This page lists part of the section. The page links below work without JavaScript. With JavaScript, a filter searches the whole section and pages the matches.
Squeezing index work into limited time, Dave has a local model write a piece and figures out later whether it's useful. His finding: a coding LLM does far better when asked for a single function than for a whole piece of software or script.
because here's one thing I have found out with these coding LLMs is if you tell it to write a function, it is way better than if you tell
Not in the audio. The end of the sentence ('it to write a whole piece of software or a whole script') is in the following Unknown cue, so it's paraphrased in the summary, not quoted.
When Adam wonders why a seemingly simple project is so hard, Dave explains that we see a complete vision of the thing with all its parts working at once. If you could flowchart it in your head and never write the software, it would be perfect. Once you break it into a linear series of steps, you find the vision is valid but getting there is far harder.
problem though is that I don't think humans are very good at determining how hard a problem actually is.
After an LLM gave up on writing a Rust feed parser, Dave concludes that these tools are really useful for checking your own code against popular ways of doing things (he hesitates to call them best practices), because the bots are trained on everybody else's code.
Code checking yourself against some. I hesitate to say best practices. But against sort of some popular ways of doing a thing. Because what this, what these code bots are trained on. Is everybody else's code.
On Apple's response to the Epic ruling (requiring an entitlement to link out, then a scare screen), Dave notes that the word itself says it all. Anything granted as an entitlement can be taken away at any moment. Developers on the App Store are digital sharecroppers.
Baked into the word entitlement is the threat of removal of that entitlement at any moment.
Advising Adam on a script that fetches podcast audio, Dave says faking a browser user agent seems smart but ultimately causes problems. Just give your app its own name. The Podcast Index itself instantly blocks traffic from hyperscaler networks (Amazon, Linode and so on) that claims to be a web browser, because nobody runs Mozilla from a data center.
I will say that you need to not impersonate a browser user agent. It can seem like the right thing to do, but ultimately you will run into problems.
We associate radio with a consumption habit (linear, scheduled listening), but Dave says that wasn't by design. Linear terrestrial broadcast was the only way it could work. Radio doesn't have to be that, and now there can be a digital version of the same thing without the constraint.
radio, radio, but radio doesn't actually have to be that. That it was only, it was only that because you didn't have any other way to just to do it.
Puzzled that apps fear adding Podcasting 2.0 payment features because Apple might reject them, Dave, a mobile app developer himself, says rejections happen to builds every day for arbitrary reasons. You add the feature, and if it's rejected you roll it back out. Your developer account won't get banned over it.
with all kinds of features as an app developer, as a mobile app developer. Every time you put a feature in you, you never, you never know what the whims of app review are going to be.
Whether it's a logo or a Laravel app, Dave says these models get you most of the way to an idea but never to exactly the thing you want. Their real use is producing something close that you hand to an actual artist or developer to finish.
they, they all really allow you to get most, you know, most of the way to an idea that you can then hand off to somebody to finish it.
From his IT day job, Dave explains that you used to spot phishing by poor translation. LLMs ended that. Links and attachments get blocked quickly, so attackers now send a harmless-sounding note to bait a reply, build trust, and only later send the link. You could get ten such emails, two legit and eight phishing, and never tell the difference.
I mean, phishing has exploded in, in like new and impressive ways since the LLM stuff, because used to, you could really depend on poor translation to sort of identify, but now
Not in the audio. Guest episode (Matt Medeiros; 0.35 threshold). This segment comes before the guest joins and is Dave describing his corporate IT work.
After a run of Podcasting 2.0 discussion sparked entirely by articles (Tom's piece, Matt's response, the episode that followed), Dave realized he'd forgotten how powerful writing is. Writing a full blog post about the live tag was also therapeutic. It took months of scattered misunderstandings out of his system and wrapped them into one definitive answer.
I realized that I sort of forgotten how powerful writing is.
It struck Dave mid-show: AI chatbots bait you in because they seem brilliant, and then they say something unhinged. Talking to one is like talking to a crazy person who is also a genius.
It just, it hits me right now. Talking to one of these AI agents is, or AI chatbots is like, it's like talking to a crazy person. Genius.
Dave imagines a different history: if live, funding, L402 and micropayments had existed around 2000-2001, we might not have Hulu or Spotify. Instead there might be 40 apps to choose from, all podcasts behind the scenes, where you buy content, send payments and listen to live events. The silos happened partly because the open pieces weren't there.
if some of those things had been in place, micropayments back in 2000, 2001, we may not have things like Hulu, Spotify, you know, these, we may not have those things today.
The radio-dial idea of discovery, where you just spin until you find something, breaks down when the dial is four million ticks long. Starting from scratch, it would take 20 years to find what you want, which is why the entry point into podcasts is almost always a recommendation, not browsing.
to go find the content you want within your app. You know, part of the problem now is the, the dot, this dial is now 4 million ticks long, you know?
Not in the audio. Guest episode (Tom Webster; 0.35 threshold). The cues read as Dave the host responding to Tom (references to Matt, episodes.fm, 'you probably get hate').
Tom says creators want the biggest potential audience. Dave says the killer is the last step: what link do you share? A YouTube creator posts one YouTube link. A podcaster is stuck choosing between episodes.fm, an Apple web link, or a row of app badges. Everything in a closed system goes through one entity, so there's always one link, and open podcasting has no single distribution link.
That's the benefit of the closed system is that it all goes through one entity. And so you always have one link. It's always a link to Twitter, a link to YouTube, a link to Spotify.
Not in the audio. Guest episode (Tom Webster; 0.35 threshold). The cues read as Dave the host responding to Tom (references to Matt, episodes.fm, 'you probably get hate').
Talking with ad-research veteran Tom Webster, Dave compares the advertising industry to tax collecting. The people who benefit complain about how complex and difficult buying is, and the people who actually pay for it with their attention hate the experience. You get hate from both sides.
It's like, it feels like being, um, like a tax collector, like the people benefiting from what you're doing, complain about the difficulty of the process and the complexity of, you know, the buying and all that kind of stuff.
Not in the audio. Guest episode (Tom Webster; 0.35 threshold). The cues read as Dave the host responding to Tom (references to Matt, episodes.fm, 'you probably get hate').
After a critical but thoughtful article from industry analyst Tom Webster, Dave says it's always good to hear different perspectives. Every time someone who isn't steeped in Podcasting 2.0 comes on, it's a learning experience. You can't just navel-gaze in the silo; you have to talk outside the bubble.
I mean, every time we have somebody on who's not just steeped in this stuff, it's it's always a learning experience.
Dave loves RSS for what it gives us but admits it's a pretty terrible spec. Nobody would design it the same way today: it's loose and vague, and half the tags make no sense. All of podcasting sits on the shakiest of specifications, so any improvement is an improvement, and somehow it still works.
And, and the restrictions on, I mean, it just, what I'm trying to say is that all of podcasting is built on the shakiest of, of specifications to begin with. So anything, any improvement is an improvement
Responding to Tom Webster's article saying Podcasting 2.0 has had spotty success, Dave agrees, and adds that spotty success is all most people can hope for. Podcasting is hard and messy.
spotty success is all that most of us can hope for.
Dave wonders whether listeners get antsy when a favorite show (Curry and the Keeper, No Agenda, America This Week) is late because there's so little truly good content: not time-waster content, but something people look forward to that makes a meaningful impact on their lives. Even as a heavy YouTube user, he's as disappointed by its algorithm as by his podcast app.
The, the reason why people get so antsy about something like curry in the keeper, not coming out at, you know, at an expected time or no gender, that kind of thing. It might be just because there is so little good content
Everyone on Podcasting 2.0, the people Dave disagrees with included, is donating time on top of busy lives. So he asks that people give each other a little grace when things don't go the way they think they should.
All of us, everybody, everybody, James included, we're all donating our time. So can we please like give a little bit of, um, of grace to each other when things don't go exactly the way that we think that they should?
Recalling how the World Outline project with Dave Winer ended after he was made to feel small, Dave names his fear for every open source project: if you're mean-spirited and try to make people feel bad about their work, you get people who don't want to participate and just go away. He says he won't leave, but he has to work through it so it doesn't come to that.
because if you if you make people if you are mean spirited and you try to hurt people's feelings or make them feel bad about the work they're doing. That is the result that you're going to get. You're going to get people who don't want to participate
After a hurtful exchange over the image tag, Dave explains that his thick skin on the project is self-defense. If he lets criticism get to him, it hits hard and spills into his life, so sometimes he keeps an emotional and mental distance, which can look like ignoring people. He doesn't mute or block people; he chooses to be open about how things make him feel, because even thick skin has a breaking point.
The reason the reason that I have thick skin, 99 percent of the time on. On this project is because if it's a self-defense mechanism, if I don't, if I let this stuff get to me, it really gets me hard
On image-tag design, Dave isn't surprised it's hard. There are two opposing approaches: define a banner as a fixed metric, or infer the use from metrics. One makes for an easier but error-prone spec; the other is more difficult but more rigid, like strict typing in programming. Either way you fall a little afoul on one side.
There's a tension between those two things and one is much easier, but one is an easier spec, but it's error prone. The other one is more difficult, but it's more rigid
Dave praises the people working through the images tag with him: they stick to the point, get everybody's cards on the table, and keep ego out of it. He also asks listeners to treat his on-air brainstorming as thinking out loud about new ideas, not a decision.
And I want to say this. Stephen and Nathan are so pleasant to talk through this stuff with because it's just like it's sticking to the point and getting everybody's cards out on the table, and there's very little ego that gets involved
Called 'the ignorer in chief', Dave explains that no human can address everything all the time. The mind doesn't have the bandwidth to solve problems on 25 fronts at once, so things have to be ignored, set aside or picked up later. He hates that people sometimes feel ignored, but that's life.
No human being can address everything at all times. I mean, it's just not possible. There's not enough bandwidth to the human mind to solve problems on 25 different fronts at once. Things have to be ignored or set aside or come back to later.
Dave's wife never uses a podcast app; she just finds episodes some other way. He sees why. Browsing Hulu or Prime carries no stress, but podcast apps full of unplayed episodes make you feel you were supposed to do something and didn't. Most people have only two or three shows they never want to miss, and it's easy to just go to those directly. He thinks it may be time to rethink the inbox-style app.
But podcast apps stress you out because they make you feel like, oh, I was supposed to do this thing and I didn't.
Not in the audio. The span starts with one Unknown cue ('But podcast apps stress you out because they') that sits inside Dave's 8-cue runs and clearly carries his sentence.
Dave has always been uncomfortable deprecating namespace tags, and this week showed why. Transistor added Bluesky support to the social interact tag that everyone had called dead. Had it been deprecated three months earlier for lack of adoption, Transistor would have had to build something proprietary with no chance of the cross-app adoption the tag allows.
Yeah. And we're, you know, so a good example of that is like deprecating tags out of the namespace. I've never I've never wanted to do this. I've always been opposed to this.
With people using the Index API for research and AI training, Dave stresses that the API is a full member of Podcasting 2.0. On a technical level, Podcasting 2.0 is the index, the namespace, Podping, and any other spec someone brings that answers: here's what's wrong with podcasting, what can we do to fix it?
Podcasting 2.0 on a technical level is the index, the namespace, pod ping and any other spec that people come along with and say, here's a thing that's wrong with podcasting. What can we do to fix it?
If the future of podcasting is hyperlocal, Dave says it benefits everyone in the actual infrastructure (podcasters, hosting companies, apps), not the ad-driven industrial complex. Hyperlocal explodes the number of people podcasting, which means more hosting accounts and more links out to new apps.
term the future of podcasting in the future. It just sort of started stirring my head. I was trying to keep my mouth shut because you were on a roll. But the future of podcasting, if it's hyperlocal, that is beneficial to everybody.
Approaching decentralizing the podcast directory, Dave warns that you can think you're solving a problem while creating a bigger one down the line. Directories matter (you can't be found without one), and badly done decentralization just creates multiple points of centralization, with competing claims of authority.
I mean, you, we, because if we're not careful, we will do the exact opposite of what we intend.
From Ellul: propaganda doesn't care about convincing the other side. It's about getting people to act, and if you're coming up with arguments you've already failed. Dave sees short-form video platforms built entirely on that principle.
Propaganda is not about convincing people of things. Propaganda, effective propaganda does not care about convincing the other side of your position. Propaganda is about getting people to act. It's about provoking action.
Security headlines often cite scans showing hundreds of thousands of exposed vulnerable servers. Dave notes that many admins block scanners like Shodan, and the internet is full of honeypots faking those services to study exploits. That makes the counts about as useful as IAB download metrics.
Many people block those scanners IP addresses. And also there's tons of honeypots out there that fake being those things just to attract the traffic so they can analyze the exploits.
Dave tells everyone he has ever hired for help desk and tech support the same thing on day one: the very first rule of help desk is that users lie. He said it while troubleshooting a VPN and geolocation problem.
I tell them on day one, the very first rule of help desk, users lie.
Clubhouse, AI-generated podcasts: Dave says the podcast industry gets duped by every new thing that comes down the pipe. It's Lucy and the football, over and over.
But podcasting seems to be prone to this. It seems to be prone to this sort of FOMO, jump on every single bandwagon thing.
Spotify's 2018–2022 podcast bubble showed it: unsustainable financial games always reach a payday. Dave says Silicon Valley's 20 years of financial excess, and now the AI bubble, are headed for the same Dorian Gray–style confrontation with their ugly underbelly.
And we saw this with the podcast bubble that Spotify created from 2018 to 2022. Eventually all the games, the financial games end because they're not sustainable.
Enterprise Copilot told Dave it had created a pie chart in Excel that never existed, twice. His point: the hard things are what we most need help with (complicated pivot tables, booking flights, trip reservations), and those are exactly where these tools fall down.
it's the hard things that we need the most help with things like creating a complicated pivot tables or booking a flight and making trip reservations.
A cartoon character is fine because it's obviously not human. As something gets closer to real, every small flaw becomes more unsettling. Dave says LLMs are the same: being close to accurate is almost worse than being completely wrong, because they tell you the wrong thing so convincingly.
But then the closer something gets to actually looking real human, you begin to note all the little flaws become more pronounced and you become more and more uncomfortable.
After libp2p's gossip support was ripped out of IPFS, leaving everyone to roll their own, Dave says the missing piece for DHT-bootstrapped pub/sub is simplicity: libraries a normal developer can pick up and build something with quickly.
what has been needed in this area, this sort of a DHT backed bootstrapped, um, pub sub arena, what's been needed is simplicity.
On a high-traffic production server, Dave would move per-request bookkeeping, like validating auth tokens, onto an in-memory database such as Redis. That keeps the main database free for the heavy-hitter traffic.
That way you're not hitting your main database for just a little nothing sort of insignificant traffic You're keeping it reserved for the heavy hitter traffic
Dave pleads with web developers to skip server-side rendering with Next.js-style React frameworks. It breaks web extensions like LastPass, and a third-party developer has no clear way to know when the page has finished building itself.
dev, please don't do server-side Rendering with next J with with these react frameworks Please