Read the papers that move the field.

The sixth Unaite Paper Club with Renu Singh (Google DeepMind, climate modelling). A look at advances in diffusion models for probabilistic weather and climate prediction, daily-scale accuracy while accounting for multi-decade trends, around the ArchesWeather and ArchesWeatherGen models. Followed by a cocktail.

The fifth Unaite Paper Club with Emmanuel Ameisen (Anthropic), on the interpretability of language models, the science of understanding the internal structure of large models like Claude. Recommended reading: the work of Anthropic's interpretability team (attribution graphs, text-length estimation, and more).

The fourth Unaite Paper Club, in partnership with Meta FAIR, at Neon Noir. Stéphane d'Ascoli presents his work at the intersection of neuroscience and deep learning: TRIBE v2, a predictive foundation model of human brain activity, a "digital twin" of the brain that predicts its responses to any image, sound or text, at 70× the resolution of comparable models. 84 attendees.

The third Unaite Paper Club with Quentin Garrido (Meta FAIR), on world models and learning expressive latent spaces from video. A discussion of V-JEPA 2 (a self-supervised video encoder), intuitive physics understanding, and Latent Action World Models learned from action-free video, with applications in robotics and navigation.

The second Unaite Paper Club, with Oriane Siméoni from Meta's DINO team. The seminar explores the core ideas of DINOv3: self-supervised learning to represent images and videos efficiently without annotated data, with technical insights and open research directions. Followed by a cocktail. 75 attendees.

The inaugural Unaite Paper Club, with Hugo Touvron from Meta FAIR's Llama team. The talk traces the evolution from the original Llama paper to Llama 4: how Meta's researchers turned the Transformer architecture into one of the most consequential open-weights LLMs, insights into training strategies, design choices and the discovery process. Followed by a cocktail.