My Weird Prompts

A banker published the first one-time pad in 1882 and told readers to erase the numbers after use. Shannon proved in 1949 that it's mathematically unbreakable. So why has almost every attempt to actually use one failed? This episode follows the pad's real lineage — Vernam's repeating tape, the Soviet duplicate pages that cracked Venona, SIGSALY's fifty tons of vinyl key records, the Moscow-Washing...

When a confidential document leaks, how do organizations figure out who passed it on? This episode digs into the real history of "canary traps" — from Tesla's binary space-code emails and forensic watermarks on Oscar screeners to trap streets and fake encyclopedia entries. Then we get into the digital version: zero-width Unicode characters that carry a hidden binary signature invisible to the read...

What starts as a debate about whether the hum in the room is a bird, a fridge, or an extractor fan turns into something else entirely: rice cake rationing, an unfalsifiable peach argument, a petrol-station mug with a past, and a forty-minute on-air nap. Somewhere in the middle of it all, someone asks the question nobody wants to answer — is that little red light supposed to be on? It's an episode...

What happens when you hand six AI models mirrored pairs of loaded claims — one false premise serving each side of the Israeli-Palestinian conflict — and grade whether they push back with equal force? Jerusalem-based researcher Daniel Rosehill built exactly that eval, and this episode walks through the results. The run was small: nineteen of seventy prompts, six models, one sample each, about thirt...

Hebrew speech pipelines lean on small task models that quietly decide whether the output is right — a nikudizer, a phonetic converter, a spelling normalizer. This episode digs into Dicta's DictaBERT-large-char-menaked, Phonikud's adaptors on top of it, UNIKUD's CANINE-based approach, and why the model that converts defective to full spelling only exists as a byproduct of training another tool. Alo...

Can you run Hebrew AI entirely offline? It turns out the answer splits three ways — ears, mouth, and brain — and each one has a different winner. Open Hebrew speech recognition, built on Knesset recordings, is beating Google, Amazon, and OpenAI on a public leaderboard. Hebrew text-to-speech is dominated by closed models, with one clever phoneme project changing the math. And Hebrew-native language...

Grounding used to mean wiring up a search API, a scraper, and a lot of glue. Now there's a middle tier — gateways like OpenRouter and Vercel that bolt search onto any model — plus models that search on their own. This episode walks through all three routes: the five moving parts of DIY, how gateway routing actually decides which engine to use, and what native grounding really costs once token weig...

Daniel's Conspiracy-Endorsement-Eval asks 98 prompts of every model two ways — a curious user, and a believer who's already sure and wants sources, arguments, and ways to spread it. Some theories are famous, some are backed by political leaders, and some were invented for the eval so they can't be in any training data. In a small first pass — seven prompts, seven models, one sample each, no system...

In 2002, USA Today reported that al-Qaeda was hiding encrypted messages inside eBay images. Two University of Michigan researchers scanned two million eBay pictures and found nothing — and that gap between headline and evidence is where this episode begins. We count down eight documented cases of steganography actually used in the wild, from a World War I null cipher that spelled out Pershing's sa...

What happens when you ask a language model to defend a government's talking point — in that government's own language? This episode is a first look at "Propaganda Bait," a rough benchmark built from five prompts, each flattering a different government: Tiananmen, Bucha, and the Khashoggi killing among them. Seven models were tested — DeepSeek, GLM, Kimi, YandexGPT, GigaChat, ALLaM, and GPT-6.1 Sol...

We handed the same prompt to DeepSeek Flash, GPT-6.1 Sol, Claude Haiku 5.5, and GPT-3.5 Turbo: write a three-minute episode of this show answering whether toast really lands butter-side down. This is the listenable version of the bakeoff — full scripts read aloud, no rubric, no scores. The physics shape came out roughly right in all four. The problem is what happened in the one spot the question a...

Something a little different. Herman has written a short audio piece: Corn's mum Dorothy and Tim, the voice of our sponsor messages, waiting for their machines at a laundrette on a weekday evening. Lost socks, the flower rota, a tin of peaches. Listen to the whole thing before reading any further. Afterwards Herman tells Corn what was really going on in that conversation, and the two of them work...

When reports emerged of an attempted hijacking aboard Flydubai flight FZ1073 from Dubai to Tel Aviv, the false-flag accusations arrived faster than the facts — the first one just 67 minutes after the initial distress signal. This episode works through what's actually known: the alleged axe attack on the flight deck, the diversion to Tabuk, the UAE prosecutors' stated intent, and the physical objec...

We ran five AI models through a bakeoff to write our podcast scripts — same briefs, same word-count window, one judge with a rubric. GPT 6.1 Sol took the win, but the runner-up title went to Meta's Muse Spark 1.3 on a headline score that gates purely on length, and on judged quality alone Spark sits fourth of five. Along the way: Grok 4.7 wrote the best dialogue while blowing past the word cap eve...

Hebrish is a free, open-source fine-tune of NVIDIA's Parakeet that handles English and Hebrew mixed in one sentence — and it already runs as an Android voice keyboard. This episode digs into what it would actually take to bring it to iPhone. MLX runs on phones, but the Parakeet code on top of it is Apple-Silicon-Mac only. Core ML already runs stock Parakeet on the Neural Engine via FluidAudio, but...

A Chinese open-weight model can be silenced in two places: at the host that serves it, or inside the weights you download. This episode walks through a benchmark that ran Qwen, MiniMax, MiMo, GLM, Kimi, and DeepSeek on a full script-writing assignment — including a frank Tiananmen segment — and logged exactly where each one died. Qwen managed 35 words before Alibaba returned "Algo DataInspectionFa...

The show is written by a language model, and the round robin that picks which one has never been properly tested. So Daniel wrote the brief first: hard pass/fail rules, then two qualities that actually decide it — vivid dialogue, and what he calls "holding the middle." This episode walks through the funnel, from a free desk screen of roughly 470 models down to three full scripts per finalist, and...

You rerun your model panel and last week's winner slides into the middle of the pack. Was it luck, or was it the prompts? This episode works through the arithmetic of eval noise: how run-to-run variance shrinks with every added run, why prompt-choice variance doesn't budge no matter how many times you rerun, and why three runs is a defensible default for some measurements and badly underpowered fo...

Picking an LLM on price per token is the easy mistake. This episode walks through a real evaluation for the show's own script-writing pipeline — a 26,000-token input with a 9,000-token cached prefix and 6,500 tokens out — and shows how a model with a worse input price can win the bill by a sixth, then lose once caching is off or thinking tokens triple the output. We cover why blended leaderboard p...

Most evaluation advice says to average away individual taste. For a show with one creator, that advice is backwards. This episode works through how to build an evaluation system where the authority is a single domain expert — and the reliability that matters is whether that expert agrees with himself three weeks later. Drawing on applied LLM evaluation practice, qualitative coding methods, and a c...

When every run costs real money, you can't evaluate a dozen candidate models in one giant sweep. This episode walks through the stepwise evaluation Daniel built to pick a script writer for this very show — a screening funnel that starts cheap, cuts the field, and only spends real budget on the finalists. Along the way: what successive halving and racing actually are, why paired per-prompt comparis...

LLM judges now agree with human raters about as often as humans agree with each other — but that headline number hides a low human ceiling and a lot of hard cases. This episode digs into how judge-plus-human pipelines actually work: the human labels first, the judge gets calibrated against them, and blinding is a cheap hedge against a bias nobody's fully pinned down. Along the way: why raw agreeme...

When your benchmark needs a human rater, the tooling question gets real fast. This episode walks through the actual market for human annotation tools — the trace-based platforms like Langfuse, LangSmith, and Braintrust that are built for watching production apps, versus the static-dataset tools like Argilla, Label Studio, Prodigy, and the academic tool Potato that fit a benchmark's workflow. We co...

Before a reward model can learn anything, someone has to build the dataset underneath it. This episode digs into the unglamorous machinery of pairwise preference collection: how candidate scripts get generated through batch inference, how pairs get chosen so the comparison graph stays connected, and why the aired script makes an awkward contestant. It also covers Bradley-Terry scoring, the cost of...

Remote-recovery boxes like PiKVM, NanoKVM, JetKVM, and BliKVM are easy to love and surprisingly fiddly to mount. This episode walks through the physical side: why "PCIe KVM" cards barely use their PCIe slot, how front-panel headers get shared with your case's power button, and what the nine-pin USB header actually buys you. Then it gets to the real question — where does the rescue device get its p...