I mean, it’s so much shit at the moment that it’s nice the spring is giving us some positive energy, even though the sakura is way too early for Stockholm.

There are sooo many sakura 🌸 at the moment in Stockholm. Very nice!

3/ Post-hoc Probabilistic Vision-Language Models
@antonbaumann.bsky.social , @ruili-pml.bsky.social pml.bsky.social, @marcusklasson.bsky.social @smentu.bsky.social,@shyamgopal.bsky.social , @zeynepakata.bsky.social, @arnosolin.bsky.social @trappmartin.bsky.social

Ostsee Eutrophierungsgrad von 97%.

Was da so grün-gelb schimmert ist unsere Ostsee oder vielmehr: Die Eutrophierung – also Überdüngung. 97 % der Ostsee ist davon betroffen.

Folgen: Wenig Sauerstoff, weniger Fische, weniger Leben und anfälligere Ökosysteme. Gründe: Unter anderem die intensive Landwirtschaft.

(Grafik von 2023)

Sophie’s paper is a bit different: you don’t need to tweak the prompt with typos to make the LLM go crazy; you just need to scramble the tokenisation. But there is an easy fix, fine-tuning with scramble tokenisation makes your LLM robust against this. That’s what our paper is studying.

I guess we are. 😅 I think it’s inspiring to see what crazy things people do nowadays. Proper research takes time & effort and is inevitably a bit slower.

Stochasticity in Tokenisation Improves Robustness The widespread adoption of large language models (LLMs) has increased concerns about their robustness. Vulnerabilities in perturbations of tokenisation of the input indicate that models trained with a...

Stochastic tokenization is one of my favorite topics. Here is a recent preprint on arXiv.

h/t @trappmartin.bsky.social

Die CSU feiert das Aus des Verbrenner-Aus: "Ab jetzt gilt wieder der Wettbewerb der besten Antriebstechnologien."

Parallel schreibt Michael Kretschmer, CDU: "Wir müssen unsere heimischen Ressourcen Gas & Braunkohle halten & erschließen, um Abhängigkeiten zu reduzieren."

Mit Vollgas in den Abgrund.

Screenshot of Eurosky fundraising page showing amount raised: €5,226 out of a €100,000 target and number of supporters (282).

We’re fundraising to build everything we need to operate without Bluesky. Our €100,000 target will accelerate our goal of building a European app backend (Appview) and a microblogging app.

Since last night over 282 of you contributed more than €5,000.

Let's do this!

fund.eurosky.tech

Apply now: AI for Research Summer School in August at @aalto.fi. Focusing on fundamental and applied AI methods for accelerating discovery across scientific domains. For students, postdocs, academics in AI+ML, especially @ellis.eu members.
Apply by May 20: www.ellisinstitute.fi/ellis-summer...

For some reason, I always feel like I am the only ML person on social media who doesn’t like LSTMs. 😅 Cool shirt though.

Post Doc at Uni Tübingen! 100% position for 3 (+3) years; they're looking for somebody to analyze large-scale longitudinal datasets in education research.

Expertise in machine learning is an advantage, commitment to research transparency desirable 😌 proficiency in German beneficial but not required

LOGML 2026 London Geometry and Machine Learning Summer School, July 13-17 2026

If you are interested in attending LOGML, then check out our project: www.logml.ai/projects/Jia...

From the papers I have recently seen (eg, during the review process), a solid evaluation is too often lacking IMO. I think there is a lot of potential in using LLMs, eg to get some weak signals that are better than random noise, but it’s diluted by weak experimentation of people atm.

Thanks for sharing.

US international politics nowadays feels like the quest for the question: What is the dumbest decision one can make in a given situation?

🛠️ The #uai2026 workshop lineup is out!

Check out the full list of accepted workshops 👉auai.org/uai2026/workshops

Don't forget to register and see you in Amsterdam 🇳🇱 on August 21!

Looking for help 👇

“A whole civilisation will die tonight…” is always a good way to start your social media messages.

Are there any good benchmarks for evaluating LLM-based agentic systems for scientific hypothesis formation and experimental design? I found a bunch of half-convincing things.

Any suggestions?

The real reason Trump is mad at NATO.

14:52 Uhr Trump droht mit Angriffen auf Kraftwerke am Dienstag US-Präsident Donald Trump hat Iran erneut mit verheerenden Angriffen auf Kraftwerke und Brücken gedroht, sollte Teheran die weitgehend für den Schiffsverkehr blockierte Straße von Hormus nicht freigeben. "Am Dienstag wird es in Iran den 'Kraftwerkstag' und den 'Brückentag' in einem geben", schrieb Trump auf seiner Plattform Truth Social. So etwas habe es "noch nie gegeben". Er forderte erneut, die Meerenge zu öffnen - "oder ihr werdet in der Hölle leben". Trump hatte am Samstag bekräftigt, dass sein mehrfach verschobenes Ultimatum an den Iran am Montag ablaufe. "Erinnert euch daran, als ich dem Iran zehn Tage Zeit gab, ein Abkommen zu schließen oder die Straße von Hormus freizugeben. Die Zeit läuft ab - noch 48 Stunden, dann bricht die Hölle über sie herein", schrieb er auf Truth Social.

Aktuelles Beispiel für Trump-Verharmlosung der #Tagesschau. Was sich im Original wie ein unbeherrschter, unflätiger Ausbruch liest, wird bei der Tagesschau dazu, dass Trump dem Iran mit "verheerenden Angriffen [d]roht". Zur Erinnerung: das Original war ""Open the Fuckin' Strait, you crazy 1/

Welcome to the MrBeastification of British politics: the latest trick up Nigel Farage's sleeve | Kirsty Major The Reform UK leader’s energy bill giveaway certainly grabs our attention – but it’s all a distraction from the real winners and losers, says Guardian deputy Opinion editor Kirsty Major

Welcome to the MrBeastification of British politics: the latest trick up Nigel Farage's sleeve | Kirsty Major

The goal of science is to make new knowledge, the goal of a PhD programme is to make new scientists.

Should #ProbML (former AABI) become a full, multi-day conference in 2027?

We would appreciate it if you could fill out this quick 2-minute survey (4 short questions) and share your opinions with us!

🔗 forms.gle/tzwVDjtwam8j...

90s (or early 2ks) game question. I vaguely remember playing a space colonisation type of game in which you mine materials, develop technology, and colonise planets (I think). Any ideas what this was? I played it a few times only and for years tried to retrieve what the name was.

OMG, I have not seen/played this for ages.

Skiing in northern Lapland is fun!

We’re launching a new AI Methods & Software Hub in the ML cluster and are looking for a Director to build and lead it!

Shape how ML/AI and software drive scientific discovery—in close collaboration with AI and domain scientists—within an extremely vibrant and collaborative ecosystem!

Will do!