The Layer2C Labs Podcast Is Live. It Only Exists Because of the AI Voice.
There’s a new show. Layer2C Labs takes each validation lab I run and turns it into about ten minutes of audio. Three episodes are up.
Every episode opens with the same sentence, before anything else:
Written by Keith Townsend. Narrated by an AI, because Keith doesn’t work from scripts, and this voice isn’t pretending to be him.
Two questions follow from that, and they’re the ones I’d ask.
Why not just read it yourself?
Because I’d be bad at it, in a specific and boring way.
I don’t work from scripts. Everything you’ve ever heard me say out loud was unscripted: a conversation, a hallway argument, a live reaction to something a vendor just claimed. Hand me a page and the delivery goes flat and careful. It stops sounding like me somewhere in the second paragraph.
That would be survivable for an opinion piece. It isn’t survivable here, because a lab podcast lives on precision. Nine point one eight cents a verified fix. Seventeen of twenty-two. Gemma four twenty-six-B. Those numbers have to land exactly right, every time, and reading numbers accurately off a page is not something I’m organically good at. My strength is the unscripted argument. This format needs the opposite.
Why not clone your voice, then?
Because it would sound like me reading from a script, which defeats the entire purpose.
A voice trained on my public audio would carry my cadence over content I never spoke. It would sound like a recording of me talking, when it’s actually me writing. And the tells that normally separate those two, the hesitation, the mid-sentence correction, the moment I decide to go somewhere else, would be gone.
So I picked a stock voice that sounds nothing like me. You can’t mistake it for a recording of me, which means the disclosure barely has to work. Your ears do it.
The honest version of who wrote it
A model drafts the spoken script from the lab’s structured record. Then I spend about an hour on it: rewriting, cutting, re-recording lines that sound wrong, until it says what the lab actually found. So “written by Keith Townsend” is true in the way a bylined piece is true. The findings, the verdict, and the argument were mine before any model touched them, because I produced them on hardware I ran. The first draft of the spoken form wasn’t.
That’s the seam. I’d rather show it than blur it.
Between the draft and my hour sits a deterministic gate: number fidelity checked against the source record, plus a speakability pass that catches what reads fine and sounds wrong. Markdown artifacts. Slashes. Unit abbreviations. “2x”-style multipliers. Bare URLs. The gate is report-only and explicitly necessary but not sufficient. Nothing renders or gets spent until I’ve read the whole thing.
Same claim I made about deterministic AI: let the model work inside bounded, checkable steps, and let a test decide when a step is done. The judgment never moves.
This is content that wouldn’t exist otherwise
The AI voice here isn’t a shortcut on work I’d have done anyway. It’s the reason the work happens at all. Lab 001 is about borrowing a vendor’s plumbing without borrowing its judgment. Lab 002 is about owning the weights. Lab 003 is where the validator, not the loop, decides done. Each is a long read with a chart and a cost table, and the people who most need the finding have the least time to sit with it.
Booking studio time per lab, on top of running the labs, was never going to happen. The real choice was a synthetic read or no audio. I picked the synthetic read and I say so in sentence one of every episode.
I think that’s the more interesting use of this technology than the one everybody demos. Not doing your existing work faster. Doing work that wasn’t viable before.
What this show is, in DAPM terms
I built a model for exactly this question. Every decision in a system is Retained, Delegated, or Ceded. I’ve been asking CTOs to map their AI systems that way, so here’s mine.
Retained. What the lab found. The verdict. Whether a script ships. Which voice reads it. None of that moves.
Delegated. The first draft of the spoken form, to a model. Number fidelity and speakability, to a deterministic gate. The read itself, to a synthetic narrator. Delegated is not ceded: each one comes back to me before anything renders or publishes.
Ceded. Whether Apple and Spotify carry the show. That decision was never mine and never will be, which is precisely why the feed is.
Mapping it that way is what made the last rule below easy to categorize, and easy to be honest about.
The rules I’m holding myself to
Three of these are editorial. You have a stake in them, and you should hold me to them.
The findings are always mine. The moment a model generates findings rather than phrasing, this stops being disclosure and starts being laundering. That line doesn’t move.
The disclosure is the first thing you hear. Not a line in the feed description where nobody looks. Sentence one, every episode.
I read every script and listen to every file before it merges. If it lands wrong, it doesn’t ship.
The fourth one is a business rule, and I’m naming it as such because you shouldn’t have to care about it.
The feed is mine. Same design as the main podcast: I own the RSS at labs.layer2c.com/podcast.xml and the media host is swappable. A subscriber list living inside a platform is that platform’s asset, not yours. That protects me, not you. It only becomes your problem the day a platform decides my show is no longer worth carrying.
Where to get it
Apple Podcasts · Spotify · or point any player at labs.layer2c.com/podcast.xml.
This doesn’t replace The CTO Advisor podcast, which is still unscripted conversations with real people and always will be. That’s where you get my actual voice, precisely because nothing there is written down first. Layer2C Labs is the lab notebook, read aloud by someone else, because the notebook needs to be read exactly.