Latent Space Podcast

What is Latent Space Podcast?

Latent Space Podcast is a long-form AI engineering show hosted by Shawn Wang, better known as swyx, and Alessio Fanelli. It sits at the center of a broader Latent Space media brand that includes a newsletter, YouTube channel, live events, and an active community. We found it repeatedly described as a place for people who are actually building with AI, not just watching the industry from afar. The show focuses on foundation models, agents, code generation, multimodal systems, GPU infrastructure, evals, and the messy product decisions that happen when AI leaves the demo stage and lands in real software.

The story behind it matters. Swyx built a reputation through his writing on the "Rise of the AI Engineer" and through smol.ai, which helped curate fast-moving AI developments for developers and founders. Latent Space grew out of that same instinct, to create a home for practitioners who needed more than headlines. Alessio Fanelli brought a founder and operator perspective, which gives the show a different texture from research-only podcasts. Episodes often feel like informed shop talk between people who have shipped things, broken things, and changed their minds in public.

By 2025, Latent Space said it had reached more than 10 million readers and listeners across formats. The newsletter has grown to roughly 178,000 subscribers, the YouTube channel has passed 170,000 subscribers, and the show has cited 1.5 million-plus listener minutes and live events drawing tens of thousands. Those numbers help explain why founders, model labs, infrastructure companies, and product teams keep showing up. If you want to understand how the AI agent ecosystem is being discussed by the people building it, Latent Space is one of the clearest windows into that world.

Key Features

• Long-form founder and engineer interviews: Episodes regularly run well past an hour, often closer to 2 to 3 hours. That length matters because guests can move beyond launch-day talking points and explain tradeoffs, failed attempts, and why a product or architecture ended up the way it did.

• Direct access to major AI builders: Guests have included people from OpenAI, Anthropic, Notion, Box, Mistral, Artificial Analysis, and infrastructure startups. For listeners, this means hearing product and technical decisions from the teams making them, instead of getting everything filtered through secondhand commentary.

• Strong focus on AI agents: Latent Space has spent real time defining what an agent is, including its TRIM framing, tools, runtime, instructions, models. That matters because "agent" gets used loosely across the industry, and the show tries to pin it down with language builders can actually use.

• Coverage of the full AI engineering stack: The show does not stop at models. It also covers evals, structured outputs, tool calling, GPU supply, observability, security, and deployment concerns, which reflects how AI products are really built in practice.

• Newsletter plus podcast plus YouTube ecosystem: Latent Space is not only an audio feed. Episodes, analysis, clips, and summaries travel through Substack and YouTube, which gives visitors multiple ways to follow along depending on whether they prefer reading, listening, or watching.

• Large technical audience: The newsletter has about 178,000 subscribers, and the YouTube channel has crossed 170,000 subscribers. For guests and sponsors, that reach means concentrated exposure to developers, founders, and AI decision-makers rather than a broad consumer audience.

• Fast reaction episodes on major AI releases: Alongside deep interviews, the hosts also produce quicker breakdowns of major launches and market shifts. This helps listeners stay current without having to track every model release, benchmark update, or API announcement themselves.

• Useful show notes and references: Episodes often come with links to papers, repos, products, and related resources. For technical listeners, that turns the show into a starting point for deeper research rather than a one-time listen.

Use Cases

Latent Space is not a tool people "use" in the same way they use an API, but it does shape real projects. One of the clearest examples comes from its episode with Notion's team, where Sarah Sachs and Simon Last walked through what it took to build Notion AI and more agent-like workflows inside a product millions already use. They discussed why earlier attempts struggled, including weak tool-calling standards, short context windows, and too much complexity pushed onto the model. For a founder or product engineer trying to build an agent into an existing app, that episode is less inspiration and more field report.

The show also serves teams making model and infrastructure decisions. In its conversation with Artificial Analysis, the discussion went deep into benchmark design, hallucination tracking, and speed-cost tradeoffs between models. That is useful for companies choosing between OpenAI, Anthropic, open-weight models, or smaller providers. Instead of hearing "this model is better," listeners get a more grounded story about what "better" means for latency, price, reasoning, and reliability.

For API builders, the OpenAI-focused episodes are another practical case. Latent Space covered structured outputs and the push toward reliable schema adherence, a feature that matters a lot for agent systems, workflow automation, and any application that cannot tolerate malformed output. If your team is building a tool-using assistant, a customer support workflow, or an internal copilot that writes into business systems, that kind of episode can change what you build next week.

There is also a career use case. Swyx's broader work around the "AI Engineer" role has made Latent Space a recurring checkpoint for developers deciding how to reposition themselves. We found the show repeatedly used as a way to understand where the jobs, startups, and technical bottlenecks are moving. In that sense, people are not only learning about products, they are using the podcast to decide what to study, what to prototype, and in some cases what company to start.

Strengths and Weaknesses

Strengths:

• The biggest strength is depth. Compared with general AI podcasts that optimize for broad appeal, Latent Space is willing to stay on one technical or product topic long enough for the interesting details to emerge. When Notion explains why its first agent attempts failed, or when OpenAI API teams explain structured outputs, listeners get specifics they can reuse.

• It has unusually strong guest access. Many podcasts discuss OpenAI, Anthropic, or Mistral from the outside. Latent Space often gets people from inside those companies, which changes the quality of the conversation. You hear less speculation and more firsthand explanation.

• It covers the practical middle layer of AI that many outlets miss. Research podcasts may focus on papers, and startup podcasts may focus on fundraising and growth. Latent Space spends time on the engineering reality in between, evals, runtimes, tool use, latency, GPU costs, and product design.

• The hosts have credibility with the audience they serve. Swyx's writing and community work gave him a strong position with developers before the show scaled, and Alessio adds operator instincts that keep episodes from becoming abstract. That combination helps the show feel informed without sounding academic.

Weaknesses:

• It can be too insider-heavy for newcomers. If you are just starting to learn about AI agents, some episodes assume a lot of context about model providers, benchmarks, frameworks, and startup dynamics. Compared with beginner-friendly shows, Latent Space asks more from the listener.

• The show is stronger on builders and products than on ethics or policy. Those topics do appear, especially around governance and deployment risks, but they are not the center of gravity. Visitors who want broader social or regulatory analysis may find other podcasts more balanced for that purpose.

• The format favors long attention spans. That is part of the appeal, but it is also a real tradeoff. A 2-hour episode with dense technical discussion is valuable if you are deeply in the field, less so if you want a quick overview on your commute.

• It is closely tied to the Bay Area and startup AI worldview. That gives it excellent access to labs and founders, but it can narrow the lens. Compared with more academic or international AI media, the show leans toward the concerns of venture-backed builders and product teams.

Pricing

• Podcast episodes: $0
• YouTube content: $0
• Newsletter subscription: Free tier available
• Paid Substack membership: Varies by current Substack offer

For most visitors, Latent Space is effectively free to start. You can listen on standard podcast apps, watch on YouTube, and read at least part of the newsletter without paying. That makes it easier to treat as a regular research source rather than a budget decision.

The real cost is time. A single episode can run 90 minutes to 3 hours, and the value depends on whether that time replaces scattered reading across X, release notes, benchmarks, and blog posts. For many technical listeners, it probably does. If you subscribe to the paid newsletter tier, check current Substack pricing and whether the extra posts or community access are worth it for your workflow. Compared with paid analyst subscriptions or conference tickets, Latent Space is inexpensive, but compared with skimming a free blog post, it asks for much more attention.

Alternatives

Eye on AI Eye on AI is a better fit for listeners who want broader AI industry coverage, including business, policy, and societal implications. If Latent Space feels too deep in the weeds on tool calling or evals, Eye on AI may be easier to follow. Meanwhile, if you are building agents or choosing infrastructure, Latent Space usually gets closer to the engineering details.

Practical AI Practical AI serves developers too, but it tends to be more approachable and educational. It is useful for people earlier in their journey who want concepts explained clearly and applied in a developer-friendly way. Latent Space is the stronger choice when you want to hear from the teams at the center of current AI product development.

TWIML AI Podcast TWIML has long been a solid resource for machine learning practitioners and researchers. It often goes deeper into ML methods and research framing than Latent Space does. If your work leans toward model development or academic ML, TWIML may fit better. If your work leans toward application engineering, product architecture, and agents, Latent Space usually feels more current.

a16z AI podcasts and videos Andreessen Horowitz produces polished AI conversations with founders, investors, and operators. Those episodes are often strong on market framing and startup strategy. Latent Space overlaps on guests and topics, but usually feels less investor-mediated and more builder-to-builder in tone.

The Cognitive Revolution The Cognitive Revolution is a good alternative for listeners who want a wider intellectual frame around AI, including science, policy, and long-term implications. It is often more reflective and less tied to immediate product shipping. Latent Space is the better choice when your question is "how are teams building this right now?"

FAQ

What is Latent Space Podcast about?

It covers AI engineering, especially agents, foundation models, evals, infrastructure, and product decisions. The emphasis is on how real teams build and ship AI systems.

Who hosts Latent Space Podcast?

The show is hosted by Shawn Wang, known as swyx, and Alessio Fanelli. Both are well known in AI builder circles, with swyx especially influential through his writing on AI engineers.

Is Latent Space Podcast a company or just a podcast?

It is a podcast, but also part of a broader media brand with a newsletter, YouTube channel, community, and live events. Most people discover it through the podcast first.

Who should listen to Latent Space?

It is best for AI engineers, startup founders, product builders, technical leaders, and anyone trying to keep up with how modern AI products are actually made. Beginners can still learn from it, but some episodes assume prior context.

Does Latent Space cover AI agents specifically?

Yes. Agents are one of its core themes, and the hosts have spent time defining agent architecture, discussing tools and runtimes, and interviewing companies building agent-like products.

How do I get started?

Start with an episode on a company or topic you already care about, such as OpenAI APIs, Notion AI, Anthropic, or model evals. That is usually easier than beginning with the newest episode and trying to catch up on all the references.

How long does it take to get value from it?

Usually one or two episodes are enough to tell whether the style fits you. If you work in AI already, you will probably get useful ideas immediately. If you are newer, give it a few episodes and keep the show notes open.

How long does it take to set up?

There is no setup in the software sense. You can subscribe in any podcast app or on YouTube in a minute or two.

Is Latent Space free?

Mostly, yes. The podcast and YouTube content are free, and the newsletter has a free layer. Some Substack content may sit behind a paid subscription depending on the current offering.

Is it good for non-technical listeners?

Sometimes, but not always. The hosts do explain concepts, yet many conversations assume familiarity with models, APIs, and startup AI vocabulary.

How often does Latent Space publish?

It publishes regularly, though the exact cadence can vary. There are both deep interview episodes and quicker reactions to major AI news.

What makes Latent Space different from other AI podcasts?

The combination of insider guests, long-form technical depth, and a clear focus on AI engineers. Many shows talk about AI. Latent Space tends to talk with the people building it, at the level where implementation choices matter.