Listening guide
Latent Space is the show a lot of people in AI engineering check first, and deservedly. swyx and Vibhu get builders on early and ask them engineering questions instead of vision questions. If you have been through the back catalogue and want somewhere to go next, here is one answer: 5 Latent Space episodes, each paired with a Chain of Thought episode that picks up the question next door.
I host Chain of Thought. We have no affiliation with Latent Space: no partnership, no arrangement, nothing either of us gets from this page. We are just fans, and a fair amount of our work builds on theirs.
Latent Space covers the frontier as it arrives: the labs, the inference stack, the model releases, and the engineers shipping them. The hosts do the reading, so the questions land at the level of someone who has used the thing rather than read the launch post. The episodes that stay with you are the ones where an engineer walks through a decision and the tradeoff they made.
Chain of Thought sits a step downstream, and that difference is the whole basis for the pairings below. Latent Space is often first to the thing itself. We tend to arrive once a company has been running it for a couple of quarters and can say what it cost, what broke, and what they would do differently. An episode about how something works, then an episode about what happened when someone put it in production. Every Chain of Thought episode carries its full transcript on the page, so you can read one instead of listening if that suits you better.
Akshat Bubna takes apart what a cloud has to change when the thing calling it is an agent rather than a person: sandboxes to run untrusted code in, and GPU snapshotting so a machine resumes instead of cold-starting every time. The most concrete hour on agent infrastructure we have come across.
Then play
Vercel's Playbook for AI Agents: From Vibe Check to Production | Malte Ubl Malte Ubl, Vercel Modal's answer is about the primitives underneath. Malte Ubl's is about what a team builds on top of them once infrastructure stops being the hard part: Vercel's definition of an agent as software for tasks too flexible to specify up front, and the route from a proof of concept to something a company will put its name on. Same stack, one layer up.Akshay Nathan leads core product engineering at OpenAI, and this is the view from inside the team that took Codex and ChatGPT Work to ten million users in two weeks. It also covers the part nobody planned for: a coding agent getting pointed at finance and marketing work by people who do not write code.
Then play
How Block Deployed AI Agents to 12,000 Employees in 8 Weeks w/ MCP | Angie Jones Angie Jones, Block That is the vendor side of adoption. Angie Jones has the buyer side: what it took to put agents in front of 12,000 people at Block in eight weeks, built on MCP and on Goose, the open-source agent her team released. If the Codex curve made you wonder what the receiving end of it looks like inside a company, this is that hour.Baseten’s engineers follow a long-context request all the way through the system: cache-aware routing, prefill and decode split apart, speculative decoding, and the quantization work behind putting a new open model on a production API. A technical hour that does not hand-wave the hard parts.
Then play
How Intercom Cut $250K/Month by Ditching GPT for Qwen Fergal Reid, Intercom Baseten’s episode is the engineering. Fergal Reid’s is the invoice. Intercom was spending $250K a month on a single summarization task against GPT, replaced it with a fine-tuned 14B Qwen model, and kept nearly all of the difference. Played back to back, the inference techniques stop being abstract and start having a number attached.Databricks’ two co-founders make the open-ecosystem argument in detail, including Omnigent, their open-source harness for coordinating several coding agents behind one set of APIs for sessions, security policy and spend.
Then play
AMD's Challenge to NVIDIA: The Open Ecosystem Bet | Anush Elangovan & Sharon Zhou Anush Elangovan & Sharon Zhou The same argument, one layer down the stack. Anush Elangovan and Sharon Zhou make AMD’s version of the open bet from the silicon side, recorded at Advancing AI. Zaharia and Xin are arguing for openness where the data and the agents live; AMD is arguing for it where the compute does. Together they show what open is actually buying at each layer.Zico Kolter and Matt Fredrikson on why AI security does not behave like the security we already know: indirect prompt injection, why bigger models do not automatically become harder to break, and the lethal trifecta of untrusted input, access to private data, and a way to get data back out.
Then play
You Can't Secure an AI Agent with Software Charles Guillemet, Ledger Gray Swan’s answer is better software: an automated red-teamer, and a guardrail filter for policy. Charles Guillemet runs the offensive security lab at Ledger, the one that breaks Ledger’s own products before anyone else can, and he thinks software permissions cannot secure an agent that moves money, so hardware has to be in the loop. Two serious answers to one problem, and hearing them in order is worth more than either on its own.If you want a wider net than a single show, the cross-show roundup collects agent episodes from five different podcasts, Latent Space among them.
New to our show? Start with five → Everything on AI agents → Latent Space ↗