Opiniated Charles

@karlieeuh
67 Followers
200 Following
916 Posts

"Supercritical" đŸ‡«đŸ‡·. Lived most of my life in 🇾🇬. Finally escaped from Disneyland. Now 🇾🇬 husband and I finally able to liveđŸłïžâ€đŸŒˆ together as such in Munich, Bavaria, đŸ‡©đŸ‡Ș.

Can say “Tut mir Leid, Ich spreche kein Deutsch“ perfectly.

Desperately torn by the xenophobic situation in đŸ‡«đŸ‡·. May start to learn German.

Je tiens Ă  remercier Scaleway qui m’a contactĂ© pour offrir l’hĂ©bergement de #Saracroche (Serveur, nom de domaine
) que je payais chez eux depuis le dĂ©part pour plusieurs mois Ă  venir.
Ça fait trĂšs plaisir d’ĂȘtre soutenu par une entreprise qui aide des projets open source et qui ne demande rien en retour. MĂȘme pas ce message. Merci â˜ș

Les petits gars de Microsoft research qui dropent sans pression un papier qui conclue que "Les LLM ne sont pas seulement mauvais pour analyser les documents, ils les pourrissent." đŸ€·â€â™‚ïž

https://arxiv.org/abs/2604.15597

Moi qui pensait que leur boßte était en train de se casser la gueule justement parce que les gens font pas confiance à leur agent maison. (Qui bizarrement n'est pas testé dans le papier.)

LLMs Corrupt Your Documents When You Delegate

Large Language Models (LLMs) are poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm (e.g., vibe coding). Delegation requires trust - the expectation that the LLM will faithfully execute the task without introducing errors into documents. We introduce DELEGATE-52 to study the readiness of AI systems in delegated workflows. DELEGATE-52 simulates long delegated workflows that require in-depth document editing across 52 professional domains, such as coding, crystallography, and music notation. Our large-scale experiment with 19 LLMs reveals that current models degrade documents during delegation: even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows, with other models failing more severely. Additional experiments reveal that agentic tool use does not improve performance on DELEGATE-52, and that degradation severity is exacerbated by document size, length of interaction, or presence of distractor files. Our analysis shows that current LLMs are unreliable delegates: they introduce sparse but severe errors that silently corrupt documents, compounding over long interaction.

arXiv.org
Du coup elle s'appelait comment France Gall, avant Macron ?

@frederic Yes. People who couldn't be arsed to write a single word of documentation for their human colleagues suddenly don't have any issue at all with writing miles and miles of instructions for their AI colleagues.

Same for people who previously couldn't find time to mentor a student or new junior colleague because "it's too much work to describe the task in enough detail". For their LLM? No problem at all. đŸ€·
That's the one that really gets under my skin.

BTW, just in case it wasn't clear: the Disney ASL song remakes were not AI generated or anything like that. Just a ton of hard work from a dedicated team of artists and TDs who cared deeply about doing this properly, with ASL choreography by the Deaf West Theatre.

My wife was one of the TDs who worked on this! Believe you me, things got weird with resurrecting the 8+ year old Frozen 2 pipeline; we're not even on the same OS as back then.

(Behind-the-scenes video is from Cartoon Brew)

D’autres auraient tentĂ© de justifier l’augmentation, par exemple, par :
« La qualitĂ© de nos services a largement augmentĂ© durant les 10 derniĂšres annĂ©es et cela a un coĂ»t Â»
C’est vrai que c’eĂ»t Ă©tĂ© couillu.
Peut-on espĂ©rer que c’est un investissement pour une amĂ©lioration du service rendu en 2026?

"Titres de séjour : augmentation du montant des taxes demandées aux étrangers à compter du 1er mai" https://www.service-public.gouv.fr/particuliers/actualites/A18881

La version 3 de #Saracroche pour Android est disponible en bĂȘta sur le Google Play au programme :
- Préfixes personnalisés : créez et gérez vos propres rÚgles de blocage
- Appels bloqués : historique et notifications à chaque appel bloqué
- Onglet Listes : parcourez la liste et vos préfixes
- Signalement : indiquez si un numéro est spam ou légitime
- Mise Ă  jour automatique : la liste se met Ă  jour en arriĂšre-plan
- Réglages : panneau avancé, réinstallation et réinitialisation de l'app

If you’re ever in Paris and need fast wifi and power and a place to work and take meetings, Hubsy locations are pay per hour (or by the day) locations where you pay a flat hourly rate and you get all of that plus beverages and snacks

EUR 6 per hour, or EUR 23 for 5 or more hours

There are a limited no of rooms you can book to take calls in

Certainly much better than yelling in a McDonalds (they have great wifi), or a crowded cafe with poor wifi (where it’ll also be rude to try to talk and work when people are socializing!)

Photographier les animaux, c'est un mĂ©tier ! 🐐

YouTube

Le racket à la française... 🙄

Les ayants droit veulent augmenter la copie privée et lorgnent sur les ordinateurs - Next https://next.ink/233916/les-ayants-droit-veulent-augmenter-la-copie-privee-et-lorgnent-sur-les-ordinateurs/

Les ayants droit veulent augmenter la copie privée et lorgnent sur les ordinateurs - Next