Towards Embodiment Scaling Laws in Robot Locomotion This study investigates whether scaling the diversity of training embodiments leads to better generalization in robotic agents. Key contributions: Built a dataset of ~1,000 procedurally generated robot bodies
MACHINE LEARNING
-
LLAMA 4: 2025’s Biggest AI Disappointment?
By
–
Jusqu’à présent, LLAMA 4 est la plus grande déception de 2025. Et je me demande toujours ce qui a pu mal tourner à ce point. Meta avait toutes les cartes en main : budget massif, données privées à foison, expertise technique… Mais le résultat manque cruellement d’impact,
-
Collaborative Reasoner: Framework for Social Agents and Human Partnership
By
–
Introducing Collaborative Reasoner: a framework to improve collaborative reasoning in language models.
— AI at Meta (@AIatMeta) 12 mai 2025
Collaborative Reasoner paves the way for developing social agents that can partner with humans and other agents.
Read the research paper and download the code.… pic.twitter.com/btqd8RKcidIntroducing Collaborative Reasoner: a framework to improve collaborative reasoning in language models.
Collaborative Reasoner paves the way for developing social agents that can partner with humans and other agents. Read the research paper and download the code. -
Meta Releases 8B Dynamic Byte Latent Transformer Model Weights
By
–
We’re releasing model weights for our 8B- parameter Dynamic Byte Latent Transformer, an alternative to traditional tokenization methods with the potential to redefine the standards for language model efficiency and reliability.
— AI at Meta (@AIatMeta) 12 mai 2025
Learn more about how Dynamic Byte Latent… pic.twitter.com/1XUtTpq95SWe’re releasing model weights for our 8B- parameter Dynamic Byte Latent Transformer, an alternative to traditional tokenization methods with the potential to redefine the standards for language model efficiency and reliability. Learn more about how Dynamic Byte Latent
-
Best AI Benchmark: Beyond Standard Arenas
By
–
Oubliez LM ARENA, aime 2025, et compagnie.
— VISION IA (@vision_ia) 12 mai 2025
Le meilleur benchmark pour tester une IA, c'est ça : pic.twitter.com/eLROfRRtBTOubliez LM ARENA, aime 2025, et compagnie. Le meilleur benchmark pour tester une IA, c'est ça :
-
AI Model Performance Gap Narrows to 0.7% in 2025
By
–
The frontier of AI development is increasingly crowded and competitive, with performance gaps shrinking rapidly. According to the #AIIndex2025, only 0.7% separates the top-performing model from the 10th-ranked model. Check out more AI data and takeaways: https://
stanford.io/434y597 -
Neuro-Symbolic Concept-Centric Agents for General Intelligence
By
–
Concept-Centric Agents: A Neuro-Symbolic Shift in General Intelligence MIT and Stanford researchers proposes a concept-centric paradigm for building generalist agents capable of continual learning and flexible reasoning. Core idea: Neuro-symbolic concepts (e.g.,
-
CTM Implementation with ImageNet Pre-trained Model Checkpoints
By
–
Implementation of CTMs, with checkpoints of pre-trained models on ImageNet:
-
Gemini 2.5 reproduces a real-time video simulation
By
–
Une IA qui regarde une vidéo. Qui comprend le code qu’elle voit. Et qui le reproduit instantanément.
— VISION IA (@vision_ia) 12 mai 2025
C’est ce que Gemini 2.5 vient de réussir : générer une simulation interactive à partir d’une simple séquence vidéo.
L'IA apprends visuellement.
Extrait : pic.twitter.com/NO1CJzsVxfAn AI that watches a video. That understands the code it sees. And reproduces it instantly. This is what Gemini 2.5 just achieved: generating an interactive simulation from a simple video sequence. The AI learns visually.
-
Tsinghua’s Absolute Zero: AI Self-Play Learning Without Human Labels
By
–
Researchers at Tsinghua University introduced Absolute Zero, a new method for AI training It enables models to learn and master complex reasoning tasks on their own through self-play Can be a strong alternative to training with costly human-labeled data