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
How One Matrix Trick Powers Text, Image, and Voice LoRAs 16.11.2026 25:45
A single matrix equation—freezing a model's weights and learning a low-rank update as two tiny matrices—powers custom voices, image styles, and language models alike. The mechanism adds zero inference latency when merged, yet shrinks a 350-gigabyte GPT-3 checkpoint to 35 megabytes and a Stable Diffusion fine-tune to just over three. From rank-two bottlenecks and alpha scaling to why text, image, a...
Teaching Five AI Voices Israeli Hebrew 16.11.2026 28:53
A speech model can nail an English sentence and still butcher a single Hebrew word in the middle of it. Meorav Multi Voice solves that join by training Qwen3-TTS on 461 code-switched takes where only the Hebrew spans are graded, keeping the pronunciation without leaking the speaker's voice. Five synthetic accents from OmniVoice—American, British, and Australian—share that Hebrew training for about...
How Far Postgres Actually Scales in Production 16.11.2026 26:35
ChatGPT's backend serves 800 million users on a single unsharded Postgres primary, proving that database ceilings are rarely about raw capacity. Instead, production teams hit operational limits like vacuum, connection counts, and lock queues. This episode examines how OpenAI keeps one writer alive with nearly fifty read replicas and five-second schema timeouts, alongside why Instagram and Notion c...
How the Kaldi Team Pivoted to OmniVoice 16.11.2026 25:18
Speech recognition and speech synthesis have historically been separate worlds, yet the k2-fsa team at Xiaomi crossed over to build OmniVoice in just ten months. Tracing the lineage from Daniel Povey's Kaldi toolkit and Next-gen Kaldi's PyTorch stack to Han Zhu's ZipVoice, this deep dive unpacks how data pipelines proved direction-agnostic. It also breaks down OmniVoice's single-stage masked diffu...
When Speech Enhancement Erases Podcast Voice Character 16.11.2026 30:25
After Resemble AI retired its Chatterbox text-to-speech model in August, migrating the show's voices to OmniVoice seemed straightforward. But running the audio through NVIDIA's RE-USE speech enhancement model before training LoRA adapters backfired, stripping away the rough edges that gave the characters their personalities. The result was technically clean, completely forgettable, and a reminder...
When Merging LoRA Adapters Stops Being Exact 16.11.2026 23:40
Merging a LoRA adapter into its base model is exact in algebra, but finite-precision hardware tells a different story. Folding a half-mebibyte adapter into a 32-mebibyte matrix removes a serving branch rather than shrinking the model, and trades away swappability. The conversation unpacks how bf16 rounding, PEFT's hidden fp32 upcasting, and quantization introduce silent drift—as seen in a document...
How Label Studio's XML Configs and Storage Work 15.11.2026 26:48
Label Studio powers everything from bounding boxes to audio transcription with one design decision: an XML file that wires data types to annotation controls. This walkthrough covers how HumanSignal's open-source tool separates free individual labeling from paid team coordination, and what self-hosting to the Hugging Face Hub actually looks like. Expect the real architecture behind cloud storage sy...
Why CVAT Beats Generalists at Video Annotation 15.11.2026 21:55
CVAT has fewer GitHub stars than Label Studio, yet it remains the reference tool for video annotation. Originating inside Intel in 2017, its narrow focus on computer vision delivers keyframe interpolation, Outside-Merge-Split track primitives, and 3D point-cloud support that generalists lack. This deep dive examines how CVAT's architecture handles model-assisted labelling via Nuclio and CVAT AI Ag...
Small Labeling Tools That Outship Big Platforms 15.11.2026 22:27
Small, single-machine annotation tools are shipping faster than heavyweight platforms like Label Studio and CVAT. While LabelImg sits archived with 25,000 stars, desktop apps like labelme and X-AnyLabeling are adding SAM 3.1 video tracking, document parsing, and 3D point clouds in weeks. We explore six reasons solo developers and small teams skip big platforms—from local privacy and plain-text fil...
Quantization Quietly Deletes Fine-Tuned Skills 15.11.2026 25:17
A fine-tuned text-to-speech model called Meorav handled English-Hebrew code-switching at full precision, then lost the skill completely after quantization while still sounding normal. Exploring why compression quietly erases low-margin fine-tuned behaviors, this conversation unpacks outlier weights, signal-to-noise ratios, and why aggregate metrics like perplexity miss the damage. From Llama-3.1-8...
How Humans Actually Grade Generated Video 15.11.2026 26:07
Computer vision platforms like CVAT, Encord, and Label Studio offer keyframe interpolation, SAM 2 object tracking, and frame-level timelines for real footage. Yet nobody has built a widely-adopted annotation environment for grading generated video. Instead, AI video evaluation relies on thin arena voting pages like Artificial Analysis's Video Arena with its five-second watch gate, rubric experimen...
Where RLHF Preference Annotation Tools Stand 15.11.2026 24:48
Training-data tooling for RLHF and DPO preference labels is split across three shapes, and the open-source specialist slot is surprisingly quiet. Argilla is in maintenance mode after v2.8.0, while Label Studio reserves its safety and production RLHF workflows for Enterprise. Meanwhile, commercial platforms like Encord, SuperAnnotate, Labelbox, and Kili are shipping multi-turn conversation UIs, LLM...
How the Interface Shapes Data Annotation Tools 15.11.2026 20:29
Data annotation is repetitive work where the interface is the entire product, and the UI collapses as tasks grow more abstract—from heavy audio waveforms and bounding-box canvases down to two text panes and a radio button for LLM preference data. This tour maps the self-hosted landscape, including Label Studio, Argilla, doccano, Prodigy, and the archived labelImg, alongside cloud platforms like Sc...
Image Annotation When Experts Are the Bottleneck 15.11.2026 21:56
Annotating architectural styles in photos is far harder than audio or text because a label can hinge on a 40-pixel carved fleuron, and styles overlap across centuries. Drawing on the WikiChurches paper, the discussion explores why domain experts should curate prototypical examples—narrowing 117 classes to four—and pair bounding boxes with free-text descriptions. It also compares CVAT, Label Studio...
What Goes Into a Character-Voice LoRA Adapter 15.11.2026 29:31
A character-voice LoRA adapter can weigh just twelve megabytes, yet it is only half of a voice—reference-conditioned models like MOSS and F5-TTS still need a reference audio clip on every generation. This episode unpacks the de facto PEFT LoRA format, how rank controls capacity, and why training data requirements span from seven-second NanoVoice adaptations to twenty-hour F5-TTS runs. It also cove...
Merging LoRA Voice Adapters vs Keeping Them Separate 15.11.2026 21:54
Merging a LoRA voice adapter into a base model promises one standalone file, but merging into a 4-bit quantized base quadruples loss in practice. The math is exact in 16-bit full precision, yet scaling to fifty voices flips the economics: fifty merged models require 800 gigabytes, while one base plus fifty adapters needs just 24 gigabytes. We look at Hugging Face PEFT's merge_and_unload, why GGUF...
How Multi-Speaker TTS Models Keep Voices From Blending 15.11.2026 21:07
A native multi-speaker text-to-speech model stores every voice in one shared set of weights, using conditioning vectors and speaker-conditional normalization to keep them from blending into an average voice. This episode compares that coupled architecture—from Piper and Kokoro to Microsoft's VibeVoice—against training separate LoRA adapters for each speaker. It covers how vocoders, pause predictio...
How Water Dowsing Works Without Finding Water 15.11.2026 23:59
Water dowsers aren't faking it—the rod really moves, driven by the ideomotor effect, an unconscious muscle reflex that turns tiny wrist tremors into visible swings. Yet double-blind trials in Munich and Kassel proved dowsing performs no better than chance. So why has the practice survived for over four centuries? Because in most places, underground water is almost everywhere—and in karst landscape...
Why Optimizing Voice Quality Erases Identity 14.11.2026 21:04
Optimizing a voice-cloning pipeline for perceptual quality can quietly erase speaker identity. Using a failed LoRA adapter run as a case study, this breakdown traces how squared-error loss seeks the conditional mean, causing spectral over-smoothing. It examines why glottal qualifiers like rasp and breathiness are treated as defects by generative speech-restoration models, and how a four-step recip...
Four Layers Between Your Ears and Background Noise 14.11.2026 31:21
Two people can sit at the same table, hear the exact same background chatter, and have completely different experiences: one brain files it away automatically, while the other keeps re-delivering it. Rather than a single broken gate, auditory filtering is a four-layer stack. This breakdown explores the automatic "bouncer" for repeated sounds, the prefrontal cortex's active searchlight and alpha-rh...
What Breaks When You Train AI on Synthetic Taste 14.11.2026 27:18
Training a model on 200 to 500 synthetic movie plots sounds like plenty of data, given that the average Netflix Prize user rated 208 films. Yet sample size is not what breaks a personal taste model. Replacing real films with synthetic plots traps the model in a narrow pocket covering under 10 percent of real preference diversity, while personalized reward models often learn surface correlates rath...
Lossless vs. Hi-Res Audio and Whether Ears Can Tell 14.11.2026 21:08
A 24-bit, 192-kilohertz studio master streams at 4,608 kilobits per second, yet blind tests show trained listeners barely beat a coin flip when telling hi-res audio apart from CD quality. Breaking down Music Assistant's spec tables, this episode separates lossless from hi-res across FLAC, ALAC, and Tidal's abandoned MQA, and compares ceilings at Qobuz, Deezer, Apple Music, Amazon Music HD, and Spo...
Training Voice LoRAs on Filtered Zero-Shot Audio 14.11.2026 19:44
Six podcast voices now run on LoRA adapters trained for under two dollars each on OmniVoice. By generating synthetic lines with zero-shot cloning and using Whisper to discard every mispronounced take, the pipeline distills consistent speech and reliable tech acronyms without a full fine-tune. The catch is that an adapter cannot exceed its teacher: when Whisper cannot verify pronunciation, such as...
Mapping Foot Pedals by Device ID Instead of F13 14.11.2026 25:13
Mapping a USB foot pedal to an unused key like F13 sounds simple until macOS drops F21 through F24, KDE claims F13 by default, and games ignore it. A more reliable approach is per-device interception, which identifies hardware by vendor and product ID. Tools like keyd, kanata, Karabiner-Elements, and AutoHotInterception let a pedal behave differently per app without colliding with your main keyboa...
Reverse Engineering Windows Configurators on Linux 14.11.2026 27:56
Turning a Windows-only hardware configurator into a Linux equivalent splits cleanly down one line: managed .NET assemblies versus native binaries. Upstream ILSpy now ships cross-platform with ilspycmd, while dnSpyEx 6.6 detects Wine to let you debug and dump decrypted buffers at runtime. For native code, Ghidra's ReVa extension exposes small tools over MCP so an LLM can navigate cross-references,...
Similar podcasts
Replaio is not a podcast publisher; show names, artwork and audio belong to their authors and are distributed through public RSS feeds.