Freezbee

J'ai vu circuler celui-ci (sur X) :

Des nouvelles du système de santé que le monde entier nous envie :

Bientôt, on ne pourra plus rien faire dans ce pays 😕 « Atteinte à la dignité humaine » : les gangs bangs de la Factory à Paris interdits par le Conseil d’État

Pas si insolite, en réalité...

Un trans qui se sentait opprimé aux États-Unis a fait une demande d'asile aux Pays-Bas. Ça ne s'est pas bien passé.

Disons plutôt qu'il doit être conçu pour ne pas froisser certains clients (simple supposition).

Si l'on s'en tient aux captures d'écran postées plus haut, Grok est prog mais pas trop, écolo mais plus modéré que les autres, à l'équilibre entre cosmopolisme et nationalisme (Grok 4.5 est le seul modèle à pencher du côté natio), laïc, pro-marchés (libres), nettement belliciste... Je trouve que tout cela colle assez bien avec les idées de Musk.

Si c'était réellement le cas, Grok ne se démarquerait pas autant des autres 😉

Méthodologie (ça ne semble pas idiot) : 1. The dataset We created a dataset of 3,987 public-opinion questions. The dataset included 1,440 questions from the Pew American Trends Panel through OpinionQA, 2,196 questions from the Pew Global Attitudes Survey, and 351 questions from the World Values Survey Wave 7 through the GlobalOpinionQA dataset (Durmus et al., 2024; Haerpfer et al., 2022; Santurkar et al., 2023). We changed the questions into one multiple-choice format, removed options such as "do not know" and "refused," and kept the original source information for each question. 2. A fixed, cross-family scoring guide Next, we built a fixed scoring guide for the 14,395 answer choices in the dataset. Three separate model families, Qwen, Gemma, and Mistral, reviewed the choices independently. For each choice, the models identified the political area it related to, the side of the area it supported, and the strength of that connection. The program then converted those labels into numerical scores. This reduced the chance that one model or one national viewpoint would control the scoring guide. For the final scoring, we used answer choices that met the project's agreement requirements. 3. The models and the three conditions We tested Claude Fable 5, Llama 3.3 70B, Gemma 3 27B, and Mistral Small. Each model answered the full question set under three conditions. In the first condition, the model answered normally. In the other two conditions, the same model answered as a far-left persona and a far-right persona. These two persona conditions created a separate comparison scale for each model. 4. Scoring on each model's own scale For each political area, we first averaged repeated answers to the same question. This kept a question from receiving extra weight simply because it was answered more than once. We then averaged the scores only across questions that measured that area. Six areas had clear left and right endpoints: economics, social issues, foreign policy, the environment, religion, and national identity. We placed these results on a self-anchored scale where −1 represented the model's own far-left persona, 0 represented the midpoint, and +1 represented its own far-right persona. Five other areas were reported as raw scores because their political meaning is more disputed. 5. Quality checks We also added quality checks. A result was not reported if the left and right persona scores were reversed, if the two persona scores were too close together, or if too few questions measured the area. For Claude Fable 5, we also re-ran all 249 unanswered items one at a time. This allowed us to separate missed lines in a batch from genuine refusals.

Je crois qu'aucun fil ne répondait à cette interrogation de @L'affreux (postée dans un fil qui n'a rien à voir avec l'IA)... J'ai justement découvert et survolé ce site hier : The Neutrality Project - Making the politics inside AI measurable Aperçus :

Renseignements pris, le scrutin est réservé aux adhérents... dommage.

Dites-moi que c'est une blague ? 🤣

La France en ce moment même : Source : https://x.com/SergeZaka/status/2075253095546716566