We’re building an Internet where people are before profits; Carissa Veliz’s new book Prophecy shows how Big Tech’s analytics betray that mission. Discover what links CEOs and medieval astrologers this April… 👇 https://www.carissaveliz.com/prophecy

I am not anti LLM across the board. But if your use case starts with the necessary assumption that the model is a vessel of all human knowledge, then you are consulting an oracle not doing scientific research.

Je viens de découvrir cette information. Quel piètre Léonard je fais...

L'IA dans la recherche, très très loin de nous faire gagner du temps et produire du savoir, nous en fait perdre en facilitant les comportements non éthiques tout en nous obligeant à nous transformer en flics.

J'accepte d'évaluer un article, dont le sujet m'intéresse, dans une bonne revue. Je le feuillète, et très vite je tombe sur au moins deux citations qui n'existent pas, dont une à mon propre travail. Misère de "l'intelligence artificielle", une fois de plus.

Regular reminder to use R Projects and the `here` package. Future you will be happy tidyverse.org/blog/2017/12...

In my paper Sample Size Justification I discuss 6 approaches to justifying sample sizes. The focus of the paper is to teach people the tools they need to be honest when they explain where their sample size comes from. I know, so radical ;) online.ucpress.edu/collabra/art...

Oh look, AI research has it's first AI-slop Sokal affair: "A single person claims to have authored 113 academic papers on artificial intelligence this year, 89 of which will be presented this week at one of the world’s leading conference on AI and machine learning".

A couple years (!) in the making: we’re releasing a new corpus of embodied, collaborative problem solving dialogues. We paid 36 people to play Portal 2’s co-op mode and collected their speech + game recordings. Paper: arxiv.org/abs/2512.03381 Website: berkeley-nlp.github.io/portal-dialo... 1/n

You requested it, now it’s here - introducing Proton Sheets! The privacy-first alternative to Excel and Google Sheets. Spreadsheets form the framework of modern businesses, and with Proton Sheets, you can ensure your data is private and secure. 1/4

P-values do each of these things *better* than Bayes factors, through equivalence testing, sequential analysis, and meta-analysis. Better, because you get error control, so you do not make false claims too often. 2/2

3 Bayesian lies, from a preprint abstract: "Unlike p-values, Bayes factors allow for quantifying support both for and against the existence of an effect, facilitate ongoing evidence monitoring, and maintain coherent long-run behavior as additional studies are incorporated." 1/2

Blaguons sur les GPU. Vers 2000, ils étaient conçus pour faire du rendu graphique pour les jeux 3D - précision arithmétique non garantie. Puis on les a améliorés pour pouvoir faire du calcul scientifique (plus grande précision, respect de IEEE-754…). MAIS

on rappelle toutefois que ce n'est pas l'IA qui supprime des emplois, mais les patrons. www.liberation.fr/economie/eco...

Unfortunately, even when autonomous vehicles are safer, psychological factors still prevent wide-spread adoption: How safe is safe enough? Psychological mechanisms underlying extreme safety demands for self-driving cars. Transportation Research Part C: Emerging Technologies. doi.org/10.1016/j.tr...

"Our Obsession With Statistical Significance Is Ruining Science" reason.com/2025/12/01/o...

"All of this falls apart if humans don't adopt the tech. This is why you've seen Meta cram its lame chatbots into WhatsApp and Instagram. This is why Notepad and Paint now have useless Copilot buttons on Windows. This is why Google Gemini wants to "help you" read and reply to your emails."

Until tech gives more agency to citizens, smart cities are a dumb idea for democracy. "While cities may disclose how they collect data, they rarely offer ways to opt out." "Residents want agency" #privacy #AIEthics spectrum.ieee.org/smart-city-p...

"We use language to think, but that does not make language the same as thought." #AIEthics www.theverge.com/ai-artificia...

Team "all at once" here

DeepSeek just released a new state of the art math prover, DeepSeek-Math-V2, competitive with Google, OpenAI or ByteDance, while being a publicly documented open weight models. A few reading notes along the way:

And a major open science release from Prime Intellect: they don’t stress it enough but SFT part is beyond post-training. This is a fully documented mid-training with tons of insights/gems on MoE training, asynchronous infra RL, deep research. storage.googleapis.com/intellect-3-...

YES. Main reason classic pretraining dominated for so long is just that you don’t have to think so much about the data or what elicits reasoning. It’s "here". For Sutskever/Patel new podcast: www.dwarkesh.com/p/ilya-sutsk...

Many Gen AI tools are free now. If you do not train yourself (e.g. you never learn how to program without them) you will be dependent on them. In the future they will increase their prices drastically, and you will have to pay, because you can't do without them anymore.

"This lower literacy–greater receptivity link is not explained by differences in perceptions of AI's capability, ethicality, or feared impact on humanity. Instead, this link occurs because people with lower AI literacy are more likely to perceive AI as magical" journals.sagepub.com/doi/10.1177/...

[ #VeilleESR #LRU ] Les nouveaux #COMP seront généralisés à quasiment toutes les universités dès 2026 et alignés sur le mandat des présidents par @camillecordonnier.bsky.social et @anaisgerard.bsky.social pour @aefsuprecherche.bsky.social C'est merveilleux !

Signal polls are here! Find the best time for your group chat to hang out, choose a vacation destination, or ask if you should give yourself a haircut at 1 AM. Now in the latest version of Signal for iOS, Android, and Desktop. signal.org/blog/polls/

new paper by Sean Westwood: With current technology, it is impossible to tell whether survey respondents are real or bots. Among other things, makes it easy for bad actors to manipulate outcomes. No good news here for the future of online-based survey research

I'd really want to know about the guardrails being implemented here, given our recent results in @nature.com showing that people are likely to ask LLMs to help them cheat with taxes, and that LLMs are likely to accept www.nature.com/articles/s41...

It's great to see #PROPHECY gathering friends. Thank you so much for everyone who has pre-ordered. It matters. It gets noticed. It's part of being in the conversation. THANK YOU! #Book www.penguinrandomhouse.com/books/759692...

Comme notre président le souligne, #Lyon3 est sous-dotée, et cela ne date pas d'hier. www.leprogres.fr/education/20...

Soutien au président de l'Université Jean Moulin Lyon 3 qui demande la fin des inégalités de financement public dans l'enseignement supérieur. Ces inégalités, non justifiées, durent depuis des années. Mêmes conditions d'étude et de travail pour tous ! -> www.leprogres.fr/education/20...

Drumrolls, please! I’m thrilled to share the cover for my forthcoming #book from @doubledaybooks.bsky.social (Penguin Random House), #PROPHECY. This is the best book I’ve written: the boldest, the most innovative, the most personal. I have poured my soul into it like never before. 1/

From 𝗪𝗶𝗸𝗶𝗽𝗲𝗱𝗶𝗮 to 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 😬😅

Looking back, one of my main disappointment in LLM/AI research is seeing the non-commercial space shrinking, becoming more conservative, fragmented and less cooperative

The Drain of Scientific Publishing arxiv.org/pdf/2511.04820

Friends, who is doing good work/research on the impact of AI on women's employment and careers? Do you know? Please share if so!!

Gentle request that @altmetric.com and @linkedin.com resume contact to be able to track academic LinkedIn activity on AltMetric. More and more scientific discourse is happening on BlueSky (tracked) and LinkedIn (not tracked). @linkedin.com let AltMetric use your API and include you! NOW!

Google has now admitted that they will tap into users' personal data within Gmail and Drive in order for their upcoming AI search to "be more helpful". For years, Google has advertised their services as "encrypted" and "secure", which begs the question: are they really? 1/5 🧵

PSA: tips to protect yourself from scams on Signal. Every major comms platform has to contend w phishing, impersonation, & scams. Sadly. Signal is major, and as we've grown we've heard about more of these attacks--scammy people pretending to be something or someone to trick and abuse others. 1/

📣 Are you a PhD student in the quantitative social sciences interested in interdisciplnary research? Join us at the Toulouse Summer School in Quantitative Social Sciences (25 May-19 Jun 2026) at @tse-fr.eu. Apply by 15 Dec 2025 🎓🔍 👉 More info + application: www.tse-fr.eu/toulouse-sum...

@inriaparisnlp.bsky.social brought you CamemBERT, and we now bring you Gaperon (for non-cheese connoisseurs, it’s a cheese that’s flavoured with pepper and garlic 🧀 ).

If Sam Altman claimed something about a past event — say, a date or a historical episode — would you take his word for it, or would you double-check the facts elsewhere? Then why take him at his word when he talks about things that haven’t happened yet?

"Ethical due diligence should become as routine as financial due diligence. Before asking how large a technology might become, we should ask what kind of behaviour it incentivises, what dependencies it creates, and who it leaves behind." #AIEthics time.com/7332888/we-n...

The EU has let US tech giants run riot. Diluting our data law will only entrench their power | Johnny Ryan and Georg Riekeles

This is a big part of why so many data centre sites are just flat-out straight-up building their own brand new fossil fuelled power stations right next to the data centre

Rapport de l'#IGESR de cette année sur les heures d'enseignement à l'université. #ESR www.enseignementsup-recherche.gouv.fr/sites/defaul...

Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...

La Commission européenne prévoit d’affaiblir le RGPD au profit des entreprises d’IA next.ink/208221/la-co...