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...
Vizitează neapărat site-ul podcastului și susține-i creatorul: myweirdprompts.com
Autor
Daniel Rosehill
Categorie
Site-ul podcastului
Cel mai nou episod
21 nov. 2026
Unde asculți?
Podcasturi în aplicație Replaio Radio În curândPodcasturile ajung în curând în aplicație. Instaleaz-o acum și fii primul care descoperă o abordare complet nouă a podcasturilor
Episoade
Why Inference Engines Change Benchmark Scores 24.10.2026 28:41
Switching an inference runtime can drastically alter an open model's performance without touching its weights. Examining engines like vLLM, llama.cpp, and LMDeploy reveals how subtle implementation details—including tokenization quirks, chat templates, layer sliding window attention, and speculative decoding—distort both benchmark evaluations and real-world behavior. The discussion covers the hidd...
When Your Fine-Tune Isn't Done Learning Yet 24.10.2026 23:05
Your fine-tune's loss curve is dropping beautifully. So why is it still picking the wrong citation, drifting back into base-model habits halfway through an answer, or transcribing the same words wrong as the stock model? This episode works through what undertraining actually looks like in fine-tuned models — Whisper, LLMs, and speech models adapted to a single voice — and how to tell unfinished le...
Inside the Mixture of Experts Router 24.10.2026 29:57
Mixture of experts models are often pitched as an efficiency trick — activate a fraction of the network, get more capability per compute. But the interesting part is the gate itself: what happens when a token arrives and only a handful of experts get picked? This episode digs into the routing mechanism, from the small learned linear map that scores each expert to the top-k selection that decides w...
Training a Model on Your Own Taste 24.10.2026 21:55
Daniel wants a model that generates episode titles and descriptions, learns from side-by-side button presses, and improves each checkpoint. Herman walks through what that pipeline actually is: an iterative preference optimization loop built on DPO (Direct Preference Optimization), with an implicit reward derived from a frozen reference model rather than a separate reward model or RL rollout. Along...
Why Your Synthetic Data Isn't as Diverse as You Think 24.10.2026 21:44
Synthetic data generation has a hidden problem: each example looks fine, but together they teach your model a narrow slice of reality. This episode digs into cross-batch mode collapse, why temperature and personas underdeliver, and what actually expands coverage — from scenario planning to multi-stage generation pipelines. If you're generating training data at scale, the bottleneck isn't quality p...
Gold vs Silver Training Data: What Makes Labels Trustworthy 24.10.2026 21:49
Gold data means human-verified. Silver means machine-annotated. But what actually separates them, and how much human checking do you need before a generated batch becomes trustworthy? This episode digs into the staged workflow of generating data with an LLM, correcting a small sample by hand, then scaling up — and why a single correction can reveal you've been generating the wrong kind of example...
Testing Synthetic Data Generators Before You Scale 24.10.2026 22:41
Picking a synthetic data generator by model reputation is a trap — a brilliant problem-solver can produce terrible training material. This episode digs into AgoraBench, a benchmark that trains the same student model on each generator's output and measures how much of the performance gap it closes. We cover why zero-seed generation is a different problem from transforming real data, how creating ne...
Which AI Models Actually Generate Good Training Data? 24.10.2026 21:45
Synthetic training data lives or dies on the generator behind it. This episode compares the instruction models actually available for the job — Qwen, DeepSeek, Llama, Mistral, and Gemma on the open-weight side, and the Claude Opus/Sonnet/Haiku and GPT-5.6 Sol/Terra/Luna tiers on the closed side — with real pricing per million tokens and what each licence actually permits. Then it digs into the sco...
Small Models, Big Questions: When Cheap Fables Get Expensive 24.10.2026 21:43
Small language models can generate three million fables at thirteen and a half cents per thousand, but the real question isn't price — it's whether the generator understands what it's making. This episode maps the parameter tiers from three billion upward, separating tasks where small models excel (controlled transformations of verified facts) from tasks where they quietly fail (inventing and labe...
Synthetic Logs Without Leaking Your Clients 24.10.2026 24:27
Daniel has real agent-workflow logs full of his own projects and wants synthetic versions that preserve the rhythm and ambiguity without advertising what he's working on. Corn and Herman walk through what purpose-built privacy tooling actually does that a plain rewrite prompt doesn't: which libraries preserve information dependencies, where differential privacy guarantees really enter a pipeline,...
When Your AI Agent Has Too Many Skills 24.10.2026 21:23
Installing a Claude Code skill is never free: its name and description take up permanent space in the agent's context, even if it's never used. As skill libraries grow, agents end up spending their working memory on metadata instead of the actual job. This episode digs into progressive disclosure — the three-tier loading architecture behind Agent Skills — and why it postpones the cost without elim...
Phone Mic Beats a Dictation Mic: How? 24.10.2026 22:30
A dedicated dictation microphone losing to a phone is expensive feedback — so what happens when you try to keep using the phone's mic at your desktop? This episode digs into the gap between "sending audio" and "sending the audio that actually won the benchmark," why Android's recording path matters as much as the hardware, and what close-range measurement data says about distance, angle, and handl...
Fibromyalgia's 26 Genetic Clues 23.10.2026 23:55
Fibromyalgia affects millions, yet it was long dismissed as a diagnosis of exclusion. A July genetic study of 2.5 million people identified 26 associated genome regions and pointed squarely at the central nervous system, while a major drug review this month delivered far less encouraging news. This episode explores what nociplastic pain actually means, why the HTT gene link surprised researchers,...
Daycare Germs: What Immunity Actually Means 23.10.2026 22:13
Building immunity" gets said like a consolation prize every time a toddler brings home another cold. But what physically changes after an infection, and why doesn't it feel like progress when the whole household is sick? Drawing on a Clinical Microbiology Reviews paper, a Québec household transmission study, and American Academy of Pediatrics figures, this episode unpacks what immune memory actual...
Why a Cold Makes Your Whole Body Ache 23.10.2026 22:25
You caught a cold in your nose — so why do your legs ache like you ran a marathon? This episode unpacks the biology of sickness behavior: how cytokines like interleukin-6 and tumor necrosis factor turn a local respiratory infection into whole-body exhaustion, fever, and pain. We trace the signaling routes from airway to brain, explore why your muscles actually change during illness, and ask the un...
Why Synthetic Speech Needs Fake Room Noise 23.10.2026 20:38
Synthetic speech can stop completely between sentences, leaving gaps of true digital black that sound unnatural on playback. This episode digs into why engineers fill those holes with room-tone beds, how a real fix for this show landed around minus sixty dBFS on a phone, and why a constant quiet background beats a perfectly clean pause. Along the way: what room tone actually contains, why filmmake...
Voice Design: Casting Ogres and Radio Voices 23.10.2026 23:59
Anyone can make a voice say a sentence. Designing one you can listen to for hours is a different problem. This episode breaks down how modern text-to-speech voice design actually works: the two families of methods (describe-and-preview casting versus reference-audio speaker embeddings), what the ElevenLabs sliders really do, why identity and colour are separate signals, and what research on profes...
What 12GB of VRAM Can Actually Do Now 23.10.2026 22:31
A Radeon RX 7700 XT owner gave up on running models locally six months ago — the context window filled instantly and the model lost the plot. But the models changed underneath him. This episode walks through the new memory arithmetic: why a 35B mixture-of-experts model with only 3B active can run on a 12GB card, how cache costs dropped roughly seven-fold between model generations, and what that me...
Cork Chippers, Rooster Potatoes, and the Double Fry 23.10.2026 20:37
Jackie Lennox's closed on Bandon Road after 73 years, and Cork has opinions. Plumber and Cork native Roy joins the show to explain what actually makes a proper chipper chip — the right potato, the double fry, beef dripping versus vegetable oil, and why a paper bag beats a polystyrene box. Along the way: the Italian families who built the Irish chipper, the science of cold-storage sugar in potatoes...
Sarah Tuttle-Singer: The Writer Who Lives Inside the Story 23.10.2026 19:11
Sarah Tuttle-Singer grew up on Yiddish lullabies and civil rights anthems in Venice Beach, the daughter of a Freedom Rider who spied on the Klan before auditing Los Angeles City Hall. Today she lives inside the walls of Jerusalem's Old City, writing first-person dispatches from the middle of the story — including the courtyard gas explosion that killed her neighbor. This episode traces the family...
Reading the Iran Signal: Jets, Leaks, and the Midterm Deadline 23.10.2026 21:27
Dozens of Israeli fighter jets and refueling tankers reportedly flew roughly 1,300 kilometers toward Iran through Syrian and Iraqi airspace — and turned around before reaching Iranian territory. The same week, Pete Hegseth made an unannounced visit to the Negev to meet Netanyahu, and Trump's written pledge not to attack Iran before the midterms hung over everything. Every story is "reportedly," "s...
Marcus James: Jerusalem's Spiritual Mayor 23.10.2026 20:54
What does it take to become a city's "spiritual mayor" without winning a single vote? Daniel makes the case for his friend Marcus James — raised in Santa Barbara, a Kerouac-reading hitchhiker who studied philosophy, converted to Judaism in 2020, and moved to Israel in 2022. The episode traces his unlikely path from Grateful Dead shows to a Jerusalem pub called HaTaklit, and asks whether a man with...
How Many Epochs Does Fine-Tuning Actually Need? 23.10.2026 20:20
Almost nobody picks an epoch count in advance — they set a ceiling, evaluate often, keep the best checkpoint, and the epoch count is whatever that turns out to be. This episode works through two concrete cases: a Whisper speech fine-tune scored on word error rate, and an instruction-tuned language model using SFT or LoRA. Along the way: why steps and tokens beat epochs as a unit, why validation lo...
Why Fine-Tuning Runs Can't Always Resume 23.10.2026 20:38
Daniel has been fine-tuning NVIDIA's 600M-parameter Parakeet on his own read-aloud recordings, English with Hebrew mixed in. Each run costs about eighty cents and twenty-five minutes on a single L40S, so retraining from the stock model is cheap — but that raises a tangle of questions: what's actually inside a 6.9 GB training checkpoint versus a 2.5 GB finished weights file, why resuming a finished...
How Many Examples Does Fine-Tuning Actually Need? 23.10.2026 21:23
A listener has 150 hand-labelled dictated prompts and wants to know if that's a start or a third of what he needs — and his architect wife has the same question about photographs of buildings. This episode works through the real answer: count examples per label, not in total; treat vendor minimums as proof-of-concept ceilings rather than floors; and measure your own learning curve instead of trust...
Podcasturi asemănătoare
Replaio nu este editorul podcasturilor; numele emisiunilor, coperțile și materialul audio aparțin autorilor lor și sunt distribuite prin fluxuri RSS publice