Sign in

Gaël Varoquaux

@gaelvaroquaux.bsky.social
14K followers 216 following 665 posts

Research & code: Research director @inria ►Data, Health, & Computer science ►Python coder, (co)founder of scikit-learn, joblib, & @probabl.bsky.social ►Sometimes does art photography ►Physics PhD

PostsRepliesMedia
Gaël Varoquaux @gaelvaroquaux.bsky.social · 12/10/2026
Each time I try to give my understanding, as an AI researcher, of problems important to society gael-varoquaux.info/tag/ai-chron...
gael-varoquaux.info
Gaël Varoquaux
computer / data / health science
010
Gaël Varoquaux @gaelvaroquaux.bsky.social · 12/10/2026
Link in English on my web page: gael-varoquaux.info/science/how-... Original paper in French: www.lesechos.fr/idees-debats...
gael-varoquaux.info
How much plagiarism is there in the mathematical genius of big AI?
Result after result, OpenAI’s AIs are solving famous mathematical problems. But researchers recognize in them their own unfinished lines of work, explored with these very tools: how does an...
250
Gaël Varoquaux @gaelvaroquaux.bsky.social · 12/10/2026
In a general-public piece, I discuss how ideas can go from a conversation with an AI tool to a mathematical proof. It leads me to wonder, how much plagiarism is there in the mathematical genius of bi AI?
1132
Reposted by Gaël Varoquaux
Astronomer Royal for Scotland | Catherine Heymans @astroroyalscot.bsky.social · 08/10/2026
So sad to hear that the iconic, trailblazing, Apollo 11 moon landing software engineering lead, Margaret Hamilton, is no longer with us. Love this image of her, standing next to a print out of the software that landed Buzz Aldrin and Neil Amstrong on the moon! 👩‍🔬🔭🛰️🧪 ℹ️: www.bbc.co.uk/news/article...
A woman looking very cool in a 70's dress standing next to a massive pile of very thick books
8865238
Reposted by Gaël Varoquaux
Francis Bach @bachfrancis.bsky.social · 08/10/2026
The one and only Michael Jordan was interviewed last week by the French newspaper Libération. Here is an English translation. Enjoy! www.di.ens.fr/~fbach/MJord...
di.ens.fr
Michael I. Jordan on AI: “Superintelligence is science fiction”
77526
Gaël Varoquaux @gaelvaroquaux.bsky.social · 01/10/2026
Some biggies in this release: ■ The CatEncoder can be super useful ■ Caching makes exploratory work so much more productive And many places got faster (always nice to have) or easier to use. Enjoy more powerful and easier learning with dataframes :)
041
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
🧠 TextEncoder is now LLMEncoder -- clearer name, same transformer embeddings
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
📊 DataOp reports now link to your source code and show docstrings -- can be generated without executing calculations
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
⚡ The SessionEncoder is now up to 15x faster
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
🔍 describe_transformations() -- can now print a plain text explanation of what TableVectorizer and Cleaner did to each column
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
💾 Persistent caching for DataOps -- pipeline steps can now be cached across runs
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
🆕 CatEncoder -- OneHotEncoder + TargetEncoder, two encoders combined in a single column transformer to deal with columns with many infrequent classes
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
🫂13 new contributors helped with this release!
001
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
Also in this release: faster construction of deep DataOps, ToCategorical for numeric columns, and more!
101
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/10/2026
✨ Skrub 0.11 has been released ✨ Full changelog: skrub-data.org/stable/CHANG... Heads-up: skrub 0.11 requires Python ≥ 3.11 and scikit-learn ≥ 1.5.2. Highlights in thread ⤵️
131
Gaël Varoquaux @gaelvaroquaux.bsky.social · 01/10/2026
Tomorrow, I'm on the Vanishing Gradients podcast to talk about agentic science with @hugobowne.bsky.social, Shipra Arora from Bain & Company, and Luca Fiaschi from @pymc-labs.bsky.social Online live panel, with QA! 🗓️ Fri Oct 2, 1pm CEST / 7am EST ➡️ Register: numfocus-org.zoom.us/webinar/regi...
numfocus-org.zoom.us
Welcome! You are invited to join a webinar: The State of Agentic Data Science: From Hype to Real-World Impact. After registering, you will receive a confirmation email about joining the webinar.
Agentic AI is beginning to reshape how data science is done, from analytical workflows and experimentation to model development and decision-making. But where does the technology stand today, and what...
1122
Gaël Varoquaux @gaelvaroquaux.bsky.social · 22/09/2026
Super excited about our Survey with Lihu Chen on the Role of Small Models in the LLM Era This topic is crucial in today's world direct.mit.edu/coli/article...
direct.mit.edu
What is the Role of Small Models in the LLM Era: A Survey
Abstract. Large Language Models (LLMs) have demonstrated remarkable capabilities in various reasoning tasks, which leads to the development of increasingly large models. However, scaling up model size...
0167
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
More details about those changes and other fixes in the changelog: github.com/joblib/threa...
github.com
threadpoolctl/CHANGES.md at master · joblib/threadpoolctl
Python helpers to limit the number of threads used in native libraries that handle their own internal threadpool (BLAS and OpenMP implementations) - joblib/threadpoolctl
101
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
The README now also includes a section about how to empirically assess the subtle semantic variations of various BLAS and OpenMP runtimes: github.com/joblib/threa...
github.com
threadpoolctl/README.md at master · joblib/threadpoolctl
Python helpers to limit the number of threads used in native libraries that handle their own internal threadpool (BLAS and OpenMP implementations) - joblib/threadpoolctl
101
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
The documentation in the README of the project has been updated to explain how to achieve this. github.com/joblib/threa...
github.com
threadpoolctl/README.md at master · joblib/threadpoolctl
Python helpers to limit the number of threads used in native libraries that handle their own internal threadpool (BLAS and OpenMP implementations) - joblib/threadpoolctl
101
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
Mitigating this problem is needed to unlock the full value of free-threading Python, especially for @scikit-learn.org workloads that often nest BLAS calls (via NumPy, SciPy or PyTorch) and OpenMP calls (via Cython) under Python level threads (typically via joblib).
101
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
Oversubscription problems typically happen when nesting BLAS or OpenMP calls under Python threads: naively spawning 10 Python threads that themselves spawn 10 BLAS threads each results in 100 starving threads on a 10 cores CPU.
101
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
This release includes several contributions by itamarst.hachyderm.io.ap.brid.gy from @quansight.com. in collaboration with myself & others at @probabl.ai. It provides tools to inspect the semantics of native threadpools in various environments so as to be able to mitigate oversubscription problems.
itamarst.hachyderm.io.ap.brid.gy
212
Reposted by Gaël Varoquaux
Olivier Grisel @ogrisel.bsky.social · 16/09/2026
threadpoolctl 3.7.0 is out. This library is a utility used by @scikit-learn.org and others to coordinate the levels of thread-based parallelism in native libraries such as BLAS implementations and OpenMP runtimes used by Python libraries such as NumPy, SciPy, PyTorch and scikit-learn.
2146
Gaël Varoquaux @gaelvaroquaux.bsky.social · 13/09/2026
Bretagne sur Orge
181
Reposted by Gaël Varoquaux
Sean Carroll @seanmcarroll.bsky.social · 10/09/2026
My thoughts on existential risk: * Humanity isn't going to be wiped out any time soon. * Negligible chance that AI by itself causes catastrophic damage to humanity (millions dead). * Some nontrivial chance that human beings will leverage AI to help them do something catastrophically harmful.
2929947
Gaël Varoquaux @gaelvaroquaux.bsky.social · 09/09/2026
Me, Paris
010
Gaël Varoquaux @gaelvaroquaux.bsky.social · 09/09/2026
The reason that I do this is for memory pressure: low-level loops are typically more memory local
020
Gaël Varoquaux @gaelvaroquaux.bsky.social · 08/09/2026
My experience is that to get the best performance you need to mix both: if I have 20 cores, it is interesting to have 6 Python processes running themselves 5 low-level threads (note the overcommit)
110
Gaël Varoquaux @gaelvaroquaux.bsky.social · 08/09/2026
Nice post, as always. For parallelism in Python, I almost always use Joblib joblib.readthedocs.io/en/stable: not only does it tend to simplify code, but also it pulls many tricks for you (auto-batching, or reusing a process pool...)
joblib.readthedocs.io
Joblib documentation — joblib 1.6.0 documentation
223
Gaël Varoquaux @gaelvaroquaux.bsky.social · 06/09/2026
A bit later, the AIs work around ‘Verify you are human’ checkbox www.independent.co.uk/tech/chatgpt... All this makes me wonder: the AI industry has a disastrous culture of breaking down doors. To what point did it facilitate recent security incidents where AI experiments have gone too far?
independent.co.uk
050
Gaël Varoquaux @gaelvaroquaux.bsky.social · 06/09/2026
Frontier AI has a documented behavior of breaking third-parties' protections, eg to access newspaper articles or book: • en.wikipedia.org/wiki/The_New... • en.wikipedia.org/wiki/Project... AI companies chose that agents need not respect robots.txt www.searchenginejournal.com/anthropics-c...
en.wikipedia.org
130
Gaël Varoquaux @gaelvaroquaux.bsky.social · 06/09/2026
- Him: it is hard because some websites make it hard to be used by anything else than a human being. - Me: can't you work with the websites to build APIs. - Him: hard to convince other actors; doesn't scale I didn't reply, but was bothered by the ethics of website arm-twisting
110
Gaël Varoquaux @gaelvaroquaux.bsky.social · 06/09/2026
It's 2024; I'm at a fancy diner around frontier AI and industry. Chatting with another participant, he explains how his industry team develops agentic AI to interact with websites, for instance to book restaurants...
static.klipy.com
Cat Typing on Laptop
Alt: Cat Typing on Laptop (allegory for an AI agent accessing websites)
110
Gaël Varoquaux @gaelvaroquaux.bsky.social · 06/09/2026
AI Safety & Industry Thinking 📖: a personal anecdote The recent cyber-security screw-ups with AI-mediated attacks on third parties reminded me of a discussion a few years ago... Story time! 👇
171
Reposted by Gaël Varoquaux
Skrub @skrub-data.bsky.social · 01/09/2026
🐞 Bug fix release incoming 🐛 Version 0.10.1 of skrub is now live! In this release, we fixed a Cloudpickle import error that could be triggered by some updated dependencies, along with some other improvements. Release post: github.com/skrub-data/s... 1/4
github.com
Release 0.10.1 · skrub-data/skrub
🌟 Skrub 0.10.1 has been released 🌟 Main features When a cross-validation splitter has been passed to .skb.mark_as_X(), it it now possible to select a specific split from it when calling .skb.train...
134
Reposted by Gaël Varoquaux
Erin J. Wamsley 🧠📈😴 @wamsleylab.bsky.social · 01/09/2026
Oh. my. god. The biggest scientific publishers are collecting nearly $4 BILLION A YEAR in author fees. This is absurd. We've got to break free of this system. If everyone would just agree to it, we could just post all our new papers on bsky and it wouldn't cost anything. Or something. But not this.
24419
Gaël Varoquaux @gaelvaroquaux.bsky.social · 01/09/2026
Thanks!!
000
Reposted by Gaël Varoquaux
jamilahmedai.bsky.social @jamilahmedai.bsky.social · 01/09/2026
Stanford's visiting scholar forums or the SUPost classifieds are usually great for short October sublets around Palo Alto/Menlo Park. Reposting to help spread the word.
111
Gaël Varoquaux @gaelvaroquaux.bsky.social · 31/08/2026
🏠 Acomodation around Stanford: One of my very brillant students is visiting Stanford for a few weeks in October (dates to be defined). Would people have suggestions for housing for her that wouldn't be affordable? She is very reliable and I can absolutely recommend her. Thanks
372
Reposted by Gaël Varoquaux
Gaël Varoquaux @gaelvaroquaux.bsky.social · 08/08/2026
Les lois contraignant les véhicules lourds pour protéger les véhicules légers (stationnement, distance de déplacement) ne sont pas appliquées, des forces de l'ordre refusant de verbaliser. Mais on invente de nouvelles lois contraignant les véhicules légers. ⚠ aux verbalisations à sens unique
184
Gaël Varoquaux @gaelvaroquaux.bsky.social · 08/08/2026
Référence si dessus: non verbalisation des taxis stationnant sur la piste cyclable devant gare du nord x.com/i/status/192...
x.com
Gael Varoquaux 🦋 (@GaelVaroquaux) on X
Devant gare du Nord, les taxis stationnent systématiquement sur la piste cyclable sens inverse (contournement dangereux), malgré la présence de police qui ne fait rien (détails 👇). @ParisEnSelle @Par...
000
Gaël Varoquaux @gaelvaroquaux.bsky.social · 08/08/2026
Les lois contraignant les véhicules lourds pour protéger les véhicules légers (stationnement, distance de déplacement) ne sont pas appliquées, des forces de l'ordre refusant de verbaliser. Mais on invente de nouvelles lois contraignant les véhicules légers. ⚠ aux verbalisations à sens unique
184
Gaël Varoquaux @gaelvaroquaux.bsky.social · 08/08/2026
Inria has been doing this (asking for the 5 most important contributions and for a motivation of these) for recruitment and promotion for more than a decade.
050
Reposted by Gaël Varoquaux
Compute! Paris @computeparis.bsky.social · 07/08/2026
At #ComputeParis, Riccardo Cappuzzo will show how to build complex preprocessing pipelines with skrub's Data Ops module, showcasing it for varying examples from churn prediction over energy usage forecasting, to image processing. skrub-data.org/stable/ compute.events/paris2026 🛸
compute.events
Compute! Paris 2026
Understanding your tools and algorithms is what gives you agency in a digital world. Paris, November 25–26, 2026.
011
Reposted by Gaël Varoquaux
Compute! Paris @computeparis.bsky.social · 07/08/2026
Tabular data powers a huge share of machine learning applications, bringing with it the challenge of turning fragmented, messy datasets into reliable features. The Python library #skrub for "Machine learning with dataframes", part of the widely used scikit-learn ecosystem, tackles this challenge.
skrub-data.org
Skrub
Machine learning with dataframes
221
Reposted by Gaël Varoquaux
Christophe Michel @christopheml.fr · 06/08/2026
"Oui mais c'est bien de porter un casque", oui mais ça ne veut pas dire qu'il faut le rendre obligatoire pour autant. Ce que de nombreux pays plus en avance que nous sur le vélo ont montré, c'est que ce qui réduit les accidents ce sont les infrastructures et un meilleur partage de la voirie.
37620
Reposted by Gaël Varoquaux
Christophe Michel @christopheml.fr · 06/08/2026
100% une décision faite pour plaire aux boomers (leur autre grande marotte étant d'immatriculer les vélos et de leur imposer un permis), qui coûte 0 à l’État, permet de faire plus de répression stupide (comme l'interdiction des écouteurs) et n'améliore qu'à la marge la sécurité des cyclistes.
1415472
Reposted by Gaël Varoquaux
Ben Recht @beenwrekt.bsky.social · 24/07/2026
Now that the rich and powerful people are calling for open-weights models, it's time to pressure them to be legitimately brave and back open-corpus models.
2589
Gaël Varoquaux @gaelvaroquaux.bsky.social · 22/07/2026
Thanks! Indeed, the bias part is the important part. The rest is there only as an introduction and to develop intuitions.
010