晚点聊 LateTalk

晚点聊 LateTalk is a long-form Chinese technology conversation show connected to LatePost, focused on the places where AI research, infrastructure, company strategy, and capital pressure meet. It does not treat a model launch or funding shift as self-explanatory. Instead, recent episodes use those moments as entry points into the machinery underneath: attention design, KV cache offload, inference throughput, optimizer tradeoffs, GPU deployment, CPU revival, agent workloads, and the way benchmark performance can diverge from actual developer experience. The DeepSeek episodes show its range clearly. One conversation goes deep on V4’s architecture and engineering choices, while another steps back to examine DeepSeek’s organization, talent movement, AGI ambitions, product gaps, and the burden of public expectation after R1. The show also keeps a close eye on China’s AI startup field. In the crossover with 十字路口Crossing, it compares ByteDance, Alibaba, Kimi, MiniMax, and smaller application teams, asking why some companies chase technical peaks while others build lower-friction products with stronger user attachment. Business episodes widen the lens without leaving the analytical register. A conversation with Charles looks at AI-era overseas marketing, brand development, user research, data competence, and the DMES model for global brands. The Qunhe episode uses Huang Xiaohuang’s founder story to examine survival as a strategy, the limits of trend-chasing, and the transition toward spatial intelligence. Infrastructure episodes connect Musk’s TerraFab ambition, SpaceX, Nvidia’s GTC direction, heterogeneous chips, inference demand, and acquisition paths for small AI-infra companies. The pacing is patient. Some answers sprawl because the subject needs room. The questions usually summarize, narrow, and then push one layer deeper. That makes the show especially useful for listeners who already follow technology news but want the operating logic behind it: why a product matters, what constraint it reveals, and who benefits if the constraint moves.