Contenu de l'épisode
Shownotes
-------------------------------------------------------------------------------------------------------------------------------------
00 - DATABASE SURGERY WEEK — AND THE THIRD EMERGENCY OF THE MONTH
<b>The lead, and it's Dave's:</b> Saturday 22 Aug — "#api I'm in the middle of a <b>big refactor with the Podcast Index database schema</b>. If I do my job right, nobody will notice it. But, the side effect will be an overall <b>huge increase in the consistency of the data</b> (duplicates, url change lag, etc)."
@dave — the big schema refactor (22 Aug)
<b>Wednesday night:</b> "<b>Podcast Index database surgery will begin in the morning.</b>" Thursday: "Step 1. Expect a brief api outage. It should last just a few minutes." — then "Done" — then "#api <b>Adding a very large table index...</b>" — and finally, Thursday evening: "<b>Finished. Butt is intact.</b>" (6 favourites. The board's favourite status update of the month.)
<b>What Adam should ask:</b> what actually changed under the hood, what does "consistency" buy an app developer in practice, and is there anything a host or an app should do differently now that duplicates and url-change lag are being cleaned up?
<b>🔥 The other half of the week — the third emergency.</b> Tuesday 25 Aug: "#api <b>Linode emergency maintenance on our db host. API is down.</b>" Then: "No update as of yet. <b>Just hard down.</b>" Then the line that is really the story: "<b>This is the third 'emergency maintenance' event that has affected us this month. Getting kind of old.</b>" And, dryly: "I guess a reboot every few years isn't too much to ask."
@dave — Linode emergency maintenance, API is down (25 Aug)
<b>Board question:</b> three unplanned outages in one month on the host that carries the Index. Is it time to talk publicly about where the database lives, and what the fallback is? The Index is infrastructure now — other people's products go down when it does.
<b>Podping shipped twice.</b> v0.13.0 across gossip writer, watcher and monitor — "<b>a dynamic memory cap to address the Iroh bug</b>" (that is last week's memory leak, now capped). Then v0.14.0 the same day — "<b>cpu architecture in the mesh announcement messages</b> to make determining what hardware a node is running on easier." Monitor updated too: ppmonitor.podcastindex.org.
@dave — podping v0.13.0, dynamic memory cap for the Iroh bug
-------------------------------------------------------------------------------------------------------------------------------------
04 - 📌 PODCASTING'S GOLDEN OPPORTUNITY: HUMAN SPEECH
<b>Your PREP note, verbatim:</b> "<b>For PC2.0 - Podcasting's Golden opportunity! Human speech!</b>" — filed against a piece on how to spot AI writing.
Can you learn how to detect AI writing? Here are some clues
<b>The argument:</b> text is now cheap to fake, images are cheap to fake, and as of this year so is a voice. What is <b>not</b> cheap to fake is <b>a specific human, speaking unscripted, for three hours, with a transcript and a signature chain attached.</b> Podcasting is the only medium that ships that by default — and the namespace already carries the provenance parts.
<b>🔥 Evidence from this week, and it is stacked:</b>
<b>1. Apple Music is about to start labelling.</b> @js: "<b>more than a third of the music Apple receives monthly is '100% AI'</b> + usage — presumably what other humanoids would call 'listening' — of AI music on Apple Music is <b>below 0.5%</b>. = yes they are going to have to do something about it, starting with labelling." Read that ratio again: a third of the supply, half a percent of the demand.
Billboard — Apple Music to label AI-generated music
<b>2. Nobody can hear the difference in podcasting either.</b> @jaybrunet, 22 Aug: "I listened to a <b>voice clone podcast</b> last night ('Stark Reflections') and it was <b>surprisingly more concise and easier to follow</b> than his previous episodes that tend to meander off in too many directions. <b>If he hadn't said it was ElevenLabs, I wouldn't have cared or noticed the difference.</b>"
<b>3. The scale of the flood, measured.</b> Marco Arment said Overcast <b>transcribes about 30,000 hours of audio per day</b>. @js checked it against his own index: "taking a look at my 'all new episodes published' internal db for last month, <b>the world published at least 1.4m hours of new mp3 episodes for the month, or about 46k hours a day</b>... so if your 30k hours are unduplicated then you have excellent coverage — <b>the long tail is very long and, these days, full of generated audio.</b>"
@js — 1.4m hours of new mp3 in July, full of generated audio
<b>4. The shame is gone.</b> Stanley Druckenmiller, on his WSJ op-ed: "<b>I'm not embarrassed by it... I write everything using AI now for the same reason I use a calculator when I do math problems. I don't know why this is relevant. My name is on the piece. It's my message.</b>"
Notus — Druckenmiller's WSJ op-ed
<b>5. And somebody is already building the counter-stack.</b> Japanese publishers' <b>Originator Profile</b>: "OP uses <b>cryptographic digital signature technology to create tamper-proof credentials for websites</b>. Each ID card confirms <b>who published the content</b>, whether professional organizations vouch for that publisher, and information about the publisher's <b>editorial and ethical standards</b>." — that is a person tag with a signature on it.
Nieman Lab — Japanese publishers and Originator Profile
<b>⭐ The namespace question this asks:</b> we have <b>person</b> (who), <b>location</b> (where), <b>transcript</b> (what was said, word-level, timestamped). That is three quarters of a provenance stack that no AI feed can forge cheaply. <b>Do we need a fourth — a signed claim of human origin — or is the honest move the opposite: an AI-disclosure tag, because the liars won't use either one and the honest shows will?</b>
<b>The V4V angle:&