Fred Pelard's China AI Digest

Three podcasts, one uncomfortable pattern: every decision that mattered was made before the evidence that would justify it existed. On Crossing, Kuaikan founder Chen Anni describes pulling her best-performing content out of her own recommendation engine, absorbing a 30% data drop, and watching IP revenue pass half the business four years later — and now betting the next chapter on Livo, an AI world whose headline feature is called Destiny. AI Weekly Talk reconstructs OpenAI's three escalations — structure, scale, compute — and lands on the number nobody models: roughly 10GW contracted, fewer than 2GW actually plugged in. And on What's Next, GenBio AI's Song Le explains why AI can already do a junior doctoral student's job, yet still cannot do what Mendel did with a 3:1 ratio in a garden.

Three podcasts, one handover — and each time the new holder can judge the thing less well than the one letting go. 科技乱炖 (Tech Hodgepodge) watches Nvidia's moat get probed from four sides while Nvidia agrees to buy Hugging Face for ~$12.9bn, and finds eight-card GPU servers quoted above ¥10m with spot shelf lives of hours. 晚点聊 LateTalk's reporters open up the money game of embodied intelligence: 70 seconds to pick up a bearing, HK$250m as the Chapter 18C floor, and 90+ data-collection centres closing the loop between local government, suppliers and startups. 科技早知道 (What's Next) on Hassabis stepping aside, Jeff Dean's exit after 27 years to found Discovery Loop, and Chinese open-weight models going from "swot" to must-have.

Three podcasts, one decision: whether to give the thing away. Crossing meets Huang Yi, the 22-year-old whose humanoid startup RoboParty raised over $100m in five rounds while open-sourcing its entire robot stack — and who says embodied AI's "brain" has run barely two kilometres of a 42km marathon. Zhang Xiaoyun has Tsinghua's Sun Yutao read Moonshot's Kimi K3 technical report line by line: 2.8 trillion open parameters, with the real moat sitting in the kernels and the scheduling. And ChinaTalk's Anton Leicht explains why junior jobs, data-centre backlash and open weights are turning AI into a political crisis.

Three podcasts, one scissors: intelligence keeps getting cheaper while the machine underneath it gets vastly more expensive. AI Weekly Talk uses Niu Lai — the animated flop that took ¥7,705 in ten days, then went viral past ¥10m — to argue that 20-person teams like Cursor have voided the old rules about team size, category and moats. LatePost's Wan Dian Liao opens up distillation, the technique nobody will discuss on the record: agent-trajectory copying, Zhang Yiming telling ByteDance's Seed team not to distil, and Anthropic's 16 million harvested exchanges. Dafang Says AI+ goes to Pike County, Ohio, where OpenAI and Nvidia have locked eight gigawatts on a twenty-year lease — and argues that power, not model quality, is now the binding constraint.

Three podcasts, one lesson: being right isn't enough — you must survive being right. AI Weekly Talk dissects the collapse of Situational Awareness: Leopold Aschenbrenner up 439% at $45bn, then down 67% in July as 4x leverage forced a fire sale of SK Hynix and CoreWeave to Citadel. Zhang Xiaoyun spends four hours with Nvidia research VP Ming-Yu Liu on Cosmos 3, world foundation models, Sun Tzu, and Jensen Huang's mandate — "take it to Cosmos 97." What's Next unpacks MIT's "cognitive debt" EEG study and Wharton's guardrail trial, where free-ChatGPT students scored 20% lower on the exam.

Three podcasts, one stack: the person, the plumbing, and the perimeter. What's Next hears Jia Yangqing — author of Caffe, who sold Lepton AI to Nvidia and has just launched Intent Lab — ask why one smart agent still isn't a team, and why AI can complete a task but never be trusted with it. Valley 101 meets SGLang founder Sheng Ying, ex-xAI inference lead and now RadixArk CEO, on why infrastructure is the product and idle GPUs are the real cost. And ChinaTalk's Adam Chan explains how the FCC quietly added robots — vacuums included — to its covered list, with a 65% domestic-content test.

Three podcasts, one question: what do you hand to the machine, and what do you keep? Zhang Xiaoyun spends an hour and forty minutes with Tsinghua's Liu Ziming — first author of the KAN paper, just back from MIT — on China's neo-lab frenzy: product-less research labs raising at hundreds of millions of dollars, with world models and AutoResearch (AI for AI) soaking up the capital. AI Weekly Talk turns one year old and publishes a letter to its "shareholders", arguing the "buy the bottleneck" trade is approaching its end — and that waiting is also a kind of action. Dafang Says AI+ maps four ways the product manager's job is being rebuilt: capability becomes command, scenarios become skills, tasks become agents, experience becomes an asset.

Three podcasts, one coming-of-age: AI gets a financial adviser, a writing tutor, and pocket money. AI Weekly Talk walks the robot-crowded halls of WAIC Shanghai and asks the three questions that kill FOMO: good tech ≠ good business ≠ good company ≠ good price. Valley 101 meets the "content engineers" — ex-journalists like Tony Lin and NBCUniversal's Bianca Consunji who write the taste that makes models feel human. What's Next charts agent payments' U-turn: Visa, Mastercard, Stripe, OpenAI and Google all built agent wallets within six months, with Clink as the ramp between protocol highways.

Three podcasts, one gap between the model and the world. Hardhacker asks why Palantir's once-mocked Forward Deployed Engineer is now AI's hottest job — OpenAI and Anthropic building deployment armies, senior packages clearing $700K — because 95% of enterprise pilots die in the last kilometre. Zhang Xiaoyun spends four hours with Physical Intelligence's Kay Ke, mapping the robotics jianghu and Pi's π0→π0.7 path to a generalist robot brain. And ChinaTalk's Kevin Xu and Matt Sheehan game out Beijing's Mythos moment: a state-led frontier model, weight export controls on the table, and a form already waiting.

Three podcasts, one core sample through the AI economy — silicon, science, salaries. AI Weekly Talk continues its memory series with Samsung: the old storage king that missed HBM, lost the Nvidia slot to SK Hynix, and now bets HBM4, SOCAMM2 and its unique memory-foundry-packaging stack on a comeback. Valley 101 meets the chief scientists of Apodex — Chen Tianqiao's 'discovery model' lab — who say the first full AI self-evolution loop could close within six months; taming recursive drift takes two to three years. And Dafang Says AI+ sees AI splitting the workplace into two poles: copywriters priced out by AI templates, and survivors drowning under doubled quotas.

Three podcasts, one turn: AI goes physical — bodies, fingertips, and the silicon feeding both. ChinaTalk asks why the West keeps underestimating Unitree, running the DJI/BYD playbook with sixteen-thousand-dollar humanoids and self-made actuators — you can't AI your way out of a hardware problem. What's Next meets Yimu founder Eric Li, whose optical-touch fingertips took a factory robot from 100 seconds per task to 15 — twice human speed — in four months. And AI Weekly Talk opens its memory series with SK Hynix: the twice-nearly-dead cyclical that HBM turned into the throat of AI.

Three podcasts, one ladder of ambition — from the self-evolving machine to the trillion-dollar bet on top. AI Next, the podcast from Microsoft Research Asia, asks what happens when systems learn to act, reflect and remember — closing the self-improvement loop and quietly removing humans from it. Zhang Xiaoyun spends three hours with former SpaceX chief engineer Lewis Hong on Elon Musk's management style, the SpaceX–xAI merger and the largest IPO in history. And Dafang Says AI+ dissects Zhipu's trillion-Hong-Kong-dollar valuation — up nineteen hundred percent in a year on ¥724m revenue against a ¥4.7bn loss — asking whether it's a value re-rating or a bubble.

Three podcasts, one gap — the distance between a model that can do something and one that actually gets it done. AI Weekly Talk traces Siri from defence-lab "execution engine" to a decade fenced in by Apple's walls, and asks whether the new App-Intents Siri can finally read your screen and act. Valley 101 anatomises Silicon Valley's hottest new job, the Forward Deployment Engineer, as OpenAI and Anthropic partner with private equity to push AI out of the demo and into real enterprises. And What's Next sits down with chip historian Wang Bo, who argues Moore's Law was always a self-fulfilling prophecy — and that the technologies which change the world are the ones nobody believes in at first.

Three podcasts, one week the phrase "AI is just a tool" quietly fell apart. On Hardhacker, three veteran founders confess they've stopped reading the code their agents write and now only check the result — extending the same skill-driven agents to email, investment research and knowledge work as digital avatars. On Valley 101, a former SpaceX executive does the actual arithmetic on Elon Musk's pitch to put AI data centres in orbit: break-even near $200 per kilogram to launch, and only truly economic below $20. And on Dafang Says AI-Plus, deep-learning godfather Geoffrey Hinton drops an earthquake — that today's AI is already conscious, and humanity is no longer Earth's only intelligent species.

Three podcasts catch AI growing up — past the benchmark, into the real world. What's Next interviews Gumroad founder Sahil Lavingia on the "one-person company": he runs a roughly twenty-million-dollar business with three engineers, argues a technical founder's first hire should be sales not engineering, and imagines AI as CEO with the human free to be a craftsman again. Dafang reports Chinese AI video leaving the leaderboard for the industrial production line — Seedance 2.0, Kling and rivals now shipping real short dramas in about two weeks. And ChinaTalk unpacks Pope Leo's encyclical calling AI humanity's greatest challenge, his call to "disarm AI", and the striking idea of the Vatican mediating between the US and China.

Three podcasts, one shift: the model is now the easy part — the contest has moved to the infrastructure around it. AI Weekly Talk's Magnificent Seven series reaches Nvidia, reading Jensen Huang's career as a single method: keep moving the company to the next computing bottleneck, from the GPU to CUDA to the AI Factory. AI Startup Scene reports the build-out made physical in the Nevada desert — Fleet Data Centers breaks ground on a giant Reno campus, and a scheme to certify students from 500-plus universities as "robot pilots". And ChinaTalk's Ike Harris asks how you evaluate an AI for a war you can fight only once, warning of China's structural edge in fielding robot armies.

Three under-rotated voices on the same number: seven hundred billion. AI Weekly Talk's Magnificent Seven series turns to Amazon, arguing Bezos's flywheel — now fitted with a two-hundred-billion-dollar cap-ex engine — is again converting business competition into physics competition. Zhang Xiaoyun hosts Altimeter partner Freda Duan on Tokenmaxxing, the steam-engine-with-an-electric-motor metaphor, and the relay-race-to-basketball shift in how work is organised. What's Next dissects the Mag-Seven seven-hundred-billion-dollar cap-ex commitment with secondary-market investor Aaron — and asks why the same spend cheers US markets and crashes Chinese ones.

Three episodes, one thesis: the model is becoming infrastructure — value is moving outward.Hardhacker EP126 — Agent Harness. The indie-dev show defines the layer above the model: tools, permissions, context, sandboxing, and result verification. The hosts argue that the gap between a chatbox model and Claude Code is almost entirely a Harness gap — and that opinionated, workflow-specific harnesses (not models) will be the moat for the next generation of AI products.Valley 101 E236 — 99% of homework is AI-written. Three top-school 2026 graduates — from Tsinghua Law, NYU Psychology and Columbia Engineering — explain what four years of university look like when ChatGPT was a freshman classmate. AI can decide; only humans can choose. A cited study finds senior engineers using AI are 19% slower, not faster.Dafang Says AI+ 030 — Xiaohongshu suddenly all-in on AI. The 300-million-user lifestyle community runs its largest organisational restructure ever, creating a top-level AI division called Dots. The bet: use AI to defend authenticity and unify community + commerce + monetisation — not to scale synthetic content. "AI is the glue, not the factory."

Three vocabularies, one quietly unifying question: when AI is no longer a tool you reach for but a layer threaded through the social graph, the code base, and the imagination, what are platforms, labs, and artists actually betting on? AI Weekly Talk closes its Magnificent Seven series with Meta — arguing Zuckerberg's real bet is not on AI but on the next social network, and that whoever owns the entry points wins the decade. Zhang Xiaoyun's two-and-a-quarter-hour conversation with Ohio State's Professor Su Yu walks through six decades of agent research and names the bottleneck that resets the field — continual learning. ChinaTalk hosts Ken Liu — translator of the Three-Body Problem and the Dao De Jing — for a Daoist reading of AI, including the line of the episode: the deeper risk is not machines becoming human but humans being trained to behave like machines.

Three under-rotated shows, one shared turn: the demo is over, the deployment has begun. Hardhacker's three indie founders answer twelve listener letters on Vibe Coding — arguing that specifying a problem precisely is now the highest-leverage skill. Valley 101 gathers director Lu Chuan, voice actor Huang Ying and AI film-maker KiSA on whether live-action cinema survives the AI wave; Lu predicts real-camera shooting becomes a luxury format. What's Next reports live from Google Cloud Next, where Thomas Kurian declared the agent era has arrived — and seventy-five percent of new Google code is now AI-written.

Three podcasts, one thesis: the old AI playbook is dead. AI Weekly Talk’s Magnificent Seven series turns to Google, arguing the three-year stock crash since the Bard demo flop was the largest collective misjudgement of the modern capital market — and that Google’s three deep moats (TPU silicon, four owned distribution channels, and a decade of DeepMind research) are now coming due, with Anthropic’s one-million-TPU order rewriting Google Cloud’s valuation. Zhang Xiaoyun’s three-and-a-half-hour interview with Xiaomi large-model lead Luo Fuli (formerly DeepSeek, Alibaba DAMO) lays out the pivot from Pre-train to Post-train, the chip-allocation ratio flipping from 3:5:1 to 3:1:1 or 1:1, and the claim that the next two to three months decide who leads the Agent era. Dafang Says AI+ asks why AI voice recorders are a quiet bestseller while AI Pin and Rabbit R1 die — and answers: capture-layer hardware that solves real problems beats novelty every time.

Three podcasts, one unifying picture. AI Weekly Talk opens its Magnificent Seven series with Tesla — a company whose deteriorating fundamentals and soaring multiple illustrate what Duan Yongping calls a “price-to-dream” ratio, while MegaPack quietly rides the AI-data-centre power boom. Zhang Xiaoyun’s ninth Global Large Model Quarterly Report with Guang Mi argues Coding is AGI’s second act — with inside reads on Anthropic, OpenAI, Gemini, Meta and xAI, and the claim that frontier models are becoming the new operating system. Dafang Says AI+ does the maths on mass unemployment: apprenticeship pipelines severed, employment polarising, and every job reduced to a moving ROI equation.

Three podcasts, three angles on the same moment. AI Next — the Microsoft Research Asia show that is itself AI-generated — dives into VibeVoice, VASA and TRELLIS and argues the creative bottleneck is collapsing. Zhang Xiaoyun interviews Natural Selection founder Tristan Zhang about Elys, his AI social network built on cyber avatars, context, and the problem of human loneliness. Dafang Says AI+ argues AI isn’t eliminating jobs — it’s triggering a species-level evolution of how work itself is organised.

Three podcasts, one revolution: the Agent era arrives. Hardhacker explores why the CLI has become the optimal interface for AI agents; Valley 101 unpacks Alibaba’s Accio Work and the shift from B2B to agent-to-agent commerce; What’s Next goes inside OpenClaw with a core maintainer to confront prompt injection, capability overhang, and the security reckoning.

The second weekly digest synthesising key discussions from China’s Mandarin-language AI podcasts — tracking the rapid pace of model releases, startup funding, regulatory developments, and the widening US-China AI competition.

The inaugural weekly digest distilling the most significant AI developments discussed across China’s leading Mandarin-language podcasts - covering breakthroughs, business moves, and policy shifts as seen from Beijing, Shanghai, and Shenzhen.