Saturday, 14 November 2026
27 episodes · 6596 in all
#6606: Why Optimizing Voice Quality Erases Identity
Squared-error loss and generative speech enhancers treat breathiness, rasp, and vocal fry as defects, averaging distinct voices into a median.
#6605: Four Layers Between Your Ears and Background Noise
In a classic listening test, two-thirds of people with low working-memory span heard their name in an ignored ear, versus one in five with high span.
#6604: What Breaks When You Train AI on Synthetic Taste
Inverting a user's preference pairs left a personalized reward model's predictions unchanged in 86 percent of cases.
#6603: Lossless vs. Hi-Res Audio and Whether Ears Can Tell
In a 12,500-trial meta-analysis, listeners picked hi-res audio over CD quality just 52.3% of the time—barely better than a coin flip.
#6602: Training Voice LoRAs on Filtered Zero-Shot Audio
Training six OmniVoice LoRA adapters cost under two dollars per voice, using Whisper to filter out mispronounced takes.
#6601: Mapping Foot Pedals by Device ID Instead of F13
On macOS, F21 through F24 are dropped entirely, and KDE binds F13 by default—proving reserved keys aren't safe for foot pedals.
#6600: Reverse Engineering Windows Configurators on Linux
AvaloniaILSpy is archived, dnSpyEx now runs natively under Wine, and Ghidra's ReVa extension lets LLMs trace native binaries over MCP.
#6599: Two Sentences That Broke AI Reverse Engineering
Two sentences added to Claude Code's prompt to stop game cloning accidentally caused it to refuse legitimate reverse engineering mid-task.
#6598: Why Your Python Command Is Probably a Rust Binary
Running pip install uv downloads a 53-megabyte Rust binary, with a tiny Python stub whose only job is pointing to it.
#6597: Why Terminal-First Developers Embraced AI Agents
Stack Overflow's survey puts Claude Code at 65.5% adoption, driven by developers running multiple AI agents inside tmux and git worktrees.
#6596: Why Audio CLIs Are Shaped for AI Agents
SoX hasn't had a stable release since 2015, yet it still beats GUI editors at noise profiling and gives AI agents text they can actually read.
#6595: How Terminal Video Editors Give AI Agents a Timeline
In the Cassette-Editor study, giving an AI agent a typed timeline and rulebook jumped its editing success rate from 47% to 89%.
#6594: How Command-Line Image Tools Actually Compare
One misplaced flag in ImageMagick turns a 96-millisecond thumbnail into a 366-millisecond job using eight times the memory.
#6593: Same AI Model, Different Cyber Permissions
On CyScenarioBench, Opus 5.5 went from zero to thirty-four out of fifty tasks solved purely based on the operator's verification tier.
#6592: How Capture the Flag Trains Hackers and Gets Them Hired
Cybersecurity competitors regularly lose twenty minutes of a forty-eight-hour event just by formatting curly braces wrong.
#6591: Can a Twenty-Minute Fine-Tune Stabilize a Cloned Voice?
Zero-shot voice cloning reinterprets a 10-second clip on every call, compounding into vocal drift across hundreds of hours.
#6590: Passkeys on YubiKeys, Hardware Wallets, and 1Password
YubiKeys have supported passkeys since mid-2018; the real security divide isn't hardware versus software, it's copyable versus non-copyable.
#6589: Why Standing in Line Feels Longer Than Walking
Houston airport cut baggage complaints to near zero by making passengers walk six times farther, increasing total wait time.
#6588: Xpra and Per-Window Remoting for Remote Browsers
Xpra's virtual display defaults to 8,192 by 4,096 pixels—33 megapixels per frame—turning a remote Chrome window into a bandwidth bottleneck.
#6587: When to Skip the CLI Wrapper for AI Agents
A CLI wrapper can cut an agent's prompt to 80 tokens, but spawning a new process per call makes it 25 times slower than a persistent MCP server.
#6586: How OmniVoice Cloned Our Voices From Our Old AI
After Resemble AI retired Chatterbox in August 2026, k2-fsa's OmniVoice cloned the show's hosts from their own previously broadcast synthetic clips.
#6585: Four Primitives Behind Every Pairwise Eval Tool
Audio listening tests and LLM arenas share the same four primitives, yet only LLM tools swap pair order to catch position bias.
#6584: Why AI Leaderboards Dropped Elo for Bradley-Terry
Replaying 135,632 arena votes with online Elo instead of Bradley-Terry shifted Claude Sonnet 4 by 32 places.
#6583: Why Comparing Two Things Beats a Rating Scale
A 2014 study found people finish pairwise comparisons in 84 seconds versus 305 seconds for absolute scores, with far less noise.
#6582: Why Text-to-Speech and Speech-to-Text Aren't Mirrors
One second of speech takes 24,000 sequential decisions in WaveNet, while speech recognition throws away most of the audio signal.
#6581: Chatterbox Retires as Open TTS Splits in Two
Resemble AI quietly retired the 28,000-star Chatterbox voice model while Fish Audio's 35-billion-parameter Drama 3 demands data-center GPUs.
#6580: How One-Way Data Diodes Protect Power, Water, and Rail
France's cyber agency ANSSI now forbids firewalls on top-class industrial networks, mandating one-way hardware data diodes instead.