Live with Tim O’Reilly

About Live with Tim O’Reilly

A series of candid conversations designed to slow down, dig deeper, and share real insights you can build on. Get the story behind the show here: https://www.oreilly.com/radar/more-slowly/

Publisher

O'Reilly

Category

technology

Language

en

Explicit

No

First episode

4 May 2026

Latest episode

26 Sept 2026

Latest episodes

23 episodes in the feed.

1. 

   26 Sept 2026

   Give Your Agents Superpowers: Jesse Vincent Live with Tim O’Reilly

   As the former chief maintainer of Perl 5 and project manager for Perl 6, as well as a longtime programming team leader, Jesse Vincent learned early that getting good work from smart, capable people often depends less on giving orders than on providing context, clear goals, and useful feedback. That experience turned out to be unexpectedly useful for working with AI agents. Now, as the founder of applied research lab Prime Radiant, he’s created Superpowers, a framework that teaches Claude Code to work more like a disciplined senior engineer. Jesse joined Tim O’Reilly to explain what building Superpowers has taught him about agentic development and why the scarce human skills for software folks are now knowing what you want, expressing it clearly, and judging whether the result is any good. Jesse’s approach borrows as much from management as from software engineering. He and Tim discussed why agents respond better when you explain why a rule exists, why prohibitions often fail, and how the techniques Jesse learned managing junior programmers are helping him now. Jesse approaches agents less like tools that need perfect instructions and more like colleagues who need context, clear goals, and well-designed processes, so his agents have roles, journals, and even a "therapist" that helps them update their own operating instructions. But the goal of Superpowers isn’t to make agents imitate humans. It’s to give them enough structure and discretion to turn high-level direction into the intended result while still speaking up when something looks wrong. That puts more responsibility on humans, not less. Jesse makes his agents surface unknowns before they start, explain plans back in small pieces, and prove that their work actually functions. He also avoids letting the same agent that produces the work also certify it. As AI takes over more of the labor of production, Jesse argues, writing, curiosity, domain knowledge, taste, and judgment become more valuable. The programmer’s job is changing, but the need to understand what should be built, communicate that intent, and recognize good work when you see it isn’t going away.

2. 

   17 Sept 2026

   Cracking LLMs Open: Emmanuel Ameisen Live with Tim O’Reilly

   Anthropic's interpretability team studies Claude's internals the way neuroscientists study a brain, but the science is younger, messier, and far less mapped. Emmanuel Ameisen has been part of that effort for the last two years, and at O'Reilly's recent Foo Camp, he shared some of what’s happening inside of Claude when it processes information. Emmanuel joined Tim to reprise that talk, then take questions from the audience. Token prediction is often dismissed as simply “pattern matching,” Emmanuel explained, but that undersells what’s actually happening. Ask Claude to finish a sentence about a short bike ride across the “GG bridge” and it correctly infers that the bridge must be the Golden Gate, placing the user in San Francisco. From there, the model can build on this understanding to answer questions, for instance how long it will take to get to “world-class skiing” at Lake Tahoe. Emmanuel and his team study the patterns that show up when the model encounters particular ideas, and they can even intervene in those activations and see how the model’s behavior changes. Models are building a complex understanding of the world, and as Emmanuel pointed out, they “use their world model at every token.” Emmanuel and Tim discussed what Anthropic's interpretability team has learned by "pushing or pulling" on a model’s activations, what AI-generated poetry tells us about planning and world models, how models do math by placing numbers on “particularly shaped curves,” and why we should think about LLMs as actors playing a role down to the emotions they represent. They also explored what humans can do that models still can’t and the bigger question of what studying LLMs might teach us about how our own minds work.

3. 

   11 Aug 2026

   Escaping the Prompt Debt Trap: Drew Breunig Live with Tim O’Reilly

   Drew Breunig has spent years watching how AI teams grow their LLM applications, and he's identified a problem most teams don't notice until it's already slowing them down. As CEO of CMPND, a maintainer of the DSPy open source framework, and author of the forthcoming O'Reilly publication Context Engineering Handbook, he's seen the same pattern play out over and over. A team builds a quick LLM prototype, it mostly works, and then they start patching the gaps, one edge case at a time, one instruction at a time, until the prompt is so bloated and fragile that nobody wants to touch it, it stays pinned to a model from two years ago, and the person who built it is the only one who understands it. If that sounds familiar, you're probably already dealing with what Drew calls prompt debt.Watch as he joins Tim O'Reilly to explain why it's more widespread than most teams realize, and what the best builders are doing instead. Drew traced prompt debt back to two root causes. The first is that natural language is a fundamentally imprecise programming medium. Studies show the same medical question phrased as a patient versus a psychiatrist yields completely different answers, and in one example from the talk, claiming to be a Philadelphia Eagles fan was enough to bypass model safety guardrails entirely. The second is that the biggest labs are increasingly tuning models like Claude for their own products in ways that quietly work against anyone building something custom on top. The conversation also covers why this reliability-over-diversity trade-off is making frontier models worse writers even as they improve at code, why open-weight models may be the Apache to OpenAI and Anthropic's Netscape, and what you can do now, from investing in evals and decomposing tasks to treating prompts as perishable, to build AI systems that hold up. As Drew puts it, "Vibe coding makes code free, as in puppies. You can go get a puppy for free, but now you have to take care of it and it's in your house."

Reach and audience

Public platform figures. Ratings count people who left a rating, not total listeners.

Spotify

5.0 / 5

1 ratings

Contact Live with Tim O’Reilly

Guest appearances

Books guests

Based on episode analysis; this does not confirm that the show is currently accepting guests.

Shows with a similar audience

• 

  The Pragmatic Engineer Podcast

• 

  Dwarkesh Podcast

• 

  The AI Breakdown: Daily Artificial Intelligence News and Discussions

• 

  Pivot

• 

  The Rest Is Politics: US