Saturday, 31 October 2026
21 episodes · 6596 in all
#6397: Why Model Benchmarks Fail at Wit
When prompted for a joke over 1,000 times, ChatGPT repeated the same 25 jokes more than 90% of the time.
#6396: How Political Scandals Exposed Boring Email Bugs
A 2007 congressional inquiry found the Republican National Committee preserved zero emails for 51 of 88 senior White House accounts.
#6395: Why 3M Primer 94 Sticks to Everything
A tiny active payload of under six percent solids allows 3M Primer 94 to physically diffuse into low-surface-energy plastics like polypropylene.
#6394: Why API Email Still Relies on SMTP
Connection pooling cut Mailgun's delivery time from ninety seconds down to roughly six hundred milliseconds per message.
#6393: Why Sol Got Demoted as Our Script Writer
A pattern search across forty recent episodes found host emotional forecasting in 55% of scripts generated by GPT-6.1 Sol.
#6392: Verifying Generated Arabic Text as a Non-Reader
A study of DALL-E 3 across 2,808 images found its correctness on Arabic letter forms was effectively zero.
#6391: How Forensic Tools Search Android for Spyware
Amnesty's Mobile Verification Toolkit removed direct ADB shell analysis because device shell differences caused inconsistent outputs.
#6390: Why Editing Vector Databases Silently Re-Embeds Text
A one-word edit cannot nudge existing vector numbers; the model must re-embed the chunk from scratch or search answers with deleted text.
#6389: Decaying Old Memories in Vector Search
A time-decay multiplier can drop a 0.92 similarity result down to 0.41, allowing mediocre fresh matches to overtake it.
#6388: Why Swipe Decks Make Better Training Data
A Tinder-style swipe deck removes comparison to force isolated choices, turning rapid binary labels into active learning training signals.
#6387: What Really Runs When You Compact an Agent's Context
Commercial compaction relies on main models and prompts, while purpose-built fine-tunes like Compactor-Qwen3.5-4B remain rare.
#6386: Where Harmonizable Processes Actually Work
A cracked bearing hits on a schedule, turning vibration into correlated frequencies that standard stationary models average away.
#6385: Turning a Scored Matrix Into One Winner
Adding an identical copy of an existing tool can flip the entire leaderboard in standard decision analysis models.
#6384: Blending Input and Output Token Prices
Decode computes at just 1.4% utilization on an H100 because reading 140 gigabytes of weights takes 280 times longer than the arithmetic.
#6383: Choosing Class Breaks for Skewed Price Data
In one dimension, k-means clustering avoids random restarts entirely and runs in polynomial time via dynamic programming.
#6382: Choosing AI Models on Cost and Performance
Model prices vary thirty-eight-fold per index point, meaning ratio formulas love cheap options regardless of actual quality.
#6381: Turning Anonymous Diarization Clusters into Names
Shortening an enrolment sample to two seconds can triple the verification error rate compared with full-length recordings.
#6380: Choosing the Right Voice Activity Detector
A compact model file can hide a massive runtime, while Cobra bundles its proprietary detector and engine into just 450 kilobytes.
#6379: Porting Voice Models to Core ML and MLX
Core ML packages separate computation specs from raw weights, while MLX runs on CPU and GPU without touching the Neural Engine.
#6378: Running Speaker Identification Locally on Phones
Exported voice embedding models like CAM++ run offline on mobile devices in packages as small as twenty-eight megabytes.
#6377: How Audio Models Tell Voices Apart
A diarization model can tell two voices apart without understanding a single spoken word or assigning a real name.