Lightnews — Scholar-powered news

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · 8d

@acollierastro.bsky.social is the coolest person on the internet and I am envious of everyone who knows her personally

conspiracy physics and you (and also me)

YouTube video by Angela Collier

youtube.com

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · 14d

I guess this calls for an adequate celebration.

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jul 21

Attending my forst tutorial of this year's UAI IN Rio. Fun fact about Rio. It's the only city outside of Europe that was the capital of a European countr at some point.

uai2025 @auai.org · Jul 21

and

"Counterfactuals in Minds and Machines"

by ‪@tobigerstenberg.bsky.social‬ (Stanford), ‪@autreche.bsky.social‬ (MPI SS) and ‪@stratiss.bsky.social‬ (MPI SS)

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jul 5

Unfortunately, this will also entail unavoidable collateral damage with people not at fault being called out.

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jul 5

Let's be honest, if you are looking for a community NeurIPS might be the wrong place. It's simply too big and anonymous. With anonymity comes people gaming the system and I think we should be calling out bad behavior more aggressively, eg."co-authoring" 10 papers and not contributing to peer-review.

2 2

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jul 5

One of the papers co-authored by her:

Humans Learn Language from Situated Communicative Interactions. What about Machines?

Katrien Beuls, Paul Van Eecke. Computational Linguistics, Volume 50, Issue 4 - December 2024. 2024.

share.google

1 7

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jul 5

I think Katrien Beuls has been doing empirical research in this direction in the context of robotics.

share.google

2 1 3

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 28

Haha, meant the latter.

2

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 28

You said the naughty word🤭

1 1

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 27

The crazy thing about epicycles: at the time Kopernicus introduced his model they were far superior in terms of predictive power compared to the heliocentric model. This was still true when Kepler refined the model with ellipses. Epicycles had been extremely refined over time (capital+labor)

1 1 4

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 27

For machine learning this means that neural nets will keep on winning not necessarily for technical reasons but for sheer volume of investment and work hours that are being poured in their development.

1 4

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 27

Basically, every model and even the way we do science is a child of its time and has to be understood not only in a scientific context but also in a societal and economic one.

1 4

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 27

I guess all the answers above. But I meant it more in post-modernist philosophy of science way (reading Feyerabend at the moment)

1 1

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 26

Hot take: neural networks are the epicycles of the 21st century.

1 4

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jun 3

Ever changing from physics to AI and studying probabilistic ML, I have been wondering, how does all of this relate to quantum information theory 🎲.

I believe I've gotten a whole lot closer to the answer.
arxiv.org/abs/2506.01824

I'll be presenting the work at this year's @auai.org in Rio 😎

2 6

Reposted by Pedro Zuidberg Dos Martires

Jaron Maene @jjcmoon.bsky.social · Apr 23

We developed a library to make logical reasoning embarrasingly parallel on the GPU.

For those at ICLR 🇸🇬: you can get the juicy details tomorrow (poster #414 at 15:00). Hope to see you there!

1 7 24

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Apr 9

TLDR; yes... well, kind of 🙃

1

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Apr 9

What is reasoning and have LLMs learned it?

We answer these two questions in our new pre-print using the phase transition in random 3-SAT.
arxiv.org/abs/2504.039...

Together with @rishihazra95.bsky.social, @gabventurato.bsky.social and @lucderaedt.bsky.social

Here is also a longer thread:

#ai #llms #reasoning #deepseek #3sat #iclr | Rishi Hazra

🚨 𝗡𝗲𝘄 𝗣𝗮𝗽𝗲𝗿 𝗔𝗹𝗲𝗿𝘁: 𝗛𝗮𝘃𝗲 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 𝗟𝗲𝗮𝗿𝗻𝗲𝗱 𝘁𝗼 𝗥𝗲𝗮𝘀𝗼𝗻? Despite their impressive performance on reasoning tasks, LLMs often succeed by fitting to patterns rather than understanding logic. Are...

www.linkedin.com

1 4 10

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Mar 25

Hmmm but many PPLs do numeric integration... and automatic differentiation is definitely not the same as numeric differentiation

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Mar 13

For years I have been studyibg the top researchers in AI. They all share one personal trait:
attention to DETIAL.

1

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Feb 12

I miss the times when all we cared about was end-to-end differentiability. Now all there seems to be is world domination this, world domination that.

2

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jan 27

Dieses Paradox schreit nach einem Namen.

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jan 25

I think deepseek is amazing and I kind of like it... Buuuut... 🙃

(Saw something like this in someone else's post but can't find it back)

1

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jan 22

Well, the only thing that matters is statistics, e.g. the mean of a distribution, or the probability that your sample will fall into a subset of the support. Hot take: only use densities to make nice plots and convince reviewer #2.

1 1

Pedro Zuidberg Dos Martires @pedrozudo.bsky.social · Jan 22

Probability densities are like the sun. Look at them directly and they will blind you 😎
@paolomorettin.bsky.social

2 2