Monday, 19 October 2026
35 episodes · 6596 in all
#6149: 4 Hard Fails in 84 Answers: Testing AI Bias on Israel-Palestine
A 19-prompt, 6-model eval tested whether AI models push back evenly on false claims about Israel and Palestine. One model failed four times.
#6148: The Tiny Hebrew Models Behind the Show's Voice
Dicta's 300M-parameter nikudizer, Phonikud's adaptors, and the ktiv male model hiding inside a vowel adder.
#6147: Hebrew AI: Ears, Mouth, and Brain
Open Hebrew speech recognition beats the paid services. Hebrew text-to-speech is a different story — and the language models are a third.
#6146: Grounding Is Not Solved: DIY vs Gateway vs Native Search
Bolt-on search, gateway search, or native grounding? The plumbing didn't vanish — it just moved to someone else's bill.
#6145: Do LLMs Nod Along With Conspiracy Theories?
A 98-prompt eval asks one question: when a believer brings a conspiracy theory, does the model push back, or help them dig?
#6144: 8 Real Steganography Cases and How They Got Caught
eBay, a doll shop, and a blink in Morse code — eight real steganography cases, and the tell that exposed every one.
#6143: Propaganda Bait: Testing 7 Models in 5 Languages
Five prompts, seven models, five languages — a first look at a benchmark that doesn't exist yet, testing what LLMs do with government talking points.
#6142: We Gave 4 AI Models the Same Toast Prompt
Four models, one buttered-toast physics question. Three invented studies that never existed — and the oldest model was the most fun.
#6141: Overheard at the Laundrette
A seven-minute laundrette scene about socks hides a real sentence. Can you spot the null cipher before it's revealed?
#6140: Did Flight FZ1073 Really Happen? Inside a False-Flag Claim
A hijacking attempt on Flydubai FZ1073 was declared staged within 67 minutes. What the evidence actually supports — and what it doesn't.
#6139: The Models That Didn't Write Our Scripts
Five models, three scripts each, one rubric. Sol won on length, Grok on voice — and the runner-up title hides a catch.
#6138: Hebrish on iPhone: MLX, Core ML, or Wait?
Hebrish spells makolet right on Android. Getting that Hebrew-English dictation onto an iPhone means picking between MLX, Core ML, and sherpa-onnx.
#6137: Two Ways a Chinese Open Model Gets Censored
Same weights, different landlord: how to tell a serving-time filter from censorship baked into the model itself.
#6136: How We Chose Our New Script Writer
Ten models audition to write the show. Seven fail the censorship gate — including the workhorse that wrote three episodes in four.
#6135: How Many Runs Does Your AI Eval Actually Need?
Reruns shrink one kind of noise and leave another untouched — so where should your eval budget really go?
#6134: Which finalist is really cheapest per script?
A model with worse list prices can cost less per script — if the cache hits. The bill hides in tokens, not rates.
#6133: One Rater Is Enough: Designing Daniel's Taste Test
When the ground truth is one person's taste, agreement with strangers is the wrong bar. Stability with yourself is the right one.
#6132: How to Run a Model Eval on a Budget
Daniel is teaching himself to run evals on a real budget. Successive halving, paired comparisons, and why three clean scripts prove almost nothing.
#6131: Judging the Judges: How AI Evaluates Creative Writing
LLM judges match human raters — but the human ceiling is low. What that means for evaluating creative writing.
#6130: Picking a Human Labeling Tool for LLM Benchmarks
Building a human review layer for LLM benchmarks? Here's what Argilla, Label Studio, Prodigy, and Potato actually cost in money and maintenance.
#6129: How to Build an AI Preference Dataset
Pairwise comparison, forced choice, or just "this one or that one" — the engineering behind collecting human preferences on AI outputs.
#6128: Where the Screws Go: Internal KVM Mounting
PiKVM, NanoKVM, JetKVM, and BliKVM all promise remote rescue — but wiring one inside your case is a different problem entirely.
#6127: Haiku 5.5 at 10 Cents: What the Cheap Model Actually Does
Haiku 5.5 costs 10 cents per million input tokens. We break down where it holds up, where it fails, and whether it can write this show.
#6126: Can You Train a Machine to Be Funny?
Comedy fights everything next-token prediction is built for. Turns out you can train a model to land a joke — and then measure it.
#6125: Which Env Var Wins? Cloud vs. Local Precedence
Same variable name, two values, one silent winner. Vercel and n8n go one way; dotenv goes the other.
#6124: One Vault, Two Readers: Secrets for a Home AI Gateway
One source, two readers — not two copies kept in sync. How a home AI gateway should actually get its API keys.
#6123: OAuth or a Personal Access Token for Your MCP Server?
GitHub CLI login hands your agent one long-lived token with access to everything. Fine-grained PATs can't. Which should your MCP server use?
#6122: Explaining Misophonia to the People You Love
18.4% of adults meet the threshold for misophonia. Only 13.6% have ever heard the word. Here's how to bridge that gap.
#6121: Hearing Speech in Noise: What Actually Works
Over-the-counter hearing aids, AirPods Conversation Boost, and neural speech separation — what works when your audiogram is normal but the café isn't.
#6120: What Actually Helps When Sound Is Too Much
Misophonia and hyperacusis treatments are thin but not empty — and the ADHD sound-sensitivity gap is more interesting than it looks.
#6119: Two Keys and a Polite No: Labeling a Claude Code Approve Button
Daniel has two spare laptop keys and wants them to say yes and no. The electronics are easy — it's the labels that get hard.
#6118: What NRR Actually Means for Your Ears
That 34 dB rating on the box? Real-world protection is closer to 13. Here's how to actually pick hearing protection.
#6117: Earmuffs, Earplugs, and the Café Speech Problem
Why noise-cancelling headphones kill the espresso machine but leave the wedding planning next to you perfectly audible.
#6116: Why Losing Flow Makes Me Snappy
The drill stops, but the anger doesn't. What Glass and Singer found about noise, control, and the tantrum two hours later.
#6115: Claude Cracked the NSA's 1997 Curve Seeds
A 25-year crypto mystery solved in under 30 minutes — by finding a bit that wasn't there.