Podcast Atlas
entity Updated 2026-08-17 Topics: Culture

LateTalk

179: 蒸馏风暴：一场无人公开谈论的技术竞赛 adds a no-guest industry-explainer episode on Model Distillation / 模型蒸馏. It connects LateTalk’s recent Kimi K3, Tencent Hunyuan, and RSI coverage into one competition frame: model progress now depends on teacher outputs, agent environments, evidence standards, terms-of-service risk, and whether teams preserve independent research capability instead of only chasing fast follower gains.

178: 与田渊栋聊 RSI：模型自进化如何到来？ adds 田渊栋 and Recursive Superintelligence to LateTalk’s AI-research branch. The episode connects Recursive Self-Improvement, AI Research Feedback Compression, Research Taste, Mechanistic Interpretability, Frontier Model Scaling, and AI Organization Design around the question of whether model self-improvement will be smooth scaling, punctuated breakthroughs, or a mix of coding, research taste, and new principles.

LateTalk, or 晚点聊, is the podcast/show context for 170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人, 171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？, and 167: 洋葱学园杨临风：用AI制造捷径，是在杀死真学习. The first source hosts a quarterly embodied-intelligence review with Chen Zhe Peter, using industry events, robot demos, model releases, and commercialization signals to map the state of Physical AI.

The Q2 2026 AI-quarter source adds Henry Yin as a guest and shifts from embodied systems to frontier-model competition, Codex versus Claude Code, Auto Research, Recursive Self-Improvement, enterprise-owned models, open-model substitution, computer use, and real-time voice.

166: 许华哲再次具身创业：不想错过最大的西瓜 adds a founder-interview version of the same embodied-AI thread through Xu Huazhe and Poke Robotics. The episode focuses less on quarterly industry mapping and more on why Xu believes Physical AGI, household robots, AI Native Robotics, and Robot Active Use Metrics should define the next phase.

167: 洋葱学园杨临风：用AI制造捷径，是在杀死真学习 adds an education and AI-learning branch through Yang Lingfeng / 杨凌峰 and Yangcong Xueyuan / 洋葱学园. The episode connects Self-Directed Learning, Learning Experience Design, AI As Tutor, and AI Shortcut Risk around the question of whether AI helps students think or merely gives them faster answers.

汉洋：为什么做《蜉蝣天地》 mentions LateTalk as one of 汉洋 / Han Yang’s earlier content practices. In that source, the show functions less as a technical subject and more as part of the backstory for why 蜉蝣天地 / Fuyou Tiandi tries to loosen Media Form Constraint through Long-Form Conversation.

Momenta IPO后再访曹旭东：就是想做没有尽头的AI adds an IPO-era founder interview with Cao Xudong of Momenta. It connects LateTalk’s embodied-AI thread to Autonomous Driving Data Flywheel, World Models, Physical AI, and Low-Cost Short-Cycle Validation, using autonomous-driving commercialization as the bridge toward robotics.

173: 对话姚颂：深鉴、东方空间、再出发，「天才少年」十年后 adds Yao Song / 姚颂 as a continuous hard-tech founder case. The episode ties LateTalk’s AI and embodied-intelligence coverage back to DeePhi Tech / 深鉴科技, Xilinx / 赛灵思, Orienspace / 东方空间, Gravity-1 Rocket / 引力一号, and Striding AI / 正行创新, making the show a place where Chinese AI-chip, commercial-space, and physical-intelligence routes can be compared through one founder’s operating history.

174: AI冲击企业软件巨头？与SAP原欣聊大模型to B的颠覆与边界 adds Yuan Xin / 原欣 of SAP and moves LateTalk’s AI thread from frontier-model and embodied-intelligence reviews into enterprise software. The episode connects Enterprise Resource Planning, ERP Trust Moat, Forward Deployed Engineer, Autonomous Enterprise, China Enterprise AI System Debt, and SAP’s Alibaba Cloud / Qwen / DingTalk China ecosystem route.

175: 对话Liblib陈冕：关于活下来，以及所有接近死亡的时刻 adds Chen Mian / 陈冕 of Evoken / 言语科技 as an AI application-founder case. The episode connects Liblib, Lavod, and Lib TV to AI Application Survival Strategy, AI Startup Unit Economics, AI Subscription Economics, Model Provider Tool Competition, and AI Organization Design, making LateTalk’s AI coverage more directly about application-layer survival under model pressure.

176: 姚顺宇，来到腾讯300天 adds Gao Honghao / 高宏昊’s reporting-based discussion of Yao Shunyu / 姚顺宇’s first 300 days at Tencent Hunyuan / 腾讯混元. It extends LateTalk’s AI coverage from application survival and quarterly model reviews into large-company model-team reform, Federated AI Organization, WeChat VLM / 微信 VLM, and Tencent’s tension between product-unit autonomy and centralized AI infrastructure.

177: 详解Kimi K3：强到冲击Anthropic估值的模型什么样？ adds a technical-report deep dive on Kimi K3 with 赵晨阳 and 曾志远. The episode pushes LateTalk’s AI branch from organization and product strategy into KDA, Attention Residues, Quantile Balancing, Per-Head Muon, Kernel Development Agents, AgentIn, and the boundary between open weights and a repeatable model-development pipeline.

Connections