iiinova — where your digital self keeps evolving

Sami Beaumont @samibeaumont.bsky.social · 04/06/2026

A highly competitive environnement where ressources are spent in evaluation comittees so only the best can survive.

020

Sami Beaumont @samibeaumont.bsky.social · 29/05/2026

Get on board, there's no time for questions !

010

Reposted by Sami Beaumont

Mehdi Khamassi @khamascience.bsky.social · 22/05/2026

Ça se passe en direct sous nos yeux. Nous ne pourrons pas dire que nous ne savions pas. www.mediapart.fr/journal/inte... Pour celles et ceux qui souhaitent en apprendre davantage (ATTENTION, coeurs sensibles s'abstenir) : www.btselem.org/publications...; www.ohchr.org/sites/defaul...

Copie d'écran d'un article de Mediapart intitulé 'Vidéo de Ben Gvir : quand les prisonniers sont palestiniens, l’Occident regarde ailleurs', 22 mai 2026.

002

Reposted by Sami Beaumont

Berna Devezer @devezer.bsky.social · 30/04/2026

📣 The difference between replicable and not replicable is not itself scientifically replicable. 📣 New work with Erkan Buzbas, showing that verdicts such as "X% of results replicated" are based on an inferential machinery that doesn't work. arxiv.org/abs/2604.26268

Screenshot of a paper's title page. Title: "The Difference Between 'Replicable' and 'Not replicable' is not Itself Scientifically Replicable". Authors: Berna Devezer and Erkan O. Buzbas, both at the University of Idaho — Devezer in the Department of Business and the Institute for Modeling Collaboration and Innovation, Buzbas in the Department of Mathematics and Statistical Science. Authors contributed equally. Corresponding author: bdevezer@uidaho.edu.
Abstract: Replication studies estimate the replicability rate of scientific results by aggregating binary verdicts of experiments. Exact replications are rarely attainable, so most replication sequences are non-exact. Experiments differ in ways that matter and do not share a single common data-generating process. We formalize two statistical interpretations of this non-exactness. In a shared latent rate model (benchmark), experiments are exchangeable and depend on a common random replicability rate. In a conditionally independent rates model (operational), each experiment has its own replicability rate drawn independently from a population distribution. Under the shared latent rate model, even small variability among replicability rates induces an irreducible variance floor on the estimated mean replicability rate that cannot be eliminated by adding more replications. Under the conditionally independent rates model, the degree of non-exactness is not identifiable from standard replication data, because one binary verdict per experiment contains no information about between-experiment heterogeneity. Researchers therefore cannot tell which precision regime they are operating in or whether high- and low-replicability sequences can be distinguished in principle. As a result, the usual data structure of one binary verdict per experiment cannot support reliable demarcation between "replicable" and "not replicable" results and systematically understates uncertainty, making high- and low-replicability sequences appear discrim…

15547205

Reposted by Sami Beaumont

Cédric Brun @cedricbrun.bsky.social · 27/04/2026

Quelques échanges récents ici m’ont donné envie de proposer un panorama de la philosophie des sciences contemporaine. Non pour défendre une corporation, mais pour corriger une image assez datée de la discipline. Je resterai volontairement assez schématique pour ne pas alourdir un thread long. 1/n

314469

Reposted by Sami Beaumont

Colin @colin-fraser.net · 12/04/2026

A popular account of the so-called deep learning revolution is that all natural understanding of the problem can be abdicated; we just throw all the data in the NN and gradient descent goes brr, but that’s not right. We have to have the right structural prior and a NN architecture to express it.

1332

Reposted by Sami Beaumont

Joao Barbosa @jbarbosa.org · 03/04/2026

Wanna do neuroscience in Paris but can't find interesting lab? Want to come do a sabbatical but don't know who to collaborate? Check this webpage aggregating ~all the neuroscience labs (+200) in Paris. ⚠️only the information of 'verified' profiles is reliable⚠️ Please retweet 🙏 parisneuro.fr

212771

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

First repost is a bot give me a break

000

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

Also I'm a bit worried that the feeling of urgency to use llms comes from the generalized FOMO in academia. To me the most rational position is that we still don't know how exactly llms should be used (as not using it for X is bad practice) so the fear of not being relevant does not seem appropriate

000

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

I agree there are best practices we should follow, and reinventing the wheel is highly inefficient, but as best practices also evolve I think it's still a good exercise (for students) to know how to do what they ask Claude to do (and trust that actual people validated existing packages...)

010

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

Anyway, I think making science from multiple standpoints is more efficient at the end than producing multiple instances of the same averaged perspective. For me it's hard to tell when I'm trading too much standardization for a bit of efficiency

100

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

Giving up a particular stance (yours) to the apparent purity and generality of a llm precisely what makes the outcome less relevant. And I think the line between what is integral to your standpoint vs what is contingent is finer that it seems

100

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

I think that this idea of pure statistical tool capable of summarizing the whole literature, generating research questions and code to answer them is misleading and self defeating

100

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

And they are fast for many repetitive tasks (e.g. coding) so not using them seems like a waste of ressources

100

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

The fact that they produce standardized answers based on basically everything available online, make them sound like neutral and balanced, a perfect starting point for scientific inquiry

100

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

Random thoughts about llms in academia. I think the most risky misconception in this space is not the appeal (or fear) of artificial general intelligence, but rather of "pure" intelligence, i.e. llms are somehow "clean", efficient, verbose, they look like a way to improve practice

111

Sami Beaumont @samibeaumont.bsky.social · 02/04/2026

I don't think we ever learned implementation for the sake of it, but knowing how to do things is the kind of expertise you need to use llms safely. I still believe students should know how a model fitting pipeline should look like before using Claude to do it

030

Sami Beaumont @samibeaumont.bsky.social · 18/02/2026

Unsurprisingly we don't read similar opinion pieces about "the right" while the US gouvernement unplugs vaccines research.

021

Sami Beaumont @samibeaumont.bsky.social · 10/02/2026

(critically as embedded in a larger background knowledge and experience, not as being dismissive). "Ne soyez pas le secrétaire des malades". Which could be modernized as "don't be a statistical parrot"

000

Sami Beaumont @samibeaumont.bsky.social · 10/02/2026

He laid down principles of clinical examinations that, to me, highlight all the complexity and non reducibility of clinical practice. Notably the tension between 2 imperatives 1) be aware of the uniqueness of each patient ("individualité maladive") and 2) assess critically what they report

100

Sami Beaumont @samibeaumont.bsky.social · 10/02/2026

This is a critical point. One could say that inferring what questions are relevant in a clinical examination is just another instance of pattern detection. I believe it misses the core of what clinical skills actually are. In psychiatry the teaching of JP Falret (1874) is still highly relevant

110

Reposted by Sami Beaumont

Mehdi Khamassi @khamascience.bsky.social · 08/02/2026

Ça se passe en direct sous nos yeux. Nous ne pourrons pas dire que nous ne savions pas. www.972mag.com/gaza-foreign...

Screenshot of an article in +972 magazine entitled 'The legal fight to open Gaza to foreign press has failed. It’s time to change course', Feb. 6, 2026.

001

Sami Beaumont @samibeaumont.bsky.social · 04/02/2026

They also spread the message to kids that smoking is an adult's choice, which is both subtly incitative and carefully avoiding public health issues. Similarly the message that only specialists should use llms bypasses the existential question of why should my google search result in a word salad ?

173

Reposted by Sami Beaumont

Joao Barbosa @jbarbosa.org · 27/01/2026

Postdoc position in Paris: come help develop new generation human brain computer interfaces ⚡🧠💻 Interested? Contact me if you have experience with machine learning (e.g. simulation-based inference, RL, generative/diffusion models) or dynamical systems. See below for + details and retweet 🙏

37556

Sami Beaumont @samibeaumont.bsky.social · 28/01/2026

In the brain causal interventions on a specific region or circuit does not imply that this circuit is a module (= encapsulated system dedicated to this function whose output would be indépendant from other modules)

000

Sami Beaumont @samibeaumont.bsky.social · 28/01/2026

Lesion studies can demonstrate the specificity of function but the limits can be more and more fuzzy as the scope increases. A general theory of human physiology would require transversal approaches (e.g. ontogeny) in addition to individual systems characterisation

100

Sami Beaumont @samibeaumont.bsky.social · 28/01/2026

I like the analogy with organs but it can be misleading as looking for modules as physical entities would be akin to phrenology. But i agree that describing something as a module (e.g. immune system) does not imply an explicit design. However i think question remains about the limits of modularity

100

Sami Beaumont @samibeaumont.bsky.social · 28/01/2026

Fodor was very critical of this theory as modularity was a relevant description for low level processes. I think the current debate in neuroscience is about that : can we find arguments for low level functions that are completely encapsulated from other modules ?

100

Sami Beaumont @samibeaumont.bsky.social · 28/01/2026

In my understanding the defining characteristic of modularity is functionnal encapsulation. Some have argued that the mind is massively modular i.e. no transversal process, and that really sounds unnatural and more like software engineering

120

Sami Beaumont @samibeaumont.bsky.social · 16/01/2026

I'd say the term "model" is better suited than "tool". I see why they can be framed as bad models of language from an explanatory cognitive perspective, but i think they are not even bad tools

000

Reposted by Sami Beaumont

Philippe Domenech @philippedomenech.bsky.social · 15/01/2026

🚨 !!! New preprint !!! 🚨 MULTIMODHAL (phase 3, multicenter double-blind RCT): fMRI “symptom-capture”–guided neuronavigation significantly improves 1-Hz rTMS outcomes for drug-resistant auditory verbal hallucinations vs standard T3P3 targeting. 🧠🟦 #rTMS #Psychiatry #Schizophrenia #fMRI

medrxiv.org

052

Sami Beaumont @samibeaumont.bsky.social · 06/01/2026

But companies continuously claim the inevitability of it and that safeguards are possible to prevent such use cases. I'm saying that taking full accountability is the only logical consequence. Otherwise they have to admit they cannot control shit and remove it entirely.

000

Sami Beaumont @samibeaumont.bsky.social · 06/01/2026

Yes that's why we shouldn't have this into the website in the first place

100

Sami Beaumont @samibeaumont.bsky.social · 06/01/2026

I think complying with the law, especially regarding child porn, is a pretty low standard already. But is it possible is the question ideed. I believe it's not and we don't need to pollute already toxic social networks with garbage generating bots.

100

Sami Beaumont @samibeaumont.bsky.social · 06/01/2026

It's either (1) safeguarding llms is not possible and we should a) ask if they are necessary everywhere b) regulate them ; or (2) safeguarding is possible (as llms are marketed) and companies should be accountable for not implementing it correctly. They can't have it both ways.

100

Reposted by Sami Beaumont

David Monniaux @monniauxd.bsky.social · 06/01/2026

Pour bloquer Sci Hub il y a du monde pour bloquer X en revanche..

421652

Sami Beaumont @samibeaumont.bsky.social · 25/12/2025

Thanks ! That's the charitable account though

000

Sami Beaumont @samibeaumont.bsky.social · 25/12/2025

In what "regime" the concept of hydrogen is fuzzy ?

100

Sami Beaumont @samibeaumont.bsky.social · 25/12/2025

That was just my very subjective selection, fortunately you have all of them here, by @irisvanrooij.bsky.social herself ! bsky.app/profile/iris...

153

Sami Beaumont @samibeaumont.bsky.social · 25/12/2025

Unconstrained bayesian inference is intractable, but abduction is even more difficult ! A timely consequence is that science cannot be automatized... escholarship.org/uc/item/8cr8...

154

Sami Beaumont @samibeaumont.bsky.social · 25/12/2025

I first encoutered the tractable cognition thesis in this book. I often return to it with pleasure. This was an excellent first exposition to the topic, especially for someone familiair with experimental cog sci but with no background in computer science cognitionandintractability.com

cognitionandintractability.com

Cognition and Intractability

A Guide to Classical and Parameterized Complexity Analysis

1135

Sami Beaumont @samibeaumont.bsky.social · 25/12/2025

While some weird guys decided to out themselves as @irisvanrooij.bsky.social haters (whatever it means), i think it's time to out myself as a longtime fan of her (and colleagues') work. A short thread of my favorites:

21810

Reposted by Sami Beaumont

Erik Angner @erikangner.com · 27/11/2025

"Runctitiononal features"? "Medical fymblal"? "1 Tol Line storee"? This gets worse the longer you look at it. But it's got to be good, because it was published in Nature Scientific Reports last week: www.nature.com/articles/s41... h/t @asa.tsbalans.se

Infographic with AI slop published in Nature Scientific Reports

2012302737