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.
Extracting dates, invoice numbers, vendors and amounts from 1,500 pages of invoices in different formats, Dave found that feeding everything straight to an LLM was a disaster. What worked: OCR with existing software (Apache Tika), pre-process the text into a predictable format, and only then hand it to a well-prompted LLM for one very specific job. They tried about ten models before one fit.
I think the learning curve in all of this stuff is, is, is learning when to bring in the LLM. To do a specific, a very specific thing.
When Adam says AI is taking people's jobs, Dave relays a conversation with someone whose company's layoffs had nothing to do with AI and everything to do with a stock price at a third of last year's. The company mandates AI because its backlog of tickets and feature requests is endless, and for the first time agents are letting staff make progress on it.
That has nothing to do with AI and everything to do with their stock prices. One third what it was a year ago.
A hosting company saw spiky downloads on one show and found it was using a traffic-generation company that 'boosts Apple rankings' by paying mobile-game players (with loot crates) to follow the show. It's a temporary, non-organic bump from people who never intend to listen.
These are people that never intend to listen to your podcast. They don't give a crap about your podcast. They are just there to get their next loot crate in their mobile game or whatever this is.
Dave explains why he predicted a 2026 flood of AI podcasts. IAB specs can't filter out all download noise, and it's no coincidence that DAI-supported AI podcasts rose alongside a huge jump in browser downloads. AI operators exploit hosting free tiers for ad revenue on non-human downloads. Hosts have to get in front of it, or advertisers will drop DAI, and free tiers (which only work ad-supported) will vanish.
It can't be a coincidence that at the same time we've seen the rise of AI, of, of dynamic ad insert, insertion supported AI podcasts. We've also seen a huge increase in the number of browser downloads.
In the AI-podcast debate, Dave partly agrees with Adam: some content, like traffic and weather, is very automatable. The 90s radio guy doing traffic and weather on the tens might as well have been a tape, and you'd never know. There's no inherent problem with automating information delivery in a verbal way.
you hear some dude on the radio giving you traffic and weather on the tens. Well, you, that guy might as well be automated on a, on a tape. You have no idea.
No one has implemented an AI-content label because defining the content is impossible. So many podcasts use AI somewhere that 'made with AI' says nothing. Like the endless debates over a more granular explicit tag, such labels really exist only for filtering. Still, if people in the industry think a label is necessary, Dave will support it.
let's define at least what the problem is because defining the defining the content is the hard is impossible. This is, this is why nothing has happened until, until this point.
Podcasting 2.0 is a collection of people doing what interests them, like Valve's famously flat structure. Sometimes things get dropped when people lose interest, but with an all-volunteer group that's how it has to work. Dave has stepped back from the value/Bitcoin work because he doesn't know the answers and would rather watch people build than argue.
I mean, there's no, it's all volunteer and you only, you can only expect volunteers to work on the things that they're passionate on, passionate about.
Podping runs on the Hive blockchain, but a blockchain is just data on servers somewhere and can fail or fork, however distributed it is. Dave and Alex's plan is to leave Hive as is and add a second redundant layer: a peer-to-peer gossip system bootstrapped over DHT, like BitTorrent's.
Because blockchain, you know, blockchains are just data that exists in a server somewhere. And so what the idea that Alex and I've been kicking around for a couple of years now is having a second layer.
Radio stations want all their content in the station's own app, but siloing doesn't work. 'Everywhere you get your podcasts' is powerful because podcasting collectively is one giant radio receiver: listeners use the app style that fits them. Dave remembers early-2000s radio shows each locking streams into their own website or desktop program.
That sort of siloing does not work well. You have to be available everywhere. That's what makes that phrase, everywhere you get your podcast so powerful is because collectively you can see podcasting as one giant radio receiver.
Radio now really only exists in the car. As car makers stop seeing it as necessary and it slowly erodes from dashboards, it won't matter whether a station has a transmitter; it'll be transmitting to nobody. Dave sees this as less a demographic shift than a slow removal.
exists anymore. And if it slowly erodes out of the car, it doesn't matter if you have a transmitter or not, you're transmitting to nobody.
Asked to define AI slop, Dave borrows the gardener's definition: a weed is just a plant in a place you don't want it. Slop is defined by context and unwantedness, not by any intrinsic property, which is why it resists a technical definition.
Okay, slop is the same as, defining slop is like defining what a weed is. A weed is just a plant that's in a place you don't want it.
Dave quotes a line from his reading: mechanical efficiency always externalizes its inefficiencies, the way a car pushes heat and exhaust out the tailpipe. Applied to AI coding, the inefficiency gets pushed onto the human: the mental load, mania and exhaustion of overseeing it.
And so I think that the payload that comes along with AI coding is the AI's inefficiency is us. We are the tailpipe.
Not in the audio. The book and author of 'mechanical efficiency always externalizes its inefficiencies' aren't named in this passage (Dave says only 'he makes this one comment' in something he read last night).
If the Index gets swamped with junk podcasts, it has to play self-defense to keep it from costing real money. 'Storage is cheap' is true of object storage, not block storage. Junk feeds don't sit somewhere for a fraction of a cent; they go into the database, which only ever expands in CPU cycles and storage.
Block storage is disk storage is expensive. And these are, remember, you got to remember these are these, these podcasts, these junk podcasts are going into the database.
Dave hates solving a problem just for himself with a private rig nobody else can use. If he has the problem, others have it too, so he always wants a bigger, open, public solution that other people can use themselves.
I hate solving problems for me in a one-off way and not like it. I always want it to be like a bigger open public solution to the problem that
When a show strips chapters, the obvious fix is a second feed that points at the original enclosures and adds your own chapters and transcripts, a FeedBurner-style proxy. Dave calls it a bad solution: that feed gets out in the wild, people subscribe to it, and it becomes a mess. He'd rather have a private way to amend a feed.
in the wild. Now there's going to be another feed that people are subscribing to. It's just going to be a mess and you don't want to get into that.
Not in the audio. Parts of the reasoning ('your own chapters and transcripts', 'that feed is going to get out') are in Unknown-labeled cues and are summarized, not quoted.
Dave thinks the future of AI is local. Hardware is moving from a CPU plus a discrete GPU on PCIe to a system-on-chip with unified memory shared by CPU and GPU/NPU, as in Apple silicon, new Ryzen parts and Nvidia's DGX Spark. Expect machines with 128 GB of unified memory capable of running strong local models.
a discrete GPU connected by PCIe. It's moving to unified memory with an, with an APU or NPU built on the package.
Not in the audio. The setup ('system on a chip with') is in an Unknown cue just before. Adam's 'Like Apple stuff, you mean?' (Unknown cue) right after shows the other host is asking, so this speaker is Dave.
Dave uses AI to spin up new products that otherwise wouldn't exist, but he's afraid to let it touch important existing code bases. He hadn't used any AI on the Podcast Index API server itself.
I think, like, I'm afraid to let it touch existing code bases that are important. Like, I've not used any AI on the podcast index API backend.
Not in the audio. Voice labels in this episode put some of Adam's speech in Dave runs (e.g. the 00:18:45 app-building story and around 00:44). This passage is confirmed as Dave's by content: it's his Podcast Index API backend.
Getting briefly spiritual, Dave says he feels God is doing something important in the lives of everyone involved in Podcasting 2.0. Time is a factor he's afraid to tamper with, because spiritual truths often need time to manifest in our minds and can't be short-circuited.
I feel like spiritual truths a lot of times need time to manifest themselves in our minds that you can't short circuit.
Dave values Podcasting 2.0 as a rare pure-play open source project with no corporate ownership, one he'd keep running himself if the money disappeared. When donations cover hosting but you aren't living off them, your timeframes get longer. A constant stream of support money makes everything feel urgent and pressures you to drop what you're doing.
But when you have a truly open source project where you're getting money and it's meeting your needs, you know, for the hosting and that kind of thing, it's paying the bills, but you're not, you know, you're not living off of it. Well, then your timeframes get longer.
When aggregator troubles prompt talk of raising more money, Dave pushes back: the issue isn't money, it's that the thing isn't built right. For technical engineering problems, a big cash infusion (startup culture) just pressures you to accelerate solutions instead of building them properly.
I think my exact words were, it's not a money problem, it's an engineering problem.
Not in the audio. 'an engineering problem.' is in an Unknown-labeled cue. It's kept because it plainly completes Dave's sentence from the 9-cue Dave run.
Reviewing Podverse's new schema with a separate table for every Podcasting 2.0 tag type, Dave flags the trade-off. All the linkage across tables means updates and deletes in one table ripple outward, causing more database locking. Proper indexing alone may not solve it.
The one, um, the one drawback to having separate tables per sort of like a item type is you tend to have more locking, you know, database locking as things cause, cause that updating or deleting content in one table since you have just tons of linkage across tables
Not in the audio. Guest episode (Mitch Downey and Archie of Podverse; 0.35 voice threshold). Dave is the one questioning their schema here; the 7-cue Dave run asks them 'Have you seen any of those kinds of' issues.
Comparing it to strict email filtering, where mail without good hygiene just doesn't get through, Dave says the Index won't block any particular person. Instead it will keep adjusting the knobs and sliders until the filters are tight enough that garbage, spam and slop content simply can't get in.
We're not going to block any particular person, but we're going to have to jiggle the knobs and like jiggle the knobs, move the sliders around to make the filters so tight that if you have garbage content, spam slop content, that it just can't get in.
Not in the audio. Guest episode (Alecks Gates; 0.35 voice threshold). The 6-cue Dave run responds to Adam's own email-filtering story, so it's the other host, not the guest.
Dave believes there's a legitimate core to people's discomfort with AI that shouldn't be dismissed, even if no one can fully explain it yet. Technology does this all the time: it makes us uncomfortable, and sometimes it takes years to see why, as with the phone in our pocket. Not being able to explain it right now doesn't mean it isn't there.
just because you can't explain why right this second doesn't mean that it's not there.
Dave highlights Cory Doctorow's University of Washington speech. A 'centaur' is a human in control driving a machine; AI flips that into a reverse centaur, a human walking around with a machine head in charge. It explains his discomfort with code he generates but doesn't understand. You can learn something from everybody, even people you don't always agree with.
And what he's saying is a centaur is something that is like a, a centaur in technology terms would be like a human that's driving some sort of machine.
Every argument against AI that Dave tries (jobs, web garbage) he can partly talk himself out of. What he sees in each is a problem that existed before AI, which AI is simply accelerating: 'a garbage accelerant.'
for all these arguments, you kind of already see that, that there was something there that pre-existed the AI that was already a problem and the AI is just accelerating it. It's, it's an accelerant.
Dave loves having an agent do boilerplate drudgery but despises not understanding what just happened. With Godcaster and Podcast Index code he may not remember every line, but he always has a mental map of where everything is and can get there in ten seconds. AI-generated code feels like wearing glasses someone smeared Vaseline on.
have it do a bunch of the boilerplate sort of drudgery code for me, but I, I also despise the feeling that I don't understand what, what I just did. Like, I feel so mentally disconnected from the code.
The current aggregators are fast at the cost of sometimes bringing the database to its knees. Only their underpowered VMs keep it from locking up. Moving to a queue feeds the database at a predictable rate, which means a little slower (minutes instead of seconds), and Dave thinks that trade-off is worth it.
So the queue thing is really good for being able to just feed the database with a, at a predictable rate, but predictable is also a synonym for slow.
Dave is fine with the Podcast Index's technical debt because it isn't the kind that bites you later and leaves you stuck. It just means things aren't as efficient as they could be, so scaling up means throwing money at it. Because money is the one thing the Index can't throw, that's why he's rebuilding the aggregator pipeline.
and I'm, I've been okay with our technical debt because it's not the type of debt that comes back and sort of like bites you in the rear in the future in a way where like you, you get stuck
Faced with duplicate shows and conflicting redirects, Dave states the fact everything else follows from: without clear intent from the podcast owner, it's impossible to know which feed URL is correct. Apple's only real advantage is that owners update their feed URL there. His proposed answer is to put the owner's attestation in a community-managed place everyone can hook into.
If you don't have a clear intent from the podcast creator, you will not know ever whether the feed URL you have is the right one.
The Index was started to decouple podcasting from Apple's control, and Dave still holds that Apple's directory shouldn't be authoritative (it has bad data too). But he says you can't be an idealist or utopian dreamer. You have to admit that Apple gets first-hand signals from owners that the Index doesn't.
we wanted to decouple podcasting from Apple's control. But again, I don't think you can be an idealist or some sort of utopian, you
After finally paying for bigger VMs to fix memory crashes, Dave warns himself and listeners not to make spending a habit. If you get used to it, you'll overspend like someone undisciplined with a credit card, and you'll regret it. He brings that discipline everywhere, his day job included, because cloud hosting providers want all of your money.
Do not become comfortable with solving problems with money. Do not let that become a thing that
Dave reads Apple Podcasts as a product Apple employees and executives build for themselves. They listen to and love podcasts, they use iPhones every day, and they know it won't make money. It would feel like less of a device without a top-notch podcast player, so they develop features they themselves want.
want, and they also are obviously iPhone users and they want their product that they use every day to be better, to have a top notch podcast player built into it. And they know it's not going to make a, they know it's not gonna make money
Responding to claims that Spotify is pushing Apple Podcasts, Dave disagrees. Apple Podcasts could have added nothing for 20 years and still lived on in the Apple fortress. What changed is a renewed vigor in RSS itself. Apple now checks whether an open community namespace already exists for a feature and joins it, so it stays compatible with what the community is doing.
competition from Spotify. I don't think that I do not credit Spotify with pushing Apple podcasts to make a bunch of changes or make or add a bunch of new features. I credit a reinvigoration of RSS.
Dave notes that at temperature zero, with no randomness added, an LLM is deterministic and gives the same output every time. Nobody runs it that way, he says, because that would drop the illusion. The 'magic' feeling depends on the randomness, which is smoke and mirrors.
Deterministic, like if you have the temperature at zero and you don't introduce any randomness, it should literally give you the same output every time. But nobody does it that way because it drops the illusion.
Before the value-for-value discussion with guest Chad F, Dave reflects that it's not inappropriate to hold convictions even when other people don't share them. His own: Bitcoin is a modern hard asset that shares gold's store-of-value role, but is far easier to get into and out of.
a lot about this and I feel like it's not inappropriate to have sort of convictions about things. And even if nobody, even if other people don't share those convictions.
Not in the audio. Guest episode (Chad F; 0.35 threshold). The 7-cue Dave run follows Adam's 'To the boardroom?' prompt and continues the speaker's own setup ('I knew we're gonna be talking to Chad today'), so it isn't the guest. Nearby macro-economics passages (Goldman Sachs, government spending) are also labeled Dave but were skipped as uncertain.
Not everything has to make the number go up. Dave says sometimes you build something because it's beautiful, meaningful or useful. If the world picks it up, great; if not, it's still great. One of the most fun things Podcasting 2.0 ever did was a live event at a Bitcoin conference with maybe a couple thousand participants. The numbers were tiny, but it was so much fun.
just think that we, sometimes you build something, you some, sometimes you build something just because it's beautiful or just because it's meaningful or just because it's useful.
Dave says capitalism is fine: everything else just sucks worse. That doesn't make it perfect. Its cost is a perpetual growth-at-all-costs mindset that elevates materialist, outcome-based thinking above everything, and the West applies it to every part of life, like an infection. He says the same mindset kills momentum in projects.
Everything else, capitalism is fine. It's just that everything else sucks worse. It doesn't mean that it's actually perfect in any way, shape, form, or fashion.
Discussing a piece on how people knew the earth through poetry and philosophy (quoting Dante) long before photographs, Dave says the West has a physicalist mindset: if it isn't material, it isn't real. Science is one part of human experience. We make ourselves spiritually poorer by raising any single part (economics, philosophy, the hard sciences) above everything else.
It's not every, it's not all of it. And we, we make ourselves poorer, you know, spiritually by taking one aspect of human life and putting it over all else, whether that
Faced with slow transcript stats queries on a 7.5-million-row table, Adam asks whether money can fix it. Dave says no. Scaling up the VM would only delay the problem. The right fix is for the developer to accept that elegance isn't what matters most, and to denormalize by duplicating the feed ID onto the transcripts table so it works fast for every real use case.
The right solution is to, as a developer, get over yourself and say, it's not, elegance is not the, is not the most important
Sticking around looks like life, and real lives are messy. Contributors go through busy phases, so an open source project should expect bursts of activity around a feature (like the images tag) followed by quiet. Dave says that's fine. Pushing the pedal to the floor and expecting constant output just makes people tired, and they leave.
And so you have to expect that there's ebbs and flows in the activity of an open source project. And that's, that's fine. That's to be expected.