My Weird Prompts

The received wisdom says serious AI inference needs a GPU — but that's not the full story. This episode unpacks the real bottlenecks: memory bandwidth vs. parallel compute, prefill vs. decode, and the surprising workloads where CPU actually wins. From classical ML models like XGBoost to Whisper transcription and Mixtral on a laptop, we explore when the GPU advantage compresses from 100x to 6x, and...

When you dictate a messy to-do list and want clean JSON out, the question isn't just whether to fine-tune or prompt a general model — it's whether you can tolerate fabricated deadlines or missed tasks. This episode breaks apart format adherence, judgment quality, and failure predictability, exploring where fine-tuning actually moves the needle and where the real cost lies: not in compute, but in a...

Daniel needs to extract structured data from product labels and book covers — serial numbers, titles, authors, and a wildcard field for anything unexpected. Should he use a general vision model like GPT-4V or build a purpose-trained pipeline? We break down the real tradeoffs: accuracy on tiny rotated text, cost at scale, offline capability, how each approach handles the novel-fields requirement, a...

Most tutorials skip the hard parts. This episode walks through the full process of training a single-class object detector from scratch: how many images you actually need, why scene-based splitting matters more than random splits, how to annotate consistently, and why YOLOv8n is the right architecture for small datasets. Using the concrete example of detecting a specific anti-graffiti poster desig...

What does daily life actually look like in North Korea today — not the famine-era memoirs, but the country as it exists right now? This episode goes inside the gap between Pyongyang and everywhere else, exploring how the donju (money masters) live in high-rise apartments with smartphones and smuggled K-dramas, while most of the country still struggles with intermittent electricity and chronic food...

AI gateways promise more than just model routing—PII reduction, context augmentation, and system prompt injection at a single chokepoint. But every middleware layer adds latency, debugging complexity, and maintenance burden. We examine when these features justify the overhead and when simpler solutions like cloud credential managers or direct API calls make more sense. A practical look at the self...

Ever found a perfect model on Hugging Face only to wonder if your laptop can actually run it? This episode breaks down the two main pathways for running models from the Hub — local inference and cloud deployment. Learn how the compatibility tracker calculates TOPS and memory bandwidth, how Hugging Face's content-addressable cache stores weights, and the critical difference between the Inference En...

Pop-up light boxes fatigue after 50 folds, card systems charge $45 for coated cardboard, and mini backdrop stands can't hold their own fabric. But there's a fourth path most people overlook: a modular desk studio built from an articulating arm, a small LED panel, and a roll of seamless paper. For about $115 total, it outperforms all three consumer solutions with zero permanent footprint. Episode #...

Ever tried drilling into concrete with a regular hammer drill and watched the bit just skate across the surface? This episode explains why. We break down the mechanical difference between cam-action hammer drills and electro-pneumatic rotary hammers, and why the SDS-Plus bit system was invented in 1975 to solve a fundamental physics problem: friction can't hold a bit when you're pounding it thousa...

When a company knows it will lose a lawsuit, fighting to the end can actually be the rational choice. This episode explores "tokenistic litigation" — the practice of mounting a defense not to win, but to avoid the collateral damage of admitting fault. From insurance premium spikes to regulatory investigations to copycat lawsuits, the costs of settling can dwarf a trial loss. We examine the pervers...

Jerusalem’s lampposts are plastered with posters of the Lubavitcher Rebbe captioned “King Messiah” — a saturation campaign that’s sparked a secular citizen coalition fighting back. But what does Israeli law actually say? We break down the Municipalities Ordinance, the enforcement gap, and the legal and ethical dilemmas of citizen-led removal. From fines of 5,000 shekels per poster to the paradox o...

Small talk dread is real, but most advice for overcoming it is either too vague or actively counterproductive. This episode cuts through the self-help noise to examine what conversation research actually says works. We break down why "just be curious" fails the people who need it most, why "power moves" like deliberate mirroring backfire, and what specific tactics — like the add-a-sentence rule an...

We trace the DNA of clickbait from the 1835 Great Moon Hoax through mid-century ad copy to modern YouTube thumbnails. Why does the curiosity gap still work even when we know the game? And what happens when AI learns to optimize the same psychological levers? Episode #172387 — open it directly at myweirdprompts.com/172387

When a security incident turns a shared Jerusalem street into a fault line, the coexistence that seemed so real suddenly reveals its paper-thin nature. This episode goes beyond peace treaties and municipal budgets to ask a deeper question: is tolerance something you can actually teach, or is it baked into personality? Drawing on the Big Five personality model, Gordon Allport's Contact Hypothesis,...

Anyone who uses voice-to-text knows the frustration of watching a technical term get silently replaced with something that sounds similar but means something completely different. "LORAs" becomes "Lauras," and suddenly your transcription is nonsense. In this episode, we explore a practical solution: running audio through multiple transcription services, then using an LLM as a judge to reconcile th...

Should the best teachers also be practitioners? This episode explores the theory-practitioner split across law, medicine, dentistry, and architecture. We examine how dental schools flipped their curriculum, why clinical law professors burn out at higher rates, and whether tacit knowledge can ever be taught in a classroom. The answer isn't simple — and the dual role might be asking for two differen...

Why does the world call Israeli settlements illegal, while Israel insists they're not? This episode walks through the ICJ's 2024 advisory opinion, the Fourth Geneva Convention, and Israel's official legal defense — from the Green Line to the ring neighborhoods of East Jerusalem. We trace the settlement program from 1967 to today, covering the Allon Plan, Gush Emunim, and the legal machinery behind...

Why do tens of thousands of cases still go to trial every year, even as precedent piles up and outcomes grow more predictable? And what makes a tiny handful of those cases — like the decomposed snail that birthed modern negligence law — become household names? This episode explores the gap between how precedent works in theory and how it actually functions in practice: asymmetric incentives, strat...

Daniel noticed something living in Jerusalem — Israel doesn't seem to have as many domestic fires as you'd expect. Is that perception backed by data? We pulled the USFA and NFPA numbers to find out. The answer reveals a universal truth about what actually starts fires in homes: cooking dominates by volume, but heating equipment and electrical faults are disproportionately deadly. We break down the...

Why does one country have sixty times more homicides than another? This episode digs into the comparative criminology data to separate gut feelings from evidence. We cover the strongest demographic predictor of violent crime (the youth bulge), the independent effect of alcohol consumption, and the controversial lead-crime hypothesis that may explain the 1990s crime decline. We also tackle the mode...

Hostage situations dominate headlines but are statistically negligible. So what actually determines how safe a country feels? We dive into the data comparing Israel and the United States on homicide rates, property crime, policing density, and perception. The surprising finding: police headcount doesn't predict safety nearly as well as economic equality, social trust, and institutional legitimacy....

When you hear "hostage situation," you probably picture a bank robbery gone wrong with SWAT teams and negotiators. But the reality is dramatically different. In this episode, we break down the actual statistics — from the FBI's defunct HOBAS database to global UN data — and reveal that most hostage incidents in the US are domestic disputes that resolve in under two hours. Then we compare that to c...

This episode traces the brutal history of sieges from ancient Rome and Assyria through to modern conflicts in Sarajevo, Aleppo, and Gaza. We explore how what was once standard military practice — surrounding a city and starving its population — is now explicitly prohibited under international humanitarian law. Yet despite Geneva Conventions and ICC statutes, the tactic persists. Why does the legal...

Is magnetism actually involved in honing a knife edge? No — but the myth persists. In this episode, we break down the mechanical reality of what happens at the microscopic edge of a blade, from rolled steel to abrasive sharpening. We cover how much material sharpening actually removes (spoiler: very little over decades), why tiny utility blades are harder to sharpen than chef's knives, and the fou...

Private military and security companies (PMSCs) are a $240 billion industry, and their primary clients are states. This episode explores why these groups are structurally invested in perpetuating conflicts, not just fighting in them. We examine the regulatory black hole that makes them useful to governments and impossible to control, from the Wagner Group's mining operations in the Central African...