Live with Tim O’Reilly

Longtime software engineer Steve Yegge has lately been exploring the limits of vibe coding with projects like Beads, his coding agent memory system, and Gas Town, a proof-of-concept agent orchestrator so complicated, expensive, and chaotic that he warned those interested NOT to use it. As Tim O’Reilly points out in his takeaways from this episode, “Steve has always been one of the most provocative thinkers in our industry.” Steve joined Tim for an insightful and entertaining conversation on coding with AI that touched on everything from computer graphics in the ’90s to desire paths and why getting riled up is the key to writing a good blog post.

Steve and Tim spent a lot of time discussing the wider context around Gas Town, including why Steve sees it as an enterprise tool—and as an executive assistant who takes on the mundane work so you can focus on the important problems. They also covered agent orchestration and the evolution of coding, using Steve’s eight-level framework; AI vampires and how exhausting it is to work with a stable of agents (Steve’s taking two naps a day!); why you might find yourself on the wrong side of the bitter lesson, and what to do about it; the reason Steve “wouldn’t touch [OpenClaw] with somebody else’s 10-foot pole”; why he’s adamant that developers need to overhaul their mental models from structure and framework cognition to just letting AI do its thing—and why he’s not even looking at the code anymore; and much, much more.

“The big takeaway,” Steve told Tim, “is that there’s always more work. It doesn’t matter how superhumanly good your helpers get; you’re just going to want to do something bigger. Our ambition will always outstrip our compute.”

Check out Tim’s takeaways from the conversation, plus clips, on Radar.

Addy Osmani should be a familiar face to the O’Reilly community. He’s the cohost of our AI Codecon events, the author of Beyond Vibe Coding, Leading Effective Engineering Teams, The Effective Software Engineer, Web Performance Engineering in the Age of AI, Learning JavaScript Design Patterns, and Building Web Apps with Bolt, and a prolific blogger on Radar and with his own newsletter, Elevate. He’s also a longtime Googler who’s spent nearly 14 years building developer experiences in Chrome and is now helping developers and businesses succeed with Gemini. As Tim O’Reilly put it in the introduction to this episode, “Addy Osmani is one of those people who is really grounded but also really able to think big and see the future.”

Addy sat down with Tim to chat about the state of the industry as it moves toward the orchestration of multi-agent workloads. In their wide-ranging conversation, they covered the tension between creativity and productivity, and balancing velocity with long-term technical maintenance and reliability—particularly from the enterprise perspective. Larger organizations can’t just let the agents rip. As Addy explained, “The real frontier for business is not necessarily having hundreds of agents for a task just for its own sake. It’s about orchestrating a modest set of agents that solve real problems while maintaining control and traceability.” And then there’s the as-of-yet unsolved problem of making everything work together as smoothly as possible.

Along the way, they considered the distinction between “feeling” productive and “being” productive (h/t Will Manidis); how to keep up-to-date on the latest trends and conversations; why being able to explicitly define the architecture and the purpose of what you’re building will matter more than how fast an AI tool can build it; why it’s still a good moment for young students to become software engineers; how MCP and A2A complement each other; why YOLOing token use is probably not the best strategy for most people; and more.

Watch now, or read Tim’s takeaways from the conversation (with clips) on Radar.

In the year’s final Live with Tim O’Reilly, Tim hosted an AMA to pull together some of the threads that have emerged during his conversations with guests who, as he put it, “know way more than me”—and think through how they may play out next year. “With AI, regardless of what happens with scaling, we now have a set of new capabilities that really challenge our thinking and our understanding about the role of human expertise,” he explained in his introduction. In his audience-guided talk, Tim explored the ramifications on our jobs and responsibilities, open source, and building participatory markets writ large.

Watch now to hear Tim’s perspective—illustrated by examples and anecdotes drawn from his long involvement in the tech industry—on why we’re still waiting to see an explosion of software productivity; developing (and sharing) skills and products as a side-hustle; what’s surprised him about AI this year; whether we’re moving toward personalized software; how to keep up-to-date with the speed of innovation; the growing AI backlash; how new technologies like MCP are informing the evolution of the AI industry; grappling with the bitter lesson; and much more.

Kent Beck’s career-long mission has been to improve software development. He created Extreme Programming back in the ’90s to address some of the issues that were slowing productivity. More recently, he’s been working on a series to help “tidy” development. Kent joined Tim to share his philosophy on tidying not just code but also the very human tasks of collaboration and teamwork, as well as his approach to coding with AI. Kent offered takeaways from collaborating with the “genie”—his term for AI tools, which may grant your wish but usually not in the way you wanted—focusing on the benefits beyond the ability to generate more code faster.

Check it out to learn why Kent thinks inhibition is the key to building reliable systems out of unreliable components like generative AI; what carries over from Agile and XP and pair programming to AI; the things old programmers know that new programmers still need to learn; why the deepest insights come the furthest into maintenance (hat tip to Ward Cunningham); and the importance of fostering a sense of shared responsibility throughout the organization as AI coding tools accelerate the pace at which code can be produced. Kent wrapped things up by discussing the joy augmented coding brings him and the knock-on effects that come from “engineer playtime” with GenAI tools. As he explained, “It’s not what [developers] produce; it’s how much faster…they can learn and how much richer their thought processes can be. And that benefits everybody.”

If the assumption that we’ve lived with for years that scale brings lower marginal costs isn’t true, we really have to rethink a lot of things. Anjali Shrivastava is an independent researcher and data scientist looking into the unit economics of AI’probing how the seemingly simple token unit of cost for generative AI services is actually highly variable, leading to hidden margin risks that undermine conventional subscription or usage pricing models. Anjali joined Tim to talk through her recent work, particularly as it relates to code-generation tools like Cursor and Claude Code.

Over the course of a very interesting hour, they got into the the distinctions between input tokens and output tokens and between reasoning tokens and output tokens, and why they matter; the challenges providers face in instituting systems to offer visibility for users, and even for themselves’and why it’s not only a business problem to solve but one at the heart of AI products as they currently exist; how AI service providers differ from traditional SaaS businesses in ways that make the latter’s pricing model obsolete (even though as of now, it’s the model that AI providers continue to use); and how AI providers can balance, and even temper, demand as use scales. They also discussed the efficiency opportunities new hardware may provide, one potential future where AI gets much pricier, why efficiency alone won’t solve the pricing issue, how external bottlenecks, like electricity, influence pricing, and much more.

There may be a few people who are still unfamiliar with at least some aspect of Reid Hoffman’s long career—among other things, he’s a cofounder of LinkedIn, a partner at Greylock, the author of Blitzscaling and Superagency, and the cohost of the podcasts Masters of Scale and Possible. As Tim O’Reilly put it in his introductory remarks in this episode, “Reid has really been a force for good in the world.” Reid joined Tim for a conversation that touched on topics from across his career but was focused squarely on how AI is transforming business.

Drawing on Reid’s investment work (he led OpenAI’s first commercial round), they first looked toward important areas for further research and funding—Reid argues that one promising area will be reinventing current industrial processes, like that for drug discovery, to be AI native. Then, in their wide-ranging discussion, Reid and Tim considered the implications on business as AI becomes enmeshed in business and society, grappling with weighty questions that will guide whatever the future may hold:

• Will the future of AI look more like Windows, with OpenAI in place of Microsoft, or the open web, with a proliferation of standard protocols like MCP?
• How will job roles and skills evolve in fields as varied as programming, medicine, marketing, and customer services? And more importantly, what choices should businesses make as AI disrupts the workforce to continue valuing the skills that are uniquely human?
• And why should people just entering the workforce optimize their career for AI—and have fun doing it?

Stick around until the end to watch Reid chat with his AI-generated 1985 self about his fears and hopes for the future.

Claire Vo, the founder of ChatPRD and host of the How I AI podcast, has a real knack for extracting and sharing practical tools, tips, and ideas that anyone can put to use today. In this episode, Claire sat down with Tim to share her perspective as a longtime product leader who began using AI to boost her own productivity and ended up turning it into a successful business.

Claire sketched the origin story of ChatPRD and got into what it’s like to be the only full-time person working there (alongside a few fractional employees and a host of agents). She also weighed in on how to balance traditional software with AI tools; why AI product leaders should think beyond good prompts to consider an architectural view of where your model fits in your application; and how, even if product management is disrupted by AI, product managers can still add value by coming up with creative, innovative, and nonobvious solutions to tough challenges—and, of course, by talking with customers to understand their pain points. As Claire noted, “It’s just not enjoyable to be interviewed by AI unfortunately. So I think your job is safe for right now.”

For years, technologist and consultant Sam Newman has helped enterprises come to grips with sticky issues in software architecture, and he’s been thinking deeply about how AI will reshape software engineering. He joined Tim for this episode of Live with Tim O’Reilly to explore the technological and organizational changes AI is already bringing about. As Sam pointed out, it’s chaotic out there.

Sam chatted with Tim about the consequences of the industry’s narrow focus on code completion rather than experimenting with where else in the software development life cycle AI might bring value; how the economics of AI are shaping up—and what that means for our systems (Sam guesses that we won’t see companies undertaking big projects like rearchitecting until they figure out how to actually make money with AI); and why you should take a model-agnostic approach to developing products built on LLMs. In their hour-long conversation they also touched on security, building data centers for AI workloads, the tools Sam’s using in his day-to-day, and the positive developments Sam’s seeing with MCP.

When Sam Schillace, Steve Newman, and Claudia Carpenter built Writely—their browser-based word processor, which became the basis for Google Docs—they saw a future that most of us didn’t yet appreciate. It’s been 20 years since Sam, Steve, and Claudia launched the Writely beta. To celebrate that achievement, they joined Tim to reflect on the success of their revolutionary product and discuss what it might teach us about the current moment.

Sam, Steve, and Claudia solved a problem people had been complaining about for years (version control in shared docs), and did so by putting together available technology in a new way where suddenly everything just clicked. (They also did it quite quickly—as Steve pointed out, it was only a year from Writerly’s inception to its acquisition.) They shared some of that history with Tim, with Sam explaining why fast experimentation and “learning by doing” was so valuable to their project, and remains so in the AI era.

They also got into tackling the hard problem of collaboration (which is bigger and more human than a single tool can account for), and what AI might make possible that we couldn’t do before; designing paradigm-shifting products that are still easy to use, and why product designers should always keep laziness top of mind; the new economics of AI software development; and why current AI products may be the equivalent of CD-ROMS. And they had some advice for those looking to build products with AI. As Claudia noted, you can’t just try to “bolt on” AI; you have to find the use cases that you previously “just couldn’t even think about doing that AI lets us do now.”

Tim also provided some context in advance of their conversation. Check it out on Radar.

AI boosters would have you believe that AI is such an unprecedented leap forward that it will change everything almost immediately, that the singularity is upon us. Princeton professor and author Arvind Narayanan begs to differ.

Arvind and his coauthor Sayash Kapoor make the case in their recent paper “AI as Normal Technology” that, while AI is indeed a transformational technology, it’s far from unprecedented. In this episode, Arvind joined Tim to explain why AI is likely to follow much the same patterns as other profound technology revolutions, such as electrification, the automobile, and the internet. In their hour-long conversation, they thought through the social impact of AI, from how AI might change software to broader concerns such as why fears of AI running amok are overblown; why labor law, not copyright, may be the key to AI justice; and how AI might help address some of the problems baked into our education system.

Get an overview of the conversation and Tim’s takeaways from it on Radar.