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...

Be sure to visit the podcast's website and support the creator: myweirdprompts.com

Author

Daniel Rosehill

Category

Technology

Podcast website

myweirdprompts.com

Latest episode

Nov 19, 2026

Where to listen?

Podcasts in the app Replaio Radio Coming soon

Podcasts are coming to the app soon. Install now and be the first to see a whole new take on podcasts

Get it on Google Play Install for free Android 10M+ downloads · 4.7+ rating iOS soon

Episodes

Plugging a Custom Voice Model into Wayland 13.11.2026

Most open-source voice projects assume you need speech recognition, leaving a developer who already trained their own models with nowhere to plug them in. Exploring tools like VoiceMode, Fono, VoiceUse, and the OpenAI Agents SDK reveals a gap between voice channels and actual Wayland desktop control. The conversation covers how to pair a fast command classifier with an LLM fallback, and why Waylan...

Why Home Voice Agents Still Need Home Assistant 13.11.2026

Home Assistant already speaks MCP in both directions, so demoting it to one tool behind a server-hosted voice agent is technically possible with frameworks like LiveKit Agents, Pipecat, and Speaches. The real bottleneck is the hardware on the ceiling: ESPHome satellites and the Voice Preview Edition expect Home Assistant to drive the pipeline. This look at the architecture compares full server-sid...

USB Far-Field Mics and the Echo Cancellation Catch 13.11.2026

A far-field microphone array is really a small computer with microphones attached, and the DSP chip behind them is the actual product. For a DIY Home Assistant and Whisper satellite, USB Audio Class boards like Seeed's ReSpeaker XVF3800, the older ReSpeaker v2.0, and the ReSpeaker Lite offer onboard beamforming and noise suppression, unlike generic AliExpress arrays or the PlayStation 3 Eye. The c...

The Wake-Word Race Inside Home Assistant Satellites 13.11.2026

Multi-room Home Assistant voice setups rely on cheap satellites—like the Voice Preview Edition or ESP32-S3-BOX-3—streaming audio to one central hub. Because wake-word arbitration is a pure network race with no volume comparison, the farthest satellite can win and silently drop your command. We unpack the real-world pitfalls of multi-room voice: why turning down sensitivity backfires, how power-sav...

Why Kitchen Smells Hijack ADHD Focus Like Noise 13.11.2026

Adults with ADHD often smell just as well as anyone else, yet brain imaging reveals their olfaction-associated regions handle those signals differently. This episode examines whether kitchen aromas like brewing coffee can hijack focus the same way background noise does. From the contested clinical status of sensory processing disorder and sensory over-responsivity to the changing-state effect in s...

Inside Laptop Microphones and Why They Hurt Dictation 13.11.2026

Almost every mainstream laptop—from the ThinkPad E15 to the Dell XPS 13 and Surface Laptop 7—ships with just two microphones, leaving only the MacBook Pro 14 with three. This deep dive explores whether you can spec a better, dictation-grade beamforming mic on a Framework laptop, why premium MEMS capsules cost just a couple of euros, and how call-focused audio processing like Voice Focus actually t...

When Fine-Tuning Over-Generalizes Instead of Forgetting 13.11.2026

Fine-tuning a speech model on IP addresses can make ordinary phrases like "ten one" transcribe as "10.1." Rather than catastrophic forgetting, this quieter failure is over-generalization, where a model applies a new surface pattern everywhere. The discussion unpacks why labels act as a teaching surface, how to catch formatting bleed using contrast sets and B-WER/U-WER splits, and what replay-share...

Five Projects to Build a Language Model From Scratch 13.11.2026

Training a language model from scratch used to require massive budgets, but a thousand-fold drop in compute costs has changed the math. A de facto five-step ladder—centered on Andrej Karpathy's micrograd, Zero to Hero lectures, minbpe tokenizer, microgpt, and nanochat—takes learners from a 100-line autograd engine to a working GPT. Along the way, building a byte-pair tokenizer explains why models...

Donating Bandwidth Without Becoming an Exit Node 13.11.2026

Donating bandwidth to help people bypass internet censorship doesn't mean turning your home connection into a risky exit node. Tools like Tor's Snowflake, Psiphon's Conduit, Lantern's Unbounded, MoaV, and the WEPN pod all act as entry relays, passing encrypted traffic to external servers. We break down how each tool works, why there's no native router package, and the real-world tradeoffs between...

Serving Parakeet Speech Models on Cheap CPUs 12.11.2026

Deploying custom speech models does not require keeping an expensive GPU running around the clock. Exploring NVIDIA's Parakeet TDT 0.6B architecture reveals why quantized INT8 weights run exceptionally well on everyday CPUs, how serverless cold starts shrink from minutes to two seconds using GPU memory snapshots, and where fine-tuning accents risk running into quantization errors. The trade-offs b...

When Real Engineering Teams Still Choose Encoders 12.11.2026

While decoder models dominate headlines, production engineering teams frequently rely on bidirectional encoders for classification, tagging, and retrieval. A close look at real deployments explains the economics: Roblox reached 25,000 moderation checks per second on CPUs, Itaú built BERTaú from scratch to gain 22 percent in retrieval, and PayPal weighed custom pretraining against standard fine-tun...

How Models Score Next Tokens All at Once 12.11.2026

Scoring every word in a vocabulary seems computationally ruinous, but language models and speech recognizers compute every token's probability simultaneously in a single matrix multiplication. Softmax merely normalizes raw scores, while the real cost scales with vocabulary size rather than repeated model evaluations. Examining architectures like NVIDIA's Parakeet reveals how token-and-duration tra...

Expecting Gidi, Act Three: Twenty-Eight Days 11.11.2026

An audio drama returns to the bench outside a shuttered café to track the real-life aftermath of Israel's election as of November 11. With the presidential mandate officially granted, a twenty-eight-day countdown begins to cobble together sixty-one parliamentary seats. The breakdown explores four precarious paths forward: minority rule reliant on abstentions, uneasy alliances with religious factio...

Why Speech Models Need More than Fifty Hours 11.11.2026

Folk wisdom suggests automated speech recognition fine-tuning stops paying off after fifty hours of audio, but model architecture tells a different story. The fifty-hour mark marks the ceiling for inexpensive language model adjustments, but stronger acoustic adaptation requires hundreds of hours. An exploration of code-mixed datasets, rare-word synthesis, Zipf's law in training data, and simple te...

Why AI Software Creation Threatens the SaaS Model 11.11.2026

Software stocks recently shed two trillion dollars over expectations that agentic AI will break traditional per-seat business models. At the same time, tools like Microsoft Copilot Code let non-programmers build working apps in hours, moving the real barrier to entry from coding syntax to system judgment. An examination of whether widespread creation outweighs SaaS disruption explores how open-sou...

Why People Pay to Drink Terrible Spirits 11.11.2026

Fermentation takes human ingenuity to unusual places, from industrial oddities to ancient living traditions. Across the globe, drinks like Chicago's famously bitter Malört thrive because grimacing became the product itself. In Thailand, SangSom dominates the domestic spirit market while masquerading as whiskey despite actually being rum. Meanwhile, Andean chicha de jora relies on saliva enzymes to...

How Economics Broke and Rebuilt Israeli Wine 11.11.2026

Israeli viticulture stretches back to biblical antiquity, but modern winemaking was built by Baron Edmond de Rothschild and the Carmel cooperative in the late nineteenth century. When geopolitical shocks erased export markets after World War I, the industry turned toward high-yield sweet sacramental wine to survive. A look at how monopolistic market structures dictated decades of poor wine, and ho...

Train Your Wine Palate with Side-by-Side Tasting 11.11.2026

Developing an intuitive palate for wine does not require sommelier certification, complex vocabulary, or expensive cellar management tools. The most effective approach relies on side-by-side comparative tastings, focusing on basic structural elements like acidity, tannin, body, and sweetness before attempting to identify obscure aromatic notes. By maintaining minimal, sustainable tasting notes and...

How Drip Irrigation Rebuilt Israeli Wine 11.11.2026

Israeli viticulture transformed from nineteenth-century sacramental bottles into a landscape of over three hundred boutique wineries. This guide tours key producers across diverse microclimates, from high-altitude volcanic basalt soils in the Golan Heights to ancient Judean Hill terraces. Special attention is given to the arid Negev Highlands, where wineries rely on precision drip irrigation, brac...

Why Hospital Wayfinding Relies on Environmental Design 11.11.2026

Poor navigation design carries hidden human and financial costs across complex spaces like healthcare facilities. Examining environmental wayfinding—from the colored floor lines and native animal motifs at Wolfson Medical Center's children's hospital to agent-based architectural modeling—reveals how spatial layout relieves cognitive load. The discussion highlights how landmarks, visible greenspace...

How Typesetting Shifted From Metal to Software 11.11.2026

For over five centuries, typesetting was a specialized manual craft defined by physical lead sorts, meticulous redistribution, and later, towering line-casting machines like the Linotype. But the arrival of PostScript, Aldus PageMaker, and digital fonts collapsed thousands of trade shops in just a decade. The trade's disappearance reveals what happens when a protected mechanical skill dissolves in...

Why Typography Science Keeps Disproving Type Rules 11.11.2026

Every word read in public spaces or on screens was drawn by a type designer, yet legibility science and typographic craft remain at an impasse. A look into how typefaces are built reveals that century-old rules about serifs often fail basic methodological scrutiny. When Applied Design Works created Atkinson Hyperlegible for the Braille Institute, they abandoned classical harmony entirely, intentio...

Why Prescribed Stimulant Withdrawal Stays Unstudied 11.11.2026

Immediate-release stimulants clear the body within hours, but the neurochemical adaptations built around daily dosing can linger for weeks. Despite clear label warnings for fatigue, dysphoria, and sleep disruption upon cessation, medicine treats therapeutic stimulant discontinuation as negligible compared to SSRIs. Exploring the pharmacokinetics of methylphenidate, prodrug ramps like lisdexamfetam...

Why wav2vec 2.0 Still Dominates Speech Downloads 11.11.2026

Modern speech leaderboards are dominated by giant models like NVIDIA Parakeet and Whisper, yet Meta's wav2vec 2.0 base checkpoint still commands roughly 1.8 million downloads a month. The architecture turned raw speech into discrete, tokenized units using clever self-supervised pre-training with minimal labeled audio. Community fine-tunes like XLSR variants and paralinguistic encoders explain how...

Why Speech Recognition Models Struggle With Accents 11.11.2026

Speech-to-text models still benchmark against standard American and British English, treating other accents as margins of error. Longitudinal audits show newer systems often widen performance disparities, while decoder models like Whisper frequently insert hallucinated text rather than omit difficult words. Meanwhile, consumer interface options like locale dropdowns and prompt instructions fail to...

Listen to the My Weird Prompts podcast in Replaio

Radio and podcasts in one app - free, with no sign-up. Install today and do not miss the launch

Get it on Google Play

Replaio is not a podcast publisher; show names, artwork and audio belong to their authors and are distributed through public RSS feeds.