 27

Can we understand the mechanisms of a frontier AI model?

📝 Blog post: www.anthropic.com/research/tra...
🧪 "Biology" paper: transformer-circuits.pub/2025/attribu...
⚙️ Methods paper: transformer-circuits.pub/2025/attribu...

Featuring basic multi-step reasoning, planning, introspection and more!

On the Biology of a Large Language Model

transformer-circuits.pub

4 29 130

Sourav @souravmishra.bsky.social · Mar 27

In the Ghibli world I look like this. Not bad. Not bad at all! :-)

2

Sourav @souravmishra.bsky.social · Mar 25

Is "vibe coding" same as "prompt engineering" from 6 mo ago?

I am having trouble keeping up with names and acronyms now, forget research papers

1

Reposted by Sourav

Steven Heidel @stevenheidel.com · Mar 23

don’t learn to write code, learn to read code

1 4

Sourav @souravmishra.bsky.social · Mar 25

Too bad it's a little late for Team Australia for the Olympics

Chris Paxton @cpaxton.bsky.social · Mar 25

Breakdancing electric atlas

1 1

Sourav @souravmishra.bsky.social · Mar 20

Such a beautiful representation of the vernal equinox 😍

European Space Agency @esa.int · Mar 20

🌍 Happy #Equinox! On 20 March 2025 at 09:01 GMT/10:01 CET, just a few moments ago, the Sun appeared directly above Earth's equator - with our planet neither tilting towards nor away from the Sun!

(pic: @eumetsat.int, processed by ESA/S. Proud)
See flic.kr/p/2qT1ZUu

At the equinox of 20 March 2025, our planet is neither tilted towards nor away from the Sun, creating nearly equal day/night. This stunning Meteosat-12 image taken earlier this morning showcases the boundary between day and night running straight through the poles.

3

Sourav @souravmishra.bsky.social · Mar 20

"There was a girl from India who had overstayed her student visa for 3 days before heading back home. She then came back to the US on a new, valid visa to finish her master’s degree and was handed over to ICE due to the 3 days she had overstayed"

WOW. Just wow. Cruelty is in vogue for US govt. now

Jay Owens @jayowens.bsky.social · Mar 19

I am seriously questioning whether it will be safe to visit the US for three weeks this summer, as I had planned.

A powerful account from Jasmine Mooney, a Canadian detained in ICE facilities for 12 days for no reason. (The cruelty is the point.)

www.theguardian.com/us-news/2025...

I’m the Canadian who was detained by Ice for two weeks. It felt like I had been kidnapped

I was stuck in a freezing cell without explanation despite eventually having lawyers and media attention. Yet, compared with others, I was lucky

www.theguardian.com

2

Sourav @souravmishra.bsky.social · Mar 20

Best LLMs are decided by evals on leaderboards
(I am looking at you, Gemini)

mcc @dryad.technology · Mar 19

Quote this post with a fact from your field which is a complete total lie, but which people outside your field would completely believe if you posted it confidently on social media

1

Sourav @souravmishra.bsky.social · Mar 18

Like Notion? Well this is a open source clone sponsored by FR-DE governments. And frankly, it looks great

docs.numerique.gouv.fr/login/

Docs

Docs: Your new companion to collaborate on documents efficiently, intuitively, and securely.

docs.numerique.gouv.fr

Sourav @souravmishra.bsky.social · Mar 17

The arrival of early spring *chef's kiss* Nice shots Sid. Get yourself a proper camera

Sid @sidgairo18.bsky.social · Mar 16

Saarbrücken riverside 🚴‍♂️

1 2

Sourav @souravmishra.bsky.social · Mar 14

OLMo2 goodies ↓

Nathan Lambert @natolambert.bsky.social · Mar 13

The longer version of this story and how it relates to Gemma 3 too is here: buff.ly/5OZYGrq
For more information, read the technical blog post: buff.ly/lJxPodz
Or try the model: buff.ly/6lAGGkw
Get the artifacts here: buff.ly/Wz2VjUC

buff.ly

1

Reposted by Sourav

Ai2 @ai2.bsky.social · Mar 13

Announcing OLMo 2 32B: the first fully open model to beat GPT 3.5 & GPT-4o mini on a suite of popular, multi-skill benchmarks.

Comparable to best open-weight models, but a fraction of training compute. When you have a good recipe, ✨ magical things happen when you scale it up!

3 15 58

Sourav @souravmishra.bsky.social · Mar 14

True open source looks like OLMO2 😅 codes weights, recipes all available to inspect

Kyle Lo @ COLM 2025 🍁 @kylelo.bsky.social · Mar 13

ofc benchmarks are one thing, vibes checks are another! encourage folks to play w it & read the post

🍣 playground.allenai.org
🐙 allenai.org/blog/olmo2-32B

lookin forward to all the cool research - interpretability, training dynamics, generalization, etc - this hopefully enables 🤩

playground.allenai.org

1

Reposted by Sourav

Christoph Molnar @christophmolnar.bsky.social · Mar 13

The 3rd edition of Interpretable Machine Learning is out! 🎉 Major cleanup, better examples, and new chapters on Data & Models, Interpretability Goals, Ceteris Paribus, and LOFO Importance.

The book remains free to read for everyone. But you can also buy ebook or paperback.

4 16

Reposted by Sourav

Jeff Dean @jeffdean.bsky.social · Mar 13

Introducing our Gemma 3 open models, the most capable models that you can run on a single GPU or TPU. Multimodal, multilingual, 128k context length, and exceeds quality of other open models that are an order of magnitude larger in terms of hardware footprint. 🎉

blog.google/technology/d...

Introducing Gemma 3: The most capable model you can run on a single GPU or TPU

Today, we're introducing Gemma 3, our most capable, portable and responsible open model yet.

blog.google

2 21 140

Sourav @souravmishra.bsky.social · Mar 12

Nice. Congratulations Pierre.

(I am pretty sure I won't understand it. It looks pretty dense 😂😅 so I am honest with you. But congratulations for the acceptance 🥳)

Pierre Alquier @pierrealquier.bsky.social · Mar 12

Our joint paper with Geoffrey Wolfer @gwolfer.bsky.social "Variance-Aware Estimation of the Kernel Mean Embedding" accepted for publication in the Journal of Machine Learning Research 🥳

arxiv.org/abs/2210.06672

Variance-Aware Estimation of Kernel Mean Embedding

An important feature of kernel mean embeddings (KME) is that the rate of convergence of the empirical KME to the true distribution KME can be bounded independently of the dimension of the space, prope...

arxiv.org

1 2

Sourav @souravmishra.bsky.social · Mar 11

ML project naming is getting out of hands 😀

Chris Offner @chrisoffner3d.bsky.social · Mar 10

hmm...
github.com/nikmcfly/ANUS

1

Sourav @souravmishra.bsky.social · Mar 9

Seems like it. I saw the chapters themselves have some typos.

However that 'gradient tape' example is a bit... bizarre ? Things could have been left at static vs dynamic computation graph (define-and-run vs. define-while-run for easier speaking)

I hope the book doesn't turn out a mish-mash

Sourav @souravmishra.bsky.social · Mar 9

Ha! You are waiting on it too with preorder ! 😀 It was originally slated January & it has slipped behind schedule

1 1

Sourav @souravmishra.bsky.social · Mar 9

Any news about the book release @fchollet.bsky.social ? The new chapters look promising especially

1