My Weird Prompts

Google DeepMind's Gemini Omni landing page claims the model has "an intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics." But video models have been walking people through walls for years — and a landmark ICML 2025 paper found they generalize by case-based retrieval, not physical law. So how did Omni seem to close the gap without the radical architectural departure t...

A bounded classifier emits a probability distribution over a fixed label set in one forward pass — no token generation, no JSON parsing, no helpfulness. This episode walks through what it takes to build one for a real pipeline: which encoder to fine-tune from (ModernBERT, DeBERTa-v3, or Liquid AI's LFM2.5 encoders), what size makes sense for a few thousand examples, and why a 150 MB weight file ru...

Your retrieval pipeline returns a document, the model cites it, and the claim checks out. Everything in the trace looks fine. But what if the model actually pulled that answer from its training data and the citation is just decoration? This episode digs into a genuine gap in AI observability: tracing shows what your agent did, never why. We walk through our own planning-agent-to-research-agent-to-...

When a label printer died mid-move, Daniel didn't open a design app — he described the problem to ChatGPT, and the thread quietly crossed from conversation into a full agentic workspace. This episode unpacks what actually happened: ChatGPT's July split into Chat and Work modes, the model swap from GPT-5.6 to Sol/Luna/Terra, and the 223 registered tools Simon Willison counted in his teardown. Then...

Daniel's AI-generated show deploys to Vercel, runs on serverless Postgres, and produces exactly the kind of wide-ranging content he wants. The problem is the feed — automatic tagging that started inventing its own categories, then duplicating them in ways nobody notices. This episode works through the fix: why unconstrained categorization never converges, how JSON Schema enums and context-free gra...

Shatnez — the Torah's prohibition on wearing wool and linen together — is the paradigm of a chok: a divine statute whose reason is explicitly unknowable. This episode unpacks the surprisingly narrow definition (sheep's wool plus flax linen, nothing else), the zero-tolerance rule that makes one linen thread a problem, and the theology that treats unknowability as the point rather than an embarrassm...

Israel took delivery of the INS Drakon, a 3,000-ton German-built submarine, on September 24, 2026 — the sixth and largest boat in its fleet. Officially, Germany denies these submarines carry nuclear weapons; Israel has never confirmed it has any. Yet SIPRI, the Nuclear Threat Initiative, and the Federation of American Scientists all describe the Dolphin fleet as Israel's sea-based deterrent. This...

Eight percent of US households own no television, but most of them still watch — they just left the furniture behind. This episode starts with a listener's accidental four-month break from TV and ends somewhere more surprising: the habit discontinuity hypothesis, a Peterborough field experiment with 800 households, and the uncomfortable finding that life's disruptions open a window for change that...

Moving house with a toddler, Daniel and Hannah went four months without watching Netflix — not as a decision, but because the landlord's TV stayed on the wall. Now that they could resume, something feels off. This episode unpacks two questions: what a screenless household actually looks like (and why the data everyone quotes is measuring something else), and what happens when a habit breaks by acc...

Daniel saw a Jerusalem high-rise marketed on its Knesset view and figured the view was a trifling concern — until he lived overlooking a parking lot. It turns out environmental psychology has been studying this since 1984, when Roger Ulrich matched gallbladder surgery patients with tree views against patients facing a brick wall. The tree-view patients went home sooner, needed fewer strong painkil...

Brian Friel's Philadelphia, Here I Come premiered at Dublin's Gaiety Theatre on September 28, 1964, and it put two actors inside one man. Public Gar faces the world; Private Gar says everything his counterpart cannot. This episode unpacks the split-self device, the kitchen scene where a father's only attempt at love is advice about sitting at the back of an airplane, and the speech about loving hi...

Most retellings of the ancient symposium lead with the philosophy and treat the drinking as set dressing. It was the other way round. A guest chosen by lot — the symposiarchēs in Greek, the magister bibendi or rex convivii in Latin — set the wine-to-water ratio that governed the entire evening, and enforced participation with social penalties that would end a modern dinner party. This episode look...

Underneath a historic house near Mahane Yehuda market, engineers are digging seven stories down — without demolishing the building above. It's called top-down excavation, and the write-ups describe it as one of the most extensive uses of the technique ever attempted. This episode breaks down how you build the walls before you dig the hole, why every floor slab doubles as a structural brace, and wh...

The Torah mandates rest on the pilgrimage festivals — and mandates that every able Israelite male appear in Jerusalem. Those two commands collide the moment eighty thousand pilgrims arrive. This episode traces the ancient infrastructure that made it work: roads cut into bedrock, innkeepers paid in sacrificial hides, wells dug with public money, and a Temple that doubled as the economic engine of t...

A Brother label printer lights up for ten seconds, then goes dark. That single symptom turns out to be a masterclass in how modern devices protect themselves — and why the gear you buy first should serve diagnosis, not repair. This episode walks through the charge controller's startup handshake, the ranked list of causes behind a ten-second cutoff, and the free thirty-second test that eliminates t...

If you magnetize a screwdriver over and over for years, does the tool eventually wear out? Daniel wrote in with that exact question, plus a practical problem: his magnetizing block won't fit a flexible drill shaft. This episode covers the three categories of magnetizing tools on the market — the block, the telescopic wand, and the powered electronic unit — how to tell a good block from a lousy one...

Daniel wrote in with four questions about IEC 60320 cables: can you terminate them yourself, do adapters matter, are splitters safe, and how long can a run actually be? The answers turn out to come from different places — a wiring standard, an ampacity table, and in at least one case, an honest "no source says that." Along the way: the odd/even naming rule that unlocks the whole connector family,...

What happens when a voice agent understands you perfectly and still can't help? This episode unpacks a real ISP support call where a customer with a single, documented request — whitelisting a third-party modem — got trapped in a connectivity diagnostic he didn't need, in a language the bot's escape hatch didn't recognize. We name the failure pattern (the single-intent state machine), explain why...

Bezeq, Israel's main fiber provider, does allow customers to bring their own modem — but only approved models, and only after you submit the device's serial number and wait to be provisioned. A listener named Daniel asked three sharp questions: why an ISP would want to control device security down to the model and serial number, how reliable a device-reported serial number is as an authentication...

OpenAI is retiring custom GPTs on December 11, 2026, with a migration path to plugins built on Skills and MCP servers. This episode walks through the real timeline, the four documented plugin shapes, and the migration mechanics — including the parts that don't carry over at all, like custom actions. We also dig into the SKILL.md open standard that Anthropic released in December 2025 and why it spr...

How did wine actually get from the vineyard to a Roman cup? No bottles, no corks, no personal cellars — instead, a three-tier vessel system that treated wine as a poured commodity rather than a packaged product. We walk the chain from the dolia — six-foot ceramic tanks, some cemented into ship hulls — through the amphorae that carried it across the empire and never came back, to the thermopolia wh...

Daniel dictates his prompts, catches typos by hand, and keeps running into the same wall: fixing a misspelling is not the same job as understanding what he meant. This episode pulls apart the pipeline into its three real tasks — disfluency removal, mechanical typo correction, and semantic inference — and asks which of them a language model should be allowed to touch. Along the way: why an "um" is...

In ancient Greece and Rome, wine was almost never drunk neat — and the person who decided how much water went in held surprising power over the evening. This episode digs into the symposiarch and the magister bibendi, the dice-chosen "king of drinking" who set the ratio, paced the cups, and kept the whole room at roughly the same level of inebriation. Along the way: Plutarch's ratio manual, the mu...

Most coverage of crowd disasters leads with "panic," but the research says otherwise: victims of crushes almost never die running from anything. This episode digs into the mechanics of crowd turbulence — the density threshold where movement becomes involuntary, and the compressive asphyxia that kills people while they're still standing. From the 2021 Lag BaOmer disaster at Mount Meron to the Holy...

A listener named Daniel asks how people actually determined whether a photograph was real or altered — in the pre-digital era and now. We trace image forensics from hand-examined negatives and spirit photography hoaxes like Cottingley and Mumler, through Photoshop-era techniques like error level analysis and clone detection, to the provenance fingerprints being baked into AI-generated images today...