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/

Live with Tim O’Reilly is a technology podcast hosted by O'Reilly, with 23 episodes on record and a Required Pod Score of 80.

O'Reilly hosts Live with Tim O’Reilly, a technology show with 23 episodes published.

Our AI reads these to draft pitches. Use them as grounding for a pitch that cites a real guest and a specific topic.

• 

  Episode #11

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

  Sep 25, 20261h 2mS2

  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.

  ListenShow Notes
• 

  Episode #10

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

  Sep 17, 20261 hourS2

  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.

  ListenShow Notes
• 

  Episode #9

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

  Aug 11, 20261h 2mS2

  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."

  ListenShow Notes
• 

  Episode #8

  A Playbook for Going AI Native: Dan Guido Live with Tim O’Reilly

  Aug 6, 20261h 5mS2

  Trail of Bits is a software security research and development firm that helps clients like Facebook and DARPA solve their hardest security challenges. In 2024 cofounder and CEO Dan Guido decided the company would go fully AI native. Most of the firm's roughly 130 employees didn't buy in right away, but as of today AI is a “core participant” not just in engineering work but across the full consulting lifecycle. (Sales and operations folks are even using GitHub everyday and contributing to the company skills repository.) Dan joined Tim for an encore of a talk he gave at the [un]prompted conference in March laying out the playbook that made this transformation possible, which Tim called “far and away the best framework and discipline for going AI native." Dan then stuck around for a Q&A to flesh out some of those ideas with the live audience. Most companies get stuck at using AI as a tool rather than turning it into an operating system that compounds expertise across the entire organization because people resist adopting new technologies. For AI specifically, Dan lays the blame on a quartet of very understandable psychological barriers: a bias toward trusting your own (human) judgment over a machine's, an identity threat that makes augmentation feel like replacement, the opacity of not knowing how an AI reached a conclusion, and a low tolerance for AI's mistakes. Trail of Bits built its transformation around those biases instead of against them by standardizing on a single toolset (Claude Code), writing an internal AI handbook, creating a maturity matrix that made AI skill a normal part of career growth, running hackathons instead of mandates, and capturing everything the company learned into internal, external, and curated skills repositories. The approach has produced real results. Dan said roughly one in five bugs the firm now finds for clients is first surfaced by AI, and its go-to-market team is outperforming typical consulting industry benchmarks. That’s not to say that Trail of Bits’ transformation is complete. Asked what still doesn't work, Dan pointed to privacy-preserving inference and to the lack of any real feedback loop for learning from how employees actually use the tools day-to-day, both of which he called open problems the industry hasn't solved yet. However, the playbook the company has developed, alongside all the skills and other tools it’s shared publicly, will get you a long way toward transforming your own company. As Dan emphasized, “This is not a strategy deck. This is what we actually did.”

  ListenShow Notes
• 

  Episode #7

  The Days When Fable Got Hit by a Bus and Disappeared with Gene Kim

  Jul 28, 202659 minS2

  Long-running Claude Fable 5 agents had become the backbone of Gene Kim's personal knowledge system, one built from thousands of screenshots, videos, and half-finished side projects, spanning 50 code repositories and tens of thousands of changes. He’d been warned that Fable might be pulled from his subscription and got to preparing for that eventuality. But it disappeared eight days earlier than he planned for, mid-session, when Anthropic suspended access over an export-control issue. As he put it, the ship was soon “on fire, tumbling in space.” Gene’s the coauthor of The Phoenix Project, The DevOps Handbook, Accelerate, and, most recently, Vibe Coding (with previous guest Steve Yegge) as well as founder of the Enterprise AI Summit, so he knows a thing or two about dealing with unplanned disasters. He joined Tim for an encore of the lightning talk he gave at O'Reilly's Foo Camp about how he recovered his system, followed by a Q&A on what the shutdown taught him. Gene described watching his systems fail once the less-capable Claude Opus took over from Fable mid-session. Commands that should have worked started breaking, the model argued with itself over which instance had done what, and every action risked making things worse. It took roughly three hours that night, and more the next morning, to stabilize things, largely by recovering documentation Fable had written along the way, and Gene took the audience through the process step-by-step. He and Tim also discussed the deeper lesson of dependency risk from AI providers. We hear a lot about pricing and availability changes, but export controls, shifting safety policy, and geopolitics can force the same kind of sudden cutoff. Gene helped us understand why the solution is the same preventive, rehearsed resilience DevOps engineers built over the last 15 years for large-scale outages, now applied to individuals running mission-critical agents of their own.

  ListenShow Notes

Recent guests on Live with Tim O’Reilly. Study who booked and why before you pitch.

Based on semantic analysis of episode topics and host coverage, this show is a strong guest fit for executives in:

Industry fit is computed by PitchCentric using vector embeddings of the show's episode catalog.