This page lists part of the collection. The page links below work without JavaScript. With JavaScript, a filter searches everything and pages the matches.
Dave splits learning a language into two parts. Syntax can be picked up in a couple of hours playing in a REPL. The real curve is learning the standard library, which takes time as you keep reaching for the same functions. Most things fall down in the details once you go off the standard playbook.
That is actually, learning the syntax is not that hard. The, what most people find the biggest learning curve is to learn the standard library of whatever the language they're in.
Not in the audio. Guest episode (Barry Luijbregts; 0.35 threshold). In this segment Dave is interviewing Barry about his vibe-coding project (he asks the follow-up questions), so these Dave-labeled runs fit the host.
Since you didn't write vibe-coded software, you won't know in advance what could be wrong with it. Dave suggests a separate suite of checks: point another agent at the running web service and have it try the top web vulnerabilities, such as cross-site scripting and SQL injection.
things that need to happen there for vibe code output where you come along and say, okay, now you use some sort of agent to then test for other things that could be wrong.
Not in the audio. Guest episode (Barry Luijbregts; 0.35 threshold). In this segment Dave is interviewing Barry about his vibe-coding project (he asks the follow-up questions), so these Dave-labeled runs fit the host.
Talking with guest Barry about learning to code, Dave says the difference between taking a class and building something is everything. You can't learn a language you aren't actively coding in for a meaningful project. Real learning happens by osmosis when you're intimate with a project and find yourself in the code all the time.
difference between taking a, taking a class and doing what you're doing now is the, I'm a firm believer that you will, that you cannot learn a language, a programming language that you are not coding in for an active
Not in the audio. Guest episode (Barry Luijbregts; 0.35 threshold). In this segment Dave is interviewing Barry about his vibe-coding project (he asks the follow-up questions), so these Dave-labeled runs fit the host.
Dave explains why the medium tag treats video as a separate medium. Watching a video podcast differs from listening to an audio one as starkly as an audiobook differs from a podcast, or from music. Each engages you at a completely different level.
But the original thinking about declaring video as a separate medium was that video is a different experience. It's the difference between watching a video podcast and listening to an audio podcast is just as stark as the difference between listening to an audio book and a podcast.
In traditional software you set a discrete value in a config file and get a predictable output. With LLMs there are no discrete inputs and outputs, only suggestions of what the output may be, based on a vague idea of the input. So you end up testing over and over and talking to the model ('you are a chapter creator master...').
You have suggestions of what outputs may be based on some vague idea of what the input may be.
Nobody knows whether LLMs give false information 15% or 40% of the time, and it keeps shifting because each retraining gets new things right and old things wrong. Dave sees a software-industry mindset shift: instead of shipping what you believe is bug-free and patching later, everyone knows the output is often bad and says 'we'll fix it later somehow.'
We don't know. And it's probably always changing because every time these models are retrained, they get new things right and old things wrong.
Dave rejects the argument that trillions burned on AI are fine because usable infrastructure will be left over, like dark fiber after the dot-com bust. He calls it the broken window fallacy taken to the largest extreme possible, noting annual AI infrastructure spending equals Singapore's GDP while revenue lags far behind.
So this economic argument goes something like this. Don't worry about the trillions of dollars of economic waste, of setting trillions of dollars on fire, because at the end of it, we'll get a few billion of usable infrastructure. I mean, that is the broken window fallacy
Not in the audio. Guest episode (Daniel J. Lewis; 0.35 threshold). The span includes one Unknown cue ('fire, because at the end of it, we'll') that clearly continues Dave's sentence between his 3-cue and 2-cue runs.
With wallets and LND dropping keysend, Dave accepts the shift to Lightning addresses. What he thinks doesn't matter, only what actually happens, and that's where the universe is moving. He isn't blaming anybody. You adapt. He also doubts payment metadata is even essential: with a PayPal you often just see a short note, and what matters is that someone gave you money.
It just matters what happens. And that's the way the universe is moving. So, I mean, I'm not blaming anybody.
The Index has said from the start that it won't pull content based on what it says. Apps can layer whatever filtering they want on top. The only reason Dave would pull that trigger is platform self-defense: being overwhelmed by a kind of garbage content, like AI-generated shows or Vietnamese poker SEO spam.
what is being said, I mean, we've said this since the very beginning, we're just not going to get involved in that. That's up to the apps. The apps can layer on top of us and decide to show whatever they want.
Dave thinks the AI content flood is partly meant to make us comfortable with it. He compares it to a new acquaintance who unsettles you for no reason you can name. The longer you deny that internal feeling, the more you numb yourself to it, and sometimes you later find out the feeling was right.
And the longer you sort of deny that, that sort of, that internal feeling you have, you numb yourself to it.
Dave's thought experiment: imagine a machine that writes 10 trillion books a day. Within a few years the earth would be covered seven feet deep. That, he says, is what LLMs give us: virtually infinite 'creative' output, which then gets pumped into places like podcast directories.
So that's essentially what we have now with LLMs. We have the ability to create virtually infinite amounts of, quote, creative.
Recalling advice he credits to the author of a famous computer science algorithms book, Dave says proper code takes three writes. The first won't work, the second works but isn't great, and the third is when you fully get it. The code isn't what changes; your understanding of the problem is, until the code maps to the problem domain itself.
three times. The first, the first time is not going to work. The second time is going to work, but it's not great. And the third time is when you fully,
Not in the audio. Dave couldn't remember who said it ('the guy that wrote the famous algorithm book'), so the original source is unattributed. The quote starts right after an Unknown cue ('get proper code, you need to rewrite it') that sets up 'three times'.
Three- or four-hour blocks are fun and productive. But when Dave codes for ten hours straight with only brief breaks, he later finds that mental fatigue he didn't notice made some of that code worthless. It has to be tossed and redone, and he thinks that's true for more people than just him.
later you find out that that long stretch of time, that 10 hours you spent on a single day coding, that some of that code, just because of like mental fatigue that you were not aware of, you
Dave's advice to Adam, who has been caught up in vibe coding: try not to code at night. It's too easy to keep going, and you sleep badly for two reasons. You don't go to bed on time, and when you finally do, your mind keeps turning the problem over and over.
One thing is I, I try, I really do try not to code at night because it's, it's too easy to just keep going. And then you, and then you get, you get terrible sleep for two reasons.
Dave's mornings have a natural stop: he codes on Godcaster for about an hour and a half before work and shuts down no matter what state it's in. He's been late too many times playing it close. Experience says the bug won't get fixed in the last 15 minutes, because the closer you get to a deadline, the harder it is for your mind to focus on the task.
No, it's 15 minutes before I need to leave right now. Experience tells me that this is not going to be fixed. The closer you get to a deadline, the more like, the more difficult your mind finds it to actually focus on the thing in
BOLT11 works and can carry a note, so getting paid isn't the hard part anymore. Dave says a podcaster really wants two things: the money and a note. Verified, attributed, cross-platform payment proofs are product decisions (Fountain's social layer, True Fans) that amount to building a real payment system rather than a way to donate to podcasts.
of the reasons this is so complicated is because these are now product decisions, not really payment delivery issues because the bolt, because bolt 11 works and you can stick a note in it.
Thanking Adam and Todd Cochrane for old stories from podcasting's boot-up days, Dave notes that people who lived through an era forget to tell those stories. To them it was just something that happened, not history worth passing on, so it's valuable when they deliberately recall it.
of your history. And you don't you don't you don't remember to to tell it to other people as if it was history, you know, because you just kind of lived through it and it was a thing.
After a podcaster demanded Apple remove a podcast app for 'distributing' their public show, Dave asks whether hosts, Apple and the community could write one plain, legally worded page. It would explain that publishing a podcast publicly on the internet is like publishing a webpage, that no rights are exchanged, and that if you don't want an app to carry your show you use the block tag. People could simply be pointed to it instead of getting bespoke emails.
in full legal terminology that you, as a, if you publish a podcast publicly on the internet, that it's just like a webpage and you, there are no rights that are exchanged in this
Podcasters often ping 10 or 15 times in a few minutes while fiddling with a feed, and each tiny change (a last-build date, a fixed typo) triggers a full reparse across thousands of feeds. Dave says that's Podping's one disadvantage: sometimes it's better not to know until the creator is actually done.
Right. And so what that's, that's sort of the, if there is any disadvantage to Podping, it is that it's that sometimes you actually, sometimes it's better not to know.
An iTunes author tag over 2,048 characters broke ingestion in code that had run fine for years. Dave takes full blame: for the second time in three weeks (after a missing index on the soundbites table), he'd failed as a diligent aggregator developer by not truncating incoming data. Never trust that it 'has never happened'.
Now, proper handling of incoming data is, I mean, this is purely on me for the second time in about three weeks, I've, I've completely borked in my responsibilities as a pro as, as a diligent, uh, RSS aggregator developer
Watching Microsoft and other giants make moves that make no sense, Dave borrows the Austrian-economics idea of the crack-up boom, where accelerating inflation makes things so expensive that cracks form and everything breaks down. AI feels like a technology version of it: the whole thing fracturing until nobody knows what they're doing anymore.
But this feels like a sort of like a technology version of the crackup boom where
Not in the audio. 'Crack-up boom' is Ludwig von Mises's term. Dave thought it came from Human Action, but the words 'human action' fall in an Unknown cue, so no book entry was created.
People used to work through new problems on forums until someone landed on an answer. Now they ask an LLM, get a partial answer and give up. Since LLMs can only regurgitate what's already been written, Dave predicts new questions will stop getting new answers, leaving knowledge frozen around 2024.
But I think this, this new paradigm, the fact that the answer is so readily available through an LLM result now means that they are going, the net effect is it's going to basically freeze everything in time.
There are roughly a hundred times more local businesses than regional or national ones. Dave says a hyper-local podcast will always find sponsors, like the local TV personality everyone remembers whose ads came from car dealers, and will have a much steadier advertising base than chasing national ads.
There's just, the numbers are just so much larger as far as the total. So if you, if you focus on local, you're always going to have a much more stable advertising base than if you try to
Radio stations have split listeners into two experiences: a 'listen live' button on the website and a separate app for past episodes. With the live item tag, an app can flip between live and on-demand in one place, so the stark distinction disappears.
As you begin to support the new tags, you begin to blur the lines between live and on demand because it all just becomes sort of one thing in the app.
Most of what's on YouTube would never make sense in a podcast app. Dave's definition of quality content is simple: somebody, somewhere actually wants to watch or listen to it. If that number is zero, it isn't quality content.
quality, actual quality content on YouTube, and I always define quality content as somebody somewhere actually wants to watch it. If that number is zero, then that's not quality content or watch or listen.
Not in the audio. Guest episode (John Spurlock; 0.35 threshold). The span includes one Unknown cue ('wants to watch it.') that plainly finishes Dave's sentence between two 2-cue Dave runs.
Podcast hosts can't and shouldn't try to be YouTube at YouTube's scale. Defined correctly, the competition is to give the mass consumer a legitimate place to watch video on demand. Dave thinks podcasting can absolutely do that, and that hosts don't want YouTube scale anyway.
And the competition is, have a legitimate place to serve video to the mass consumer and have it available on demand whenever it's needed. I think that they can absolutely do that.
Not in the audio. Guest episode (John Spurlock; 0.35 threshold). A 7-cue Dave run. The setup line ('they can compete with YouTube though, if you define it correctly') is in an Unknown cue and isn't quoted.
Some Podcasting 2.0 features, like transcripts, are simple enough to live and grow on their own. The value tag is complicated and has had to be pushed. It needs someone almost like an ombudsman to guide normal people into using it, especially as wallet changes made it harder.
And I think V4V has just been, or the value tag has been one of those features that you have to push. You know, you have to get traction. You have to, it's a complicated feature.
Responding to browser makers' plan to drop XSLT (which hosts use to make RSS feeds look nice), Dave says the web has always been content-agnostic. HTML is a first-class citizen because of linking, but it isn't the definition of the web. The web is a universally addressable document location system delivered over HTTP, which is why the Content-Type header exists and browsers already accommodate many document types.
That's why there's such a thing as the content type header in the HTTP response. The web is defined, you can define the web as a universally addressable document location system.
Dave uses Docker much more than he used to, but not for small single-purpose services. For him a full Docker build-out pays off only with continuous deployment: push to GitHub, an Action builds and deploys the container. Otherwise you're manually rebuilding and pushing images, and it's easier to just run the thing directly.
Because it's only worth it to me to go to Docker, a full Docker build out, if you have, if you're going to invest the time to also go like CD, you know, CD, you know, a full pipeline
When Adam says AI coding tools rarely get syntax wrong, Dave agrees but points to the real failure mode: they commonly make up function calls in libraries and APIs that don't exist.
It's very common that it just makes up function calls in libraries and APIs that don't exist.
The money in podcasting is insignificant to Amazon, Google or Spotify. What threatens them is losing your attention. Every walled-garden move (Wondery's shift to Audible, YouTube's push) signals that they feel more threatened and are doubling down. RSS bypasses their whole setup, which is why Google has always been terrified of it.
It is just such a small number. It's irrelevant to their bottom line, but it does represent a threat to their money in a different way. And that's through your attention.
Asking why Google has wanted RSS dead for 20 years, Dave concludes that YouTube, like the AI companies, has realized there isn't enough content. It has ingested the whole internet and still can't keep its promise to be your new TV, where there's always something good to watch. So it wants your content for free.
They want your content for free because what they've realized is is that they've realized the same thing that AI has realized that the AI companies have realized has realized. There's not enough content.
Introducing the podcast:image tag, Dave says they wouldn't be doing it if it merely duplicated something that already existed. The iTunes image tag has a single href that's assumed to be a square image meeting Apple's spec, which people often ignore. The new tag goes further, with aspect ratios, widths and a purpose attribute.
Um, so this, this tag, we would not be doing this if it just simply duplicated something that already existed.
Dave doesn't think video podcasts will succeed as a thing, but he holds two thoughts at once: he wants to see people try, because he may be wrong and it may be a smash hit.
I feel like, I feel like video podcasts are not really going to succeed as a thing, but at the same time, I want to see them try it. Like I just have a, I have a dual sort of have two things in my head at the same time.
Discussing Apple's undisclosed push-notification budget for podcast apps, Dave says it's the same reason he doesn't publish the Podcast Index API's rate limits even though he knows the numbers. Once they're public, everyone runs right up to the limit. He'd go a step further and quietly vary the limits.
really the same reason we don't publish, like I don't publish rate limits for the podcast index API.
Not in the audio. Guest episode (Franco Solerio; 0.35 threshold). The quote is a 5-cue Dave run about his own API. The reason ('everybody tries to run right up to the limit') and the 'change it slightly all the time' idea come from the adjacent Unknown cues and the following 2-cue Dave run, so they're paraphrased in the summary, not quoted.
They once assumed big platforms would never adopt Podcasting 2.0, and Spotify did. Dave admits they tend to be too pessimistic. People surprise you, and he calls himself eternally optimistic that Apple will adopt more tags like funding and live.
eventually, but I mean, like we, we never said like, we tend to just be so pessimistic and people surprise, people will surprise you.
With Spotify adopting Podcasting 2.0 tags, Dave predicts more will follow. Once a company sees the downstream benefits of open, interoperable tags, adding more is such an easy win that it becomes addictive.
seeing the, the, the downstream effect of adding these open interoperable tags, you can't, you, it just becomes addictive. I mean, there's just, there, it's just such easy win.
Podcasting 2.0 spread without real marketing. Hosts just kept mentioning things on air: try a 2.0 app, we have transcripts, we have chapters. Dave cites the saying 'never argue with the man with the microphone'. The microphone is inherently powerful, and podcasting relies on that power.
works is, you know, it's not marketing really, it's just using the power of the microphone, which is inherent in podcasting.
Agreeing that people forget where podcasting features came from, Dave says it's partly because the history isn't written down anywhere, or lives only on some ancient scripting.com URL you have to hunt for. A direct link to the podcast namespace from Apple's podcaster documentation, he argues, wouldn't get lost.
saying. Over time, people forget. But partially because partially people forget stuff because it's not written down anywhere.
Dave finds Rust one of the better languages for LLM coding because it's highly strict. The stricter the language, the more the model is just following syntax, and syntax is what LLMs are good at.
And the stricter the language, the better that LLMs can sort of handle it because it's just following a syntax at that point. But you know, and that's what they're good at. They're good at syntax.