American friends, watch m.imdb.com/title/tt3307...

Honestly this works for everything “I want to trap hungry 19c ghosts in jars to help us with historical research” ✅ “Please read our holiday card; we got a hungry ghost to write it this year” ❌

European "left" when you ask them about Iran

6/ Who should apply? Researchers with a strong computational background and a genuine interest in biology—from foundations to translation.

1/ We’re hiring PIs in AI × Biology The @astar-gis.bsky.social Genome Institute of Singapore (GIS) is expanding its AI & Computation domain and recruiting junior & senior PIs.

I sometimes wonder if people in the "dark ages" knew they lived in "dark ages", and if our descendants will think that we live in an "dark age". I feel we live in a strage mixture of progress of technology and looming society collapse due to concentration of power into ...

EU should declare that English is a second official language in all member states. It's not like Balzac, Göthe or Cervantes will turn in their graves if you allow me to ask questions and find answers about how cross-country pension insurance works, and what I need to provide and to whom in English.

*Urgently* looking for emergency reviewers for the ARR October Interpretability track 🙏🙏 ReSkies much appreciated

"Culture"

"To secure ourselves against defeat lies in our own hands, but the opportunity of defeating the enemy is provided by the enemy himself". Just a random quote from "Art of war" that I remembered when I heard about FAIR layoffs.

Napisao sam polu-zajedljiv komentar, ali sam ga pobrisao jer bi mogao biti kontroverzan 0.1 osobi na ovom sajtu. Kažu da dobar hejt umire posljednji, ali nisam više siguran. Hejt je krhka i nestabilna emocija, trenutak čekaš i ode. Kao Kairos.

AI doesn't kill jobs, people do.

I wrote a short summary with some context on the blog.

my brother underperformed Q2 so we had to let him go. you know how it is.

Are there any works that explore curriculum learning in an unsupervised setting? Does the notion of Y is easier to learn if you first learn X happen automatically in auto-regressive modelling, or diffusion models or any other generative modelling approach?

I recently read this paper: alexiglad.github.io/blog/2025/ebt/ I like the idea: first learn a verifier (a model that tells you the goodness/compatibility of a prediction given some context), and then optimize predictions with respect to this verifier.

After the bubble bursts will LinkedIn be full of posts: "This changes NOTHING"?

I think people (who use this term) mean "constraining the model by constraining hyperparameters", while "ordinary" (non-inductive?) bias means "constraining the model by constraining parameters", but 🤷🏻‍♂️

God bless Chris Bishop for being consistent with notation within and across his books.

Unmute adds ears and vocal chords to your favorite text-based language model. A seamless plug-and-play augmentation with easy personalisation through voice conditioning and text instructions. We will open-source shortly.

Clicking of PDF links opens them directly in Chrome browser on Android, the same behavior as on iOS. Hallelujah, hallelujah!

Many of society’s problems would be solved if we decided to treat greed as a mental illness. Greed is not good: it’s just easy. There are many other things that “clarify and cut through” and it’s easy to sell that greed “captures the essence of the evolutionary spirit”, but it really doesn’t.

Back from Taiwan

The thoughts that randomly appear when cleaning: why it's called "love triangle" when most of the time it's only "love angle".

Job alert: Join us in Mainz as Max Planck Research Group Leader (W2) in Molecular Design ...and make your own research dreams happen on de novo design, generative models, proteins, materials ... tinyurl.com/r2xjxnuk @mpip-mainz.mpg.de

Some days I go online and I notice that wrestling gravitational pull of bullshit is becoming an increasingly herculean task. Need to read more books and disconnect.

I feel that in times like this it's good to share this, also got my own sake: Do not feed the trolls. Time and energy are your most valuable resources. Use them wisely. Focus on what matters. Trolls don't.

@hf.co & Yaak are releasing the worlds largest open source self driving datasets. We are inviting the entire AI & robotics community to search and curate datasets for training end2end models. To search the data, we are launching Nutron. Check out the video to see how it works.

Introducing the Learning to Drive (L2D) dataset by @yaak-ai.bsky.social & @hf.co: tinyurl.com/n4rbjehz - 90 TB of data: 6 cameras, GPS, CAN - Aligned w German driving task catalog w diverse conditions 🌧️☀️❄️🚧🚦⛔️🚊 - Pwrd by NLP search: tinyurl.com/mr3hjz58 - LeRobotDataset v2.1 tinyurl.com/c4aus8f2

Let's say you start reading a paper, and then you hit a block b/c you need to dive deeper. Do you:

I've just misread "Truth Social" as "Truth Soviet".

whether out of the desire for revenge or whatever, is like a yearning for self-annihilation. The real "tough choices" in the U.S. apparatus involve breaking the stranglehold of special interest groups and raising taxes. That solves all deficit concerns + also exerts downward pressure on inflation.

This obsession with "getting a good deal"... You know who got a good deal? Winners of WWI, but that wasn't so great deal on a bit longer run. That's why Marshall's plan after WWII was even a better deal. It was crucial in enabling US post-war prosperity.

Looks like we're at "we're just following orders" stage.

The German press is analyzing everything Vance said as if the guy was thinking when he said all those things. He wasn't, he's not worth one second of our time. Focus on what matters (hint: security of the EU) and ditch that corrupt state as an ally. The sooner we accept this hard truth, the better.

I could make a strawman out of it, but I don't have to.

It's a little strange to assume that Chinese scientists and engineers are only smart if they work for a U.S. company. As if they have no agency unless guided by a brilliant vision that only US founders possess. As if they have to be directed to do anything, in this case, of course, by the CCP.

I don't understand why is TikTok ban in the US controversial, and all the bans of foreign social nets (and not only them) in PR China are not controversial.

Now is the best time to build video recommendation platform. Instagram and TikTok problems.

Life imitates art - Balkan/Berlin edition.

Seventh day without heating and hot water is just the Universe sending a reminder that I should value more the things I have 😃

Resourceful.

"The global [some tech] market is set to expand by [some percentage]% by [some year]." Would be nice to get some pitches from 10 years ago and see how accurate they were.

People sometimes make fun of science that sounds stupid and random. Meanwhile, a study of lizard saliva turned into a peptide medication, which was turned into a diabetes medication, which was turned into a GLP1 weight loss drug, that just became the first therapy every approved for … sleep apnea

ModernBERT is available as a slot-in replacement for any BERT-like model, with both 139M param and 395M param sizes. It has a 8192 sequence length, is extremely efficient, is uniquely great at analyzing code, and much more. Read this for details: huggingface.co/blog/modern...

EEML'25, our yearly machine learning summer school event, will be organised next summer in the beautiful city of Sarajevo - the place where East meets West 🇧🇦🇧🇦🇧🇦. More details coming soon, please see the link in the thread!

Bruno is a brilliant and passionate person, spearheaded categorical deep learning programme and I'm sure his contributions to the field will be invaluable. Help him continue working on what he loves and knows best!

2025 will be the year of AI for science Leveraging all the things we've recently learned training AI models for 1000x impact in science and this will need data! More details: huggingface.co/blog/lemater... Thread: ...

Now that diffusion for images and discrete diffusion for text are both giving decent results, is there a reason not to train multimodal diffusion models? What are the main challenges? Why do conditional diffusion to model p(image|text) or p(image|segmentation) when we can train p(image, text)?