
(FM-GOOGLE) Gemini 2.5: Technical Report
Échec de l'ajout au panier.
Échec de l'ajout à la liste d'envies.
Échec de la suppression de la liste d’envies.
Échec du suivi du balado
Ne plus suivre le balado a échoué
-
Narrateur(s):
-
Auteur(s):
À propos de cet audio
Tune in to explore Google DeepMind's groundbreaking Gemini 2.X model family, featuring the highly capable Gemini 2.5 Pro and the efficient Gemini 2.5 Flash. These models represent a new frontier in AI, offering natively multimodal understanding, the ability to process over one million tokens of long context, and advanced reasoning through "Thinking" capabilities across diverse domains.
Gemini 2.5 Pro stands out for its State-of-the-Art performance in coding and reasoning, alongside remarkable multimodal understanding, capable of analysing up to three hours of video content. This enables exciting applications such as building interactive web applications, comprehensive codebase understanding, and powering next-generation agentic workflows, famously demonstrated by "Gemini Plays Pokémon".
However, the sources also highlight ongoing areas for development. While excelling, the models sometimes struggle with raw pixel vision input and exhibit a tendency for agents to repeat actions with very long contexts exceeding 100k tokens. Challenges like hallucinations and "context poisoning" can also occur. Despite notable increases in some critical capabilities (e.g., cyber uplift), Gemini 2.5 Pro has not reached Critical Capability Levels that would pose a significant risk of severe harm, with Google DeepMind actively accelerating mitigations in these areas.
Paper link: https://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf