Germán Kruszewski

Interested in the emergence of living, evolutionary, and cognitive systems.

[Germán Kruszewski @germank]

What conditions should a system meet for life to emerge in it? This is the question that we tackled in this Artificial Life journal paper, available at
https://t.co/nygdvI0WMu. I am very proud of this work!👇Below, a walkthrough 🧵over its main aspects: 1/n

[Artificial Life journal @ARTL_journal]

NEW EARLY ONLINE, "Emergence of Self-Reproducing Metabolisms as Recursive Algorithms in an Artificial Chemistry" by @germank & Tomáš Mikolov ...an article tackling a classic/fundamental #alife and #evolutionary problem. @alifeofficial @mitpress https://t.co/DbVJ8YYvVx

[Germán Kruszewski @germank]

@savvyRL Sorry to hear that. Also it seems like this app is totally broken with all these assholes (or single asshole with a network of accounts) suddenly harassing you now, no? I hardly login anymore these days, but it looks like it's getting worse by the day.

[Dongyoung Go @dongyoung4091]

Preference models (PMs), used in RLHF and best-of-n, are vulnerable to reward hacking. We show how to combine LM capabilities with human insight to make PMs more interpretable and reliable. A thread about our paper at https://t.co/WEfFBk4Pbd.

[Germán Kruszewski @germank]

@yoavgo No idea, but maybe this helps?
https://t.co/nlsqD2k7Om

[Hessam Akhlaghpour @theHessam]

X is such a great website. Really. No winks.

But if you also have instagram, follow Bisan and Plestia, two brave Palestinian women in Gaza.

byplestia
wizard_bisan1

They post regularly. See what they go through. Don’t let the media dehumanize the people we are bombing.

[Germán Kruszewski @germank]

@docmilanfar @ICCVConference Crème near the Chateau de Vincennes has some of the best ice cream I have ever tried.

[laurent besacier @laurent_besacie]

I have the feeling that the toolkit created by my talented colleagues hasn't yet received the attention it deserves, so I'm reintroducing their blog post on disco, designed for controling LLMs and other generative models!
https://t.co/M9GtgUVpKH

[John Gibbons 🇵🇸 @think_or_swim]

OMFG.

Turns out The Day After Tomorrow could possibly be the Year After Next?

We're hanging by a thread from full-scale #climate collapse that will be far, far worse than even our direst imaginings.

https://t.co/9udkGfYIZg

[Germán Kruszewski @germank]

Eternal sunshine of the spotless to-do list

[Greta Thunberg @GretaThunberg]

Take that in. Then try wrapping your head around the fact that we’re only seeing the beginning of a destabilising planet, while remembering that GHG emissions are still rising and the companies largely causing this are making record profits.

A bit concerning, don’t you think?

[Dr. Aaron Thierry @ThierryAaron]

1/ Here's a🧵on why the record shattering extremes in Europe are so unsettling for climate scientists.

It's increasingly looking that we've underestimated the impacts of #ClimateBreakdown...

https://t.co/vMonwth7wP

[Germán Kruszewski @germank]

@hardmaru I would rather make it Artificial Life == Artificial Intelligence https://t.co/qKaP4Cc0Ta

[Germán Kruszewski @germank]

ALife and AI are, in fact, the same field. They just don't know it yet.

[Germán Kruszewski @germank]

@sp_monte_carlo @BryanEikema ...but nowadays we can train a neural approximator of the target using techniques such as https://t.co/M9aGGP21Ph

[Germán Kruszewski @germank]

@sp_monte_carlo @BryanEikema Thanks! I'll have a look. Indeed, the first usage we found of this relaxation was in a MCMC loop (https://t.co/gMdPqZMnB6). We think we are the first to use it to actually obtain samples. The crucial thing is to have a good proposal, which for complex spaces was unfeasible...

[Dr. 𝗝𝗼𝗮𝗻𝗮 𝗖. 𝗫𝗮𝘃𝗶𝗲𝗿 @joanarcxavier]

Life is not an accident

[Germán Kruszewski @germank]

@sp_monte_carlo @BryanEikema Here's the link to @BryanEikema 's thread for anyone who might be interested in a quick tour of our study: https://t.co/pUs9m3DQbV

[Bryan Eikema @BryanEikema]

Check out our work accepted in Transactions on Machine Learning Research: An approximate sampler for energy-based models with divergence diagnostics

https://t.co/h3xmpjwsFf

Joint work w/ @germank, Christopher Dance, @hadyelsahar and @MarcDymetman at @naverlabseurope

1/10

[Germán Kruszewski @germank]

@sp_monte_carlo @BryanEikema I'm also curious about where this screenshot is from? We were also not the first to propose this approximation, and in the paper, we discuss some earlier connections (e.g., in particle filters and in variational inference), but I don't think I've seen this specific one before.

[Germán Kruszewski @germank]

@wtgowers Maybe this helps? https://t.co/OZy8hQX1GI

[Melanie Sclar @melaniesclar]

I wasn't aware of this, I just updated my settings! Truly a bummer right to have these changes happen right during ACL.

[Germán Kruszewski @germank]

@drmichaellevin Reference to https://t.co/oVi1AzlJn5?

[laurent besacier @laurent_besacie]

Presented this week at ACL (from NLP@NLE) (Outstanding demo paper) : disco: a toolkit for Distributional Control of Generative Models - ACL Anthology https://t.co/pPgnCMjQSS