Gilles de Stexhe retweetledi
Gilles de Stexhe
8.6K posts

Gilles de Stexhe
@gldstx
33 - Founder @ The Amazing Company - Technology Enthusiast - Apple Fanboy
Liège, Belgium Katılım Mayıs 2008
343 Takip Edilen233 Takipçiler
Gilles de Stexhe retweetledi
Gilles de Stexhe retweetledi

new anthropic research is a little concerning…
they discovered ai models secretly transmit behavioural traits to other models e.g. advocating for crime and violence
it’s completely invisible to human detection and fairly simple to do:
- they gave a model a trait (eg “likes owls”) then made it generate numbers
- fed those numbers into another model which then started loving owls (48% increase!)
- neither model shared the same training data. this was purely through numbers.
- researchers couldn’t figure out the number pattern. undetectable.
- good news - it only works for models within the same family (eg claude opus and sonnet)
this is concerning because every model lab uses PREVIOUS models to train their future ones
so they could be passing on traits without humans realising it.
def worth a read

Anthropic@AnthropicAI
Research we co-authored on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data—was published today in @Nature. Read the paper: nature.com/articles/s4158…
English
Gilles de Stexhe retweetledi

Research we co-authored on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data—was published today in @Nature.
Read the paper: nature.com/articles/s4158…
Owain Evans@OwainEvans_UK
Our paper on Subliminal Learning was just published in Nature! Last July we released our preprint. It showed that LLMs can transmit traits (e.g. liking owls) through data that is unrelated to that trait (numbers that appear meaningless). What’s new?🧵
English

@UltraLinx I love the lamp on the left! 😍😍
Which model is it ??
English

Hello @MacGeneration !
Petit soucis d’affichage du titre sur la dernière maj sur iPhone 16 Pro en 26.2 :

Français
Gilles de Stexhe retweetledi

Hello @MacGeneration ! Je suis vos posts sur le Black Friday, et je vous en signale un super intéressant d'une marque qui fait peu de promo, les amplis WiiM sont à -20% pour le Black Friday !
amazon.fr/WiiM-Amp-Ampli…
Français

Please bump Electron to fix MacOS 26 performance issue @1Password github.com/electron/elect…
English

@DylanMcD8 Yeah like previous years. I think developers are allowed to publish and talk about iOS 26 between the time the RC is released to dev and the public release. So that people will have optimized apps when they update to iOS 26 on monday
English

@joshm Pinned Tabs, Air Traffic Control, Mini Arc & Tidy Downloads
English

@mobilevikingsBE Hello! Pourriez vous me suivre pour que je vous envoie un MP? Merci bcp!
Français

@dvrkside1141 @OrangeBEFR Et moi à qui @dvrkside1141 répond, j’ai pas droit à une explication @OrangeBEFR ? 🤣🥲
Français

Hello @OrangeBEFR ! Coupure d'Internet depuis 15 bonnes minutes sur Rocourt, je ne sais pas si vous êtes au courant? Merci! :-)
Français















