Results
5 articlesSort
23:39ZH → English
跟李沐学AI· Jul 23, 2025Li Mu: Adding Ten Million Hours of Speech to a Text LLM, Open-Sourcing Higgs Audio李沐 · Boson AILi Mu explains how the team added a huge amount of speech data to text LLM training, built an AudioLLM that can listen and speak, and open-sourced Higgs Audio. He covers architecture trade-offs, token compression, where the data came from, and how to run it on a 3090.
2:29:41ZH → English
Zhang Xiaojun Podcast· Apr 7, 2025Podcast · 4 speakersChen Jianyu: How Robot Foundation Models Can Become General-Purpose VLA SystemsGuest 陈建宇 · Host 张小珺From SayCan and RT-2 to Diffusion Policy, Tsinghua assistant professor and Epoch Robotics founder Chen Jianyu traces the path from borrowing language models to end-to-end VLA. The hard problems remain diverse data, cross-embodiment generalization, and high-frequency action control.
2:49:35ZH → EnglishZhang Xiaojun Podcast· Feb 4, 2025Podcast · 2 speakersPan Jiayi on R1, o1, and Kimi K1.5: How Reinforcement Learning Teaches Models to ReasonGuest 潘嘉怡 · Host 张小珺Berkeley PhD candidate Pan Jiayi walks through reasoning-model reports from DeepSeek, OpenAI, and Kimi, unpacking long chains of thought, GRPO, reward design, and distillation. R1-Zero’s “aha moment” shows how reasoning strategies can emerge through reinforcement learning.
1:33:07ZH → EnglishZhang Xiaojun Podcast· Aug 17, 2024Podcast · 2 speakersGuangmi on a Year of Global Large Models: Scaling Law, the AGI Bet, and China–U.S. CompetitionGuest 李广密 · Host 张小珺Li Guangmi, founder of Shixiang, looks back from Silicon Valley at the competition among OpenAI, Anthropic, Google, and Chinese model companies. He sees model capability and cost as the forces behind the next application boom, while the enormous investment is a bet on whether intelligence can keep scaling and reshape the economy over time.
49:41ZH → EnglishZhang Xiaojun Podcast· Aug 17, 2024Podcast · 3 speakersFu Sheng: Large Models Are a Productivity Revolution, and Robotics Has a New Window of OpportunityGuest 傅盛 · Host 张小珺Fu Sheng sees logic and reasoning as the essence of large models, and their impact as closer to the steam engine than to another routine internet upgrade. He reflects on accessible models, service robots, and how founders navigate downturns, revisiting both his past judgments and his state of mind.