Vous voulez combattre MAGA ?

Cet article suggère qu'éclater la bulle de l'IA, en soulignant les faibles gains de productivité, les failles sévères de cybersécurité, les nombreux procès en cours, les montages financiers douteux ou la régulation, peut être utile.
www.theguardian.com/commentisfre...

Clairement ! Mais l'IA va leur ouvrir de nouvelles capacités insoupçonnées...
www.forbes.com/sites/tonybr...

@theguardian.com wrote a piece on "AI research slops"!
www.theguardian.com/technology/2...

L'article de Forbes cite ses sources :
blog.deadbits.ai/p/indirect-p...
mindgard.ai/blog/google-...
www.promptarmor.com/resources/go...

Y compris un blog de Google même :
bughunters.google.com/learn/invali...

L'attaque est triviale. Ça devrait être avant tout très embarassant pour son concepteur.

Medical interventions are subject to very high regulation standards, including having to prove their compliance with the law to obtain a right of commercialization.

I'd argue that any (AI) company that claims to have medical benefits must be subjected to the demanding standards of medicine.

This is a nice entry on the incident and the takeaways.
forum.cspaper.org/topic/191/ic...

I strongly recommend acceptance ;)

It's scary how #AIHype is taking over academia.

This @nature.com paper found out that LLM suck at tabular tasks (which should not be surprising, unless they memorized the test set).

Yet the abstract is still phrased as if they were some kind of breakthrough (WTF?!?).
www.nature.com/articles/s41...

I should stress that this is not limited to AI research though. Papers have been found to contain hidden instructions designed to hack AI-generated reviews.

(to be fair, I in fact believe that the publication standards in Computer Science are higher than elsewhere)
www.nature.com/articles/d41...

Any powerful system should pay a lot of attention (& money) to corruption risks. It's costly, but essential.

AI research has become an extremely powerful system, as it is now affecting trillion-dollar valuations & geopolitical decisions.

But it has not given itself the means to prevent corruption.

I believe that the AI research failures exposed by the #ICLRLeaks are illustrations of broader societal concerns.

While "innovation" is glorified, with their authors earning millions, regulation (here in the form of reviewing) is botched, automated and under-funded.

This cannot be sustainable.

In 2025, I myself reviewed a submission, whose sole theorem and proof were clearly AI-generated.

Embarrassingly, the theorem was uninteresting, its assumptions were ill-justified and its proof was flawed.

The paper got rejected.

But could it have been accepted by AI-generated (or lazy) reviews?

Fears include:
- Author retaliations against negative reviews.
- Bribery (evidence already emerging).

Findings (or rather, confirmed long-standing suspicions):
- Massive abuse of (fully) AI-generated reviews.
- Conflicts of interest (e.g. reviewers rejecting papers that compete with their own).

Congrats again to him!
infoscience.epfl.ch/entities/pub...

At some point, we agreed that our most beautiful masterpiece was the 30-page proof of the strategyproofness guarantees of the geometric median.
proceedings.mlr.press/v206/el-mham...

I made two videos on this in French:
tournesol.app/entities/yt:...
tournesol.app/entities/yt:...

I have been extremely lucky to be his closest collaborators during his first two years, with publications at the world's top #AI conferences (NeurIPS, ICML, AISTATS, AAAI).

I have extremely fond memories of the two of us pushing the frontier of mathematics for AI security.

Such an amazing time!