速递|两名本科生3个月打造的AI语音模型,挑战谷歌NotebookLM,16亿参数实现自然对话生成

Two undergraduate students created an AI model that generates podcast-style audio similar to Google’s NotebookLM. Nari Labs’ Dia model, with 16 billion parameters, can generate dialogues from scripts and add prosody, non-verbal cues like coughs and laughs. While the tool runs well and has a simple voice cloning feature, it lacks protection against misuse of generated content.

Adam获时间检验奖!清华揭示保辛动力学本质,提出全新RAD优化器

清华大学团队提出RAD优化器,该优化器通过神经网络与共形哈密顿系统的对偶性揭示了Adam的优化动力学机理,并提出了新的Relativistic Adaptive Gradient Descent (RAD)优化算法,实验表明其在多种强化学习任务中表现优于Adam。