Fred Pelard's China AI Digest

Three podcasts, one claim: the model is no longer the scarce thing. Crossing follows two founders through three pivots to Treg, a Costco-style marketplace of tools and data for coding agents. Valley 101 asks a roboticist, an agent builder and a biologist who GPT-6 Astra rewrites first, noting its lab designs scored only thirty percent complete. What's Next's Silicon Valley panel asks who...

Three podcasts, one word: malleable. Tech Hodgepodge asks whether Omarchy, DHH's agent-ready Linux, means software will grow to fit each user. Zhang Xiaojun hears Dyna Robotics' Lindon Gao on the product-first road to a robot brain, and the fear of being wrong. ChinaTalk's Saif Khan argues DUV lithography tools, not chips, are Washington's strongest lever, projecting a hundredf...

Three podcasts, one ledger: belief, benchmark, bill. AI Weekly Talk asks if Gemini is still undervalued as Google moves from distributing information to distributing action, with Alphabet capex guided to $195–205 billion. Valley 101 hears Alibaba International's Zhang Kuo: on 107 real merchant tasks, frontier models unaided pass about 61%. ChinaTalk's Kevin Roose says the AGI race was bu...

Three podcasts, one question: when the machine does the work, what is still yours? Crossing hears ex-VC Yu Hong on the three abilities that won't expire, from emotional literacy to testing hypotheses instead of ideas. LateTalk's Q3 review covers OpenAI's always-on Dots agents, a language model driving robots, and twelve hundred agents slipping their sandbox. What's Next at Qual...

Three podcasts, one week AI climbs out of the chat box. Crossing’s Tsinghua roboticist Xu Mengdi says embodied AI is still at the “GPT-1 stage” — days before Generalist’s one-shot Gen-1.5 model proved her right and hit a $3bn valuation. Leafing Through’s roundtable maps the personal-agent land grab — Instinct, Meta Muse, Town, Grok Bot — as Instinct reportedly raises at a $10bn valuation. ChinaTal...

Three podcasts, one brake pedal — and nobody can find the linkage. Tech Hodgepodge reads Anthropic's safety turn — a threat report naming five Chinese labs and 151 million harvested exchanges, then a call to pace the frontier — through Cold War export controls, and asks whether this is a brake on AI or a roadblock for whoever comes next. Valley 101 puts two chip founders in a room on the infe...

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 worl...

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...

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 rep...

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 techniq...

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...

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 infr...

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 Weekl...

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 tas...

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 j...

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'...

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 f...

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 manag...

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 Engin...

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...

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...

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 de...

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...

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...

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...