Podcast Atlas
entity Updated 2026-08-24

DeepSeek

E249｜Token经济转点：OpenClaw、Hermes到本地自研的Agent进化之路 adds a local-agent execution case. 东旭 / Dongxu says he runs a source-named DeepSeek V4 Flash locally on a Mac Studio at roughly 30 tokens per second and uses it for repeated summaries, article-memory work, and batch paper processing. The episode treats this as a shift in AI Inference Cost Structure: local or open models reduce both marginal cost and the psychological hesitation around long agent tasks.

Inside America’s AI Strategy: Infrastructure, Regulation, and Global Competition adds DeepSeek as a geopolitical update trigger. David Sacks says the DeepSeek release made Western observers take Chinese AI capabilities more seriously, placing the company inside American AI Stack Strategy, Chinese Open-Weight AI Strategy, and U.S.-China AI Macro Asymmetry / 中美AI宏观不对称 rather than only cost, model-distillation, or open-model debates.

179: 蒸馏风暴：一场无人公开谈论的技术竞赛 adds DeepSeek as both a distillation teacher and accusation target. The source treats the January 2025 R1 release and its six small distilled models, based on Qwen 2.5 and Meta Llama 3 foundations, as a key reason industry attention shifted from compression distillation toward capability distillation; it also says OpenAI and Anthropic raised suspicions about DeepSeek while not presenting complete public proof.

巴黎水和圣培露还能赚钱，雀巢为何要剥离水业务？ adds a price-increase signal. The source says DeepSeek announced on August 6 that it planned to raise API service prices substantially, while already using peak/off-peak pricing. In the wiki, this extends AI Commercialization Pressure and AI Inference Cost Structure: a model provider can win adoption through low cost, but production serving still has to fund compute, demand peaks, and future model releases.

Vans、匡威风光不再，经典帆布鞋为什么卖不动了？ adds DeepSeek Harness as an agent-infrastructure release. The source says DeepSeek Harness developer preview launched on 2026-08-13 as a plugin-oriented scheduling system connecting models and agent task execution.

从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青 adds 孟繁青’s domestic-model competition view. The source uses DeepSeek’s Multi Latent Attention as an example of constrained-resource architecture innovation, and argues that Model Distillation / 模型蒸馏 may accelerate Chinese model catch-up but is not the decisive cause of DeepSeek-style progress.

贾扬清：我所经历的「人工智能已死」到「AI 颠覆世界」的数年巨变丨串台「声东击西」S10E24 adds an outside infrastructure-builder’s respect for DeepSeek. Jia Yangqing treats DeepSeek as impressive partly because High-Flyer / 幻方量化 invested in serious GPU infrastructure before the large-model wave looked commercially obvious.

177: 详解Kimi K3：强到冲击Anthropic估值的模型什么样？ adds DeepSeek as a technical comparison point for Kimi K3, especially around MoE routing and post-training. The source contrasts K3’s Quantile Balancing with DeepSeek V3-style bias updates, and it mentions DeepSeek-style post-training recipes in the broader MOPD and Model Distillation / 模型蒸馏 discussion.

E246｜何谓蒸馏？聊聊硅谷如何看中国开放模型逼近前沿 adds DeepSeek as the comparison point for the Model Distillation / 模型蒸馏 debate around Kimi K3. The source says DeepSeek R1-style openness makes distillation more legitimate when outputs, data artifacts, or logits are explicitly reusable, and it uses DeepSeek to show why open-model progress should be evaluated through technique, data, RL, infrastructure, and Scaling Efficiency rather than one copying accusation.

176: 姚顺宇，来到腾讯300天 adds DeepSeek as an organizational shock to Tencent. The source says DeepSeek’s 2025 breakout made Tencent executives more willing to question whether older search, advertising, and recommendation teams could carry frontier-model competition, creating the opening for Yao Shunyu / 姚顺宇 to rebuild Tencent Hunyuan / 腾讯混元 with younger research talent and stronger executive sponsorship.

174. 我们还能给算法当多久的品味老师？｜对谈亚马逊AGI查晟 adds 查晟 / Cha Sheng’s model-team view of DeepSeek. He treats DeepSeek’s open releases as surprising and strategically important because they share architectural and algorithmic advances while also giving competitors and downstream builders a faster path to learn from the work. The source uses DeepSeek to sharpen the tradeoff between Open Source AI Models and the closed-product AI Data Flywheel / AI数据飞轮.

148. 对游凯超3小时访谈：开源Infra、和模型Co-design 、“如果vLLM失败，我们会后悔一辈子” adds 游凯超’s inference-infrastructure view. He treats DeepSeek as a major 2025 open-model catalyst for the vLLM community and says its infra team shows unusually strong Model-Infra Co-Design capacity around MoE, inference optimization, and model-serving efficiency.

172.全球宏观和资本市场2026半年度复盘与展望：AI叙事的下一步 adds DeepSeek as the model side of Ricky’s China domestic-substitution map. In the source, DeepSeek is not primarily a technical case; it is a future market signal that could connect Chinese AI application demand, autonomous data-security logic, and domestic AI infrastructure to China Equity Structural Selection / 中国权益结构分化.

160.如何应对中国资产牛市的“调整期”｜新书分享会成都场实录 adds DeepSeek as the second China-asset catalyst after the 2024-09-24 policy turn. 大卫翁 treats the 2025 DeepSeek moment less as a narrow model-company topic than as evidence that China remained in the same technology-competition lane, helping explain why Chinese technology and broader China assets could be repriced despite low global trust.

152.关于2026年的四个猜想 adds DeepSeek as the analogy for a possible foreign-investor China trigger. The source argues that Western investors may need a “DeepSeek moment” for China assets: a concrete event that forces them to update stale narratives captured by Western China Misreading / 西方对中国的误读 and China Narrative Split.

DeepSeek appears in 阿里千问离职余震，在几万人的铁球里如何体面生存 as a peer reference point for Qwen. The hosts describe Qwen and DeepSeek as especially important Chinese Open Source AI Models, using DeepSeek to situate Qwen’s reputation and ecosystem contribution.

In 从QQ会员到豆包包月，中国人为什么总觉得软件该免费, DeepSeek appears as a domestic alternative that may shape Doubao’s pricing room. The hosts treat it as part of the substitute set users and products can turn to if Doubao’s paid tier does not feel worth the price.

71. 编程的内燃机时代 adds DeepSeek as a timing and international-impact marker. The hosts place the recording just after DeepSeek’s open-source week and mention Aleph Alpha’s view that DeepSeek may not be a purely groundbreaking innovation but still has substantial ecosystem impact.

EP57 美股动荡，东升西降？这回是走是留 adds DeepSeek as a market-repricing catalyst. 老麦 argues that the important effect is not a simple one-company negative for Nvidia, but a broader investor question: whether other AI companies’ heavy spending will produce enough returns and whether Chinese technology assets deserve higher valuation after the model shock.

EP58 业绩平平，也要认真"摸鱼" adds a practical workplace-use case. Magic / 杰克 uses DeepSeek to review and improve a student’s composition, while the hosts use AI transcription, editing, title drafting, meeting notes, and visual-summary tools to argue that AI can create more room for Workplace Pacing only when people still apply Human Judgment Under AI.

把 AI 吹成核武器的人，亲手拉下了新冷战铁幕 adds DeepSeek to the model-substitution side of AI Export Controls. The hosts argue that if closed frontier models become unreliable because of Frontier Model Access Restrictions, many users and companies may choose cheaper, good-enough, local, or open alternatives, with DeepSeek and GLM 5.2 as examples in that ecosystem.

Vol. 167 Token 如流水，Agent 似朝阳 adds DeepSeek as a cost-control alternative in heavy AI use. The hosts mention DeepSeek, Kimi, and local models as cheaper or more controllable options when unconstrained Codex, Claude Code, or API use becomes too expensive.

当华为抛出韬定律，我们该信它到哪一步？ uses DeepSeek as an analogy for engineering optimization under cost and resource pressure. The hosts compare DeepSeek’s pricing and cost-structure story with Huawei’s Tau Law narrative to argue that competition is not always won only by the most raw compute or the most advanced process node; method, caching, utilization, architecture, and system integration can change the economic field.

别在国内卷了，去美国看看只要产品好就有人付费的市场 adds DeepSeek as a cross-border adoption signal. Win says U.S. founders were using Chinese models such as DeepSeek and Kimi when the cost and performance fit, suggesting that practical utility can weaken nationality-first assumptions inside technical communities.

138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权 adds DeepSeek as part of Luo Fuli / 罗福莉’s background and as a marker in the post-training shift through the source’s discussion of R1-era reasoning. The episode uses that history to frame Agent Post-Training and agent-framework competition rather than to analyze DeepSeek as a company.

Vol.111 关于2025年的四个猜想 adds DeepSeek as an AI example inside China Divergent Technology Route. The source mentions it as a possible signal that Chinese AI model development may also move along a different route, while keeping that claim speculative rather than a completed industry verdict.

Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授 turns that speculative signal into the episode’s immediate 2025 catalyst. 张奇 treats DeepSeek as important for engineering efficiency, MoE-style cost structure, and public AI enthusiasm, but argues that it still sits inside the LLM Statistical Boundary and does not remove the Model Post-Training Bottleneck that makes frontier behavior expensive to reproduce.

vol.124.信息过载后如何保持冷静？ | 投资账复盘 adds DeepSeek as a market-narrative catalyst rather than a technical subject. 大卫翁 says first-quarter 2025 gains in Tencent, Alibaba, and other China internet assets were helped by DeepSeek and AI repricing, even where the prior investment thesis had emphasized shareholder returns and buybacks.

133.全球宏观和资本市场2025年中盘点：中国的三个温差和美国的三个预期差 adds DeepSeek to the U.S.-market side of the story. The source treats DeepSeek as one reason investors questioned U.S. AI commercialization, compute demand, and mega-cap technology valuation, making it part of U.S. 2025 Expectation Gaps as well as Hong Kong Tech Repricing.

AI 发展了 4 年，把应用发展没了？｜AI 年中复盘 adds DeepSeek as both a 2025 heat-renewal event and a chat-stage marker. 曲凯 / Qu Kai says DeepSeek and Manus helped renew model and application enthusiasm after worries about scaling limits, while 唐杰 / Tang Jie’s cited letter treats DeepSeek R1 as a transition point from chat exploration toward coding and reasoning.

Vol. 171 假如我们有无限 Token adds DeepSeek as another cost-aware model option in the hosts’ heavy AI workflow. The episode’s abundant-token frame does not erase cost or routing constraints; DeepSeek remains part of the practical alternative set when tasks can trade frontier subscription polish for cheaper, local, or more controllable model capacity.

Vol. 172 Codex 卖重置套餐，DeepSeek 峰谷调价，苹果重回 5 万亿等 adds the practitioner reaction to Peak-Valley AI Inference Pricing. The hosts still value DeepSeek for cheap Chinese text processing, translation, and summaries, but the episode says peak/off-peak pricing and a less exciting V4 Pro release make users recalculate cost, quality, timing, and substitution through Model Routing Cost Control rather than assuming DeepSeek is always the default cheap choice.

Connections