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
Podcast website
Latest episode
Nov 19, 2026
Where to listen?
Podcasts in the app Replaio Radio Coming soonPodcasts are coming to the app soon. Install now and be the first to see a whole new take on podcasts
Episodes
USB Audio Limits on Android for Voice Models 06.11.2026 19:51
Android caps standard USB audio at 48 kilohertz, creating technical trade-offs when recording audio for speech recognition versus voice cloning. Speech models crave authentic, messy background environments, while voice clones demand strictly noise-free source material. This deep dive inspects the mobile recording pipeline, exploring how custom USB drivers bypass operating system limits, why spectr...
Why to Call a Headset Specialist Instead of Amazon 06.11.2026 19:36
Buying a headset or dictation setup online often leads to drawers full of ill-fitting gear and incompatible cables. Independent business headset shops and dictation dealers in the US, the UK, and Austria offer direct phone lines, trial units, and tailored advice. Learning what questions to ask sellers about return policies, microphone placement, and software compatibility reveals whether a shop of...
Why Whispering Breaks Speech Recognition Physics 06.11.2026 25:41
Whispering is far more than quiet talking: the vocal folds stop vibrating, pitch completely disappears, and vowel formant frequencies shift upward while vowel space compresses. This deep acoustic change turns speech into turbulent noise, stripping key periodicity cues that separate sounds like 's' and 'z'. Speech recognition models manage the shift surprisingly well despite lacking training data,...
What a VPN Node Can Actually See 06.11.2026 26:28
VPN servers terminate incoming tunnels and originate brand new connections toward websites, decrypting the outer transport layer entirely. While the server cannot access TLS-encrypted page contents, it sees cleartext IP addresses, port numbers, DNS requests, and SNI hostnames. Examining architectural compromises, log subpoenas like the Ryan Lin case, and snapshot audits reveals why connection meta...
Why Commercial VPNs Cannot Be End-to-End 06.11.2026 22:34
Standard commercial VPNs protect against snooping on local networks, but they merely shift trust from an internet service provider to the VPN operator. The provider inherently holds session keys to route traffic to the open web. Architectural alternatives like multi-hop relays and mixnets split identifying information across independent parties, eliminating reliance on a provider's promised no-log...
Why Computer Science Publishes in Conferences 05.11.2026 27:31
Computer science treats conference proceedings, rather than traditional journals, as its permanent archival record of research. This episode examines the mechanics behind academic publishing, the reality of acceptance rates, and the noisy nature of peer review revealed by experiments at major venues like NeurIPS. It explores why fields like speech recognition accept nearly half of submissions whil...
Why Switching Off One Attention Head Changes Everything 05.11.2026 21:35
Mechanistic interpretability tools can switch off a single attention head or layer during inference to change how a model behaves. Masking specific decoder heads in Whisper dramatically reduces hallucinations, while guardrail removal tools like elder-plinius's OBLITERATUS use ablation to map and project out refusal directions. Examining work from Calm-Whisper to Foo and Lee clarifies what an atten...
Why Tiny Neural Gates Replaced Voice Activity Gates 05.11.2026 21:38
Modern voice activity detection has quietly shifted from amnesiac volume gates to sub-megabyte neural networks with running memory. The ecosystem centers around open-source contenders like Silero, TEN VAD, and FireRedVAD, alongside commercial on-device options. While standard neural detectors excel against background air conditioners, separating intended speech from ambient chatter requires specia...
Voice Activity Detectors Need Adaptive Thresholds 05.11.2026 24:58
Voice activity detectors like Silero rely on fixed sensitivity thresholds that break when room acoustics or microphones change. Default settings cause clipped word boundaries, false triggers on HVAC systems, and downstream speech recognizers that hallucinate sentences from background noise. Tuning requires adjusting six post-processing parameters—including hysteresis and minimum silence duration—r...
Conferences Rule Machine Learning, Not Journals 05.11.2026 27:32
In machine learning, flagship conferences serve as archival records rather than stepping stones to academic journals. From the massive scale of NeurIPS, ICML, and ICLR to specialized venues like CVPR and ACL, acceptance rates have converged around twenty to twenty-seven percent. This episode explores how open-access journals like JMLR and JAIR fit into the ecosystem, the astronomical fees of legac...
Why Modern AI Models Cut Most Attention Layers 05.11.2026 25:56
Standard softmax attention forces models to store every past key and value, making multi-turn chat memory balloon past model weights. Production models are quietly replacing full attention with layered hybrids instead of abandoning it completely. Gated delta rules compress history into fixed recurrent states, learned sparse indexers read only relevant tokens, and latent bottlenecks shrink what rem...
Teaching Speech Recognizers to Stay Quiet 05.11.2026 26:57
Voice activity detection usually acts as a protective fence in front of automatic speech recognition models, dropping silence before hallucinations occur. While external gates remain the default engineering fix, internal model weights already track the difference between speech and ambient sound. Discover how surgical fine-tuning of specific attention heads, activation steering, and synthetic posi...
Why Training Speech Recognition From Scratch Fails 05.11.2026 22:48
Training speech recognition models from random weights requires learning acoustics, alignment, and language simultaneously from paired audio. While large language models train on trillions of free text tokens, a speech corpus provides only about nine thousand words per recorded hour. Fine-tuning an existing base model vastly outperforms scratch training below hundreds of hours. Going entirely from...
When Adding a Reranker Hurts Search 05.11.2026 19:40
Cross-encoders are often billed as a painless, universal accuracy boost for RAG pipelines and search boxes. Yet benchmark results range from a fifteen-point lift down to an outright regression, and some experiments show latency exploding two-thousand-fold for no measurable gain. The discrepancy hinges entirely on the retriever's recall curve and token economics. Here is how to diagnose whether an...
Why Easy-to-Read Fonts Must Break Harmony 05.11.2026 22:26
Visual harmony makes modern typefaces look beautiful, but uniformity causes letters to blur together for low-vision readers. Exploring the creation of Atkinson Hyperlegible by Applied Design Works for the Braille Institute, this discussion unravels the critical difference between legibility and readability. It examines why deliberately inconsistent letterforms succeed for low acuity, how kerning m...
Why Watching a Sleeping Child Defeats Attention 05.11.2026 20:41
Continuous vigilance fails rapidly when supervising a low-activity scene, an effect documented from wartime radar operators on the Mackworth Clock to modern lifeguards. Human attention drifts within thirty minutes under conditions of cognitive underload, making uninterrupted voluntary surveillance impossible. To monitor a feverish child effectively, parents should discard continuous glancing in fa...
Why the Gaussian Distribution Fails in Plain Sight 05.11.2026 30:32
The Gaussian distribution underpins everything from least-squares regression to modern AI diffusion models, controlled entirely by just two parameters: mean and variance. But its aggressive, thin-tailed exponential decay makes extreme real-world events vanish mathematically. Exploring the mathematical power of the central limit theorem alongside the subtle dangers of confusing bell curves with hea...
Why Synthetic Fine-Tuning Data Needs Tooling 04.11.2026 27:58
Naive generation loops make synthetic fine-tuning data prone to over 30 percent duplication, judge bias, and model collapse. Examining findings on knowledge collapse and simulated user demographics reveals that synthetic generation acts as a quality multiplier rather than a data curation substitute. Dedicated frameworks like NVIDIA's Data Designer and topic-tree libraries replace prompt-and-pray A...
Where to Learn LLM Training Without Generic Courses 04.11.2026 20:01
Training and fine-tuning language models has become cheap enough to learn from scratch, but high-quality education remains surprisingly fragmented. While open-source lectures like Stanford's CS 336, Karpathy's repos, and Hugging Face offer deep technical grounding, the scarce resource is compute rather than tuition. This guide maps out the four tiers of self-education—from building transformers an...
The Unsung Builders of Modern Speech AI 04.11.2026 26:25
Speech recognition commands billions in venture funding and fills massive academic conferences, yet its foundational architects remain largely anonymous to the public. This episode maps the human ecosystem behind modern audio AI, tracing the toolkits and researchers that made ubiquitous transcription possible. From Daniel Povey's grueling work on Kaldi to Georgi Gerganov's whisper.cpp and Hervé Br...
Turning Open-Source Triage into Subscription Credit 04.11.2026 26:54
Heavy users of semi-open developer tools often act as unpaid triage teams and QA departments, diagnosing critical bugs and shaping product roadmaps while paying full subscription rates. Examining the FeedCoin proposal reveals that while companies like Loophole, Orkosi, and Nexu offer fragmented credits for reports or code, an interoperable renewal discount standard does not exist. Exploring Turso...
Choosing Between MCP Servers, Agent Skills, and CLIs 04.11.2026 27:52
Developers are wrestling with how to connect AI agents to tools, debating whether to wrap APIs in command-line interfaces, lightweight agent skills, or full Model Context Protocol servers. An evaluation by Arize reveals that while correctness stays nearly identical across all approaches, cost and latency diverge by a factor of six on complex tasks. The breakdown covers context token bloat, naive A...
Who Actually Won the Open Weight Race 04.11.2026 26:01
Public download counters tell a surprisingly different story from production deployment data. While a modest 2022 sentence embedder tops all-time downloads thanks to automated library defaults, enterprise production numbers tell an entirely different tale, led by Google's Gemma and Microsoft's Phi. The metrics behind open-weight model adoption expose the massive gap between headless infrastructure...
Why Recursive File Watching Strains Linux Memory 04.11.2026 25:18
Linux file monitors like inotify make instant automation possible, but watching entire trees reveals an expensive kernel secret. The kernel provides no native recursive watches, forcing tools to spawn individual monitors per subdirectory. Each watch consumes unswappable kernel memory, creating hidden memory pressure on large directories while write paths silently pay the cost of constant checks. E...
Which AI Models Actually Last 04.11.2026 21:02
API retirement notices are arriving at a relentless pace across major artificial intelligence labs. This exploration contrasts the lifespans of legacy models across OpenAI, Anthropic, and Google, showing how providers handle deprecations differently. From OpenAI's shifting alias strategies and mass shutdowns to Anthropic's structured schedules and preservation interviews, the discussion examines h...
Similar podcasts
Replaio is not a podcast publisher; show names, artwork and audio belong to their authors and are distributed through public RSS feeds.