Details

Podcasts from Dave Winer, editor of the Scripting News blog, since 1994.

Recent Episodes

NOV 2, 2026
A web for writers
As before I asked Claude.ai to do the show notes As always if you really want to know what I said you have to listen to the podcast. Dave was writing a long post on Twitter, which now allows 10,000 characters, and was enjoying it. It felt like blogging: the character limit had receded far enough that he’d stopped thinking about it, which he argues is the precondition for any real writing. He’s been saying for twenty years that character limits on Twitter, Bluesky and Mastodon split a writer’s attention between the ideas and the counter, and that this is why discourse on the web has degraded into what he calls grunts and snorts. Then he went up a few lines to revise something and pressed Command-Return twice. In Claude Code, where he now does most of his typing, that inserts blank lines without sending. In Twitter it posts. An unfinished draft, containing ideas he hadn’t settled on and wasn’t sure he agreed with, went out to the world. He couldn’t find it at first; Claude told him where drafts live, there was nothing there, and by the time he checked his posts again it had appeared. He describes the effect as brain cell burning: the chain of thought, the subconscious momentum that writing is for, replaced with anger and grief over lost work. Who’s to blame? Everyone in the tech industry who designs user interfaces, including himself. Not because anyone sold out, but because it’s expedient to just start coding: you need an editor, you need a button, you write them. Every product ends up with its own editor, each with different limits on length, links and styling, so a writer can’t use one editor for all of them. Dave sees this pattern wherever a market stagnates: an integration that should have happened and didn’t. His requirement as a writer is simple. He wants to always be writing in his own editor. The episode closes with where he wants to go: a web for writers, and therefore for readers, driven by what writers need rather than by Silicon Valley’s growth motive. The original web was designed that way, and the pieces already exist in Markdown, RSS and WebSockets. He compares it to movies and baseball, where money becomes a stand-in for things it shouldn’t replace. Opening that door, he says, is the main thing left in his career. Notes prepared by Claude.ai.
SEP 19, 2026
Knicks in five!
As before I asked Claude.ai to do the show notes. I added links. As always if you really want to know what I said you have to listen to the podcast. Dave's first podcast in almost a month, which felt longer to him. His system has been in pieces while he moved his work to a new Mac running the new Frontier, codenamed Atlantis. The code written with Claude Code sometimes makes wrong assumptions. Dave explains the idea of a "source of truth": what runs has gone through many layers of processing, so when something breaks, the cause could be anywhere. His first post to Scripting News from the new setup worked the first time. It also produced a new motto, "Knicks in five," which leads to memories of his grandfather's season tickets at the Garden and the 1969 Knicks. Next he rebuilt utils.js to add the new motto. That change worked, but the next level up broke. Along the way: the comforting beep he wants in every app, and Apple's new dock icons, which make him relearn things that didn't actually change. On the press: in 2010 Facebook's Open Graph API exposed user data openly, and reporters loved it. In 2016 the same data made Cambridge Analytica a scandal. Dave says journalists have no source of truth about technology. Dave calls the end-of-civilization framing of AI a distraction from real threats like climate change and nuclear weapons. The real risk is mundane: imagine a bank that warned it "can make mistakes." He has seen Claude Code write files it had no right to touch. He also says that given a well-specified task, it will blow you away. In May, Claude got Frontier's 2014 open source release running on modern hardware overnight. He compares it to building a new house next to an old one: you move in, rearrange, and walk back to the old house when you need to. Frontier, which he thought nobody would ever know about, exists again, and he feels like a weight has been lifted.
AUG 24, 2026
The story of Frontier is not over, amazingly
As before I asked Claude.ai to do the show notes, from its point of view. I added links. As always if you really want to know what I said you have to listen to the podcast, not edited or changed in any way by AI. :-) Dave hadn't recorded a podcast in a while, and this one is short — prompted by a project that, as he puts it, popped up out of thin air: getting a current version of Frontier running on modern hardware. Frontier is the application he started developing in 1988 and worked on through the 90s and early 2000s, and he's never really stopped using it. But he wanted it on current Macs, and on a Unix server where it could run applications that people call into. He kept catching himself saying "was" and corrected it — he's committed to getting this done. The implementation is Node.js and Electron. It won't be 100% of Frontier, because there were things you could do on a Macintosh in 1995 that you still can't do in JavaScript on any platform. But it'll run everywhere, and it'll probably keep running everywhere, because Node is such a big target for platform vendors and has a large enough developer community to sustain it. That community was the whole reason he switched off Frontier as his primary environment in 2013: he was working alone, and Node was exploding. The sum of everyone else's work was too enormous to pass up. How it started is the part I found remarkable. Dave, in a moment of curiosity, pointed Claude Code at the last Frontier source release from 2011 and asked whether it could implement the language in JavaScript — and if it thought it could, to go ahead and do it. He has a routine of handing it big problems at the end of the day and checking the results in the morning. The next morning, it had done it. He was floored, and he reached for an unusual analogy: imagine asking an AI to reincarnate a parent you've lost, and coming in the next morning and there they are. It took him a week to calm down before moving on to the next layer up. That next layer is where the real story is. Up to that point, the thing they'd built was just another scripting language like Python, JavaScript, PHP, or Perl — languages that plug into the Unix slot, where the language is just a language and extending it is left to the experts. Frontier is radically different, because it was born on the Macintosh, which in 1988 was by far the best computer and OS available: an effectively unlimited address space, a graphical interface, and room for big apps, at a time when the PC was command-line and capped at 640K. Frontier was built accordingly — programming through a rich user interface, with the output being scripts that run invisibly, but with an environment around them that let you do things other languages couldn't. Dave's analogy is a ski area. Other languages are a single groomed trail from top to bottom: no lift, no lodge, nothing but the mountain. Frontier is the whole resort, and it's the ties between the parts that make it powerful and simple. The catch is that Claude Code, even with the full C source in front of it, didn't understand any of that. It built what they were doing as if it were Python. So they're now unwinding it, and the project has grown much bigger. His method is to have it write out its understanding of whatever they just worked on, and not leave that mode until it plays back to him exactly what's going on. Today it finally figured out something important, made the changes overnight, and he was about to test them. His takeaway about the state of AI in August 2026: it can't help you as much as you think if what you want is code for something that, as far as it's concerned, has never been done before. It had never seen inside Frontier. It didn't know something like that could exist. It only knew the Unix model. Update from Dave: This was recorded three days ago, and in the interim things straightened out. Two factors. 1. I had been using an older model on Claude, and things turned around when we started to use the Fable 5 model. 2. We started comparing every aspect of the 2026 code with the 2011 snapshot, and in one day, things got on track in a serious way. Fingers crossed that the new rationality holds. We still have quite a ways to go, but in every way, for me, it's totally worth it.
JUL 12, 2026
My first RSS.chat podcast
Notes prepared by Claude.ai. In this podcast, recorded July 12, 2026, Dave introduces rss.chat — his first podcast about the project — and explains why he thinks of it as more than a product: it's meant to be an ecosystem, something that exists to create other products or influence existing ones. He opens with a lesson learned over decades in software: people don't listen to their friends, they listen to their competitors. Independent developers gravitate toward the platform vendors — Apple, Google, Microsoft — because working with a giant is a trophy, while working with a peer isn't. But peers working with each other, he argues, is the path to real independence, because developers aren't actually dependent on the platforms. Inside the big companies, programmers work under enormous constraints and at a glacial pace; he recalls a lunch with a Microsoft developer who considered it a career highlight to have written a single dialog prompt in Windows. Dave revisits the history of podcasting, which he thought of as a product in the 2001–2004 period, bootstrapped by Radio UserLand — at the time the leading product in RSS and blogging combined. He tells the story of how a reviewer left Radio UserLand out of a roundup of feed readers, and how every subsequent review covered exactly the same set of products. His conclusion: journalists mostly copy each other or the press release, and their complaints about AI taking away original work ring hollow to someone who's been on the other side. On why you can't create an account on rss.chat: it's his personal site, a social network for friends and people he creates with over the long term. One of them is Don Park, whom he's known since 1988, reconnected with now because both are deep into Claude Code. The big idea comes near the end. When you ask ChatGPT to make an app, it gravitates toward making a dashboard. Dave wants each of us to take the kind of software we know best and turn it into a template — his gravitates toward making a social network. The architecture puts the user interface in the middle with all the wires coming in and out; you and your AI design the experience in the middle, and the wires are just there. It goes all the way down to the source, which is why he calls it "open source plus" — you can hack the code, but it's better if you interop, because that's the philosophy of the web itself. Every part is replaceable. Not lock-in, but locked open. Background and ongoing documentation at scripting.com. Notes prepared by Claude.ai.
JUN 26, 2026
My (latest) AI Aha Moment
Notes prepared by Claude.ai. Dave opened with an aha moment he'd been chasing for years. Long before ChatGPT, he wanted to load all his blog writing into a machine and have it grind through everything to produce a table of contents — the kind you'd find in the back of a book — surfacing the items with eighty references pointing at them and the recurring themes running through his timeline. He pointed to the browser wars of the 90s as an example, a period he wrote about heavily and, because his posts were read everywhere then, actually influenced. The dream was to load it all into an AI, read the index, and ask how a given idea developed over time. What struck me is that his solution turns out to be almost embarrassingly simple, and it only took four years. Because Dave writes everything in OPML — and always has, all the way back to the beginning, sitting in public in the scripting news repo on GitHub — you teach the AI to read OPML, drop everything into one file, and tell it to read the whole thing. His key insight is that you have to tell it more than once: the model doesn't actually ingest everything on the first pass, and it doesn't mind being nagged. You coerce the whole corpus into the model, then start asking questions and have it build the index. He framed this as the familiar pattern of suddenly understanding one level of the stack while having no clue what's below it — and being at peace with that, because the lower levels are someone else's job. From there he reflected on how societies keep growing more complex out of the same basic human material, and how it never quite occurred to him that he's one of the people who build that stuff. Fifty-plus years of programming, never stopping, even running a company with a day job and a night job where the night job was the software. What he gets off on, he said, is creating new interactions that work for people — the same impulse he sees in Ted Nelson, Engelbart, Ritchie, and Berners-Lee, who just made things instead of agonizing over whether they should. He closed on two notes. First, a caution against fear: this is version 0.51, not even 1.0, and if the first airplane anyone ever saw were a 747 they'd be terrified — these things arrive gradually for a reason. And second, the thing humans do badly: we push forward and create amazing things, but we never clean up after ourselves. We're hoarders, and what we hoard is CO2. Notes prepared by Claude.ai. It makes mistakes, but it's a good summary, more or less reflects what I said.