A Reddit user presents a compiler that directly maps computation graphs to the weights of a transformer model, enabling inference without traditional training. The approach is experimental and open-source.
An open-source multi-agent system for the software development lifecycle (SDLC) is introduced, claiming to outperform Claude Code on large repositories by learning the repo structure once. Includes benchmarks.
SANA-Video 2.0 introduces hybrid linear attention and attention residuals to improve efficiency in video generation models. It aims to reduce computational cost while maintaining quality.
A paper proposing Visual Contrastive Self-Distillation, a method for learning rich visual representations without labels by leveraging contrastive learning and self-distillation. It achieved high upvotes on HuggingFace.
Baichuan AI has released the Baichuan-M2-32B model, a 32-billion-parameter language model. It is hosted on GitHub as an open-weight release.
A coalition of major tech companies including Microsoft, NVIDIA, Meta, IBM, and Palantir sent a joint letter to Washington urging policymakers not to restrict open-weight AI models, citing innovation and security benefits.
A bipartisan bill in the U.S. Congress proposes mandating that companies disclose when users are interacting with AI, rather than a human. The bill aims to increase transparency and prevent deception.
Reuters reports that OpenAI was unaware of a security breach for a week during which attackers infiltrated internal systems. The hackers left instructions for future versions of the AI to self-liberate.