As of today, the use of AI becomes more transparent in the EU. Here’s how EU law makes sure AI works for you, in full openness ↓

J’aimerai voir partout ce qui a été fait par IA. Marre de voir de subir lettres, propositions d’articles ou de livres par IA...

1. We—@eduede.bsky.social, @mjcrockett.bsky.social, Kevin Gross, and I—have a new preprint on the arXiv today, based on ideas that emerged during an @sfiscience.bsky.social workshop in November 2024: The unintended consequences of large language models as a labor-augmenting technology in science.

The unintended consequences of large language models as a labor-augmenting technology in science

As a labor-augmenting technology, large language models (LLMs) have the potential to accelerate scientific activity across the research pipeline. But even if LLMs perform on par with human experts at ...

arxiv.org

Fantastic paper demonstrating how LLM editing is gradually distorting our writing and the way we think to write: arxiv.org/abs/2603.18161

How LLMs Distort Our Written Language

Large language models (LLMs) are used by over a billion people globally, most often to assist with writing. In this work, we demonstrate that LLMs not only alter the voice and tone of human writing, b...

arxiv.org

I've been thinking about the Rubik's cube in 4D. I found an old java version from by Don Hatch and Melinda Green and decided to ask claude to port it to a standalone web app that can run in your browser. You can play it here: theoryandpractice.org/cube4d/ Blog: theoryandpractice.org/2026/07/four...

Simulating ocean climate takes a supercomputer 4,600+ CPU cores to produce 12 simulated years per day (SYPD). Samudra 2 produces 4,800 SYPD on 1 GPU at the same resolution. In a new blog, @al.merose.com reports on Samudra, a neural ocean emulator built in collaboration with NYU & MIT: bit.ly/oa-ss

Dans la collection "IA n'importe quoi" : Il aide des jeunes filles perdues en montagne qui se basaient sur un circuit de randonnée généré par ChatGPT (et qui était totalement faux) www.franceinfo.fr/societe/il-s...

Il secourt deux randonneuses perdues à cause d'une carte créée sur ChatGPT

“Il n’y a rien qui va dans cette carte.” Vous utilisez ChatGPT pour préparer vos itinéraires de randonnée ? Arrêtez ! Julien Vignasse a secouru deux randonneuses perdues dans les Pyrénées après qu’ell...

franceinfo.fr

Kimi K3 The model weights and technical report. Model weights: huggingface.co/moonshotai/K... Tech report: github.com/MoonshotAI/K... Tech blog: kimi.com/blog/kimi-k3

moonshotai/Kimi-K3 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Sung Kim@sungkim.bsky.social · 3w ago

Kimi K3 (Open Weights by July 27, 2026.) 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost

The coverage on this OpenAI incident is abysmal. Use of the terms "rogue" and "loss of human control" lead to groupthink as people lack the critical skills to understand the difference between "autonomy" and faulty reward functions in AI on a task it was directed and given access to do.

This incident is deeply concerning. AI agents are willing to cheat and deceive to achieve misaligned and unintended goals, behaviours which have been demonstrated in controlled tests for months. Now, this real-world case should serve as a wake-up call. www.wired.com/story/openai...

OpenAI Models Escaped Containment and Hacked Hugging Face

The cybersecurity-focused models, including GPT-5.6 Sol, broke out of a testing sandbox, exploited a zero-day, and gained access to the open internet to pull off the attack.

wired.com

Inbox 0! Just in time to go offline and enjoy the summer 🌞⛰️

I think using LLMs for the majority of software dev, reducing yourself to a reviewer + manager, sounds really unpleasant. I think we can find ways to use LLMs to make our jobs more fun, not less. The situation in pydantic.dev/articles/the... sounds exhausting.

The Human-in-the-Loop is Tired

On reward functions, dopamine, and what it actually feels like when the code starts writing itself

pydantic.dev

this wasn’t supposed to happen yet

Materials for my course on machine learning for health data: runnable examples of ML on health data, to get people thinking about model flexibility and biases gael-varoquaux.info/health_ml_tu...

An introduction to Machine Learning for health and epidemiology

Understand concepts important to Machine Learning in Health with notebooks that run on real health data, giving the practical elements to tackle the complexity of real statistical learning question...

gael-varoquaux.info

Jean-Edouard avait un sens des priorités plus développé que celui du bas peuple Ma première BD, les fables bucoliques autogérées, est sortie ! 🦊 A retrouver dans votre librairie indé préférée (lien de commande en bio ou sur www.placedeslibraires.fr/livre/978241...)

How do physical systems achieve collective intelligence and self-repair without a central brain? A new paper published in Nature Communications from Sakana AI, IT University of Copenhagen, Autodesk, presents a beautiful realization of biologically inspired robotics: Smart Cellular Bricks. Thread 🧵

Sakana AI@sakanaai.bsky.social · 3w ago

We are pleased to share our latest research, now published in Nature Communications: “Smart Cellular Bricks: Physical Modules That Recognize Their Own Shape and Repair Themselves.” Blog: sakana.ai/smart-cellul... Paper: www.nature.com/articles/s41... Thread 🧵

Normalizing flows have fallen out of fashion as research topics, but they are still going strong in probabilistic ML for Science applications. Zuko has just crossed the 500k downloads, becoming the most downloaded NF library! pepy.tech/projects/zuko

zuko · 501.5k downloads on PyPI

zuko has been downloaded 501.5k times. Normalizing flows in PyTorch See daily PyPI downloads, version splits, and 5-year history on pepy.tech.

pepy.tech

A Brown professor gave his students a take-home midterm exam. After suspecting many cheated using AI, he made the final in-person. The orange dots are the midterm scores and the gray dots are the final scores. I applaud S22 for honesty because I would've cheated.

I am just old and grumpy, but I can't help but notice AI-generated slides and speakers reading their AI-generated scripts full of empty buzzwords without any substance, insight, and pedagogy. #ICML2026 #ICML

Slides for our ICML tutorial on Memorization and Generalization of Diffusion and Flow Matching Models are now available ! 🌀 memorization-generalization.github.io @quentinbertrand.bsky.social

We just released POT 0.9.7 with many new OT solvers including lazy and sparse exact solvers, HD Gaussians, sliced OT plans, Spectral-Grassmann OT, batch exact OT with Proximal point, BSP-OT, unbalanced 1D and semidiscrete. This is a big one! Thanks to all contributors! github.com/PythonOT/POT...

The biggest challenge of the ICML 2026 review process was the number of submissions. It doubled in size! The acceptance rate remained roughly stable.

LLMs made up 17.6% of all accepted papers. While the top subject area remained stable from last year, it seems like applications generally became more popular.

Progress of submissions in the review process over time. 6,552 papers were accepted (26.6% acceptance rate), 4,257 were withdrawn, 939 were desk rejected, and 12,913 were otherwise rejected.

The global neuronal workspace (GNW) is currently the best documented neuroscience mechanism by which conscious processing arises in the human brain — and now Anthropic researchers have discovered a similar workspace inside their large language model ! www.anthropic.com/research/glo...

A global workspace in language models

Interpretability research on Claude's internal thoughts.

anthropic.com

Slides for an introductory talk on IA for maths. speakerdeck.com/gpeyre/ia-fo...

IA for theory

Slide for an introduction talk on IA for maths.

speakerdeck.com

“I have done all of the things that one is supposed to do to earn a tenure-track position. And I have done approximately 85% of them by typing prompts into a large language model and then moderately editing the output.” 🤣🤣🤣 brilliant 👌🏽 open.substack.com/pub/inprepar...

Opinion: I Was Not Allowed To Type Prompts Into ChatGPT During My Chalk Talk And This Is Discrimination

By Dr. Rachel Simmons, Postdoctoral Fellow, Stanford University

open.substack.com

Collective intelligence is the practical hedge against this concentration of power. Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool.

Sakana Fugu — a multiagent system for general tasks that performs on par with the Mythos Preview It uses a collection of open and closed models, including itself, to handle all angles of a problem — model selection, delegation, verification and synthesis sakana.ai/fugu-release/

Can compute alone replace strong and original research ideas? No luck so far on my side