My Weird Prompts

When ChatGPT asks you to approve a tool call, something has already scanned the outgoing data for usernames, addresses, and other PII — and that something is probably a small language model. This episode digs into the "guardian" class of compact models doing narrow, high-volume security work: PII-Tracer, OpenAI Privacy Filter, GLiNER2-PII, and the consistency problem that makes partial detection w...

Most advice about custom instructions says make them longer. The evidence says the opposite: a leaner prompt can be functionally equivalent, four and a half times cheaper, and faster to first token — because every added rule dilutes the attention paid to the others. This episode takes apart a real listener's general-purpose ChatGPT instruction line by line, separating what earns its tokens from wh...

Mobile number porting is a legally mandated right, standardized by 3GPP and enforced in Israel since the mid-2000s. The consumer version is one sentence: disconnect the old SIM and wait thirty minutes. The technical version is a signalling relay, a portability database, and a synchronization problem that has nothing to do with routing speed. We trace what actually happens when a number moves betwe...

A single 2015 forum post, one scraped phone number, and a chatbot that confidently hands it to strangers over a decade later. This episode digs into "AI doxxing" — the documented cases where Gemini, ChatGPT, Claude, and Grok surfaced real people's phone numbers — and the uncomfortable mechanics underneath. We trace how a number gets into training data, why one mention is enough, and what labs can...

Israel deregulated its mobile market to multiply competitors, but the antennas collapsed into two shared networks — PHI and S-Plus — and virtual operators aren't required to disclose which one they ride. A listener's phone migration saga raises two questions: can you detect the host network from the outside, and do MVNO subscribers get the same service as direct ones? This episode walks through th...

Daniel bought a data-only eSIM, then discovered the provider's portal only accepts SMS one-time codes — which a data-only plan can't receive. A second case: a switched phone number left him permanently locked out of his Israel Post account, because changing the number required authenticating with the old one. This episode unpacks why these lockouts aren't bugs but failures of imagination, tracing...

A country that eats chopped salad at seven in the morning, and a country where the same dish signals joyless dieting. How did raw vegetables become the backbone of the Israeli and Turkish breakfast table — and rabbit food in America? This episode traces the Israeli salad from Levantine shepherd's salad to kibbutz dining hall to contested national dish, then builds the global list of cultures that...

Tardive dyskinesia affects roughly 20-25% of patients on long-term dopamine-blocking drugs, yet the mechanism behind its permanence is still debated. This episode digs into D2 receptor supersensitivity, the maladaptive synaptic plasticity theory, and why removing the drug often doesn't remove the movement. Then we get into the numbers: how much have second-generation antipsychotics actually reduce...

A listener's email about Amazon's tightening return policy sends us down a rabbit hole of consumer overreach, from petty market grazing to industrial-scale refund fraud. We trace the spectrum from a single grape at a date stall — governed by nothing but eye contact and shame — to the woman who returned over a hundred Kindle books while reading every one of them, and the wardrobers, home-stagers, a...

A friend of the show wants to launch a printing business in New York, and his marketing plan involves an airplane towing a QR code banner along the approach path into the airport. We take the idea seriously enough to ask the obvious questions: can a passenger actually scan a code while landing, what are the legal limits around airports and highways, and where does physical advertising in extraordi...

Around 40% of people worldwide now say they sometimes or often avoid the news, up from 29% in 2017. It's not fragility — it's a rational response to a feed that filters for failure. This episode unpacks why every individual story can be true while the composite is a lie, what constructive journalism actually means (it isn't positive news), and how the ad model itself shapes what you're allowed to...

Two offhand comments — one about hygiene, one about water use — convinced Daniel that baths were pointless. He's been a shower person ever since. But when he finally climbed back into a tub with frankincense and myrrh salts, sick and achy, he wrote in to ask what the science really says. It turns out the research on immersion bathing is stronger than the folklore ever was, and he was robbed. This...

India has overtaken the US as the largest contributor base to public open source projects, and one in three new GitHub developers now joins from a country that wasn't in the global top ten in 2020. Starting from a Croatian-built offline dictation app called Whisperian, this episode digs into GitHub's Octoverse data: quadrupling developer populations in India, Brazil, and Indonesia, triple-digit gr...

A grown man reaches into a bin of Haribo at Mahane Yehuda, eats a fistful, and walks on — and almost nobody stops him. This episode uses that small, gross moment as a lens on how social norms actually work: why vendors tolerate the loss, why the rare owner who enforces hygiene gets punished for it, and how descriptive norms quietly eat injunctive ones. Along the way: food-safety microbiology on op...

Everyone building agents reaches for the biggest model and wraps every endpoint as a tool. But there's a quieter pattern: tiny text-to-text models doing one narrow structured job. Daniel wrote in asking whether small models doing predictable structured output inside MCP pipelines actually ship in production, or whether everyone just lets a general-purpose model emit tool calls. The answer runs thr...

When Hebrew words are written in Latin characters inside otherwise English text, standard language identification fails: it returns the dominant label and the Hebrew vanishes. This episode works through the terminology (LID, CS-LID, transliteration, diacritization, G2P, text and script normalization), why romanized Hebrew defeats script-based heuristics, and what the literature actually offers — M...

Daniel has a working podcast pipeline: a language model writes scripts, retrieval tools feed it, and text-to-speech reads it out. But the occasional Hebrew word keeps breaking things, and the pipeline itself is trapped in his head. This episode digs into two questions at once — how to handle Hebrew code-switching in TTS, and how to package a chained AI workflow so it survives past the project that...

Android keyboards quietly accumulate months of your typing habits — your email address, your common phrases, your personal words — and lock most of it inside vendor cloud silos you can't export or back up. This episode unpacks how the components actually work under the hood: the InputMethodService plumbing, the learned frequency model, the surprisingly portable UserDictionary, and the two very dif...

When you connect ChatGPT's email plugin to a workspace address and a Gmail address, you tick a few scope boxes and everything just works. But what's actually happening in that moment — and who holds the credential that sends your mail? This episode walks through the OAuth 2.0 authorization code flow with PKCE, explains why you never see a token (it isn't yours to see), and untangles the difference...

A listener with functional dyspepsia asks why dopamine-blocking drugs (like sulpiride and metoclopramide) kill his nausea but leave him exhausted the next day — and why his daily ADHD stimulant never makes him nauseous at all. We trace the area postrema, the brain's blood-barrier-free chemoreceptor trigger zone, and explain why nausea and vomiting aren't the same thing, why D2 blockade is both the...

You pass a language ID to a multilingual TTS model, and every word in that utterance obeys it — including the ones that belong to a different language. That's why "Shabbat" sounds fine and "pinui binui" sounds like a bad impression. This episode unpacks a listener's proposed pipeline for code-switched speech: detect the Hebrew words written in Latin characters, map them to Hebrew script, and re-sy...

Daniel has been replacing typing with voice for two years, and moving from Whisper to Parakeet V2 on Android was a massive step up. But when your dictation pipeline chains a voice activity detector, an ASR model, and one or more cleanup models, how do you know it's actually working — or is the phone just unusually idle? This episode walks the full chain stage by stage: why ASR models are deliberat...

Daniel is sick in bed and has decided the podcast needs a rebuild: bring back audio prompts, add follow-up questions, and give him a push-to-interrupt button so he can grab the floor mid-episode. He thinks the button is the hard part. The research says it's the opposite — turn-taking primitives like push-to-talk have been standardized since walkie-talkies, while the live full-duplex version he's i...

A listener named Daniel asked two questions: who models technology as something playful rather than something to be mastered, and how do you keep the magic when deep technical work stops feeling fun? This episode works through both — comparing Linus Torvalds' invisible, load-bearing infrastructure with Linus Sebastian's filmed, messy process, and examining what actually makes on-camera camaraderie...

Daniel wanted a notification whenever a pigeon landed on his balcony, and he was ready to buy a TPU and hunt for a species-ID model on Hugging Face. Turns out Frigate's default model already has bird as a built-in class, and the leftover work is entirely in configuration. This episode walks through the actual engineering: why YOLOv9-tiny at 320×320 beats a bigger model on a small distant bird, how...