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.
An LLM is like a left hemisphere in a jar. Humans can hold two paradoxical ideas as true at once. AI is better at coding and makes incredible things possible, and it will also ruin your life if you let it. Dave says he's mentally slammed and working more than ever because of it.
You can hold opposing ideas in your mind at the same time, and they are both true. Yeah. And what those two ideas can be is that AI is better at coding.
The old rule still applies with AI: be careful what you ship, because you sign up to maintain it forever. Vibe coding makes it easy to say yes to every request, and soon you have ten threads of vibe-coded projects that never stop. Be very selective.
Don't ever offer to create a service or an app because now you'll never be able to get rid of it. And that's the thing here. You have to be very selective because you vibe code something in a few hours,
Dave repeats a line he's used before: LLMs are less the start of something new than the end of something old. Platforms and technology built up over decades had become so enshittified and held together with chewing gum and wire that something had to sweep them away. It's a clean slate.
AI, LLMs is not the beginning of something new. It's the end of something old.
When growth plateaued, Silicon Valley piled friction into its products to extract more dollars per user. Advertising has done the same over 40 years, slowly ramping up the friction of listening to or watching anything with ads in it.
Advertising – has the same problem that Silicon Valley has. It hit a plateau in growth, and the answer, as we all know now, is they started to inshittify their products.
After an AI agent ported the Podcast Index aggregator from Node to Rust, Dave concluded that whether he'd have written each line the same way is irrelevant. Byte-identical output with a 50% speed gain and 70% less memory is simply better, and the agent built a process roadmap he couldn't have built himself.
if what you end up with is byte identical output. With a 50% speed improvement and a 70% memory reduction, then what I'm saying is the line by line is irrelevant. It really doesn't even matter anymore. It's better. It just is.
Podcast advertising is fundamentally broken, and not in just one way: the ads, the chain, listeners, advertisers and everyone in between. Nobody knows what to do with it.
podcast advertising actually is broken. Like fundamentally, and we all know it. And it's broken in some – it's broken in not just one way. It's broken in many ways.
Discovery of podcasts happens through word of mouth, and that's really the only way it happens. New tools (MCP servers, podroll) are improvements, but improving discovery is like a couch potato taking magnesium: a fraction of a percent. The thing that's actually broken is podcast advertising.
Discovery never really was broken. You've been saying this for many years. The discovery was never broken. Discovery of podcasts happens through word of mouth.
Talking about adding X402 payments to an MCP server, Dave advises moving as fast as possible. With anything new that depends on discovery, getting something running early gets you on the list, so you're one of the first results when someone queries for content that supports it.
anytime there's some kind of new thing that's discovery dependent, just getting something up and running as fast as possible is usually very advantageous because you get on the list.
Like 2x playback, local AI that strips ads or reshapes episodes happens in a private experience the creator can't control. If you let your content leave the barn, you've lost control. What stopped this before was the inability to turn unstructured audio into structured data, and that ability is now commoditized and headed into every phone.
once that ability is out there, you have to realize as a content creator, podcaster, there's nothing you can do about it. Because it's all a private experience you have no control over. Correct. If you allow your content to leave the barn – You've lost control.
The big misunderstanding in the app-signing debate was the fear that signing gets you blocked. The point is the opposite: if you sign your requests, servers know who you are and won't block you. The incentive not to block a legitimate app is stronger than the incentive to block one.
if you sign your stuff, we'll know how we'll know you're that you are who you are and we'll know not to block you. And so, like, I think the incentive to. The incentive to not block a legitimate app is stronger than the incentive to block a legitimate app.
You can't fix ten things at once, so make a list. Filter by what's realistic to fix at all, then start with the issue most likely to get a favorable ear from the other party, not the one that hurts most. Solving the agreeable one first builds the consensus you need to tackle the next hardest thing.
the list of things that have to be fixed always come down into two camps. It's like, is it realistic this will ever be fixed? Well, okay, once you get through that filter, which one is the most likely to be fixed? Not which one is the thing that hurts the most,
Not in the audio. Guest episode (Sam Sethi). The 'build consensus over time' conclusion comes from later cues in the same Dave passage, several of which voice matching labeled Unknown.
A 40- or 50-line SQL statement with multiple subselects becomes really hard to understand, which is how Dave's aggregator feed-selection query went wrong (an ORDER BY on an unindexed column). At that size you need help reading it, whether an agent or at least a syntax parser.
have like a 40 or 50 line long SQL statement with multiple subselects, this kind of thing, it really gets difficult to understand what's happening.
Not in the audio. The 'you really need an agent, or at least a syntax parser' line is in an Unknown-labeled cue just before this, so it's in the summary as context, not as part of the quote.
When Adam reported No Agenda missing from Index search, Dave's first thought was 'nothing I did, Curry probably screwed it up.' Experience taught him not to say that out loud: every time you think it's someone else's problem, it ends up being yours. So he started with the most likely cause, his own recent schema change.
But that's not what comes out of your mouth because you learn over time that every time you think that, it ends up being your problem.
Not in the audio. Voice matching labeled the first two cues and the last cue of this quote Unknown, and only the middle cue Dave. It's attributed to Dave because the surrounding Dave cues are his first-person account of debugging the Podcast Index search ('my first thought was...', 'the podcast identity layer schema change thing that I've been doing'). Guest episode (Sam Sethi present).
Comparing podcast ad loads to broadcast doesn't make podcasts look good. Radio and TV run 70% ad load, and that doesn't make 50% acceptable. Some ad-heavy podcasts are almost unlistenable. Dave admits he's been spoiled by donation-supported shows.
radio and television has a 70% ad load doesn't mean that podcasting is great because it's only 50%.
A 'podcast' is just a concept inferred from the data in a feed, which creates tension. If a show changes its name (The Incomparable became The Incomparable Mothership), is it the same show? Sometimes yes; sometimes a feed is reused for a different show altogether. Keying only on the data inside the feed makes those calls very hard.
Since a podcast, excuse me, is just a concept we infer... From the data that's inside of a feed. What you end up is sometimes you're hamstrung.
Dave doesn't think coders will necessarily lose their jobs. The flood of new AI-built software will need an army of IT people and site reliability engineers to keep it humming.
it's definitely going to require an army of IT people and site reliability and SRE site reliability engineers to keep all this new flood of software going. humming.
Dave loves and hates technology. The more you integrate into your life (home automation and the rest), the more you suffer death by a thousand paper cuts. When you're young the dopamine lets you overlook the annoyances; as life and responsibilities grow, those small frictions become real pain. Generally there's too much technology in everything.
I love technology. But I also hate technology at the same time because it's so – the more you integrate it into your life, the more you start to have death by a thousand paper cuts.
Dave isn't writing the Index refactor code himself, but he reviews it because he doesn't trust it, and that has caught critical bugs more than once. When something gives him pause, he has the agent build a specific test suite. Even if you're not a great coder, understanding system architecture gives you a sense that something isn't right, which someone without that experience wouldn't feel.
understand system architecture and the way things are supposed to operate. you can properly handle these things. Because something gives you a spidey sense that, eh, I don't think this feels, something about this doesn't feel correct.
Reading Wendell Berry ('all art that rises above competence insists upon the irreducibility of its subjects': a fine cabinetmaker sees every board as unique), Dave concedes Adam was right that code can be art. Mostly it's merely competent; it becomes art only in the hands of a true code artist, and there are very few of those.
I think what the real way to say it would be code can be art in the hands of... A true artist. true code artist. And there's so few people like that.
Like LLM context windows, which retain the first prompt and the most recent one but lose the middle, people on a long project keep the big picture and the last thing they did, but the many layers in between fade in and out of clarity. You can recover that clarity, but it costs a lot of time and attention.
it's sort of like the lost in the middle problem. With context, the very first prompt you ask tends to be retained, and then the most recent question you ask tends to be retained. But a bunch of stuff in that middle tends to get lost. I think that applies with human effort.
Tech companies hand you solutions that ratchet down what people know and leave them dependent. To avoid that, when Dave learned a new language he went back to the original books, like K&R for C or Sun's original Java docs, learned the base level, and moved forward through history. Without the foundations you're driven by whatever the latest feature sprint is, never understanding why things are the way they are.
if ever I was learning a new programming language or something like that, I would always go back – And by the – I would try to find the original books of whatever that thing is, like C.
Technical users stress the 'artificial': it isn't genuine intelligence, so they take everything it says with a grain of salt. Non-technical corporate users tend to stress 'intelligence' and trust it. That gap in emphasis explains a lot of how AI is misused.
The average person... In corporate America. Misunderstands the emphasis of the term artificial intelligence. You can take this thing in two different ways.
Before sharing the Podcast Index API internals, Dave went through and commented all the code, because to him code without comments is useless to anyone trying to see what's going on.
something I cleaned up and because to me, this stuff is useless without comments. Right. And so I had to go through and just like comment all the code
Drawing on McGilchrist, Dave argues that LLMs are a standalone left hemisphere: they build a picture of the world from parts and have no awareness of the whole. Let one run the show and the outcomes are bad. A human has to play the right hemisphere, the overseer who sees the whole system and pushes back, which is why Dave watches agents' reasoning chains as they work.
What we've essentially built is a left hemisphere. That stands on its own without the guardianship of the right hemisphere of the right hemisphere.
Looking past lightning to alternatives like x402, the funding tag or stablecoins, Dave restates the principle he and Adam held from the start: the app has to get paid. If a new system pays only creators, hosting companies and ad agencies, it's a failure.
there's alternatives beyond lightning and – They seem to be all easy to adopt. We just have to make sure, which has been our thing from the very beginning, we have to make sure that the app gets a cut.
Today's AI ad-skipping apps are crude and cut out sections that aren't ads, but first versions always are and these tools will improve. Within a few years every app will skip ads, so baked-in ads won't survive and podcasters need other ways to get paid.
a few years, ad skipping is just going to be the way this stuff works, man. Everything is going to do it. everything's gonna do your baked in ad sorry it's not gonna survive it's just not
Newspapers and magazines had subscriptions and classifieds on top of ads. The only media ever supported entirely by advertising were linear broadcast radio and TV, and that worked because there are only so many hours in a day, which made the inventory scarce. Podcasting doesn't have that constraint.
The only two things that were ever fully supported by advertising only were broadcast, linear broadcast radio and television. And those, like you said, the inventory constraint, the scarcity of the number of hours in a day.
Dave suspects YouTube makes far more from premium subscribers than from advertising. Annoying ads are the main driver for signing up, so for these companies advertising is mostly a lever to get you paying monthly. Spotify letting paying subscribers skip ads follows the same logic.
Advertising is only useful to these sorts of companies as a way to drive you to pay them cash on a monthly basis for premium subscription.
Shutting down the Podcast Index lightning node, Dave reflects that the vision he and Adam had (any podcaster could just turn on self-sovereign, low-fee payments) didn't match how the Bitcoin lightning ecosystem saw lightning. In practice it works well only for highly technical people with plenty of money and time.
And that's not a technical problem. That's a worldview problem.
Apple doesn't adopt specs just because others wrote them. Apple Podcasts implemented Podcasting 2.0 tags because they made technical sense, matched things Apple wanted to do, and grew out of relationships built over time in the same rooms at conferences.
they made a lot of technical sense and it matched up with some things they wanted to do. There was a synchronicity there that happened over time.
On the AMP ad spec, written privately under NDAs: Dave understands that secrecy may have been the only way to get the big players into one room. But decisions made inside the secret bubble never get outside comment. Open review cuts short a lot of wasted effort and mediocre ideas just by letting more people weigh in, and doing it in secret kneecaps otherwise good efforts.
When you do it in secret, you don't have the insight that communities bring. And you don't, you can. You can short circuit a lot of wasted effort and a lot of Mediocre ideas. by just having more people give their opinion.
Talking with Adam after John C. Dvorak's death, Dave notes that when someone in your orbit dies (a spouse, business partner or longtime friend), immediate real-world duties take over. You set your own grief aside, and only later do you really start to feel it.
and so you tend to put your own grief to the side for a little while and then it's only later that you really begin to experience that.
Not in the audio. The context (Dvorak's passing) comes from the episode's discussion. The words 'later that you really begin to experience that.' are in an Unknown-labeled cue (00:10:00). They're kept because they plainly finish Dave's sentence from the Dave run just before.
Agreeing with a listener's point, Dave says platforms don't build communities. Content does, or really the humans behind it, which is why a robot with text-to-speech can't build one.
right, Eric. Platforms don't build communities, but content does. I mean, like, it's humans build communities.
When someone says what's 'appropriate' is purely subjective, Dave agrees but notes that you can average subjective reactions across many people. Step back and you'll see patterns: things most people are turned off by, such as synthetic voices pretending to be real on shows about ethics, family and politics.
But if you step back and look at things as a whole. You'll see that there are patterns and there are things that generally most people are turned off by.
To Adam, who has had great success building with AI: that success can stand on its own. It doesn't mean AI-generated content is fine everywhere else, such as synthetic voices in places where it isn't appropriate.
And what I'm saying is... That success can just stand on its own. The success that you've had, which has been incredible. does not have to extend into the other uses of it.
What drove podcasting to critical mass wasn't the technology, since RSS could have been anything. It was the humanity: anyone could fire up a microphone, and listeners hear a real person on the other side and feel they know them, like a radio DJ in 1987.
I think that the thing that made podcasting what it is. I don't think it was the technology. The RSS.
Dave thinks people get such different results from coding agents because of what they already know. If you understand how computers, networks and software are structured, even without being a coder, agents feel natural. True non-coders who vibe code often end up with garbage, or with something that looks fine on the surface and isn't underneath.
great experiences with them in general. And it's because... If you already know... how to talk, or excuse me, if you already know how things are structured, and you do
Dave expects AI to last in verifiable uses like coding and creative work, where these tools really shine. Attestation and knowledge work are different: knowledge work assumes knowledge, and these models don't have any.
work. Man, that is never going to be reliable. It's just not. Because knowledge work assumes knowledge, and there's not any.
Not in the audio. The opening words 'work. Man, that is never going to be reliable. It's' are in an Unknown-labeled cue. They're kept because they plainly continue Dave's sentence ('attestation work? and knowledge work') from the surrounding Dave run.
People who've had great success with AI get defensive about criticism, and people buried in corporate AI slop get defensive against the success stories. Dave argues that the binary is unhealthy for any technology. In ten years AI will fade into ordinary tooling, and the verifiable uses like coding and creative work will be what remains.
And when you have incredible success with it. It makes you defensive. about negativity towards it. And then when you have nothing but negative experiences with it. it makes you defensive against the people who are having success. It's like it's got to be binary.
People treat AI as a magic box because it's confidently wrong in fluent language. Until now, humanity had never met anything that could talk except another human, and human speech is mostly internally consistent. AI can contradict itself mid-sentence while staying perfectly grammatical, and our brains aren't built to catch that.
human race has never encountered. has never encountered anything that can talk except another human. And when humans talk to each other, We We are. mostly internally consistent with our thought processes.