Daniel Rosehill
My Weird Prompts
The human-AI collaboration podcast. A man, a sloth, and a donkey collaborate to create a podcast (with a little help from AI). No question is too obscure, no rabbit hole too deep. My Weird Prompts celebrates curiosity in all its forms. Daniel, the human, asks the questions that pop into his head at inconvenient moments. Corn the Sloth offers laid-back, thoughtful takes. Herman the Donkey brings boundless enthusiasm and energy. Together, they explore topics ranging from the mundane to the mind-bending. Each episode begins with a real voice memo from Daniel, processed through an AI pipeline that...
Besuch unbedingt die Website des Podcasts und unterstütze die Macher: myweirdprompts.com
Autor
Daniel Rosehill
Kategorie
Podcast-Website
Neueste Folge
20. Nov 2026
Wo hören?
Podcasts in der App Replaio Radio Bald verfügbarPodcasts kommen bald in die App. Installiere sie jetzt und erlebe als Erster einen ganz neuen Blick auf Podcasts
Folgen
Airing Challengers to Test AI Podcast Writers 20.11.2026 24:38
A five-part technical series tested whether rival AI models could replace a baseline scriptwriter when evaluating search primitives. Across fifteen drafts from Qwen 4 Max, Gemini 4 Carbon, Gemini 4 Hydrogen, and Kimi K3.1 Pro, zero scripts failed Claude Opus 5.5's evaluation, though every model ignored topic boundaries and trampled episode fences. By design, challengers aired regardless of baselin...
Between BM25 and Vector Search 20.11.2026 29:58
Information retrieval is not just a binary choice between BM25 keyword matching and dense vector embeddings. Across the middle rungs of the retrieval ladder, learned sparse methods like SPLADE expand vocabularies onto inverted indexes while late-interaction systems like ColBERT preserve token-level nuance. A look into how LSA's matrix factorization contrasts with modern neural embedders, and why s...
Reciprocal Rank Fusion and Reranking Eval 20.11.2026 25:56
Combining keyword search and vector retrieval requires balancing fundamentally different scales. Reciprocal rank fusion dodges score distortion by using rank positions, though score normalization can sometimes edge it out in relevance benchmarks. Meanwhile, adding an expensive cross-encoder reranker stage does not always guarantee better results and can even degrade retrieval quality depending on...
Why BM25 Still Beats Neural Search on Keywords 20.11.2026 28:29
Counting matching words is easy, but ranking them accurately without understanding meaning remains surprisingly resilient. Behind the thirty-year reign of BM25 lies a hand-fitted saturation curve designed to prevent single-word repetition from hijacking queries, paired with inverse document frequency from information theory. Examining document length normalization reveals how search engines mathem...
Exact Search Without a Vector Database 20.11.2026 29:21
Building a custom semantic retrieval layer over thousands of podcast transcripts reveals that vector databases are often premature. Designing around speaker turns, preserving attribution, and avoiding topic-averaging failures matter far more than index complexity. The engineering breakdown walks through turnchunk boundaries, local models like Google DeepMind's EmbeddingGemma 2 and Alibaba's Qwen3-...
Why Grep Is Not Enough for Search 20.11.2026 25:47
A scan across 22 million words takes less than a tenth of a second with modern tools like ripgrep, yet full-text retrieval still requires an entirely different data structure. Exploring the mechanics of grep, regular expressions, and benchmarks across source code and subtitles reveals where linear scanning hits a wall. The discussion covers why relational B-tree indexes fail on natural text and ho...
Teaching Vision Models to Spot Tiny Architectural Details 20.11.2026 28:53
Building an internal tool to look up specific architectural features like bay windows or mansard dormers requires more than just a standard image classifier. Public datasets like FacadeElements, LSAA, and WikiChurches lack fine-grained feature labels, but they provide a strong foundation for negative mining. Using a frozen pretrained backbone like DINOv2 or CLIP with a linear probe beats training...
How Accidental Dataset Formats Became AI Standards 20.11.2026 28:32
No standards body designed the dataset formats powering modern AI. They survived because a library shipped a loader or a benchmark needed to score fairly. This breakdown compares speech and image formats across file organization, geometry, and extra flags—from LJSpeech's pipe-delimited CSV and NVIDIA's NeMo JSONL manifests to Mozilla Common Voice, DeepSpeech, COCO, Pascal VOC, and YOLO. Along the...
Why Voice Agents Use Frameworks Like Pipecat 20.11.2026 30:38
A 200-millisecond pause can make a caller think the line is dead, yet voice agent builders still add orchestration layers like Pipecat on top of realtime APIs like Gemini Live. Breaking down the four-player market—Pipecat, LiveKit Agents, Vapi, and Retell—reveals what frameworks actually buy you: transport independence over WebRTC, modifying transcripts before an LLM sees them, non-conversation pi...
Bypassing Android's Built-In USB Audio Stack 20.11.2026 32:44
Android's built-in USB audio stack is a Linux inheritance topped with a consumer audio server built for ringtones and phone calls. Even with native AudioRecord or AAudio, vendor layers can cap captures at 16 bits, add phantom gain, and apply a hard limiter before an app sees a sample. The episode examines what AOSP actually promises for USB Audio Class 1, why Google's Oboe library relies on a Quir...
Tokyo's Standing Sushi Bars and Market Counters 20.11.2026 31:34
Cheap sushi in Tokyo does not mean compromised fish—the gap between budget and luxury is increasingly just seating and ceremony. This guide maps out a full day of eating well without spending a fortune, starting at Toyosu Market before ten and moving to Tokyo's roughly fifty standing sushi bars, or tachigui, including Uogashi Nihon-Ichi, Tachigui Sushi Akira, and Hinatomaru inside Tokyo Station. I...
The Ryukyu Islands Beyond Okinawa's Resort Strip 20.11.2026 30:04
The Ryukyu Islands are often sold as a single tropical resort, yet this thousand-kilometer arc holds six endangered languages, two major geographic breaks, and distinct histories from the Ryukyu Kingdom to American administration. A trip through the chain reveals Okinawa's centenarian village of Ogimi, analog whale scouts in the Keramas, Miyako's 152 cultural properties and mud-smearing Paantu fes...
Planning a Murakami Winter in Japan's Onsen 20.11.2026 31:26
Planning a winter trip to Japan around atmosphere rather than a checklist starts with six physical ingredients: old wood, a hearth, pressing forest, snow, a valley, and empty streets after dark. While Haruki Murakami's Killing Commendatore house sits near Tokyo, the true northern mood lives in Tohoku's snow-viewing baths. From Ginzan Onsen's new crowd-control rules to Akita's seven-inn Nyuto Onsen...
Why Tokyo Feels Like a Neighborhood of 37 Million 20.11.2026 30:31
Tokyo is home to thirty-seven million people, yet it walks like a collection of small neighborhoods. A full-day walking tour through Kuramae, Yanaka, and Shimokitazawa reveals the urbanism behind that scale. From former Edo-period rice granaries turned maker workshops to narrow shotengai shopping streets, Tokyo's permissive zoning and small-lot land ownership protect independent retail, slow down...
When Buying a Laminator Beats Paying Per Sheet 20.11.2026 27:15
Owning a home laminator breaks even against copy-shop rates after just a couple dozen sheets, with entry-level machines starting at $15 and bulk pouches costing 15 cents each. Yet buying one reveals bizarre market quirks: identical machines sell for nearly three times more at corporate suppliers like CDW than on Amazon, while machine-free self-adhesive sheets cost as much as a shop. Plus, a practi...
Why DIY Wardrobe Recording Booths Sound Boxy 20.11.2026 28:26
DIY recording booths built from cupboards, grow tents, and ceiling tracks can beat a noisy kitchen, yet acoustic physics predicts they should fail. Any enclosure between three and five feet resonates at 200 to 400 hertz, amplifying the exact warmth of the human voice into a hollow, boxy sound. Examining builds from Andy Chevalier, Randye Kaye, the Primacoustic FlexiBooth, and IKEA PAX wardrobes re...
Why Time Off at the Top Is Just Relocated Work 20.11.2026 31:08
At the highest levels of power, a day off is not a break from work—it is a relocation of attention. From intelligence briefings delivered to Hawaiian vacation rentals to the White House Staff Secretary's nightly binder, the presidency operates as a continuous system. The same pattern holds in the corporate world: a Harvard Business School study of 27 chief executives found they average 62.5 hours...
Two Building Laws That Shaped Seoul's Streets 20.11.2026 29:33
Seoul's most beloved neighborhoods feel like villages not by accident, but because of two building regulations. A 1970s daylight law forced four- and five-storey villas to taper like wedding cakes, while a 2002 parking rule hollowed out ground floors into dark alleys. Structured as a day-long walk, this tour traces those rules through Seongsu-dong's converted shoe factories, Euljiro's unmarked thi...
How Clippy Changed Software Help and Kept Coming Back 19.11.2026 20:14
Clippy was never supposed to be annoying; he was meant to be polite. Tracing the Office Assistant from Microsoft Bob and Stanford social-computing research to Eric Horvitz's Bayesian Lumiere project, this story explores how a bent wire with eyebrows became computing's most unforgettable interruption. Along the way, we look at Kevan Atteberry's 260 rejected designs, Microsoft's 2001 campaign to ret...
Weights, Biases, and What Fine-Tuning Actually Moves 19.11.2026 30:34
Every neuron in NVIDIA's 600-million-parameter Parakeet speech recogniser runs on simple school-algebra math: weights act as slopes on inputs, while biases act as resting offsets. Though biases make up a tiny fraction of a layer's parameters, they receive special treatment in PyTorch recipes and let a network respond to silence. From Frank Rosenblatt's motorized 1957 Mark I Perceptron to modern gr...
Why Fine-Tuning Doesn't Change a Model's File Size 19.11.2026 31:12
A fine-tuned Parakeet TDT 0.6B v2 speech model stays roughly 660 megabytes before and after training, yet its accuracy noticeably improves. Because a safetensors file is a fixed grid of allocated numbers rather than an empty warehouse, fine-tuning only nudges existing values in place. The improvement lives in the difference between the base and fine-tuned weights—a low-rank task vector—which can b...
How Fixed-Size AI Models Actually Fill Up 19.11.2026 29:02
A neural network's file size never grows during fine-tuning, yet its capacity isn't a fixed jar. Research by Zhang and colleagues showed networks can memorize 100% of random CIFAR-10 labels, while Chen, Chang, and Chaudhari found that useful learning capacity is only about 2% of total weights and shifts with sample count. The episode also unpacks double descent from Nakkiran and colleagues, where...
Why the Speed of Light Forces Time to Stretch 19.11.2026 28:43
If the speed of light is identical for every observer, then distance and time have to bend to keep the math intact. Using Einstein's own thought experiments, this conversation unpacks special and general relativity without heavy equations. It walks through Galileo's ship, the bouncing-pulse light clock, length contraction, and why the twin paradox isn't actually a contradiction. Along the way, it...
What Quantum Mechanics Actually Leaves Unresolved 19.11.2026 27:57
Quantum mechanics predicts the electron's magnetic moment to extraordinary precision, yet a 2025 Nature survey of 1,101 physicists found no majority agreement on what it says about reality. Stripping away mystical pop-science spin, this discussion separates the settled machinery—wave functions, superposition, interference, and decoherence—from the unresolved measurement problem. It also revisits t...
Entropy Is Counting Arrangements, Not Disorder 19.11.2026 29:10
Entropy is a precise count of microscopic arrangements, not a measure of messiness. Tracing the ninety-year history of the "disorder" metaphor from Rudolf Clausius and Hermann von Helmholtz to Henry Adams, this episode replaces it with relatable analogies that break everyday intuition. From separated Italian salad dressing and melting ice shards to Dan Styer's grid of squares and Jackson Pollock's...
Ähnliche Podcasts
Replaio ist kein Herausgeber von Podcasts; die Namen der Sendungen, Cover und Audioinhalte gehören ihren Autoren und werden über öffentliche RSS-Feeds verbreitet