AI moves too fast. Tech media all sounds the same and rarely has its own take.

Twitter is too fragmented. You can scroll for a day and remember almost nothing.

Long-form essays are hard to get through. Official blogs are basically product launches.

The only way a normal person can keep up with AI is to listen to the people building it talk — long interviews with industry leaders, internal-feeling conversations with frontline operators.

Why podcasts?

• Sam Altman is not going to write a 5,000-word blog post telling you OpenAI’s real strategy, but he will sit in front of Dwarkesh’s microphone for three hours.
• Moonshot AI’s Yang Zhilin is not going to do a quick interview for 36Kr, but he will record a four-hour deep dive with Zhang Xiaojun.
• Karpathy is not going to write a tech blog explaining why he left Tesla, but he’ll say it in one line on No Priors.

Written interviews get compressed, edited, rewritten. Podcasts are raw material. What you hear is what they’re actually thinking.

Podcasts also fit the AI era perfectly. Commuting, washing dishes, running, cooking — these hours are already being wasted. Listening to a podcast turns them into free information. Five hours of weekly commute listening to three No Priors episodes will leave your read on Silicon Valley AI investing sharper than 90% of people who only read tech news.

Below are the five Chinese and five English podcasts I think are most worth your time.

Chinese Top 5

1. Zhang Xiaojun’s Business Talks (张小珺商业访谈录)

Host: Zhang Xiaojun, former Caixin journalist. Produced by Language Is the World Studio.

What she does is simple. She gets people like OpenAI’s Yao Shunyu, Anthropic’s Yao Shunyu, Moonshot AI’s Yang Zhilin, Manus’s Xiao Hong, and Luo Fuli in front of a microphone and talks to them for three to seven hours.

Why does this matter? Because no other channel can pull this off. A press release gives you five questions, 200 words each, all PR-polished. Talk for 200 minutes and every layer of polish wears through.

Her journalism background is the key. The three Asian Publishing Awards weren’t an accident. Her follow-up pacing draws real material out of guests. Plenty of Chinese hosts do long AI interviews. Only she does them at the level of investigative journalism.

Episodes to start with: the four-hour interview with Yao Shunyu, the 3.5-hour interview with Luo Fuli.

2. Crossing (十字路口)

Host: Yang Yuancheng (Koji), partner at ZhenFund and founder of AI Hacker House.

Koji’s position at ZhenFund means this podcast pulls first-hand interviews with early-stage AI founders that no one else can get. ZhenFund backed ByteDance, Zhihu, and Xiaohongshu in their earliest rounds — the volume of early-stage founders they’ve seen and the accuracy of their reads are top-tier in Chinese VC. Koji also runs AI Hacker House, so basically every new-generation AI founder passes through his network.

Listening to Crossing feels different from listening to OnBoard!. OnBoard! takes the Silicon Valley investor view and asks “can this company win?” Crossing takes the domestic founder-community view and asks “what are we all doing, where are we stuck?”

The recent Paperboy episode broke down the paradigm shift in AI agents really well. They even gave a Token Grant to an experimental project called YoYo Agent — being both an investor and a live experiment ground is unique in the Chinese podcast scene.

Episodes to start with: the Paperboy founder episode, and the one on Agent Harness.

3. Silicon Valley 101 (硅谷101)

Host: Jane Hongjun, founder and CEO of Silicon Valley 101, former US correspondent for Caijing magazine.

Hongjun’s advantage is that she actually lives in Silicon Valley. Five-plus years in, 213 episodes deep. Her interviews land one to two months ahead of what Chinese business media eventually picks up.

Her guests are not popular science influencers. They are former TPU engineers, AI healthcare founders, NVIDIA insiders — people who can take “why did Google build its own TPU” all the way down to chip design specifics.

The China-US dual perspective is the other rare asset. When DeepSeek-R1 dropped, her episode was the first in Chinese to give a full read of what Silicon Valley engineers were actually saying about it in private group chats.

She doesn’t only cover AI. If what you want is a tech podcast rather than just an AI podcast, Silicon Valley 101 is the better entry point.

Episodes to start with: Episode 228 on whether Google’s TPU can shake NVIDIA, told by a former TPU engineer; Episode 224, a deep teardown of OpenClaw.

4. 42 Chapters (42章经)

Host: Qu Kai (Kyle), founder of 42 Capital and 42 Chapters.

Kyle is one of the earliest VCs in the Chinese AI scene to bet on the application layer of large models. In November 2024 he said “the earlier you dare to believe in AI, the bigger your potential return” — at a time when domestic AI applications hadn’t really gotten off the ground. Later, when Manus blew up, he did a whole series of interviews with the Manus founder.

42 Chapters has a lighter rhythm than Zhang Xiaojun or Crossing. Episodes run 30 to 60 minutes, easy to finish in one commute. Kyle is an early-stage VC himself. The guests are mostly founders actually building products. The conversations focus on what happened in the AI application space this year and what new opportunities have opened up.

Its edge is “judgment plus beginner-friendly.” In 2024 Kyle predicted that the foundation model roadmap was basically settled and future differentiation would happen at the application layer. That call kept getting validated through 2025. At the same time, the show is welcoming to people new to public and private markets — you don’t need to already know a stack of terms to follow along.

Two to four episodes a month, steady release schedule.

Episodes to start with: the Manus AI founder interview series, the November 2024 wrap-up on the AI application space.

5. Slaying the Dragon (屠龙之术)

Host: Zhuang Minghao, ex-big-tech, ex-VC.

This one is the odd one out in the top 5. It’s not an interview show. It’s a monologue. Zhuang Minghao alone takes the last three to six months of AI industry developments and compresses them into 70 to 130 slides, then narrates the deck. Listening to one episode is like reading a solid industry research report.

His calls are genuinely sharp. He predicted the LLM trajectory before GPT-3 hit. He flagged the limitations of AutoGPT before the agent concept became a buzzword. In Chinese AI commentary circles, a Zhihu user once said “this academician’s grasp of LLM development is weaker than Zhuang Minghao’s” — that half-provocative compliment reflects real respect inside the community.

The format has trade-offs. The upside is information density: he covers in half an hour what others spend three hours discussing. Updates run roughly once every two to four weeks.

But for someone who needs to quickly build a panoramic view of the AI industry, nothing else replaces it. Other podcasts give you a guest’s perspective. This one gives you an analyst’s.

Episodes to start with: Vol. 45, which uses 132 slides to cover the entire AI industry in 2025; Vol. 55, 70 slides on Manus and AI agents.

Best for: investors, analysts, and operators making strategic calls who need a panoramic view fast.

English Top 5

1. Dwarkesh Podcast

Host: Dwarkesh Patel, independent podcaster, named to TIME’s 2024 list of 100 most influential people in AI. Format: very long-form deep interviews, 3 to 5 hours per episode.

Why it’s worth listening to

Dwarkesh is the most prepared AI podcaster of this generation in English.

Before every episode he reads everything his guest has published — papers, internal reports, every public interview from the last five years — and walks in with 80 questions so specific the guest has no room to “do the rehearsed take first.” He only asks what you haven’t already said elsewhere.

The result is the ceiling of guest quality and depth in English. Dario Amodei has been on twice in a year (one of them specifically on AI 2027 scenarios). Demis Hassabis recorded right after the Gemini 3 launch. Mira Murati gave Dwarkesh her first long interview after leaving OpenAI to start Thinking Machines Lab. Karpathy and Ilya have both been on.

One Dwarkesh episode beats reading ten secondhand pieces about the same guest.

One caveat: his pace is fast and the episodes are dense. Listening on a commute can be tiring. Save him for quiet listening time.

Episodes to start with: the Carl Shulman episode (nearly six hours, widely considered required listening); the Dario Amodei episode.

2. Latent Space

Hosts: swyx (Shawn Wang, AI developer evangelist) and Alessio Fanelli (CTO in Residence at Decibel Partners). Format: built for AI engineers, covering agent frameworks, inference architecture, open-source model deployment, GPU economics.

Why it’s worth listening to

The deepest AI engineering podcast in English. Dwarkesh asks “will AI change the world?” Latent Space asks “what’s the fastest way to add a cache layer to Anthropic’s API next week?” The second question is closer to the actual day-to-day of an engineer.

Its unique asset is swyx himself. He runs the “AI Engineer” conference and knows the CTOs of every AI startup in San Francisco. So guests are almost entirely frontline engineers from OpenAI, Anthropic, Meta, Databricks. Conversations go straight into “what model are we running, what token price, what latency, what gotchas.”

Every year-end they publish the “State of AI Engineering” industry report, the de facto annual benchmark in the AI engineering field.

Episodes to start with: the annual State of AI Engineering special; any episode on agent infrastructure.

3. The Cognitive Revolution

Host: Nathan Labenz, AI applications operator who personally participated in the GPT-4 internal red team. Format: AI applications, safety, and policy interviews, 2 to 3 hours per episode.

Why it’s worth listening to

Nathan’s unique edge is the GPT-4 red team — OpenAI brought him and a small group in to probe for vulnerabilities before the GPT-4 launch. That experience gave him first-hand knowledge of AI safety and capability frontiers, and lets him ask questions other hosts can’t.

But what really broke the show out was Nathan’s episode on using AI to help make medical decisions for his son’s cancer treatment. That episode pulled “AI applications” out of the engineering frame into “real-life decisions,” and is widely considered the breakthrough episode of the English AI podcast scene.

He’s done several follow-up episodes on the same theme — concrete decisions about AI in medicine, education, and family life.

If you want to understand AI not just as a tool but as a possible life companion, The Cognitive Revolution goes deepest on this in English.

Episodes to start with: Nathan’s series on his son’s cancer; the Bolt.new founder interview.

4. No Priors

Hosts: Sarah Guo (founder of Conviction VC, former Greylock partner) and Elad Gil (serial founder, investor in 40-plus unicorns). Format: VC-perspective AI startup interviews, 30 to 45 minutes per episode.

Why it’s worth listening to

Sarah Guo is one of the most active AI investors in Silicon Valley. Elad Gil has backed more than 40 unicorns. The interviews they do together are basically listening in on top-tier Silicon Valley VCs in a partner meeting — the questions they ask are at the level of “is this company worth a $50 million check?”

The guests match that level: Jensen Huang, Andrej Karpathy, Fei-Fei Li, Alexandr Wang, Bret Taylor. People at that altitude usually only share methodology in interviews, but Sarah and Elad get them down to specifics — the Jensen episode goes into how NVIDIA actually allocates GPU production internally.

The other unique advantage is pacing. 30 to 45 minutes per episode, much shorter than Dwarkesh. One commute fits one episode. If your only listening window is on the road, No Priors has the best density-to-length ratio in English.

Episodes to start with: the Jensen Huang interview; the Andrej Karpathy interview.

5. Machine Learning Street Talk

Hosts: Tim Scarfe (PhD in machine learning) and Keith Duggar (PhD in philosophy from MIT). Format: academic-leaning deep technical interviews. The only technical podcast that connects AI with cognitive science, neuroscience, and philosophy of consciousness.

Why it’s worth listening to

MLST is the most academically dense show of these ten. With one ML PhD and one philosophy PhD, the conversational style feels closer to a top graduate research group than a typical interview show. Guests include Karl Friston (originator of the free energy principle), Stephen Wolfram, Yudkowsky, Francois Chollet — researcher-level figures.

It’s also one of the few English-language podcasts that openly takes a “LLM skeptic” position. Most AI podcasts debate “how far can LLMs go?” MLST often debates “is the LLM path itself wrong?” That stance creates a counterweight to the dominant “AI inevitably leads to AGI” narrative and is valuable for building your own independent judgment.

Not the right show to start with. But after you’ve listened to nine AI podcasts and started feeling that everyone is saying the same thing, MLST will show you another possibility exists.

Episodes to start with: the Karl Friston episode on the free energy principle; the Blaise Agüera y Arcas episode on artificial life.

A few that didn’t make the top 5 but are worth knowing

These are for more specific audiences.

Chinese:

• People’s Park Talks AI (人民公园说 AI) — produced by JustSayAI. Three partners across a 12-hour time difference, a “companion-style” podcast where veterans riff on AI news. The companion product “JustSayAI Morning & Evening Briefing” is a paid subscription. For people who want to stay current on AI news without straining.
• 355 Loop (三五环) — hosted by Liu Fei. The main thread is city insiders chatting; AI is one of the topics. Liu Fei’s ability to get guests to say true things is rare in Chinese podcasts. For internet product managers and people who want a regular-user view of AI.
• Indie Hackers China (硬地骇客) — the king of Chinese indie developer podcasts, 100+ episodes in. Conversations cover indie hacking, remote work, side projects, real stories of building businesses to several million dollars in ARR, and a pragmatic take on AI. For people building indie products and wanting to use AI to run a real business.

English:

• Lex Fridman Podcast — close to 5 million YouTube subscribers. Less than half the content is AI (politics, physics, MMA all share airtime), but the Karpathy, Amodei, and Hassabis episodes hit the depth ceiling of English AI podcasts. Treat it as a one-off, episode-by-episode listen.
• Hard Fork — from the New York Times, hosted by Kevin Roose and Casey Newton. General tech news podcast with a weekly riff on AI industry events. Light, fast-paced, good for people who just want to stay current without working at it.
• TWIML AI Podcast — long-running, 700+ episodes. Content leans toward enterprise ML in production. When you need to look up “how did company X actually deploy ML?”, TWIML is an archive with no real equivalent in English.

How to get into English podcasts

A lot of people get stuck on the English barrier — can’t follow, can’t keep up, can’t find three hours for one episode. The methods below can multiply the return on English podcasts.

1. Use them as English listening practice — two birds, one stone

The guests on English AI podcasts are almost all native English speakers or researchers, engineers, and CEOs with strong English. Pronunciation is clear, vocabulary is standard, technical terms come up repeatedly. Far better than dedicated English-learning material, because you actually care about the content.

Start with No Priors (30 to 45 minutes per episode, fast pace, two hosts going back and forth). Much easier to swallow than diving straight into a six-hour Dwarkesh. Skip the parts you can’t follow. Catching 60% overall is already a win. Three months in, you’ll notice your AI-related English reading speed has come up too.

2. Use AI to compress English transcripts into Chinese summaries

Almost every major English AI podcast publishes a written transcript (Dwarkesh, Latent Space, No Priors all post full text on Substack or their own sites). The flow is:

• Find the transcript (YouTube description, show site, Substack) or extract it with AI
• Paste the full text into Claude or GPT
• Prompt: something like “summarize this AI podcast transcript in Chinese, structured as 5 key insights, each with a guest quote and a one-line value judgment”

Ten minutes to read the core of a three-hour episode. The trade-off is you lose the host’s follow-up rhythm and the guest’s micro-pauses while thinking.

3. Use NotebookLM to turn English audio into a Chinese podcast

Google’s NotebookLM now supports generating a Chinese two-host conversational podcast from any uploaded material (including audio transcripts). The flow is:

• Get the English transcript
• Upload to NotebookLM
• Generate the audio overview, select “Chinese”
• Out comes a 10 to 15 minute Chinese two-person conversation, a translated and lightly reworked version of the source

What you hear is not machine-translated captions. It’s two AI hosts riffing in natural Chinese on the original material, pacing close to the casual VC chat of OnBoard!. Great for commute time when you want to “translate” a long English interview into a Chinese digest.

4. Subscribe to the host’s Substack for the long-form summary

The standard play for English podcasts now is audio + written distribution. Dwarkesh has 80,000 paid Substack subscribers, and the reason is that his Substack is not just transcripts. It includes hand-written episode summaries, edited key quotes, and the host’s own reflection notes.

Just subscribe to the free Substack email lists for Dwarkesh, Latent Space, and The Cognitive Revolution. Even without listening to a single minute of audio, scanning the weekly emails will keep you on top of 70% of the industry’s core information.

5. YouTube chapters and timestamp-jumping

Most English podcasts mirror to YouTube. YouTube’s auto-generated English subtitles are good enough now, and combined with native chapter markers (many big podcasters add their own timestamps), you can:

• Look at the chapter list first, pick the 2-3 segments that interest you
• Jump to that timestamp and listen for 10 minutes
• Skip everything else

A three-hour episode often takes 30 minutes to skim for the essentials. Much better return than listening straight through.

Closing thought

Don’t try to follow everything at once. Pick one Chinese and one English podcast that you find most interesting, build the habit, then expand.

The information quality and density of podcasts is much higher than tech media. Two hours of a top operator talking on a podcast does more for your understanding than reading ten tech articles or scrolling 100 tweets.

Podcasts are a badly underrated information source for the AI era. Hope this list helps.