Friday, 20 November 2026
22 episodes · 6596 in all
#6717: Why Ten Seconds Clones Better Than an Hour of Training
A ten-second reference audio clip serves as a prompt address, while fine-tuning burns background hum and room echo permanently into model weights.
#6716: Choosing Open-Source Frameworks to Fine-Tune TTS
Under five hours of voice fine-tuning data in Piper requires roughly 2,000 training epochs, far more passes than a larger dataset.
#6715: How Ribbon and Boundary Microphones Actually Work
A boundary microphone's placement eliminates phase delay at rigid surfaces, providing a six-decibel boost in sound pressure for free.
#6714: Why Voice Clones Stay Flat
Fine-tuning Qwen3-TTS on ninety minutes of audio improved prosody and emotion without materially boosting speaker similarity.
#6713: Airing Challengers to Test AI Podcast Writers
Fifteen drafts across five AI models yielded six passes, nine marginals, and zero fails under Claude's judging.
#6712: Between BM25 and Vector Search
BM25 misses only 7% of queries found by modern neural embedders, but vector search misses up to 23% of keyword matches.
#6711: Reciprocal Rank Fusion and Reranking Eval
The default k=60 in reciprocal rank fusion traces back to a 2009 pilot study and remains the default across OpenSearch, Pinecone, and MongoDB.
#6710: Why BM25 Still Beats Neural Search on Keywords
Under typical search tuning, a word's fifth mention delivers over eighty percent of its total possible score before saturation flattens out.
#6709: Exact Search Without a Vector Database
A query embedding took 26 milliseconds while the exact brute-force search across the corpus took less than a third of a millisecond.
#6708: Why Grep Is Not Enough for Search
Ripgrep scans 22 million words of text in under a tenth of a second, yet cannot efficiently answer structured queries.
#6707: Teaching Vision Models to Spot Tiny Architectural Details
Doubling image resolution from 224 to 448 pixels boosted Renaissance architecture recognition by 35 percent in the WikiChurches study.
#6706: How Accidental Dataset Formats Became AI Standards
LJSpeech uses a pipe-delimited file named metadata.csv, and it became the TTS standard simply because torchaudio shipped a loader for it.
#6705: Why Voice Agents Use Frameworks Like Pipecat
Pipecat adds roughly 60 milliseconds of latency over Gemini Live, but buys transport independence and transcript control.
#6704: Bypassing Android's Built-In USB Audio Stack
On a current OnePlus phone, Android's audio layer added 12 dB of gain and a -0.98 dBFS limiter to a USB mic before the app ever saw a sample.
#6703: Tokyo's Standing Sushi Bars and Market Counters
Some Tokyo standing sushi bars serve the same Edomae fish as their pricier seated sister restaurants, starting at 70 yen a piece.
#6702: The Ryukyu Islands Beyond Okinawa's Resort Strip
Spanning a thousand kilometers, the Ryukyu Islands hold six endangered languages and distinct local traditions far beyond Okinawa's resorts.
#6701: Planning a Murakami Winter in Japan's Onsen
In Akita's Nyuto Onsenkyo, seven forest inns share ten distinct hot springs, no town, and an old-school phone-only booking system.
#6700: Why Tokyo Feels Like a Neighborhood of 37 Million
Japanese zoning specifies what you cannot build rather than what you can, making mixed-use blocks with bakeries and apartments legal by default.
#6699: When Buying a Laminator Beats Paying Per Sheet
A working home laminator costs as little as $15, meaning owning one beats shop prices after just a couple dozen sheets.
#6698: Why DIY Wardrobe Recording Booths Sound Boxy
Any booth dimension between three and five feet resonates at 200 to 400 hertz, right inside the chest of the human voice.
#6697: Why Time Off at the Top Is Just Relocated Work
A Harvard Business School study found CEOs work on 70% of their vacation days and spend 72% of their work time in meetings.
#6696: Two Building Laws That Shaped Seoul's Streets
Two Korean building laws—a 1970s daylight rule and a 2002 parking mandate—created Seoul's stepped villa rooflines and dark ground-floor alleys.