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 Word Error Rate Lies to Us 14.10.2026 26:49
Word error rate has a formula that fits on a napkin, yet it's the number that ends up in every headline about speech recognition. This episode digs into the mechanics: how edit distance at the word level gets computed, why WER has no upper bound, and when character error rate is the more honest lens. Then it goes after everything a single score buries — normalization choices that swing results by...
What Your GPU Is Actually Waiting On 14.10.2026 27:27
You've got the model loaded, VRAM to spare, and the GPU reports near-total utilization — so why is generation crawling? This episode pulls apart the single number everyone screenshots and looks at the physical mechanisms hiding underneath it. VRAM capacity is a feasibility check, not a performance prediction; decode is a streaming memory workload that runs sixty to two hundred times to the memory-...
There's No Such Thing as Tokens Per Second 14.10.2026 30:16
Everyone quotes tokens per second and time to first token as if they're the whole picture. They're not — an LLM serving system doesn't have a performance number, it has a surface. This episode breaks down prefill versus decode, why one is compute-bound and the other memory-bandwidth-bound, the latency and throughput metric families that quietly disagree with each other, and how to design a benchma...
What RTF Actually Tells You About Speech Models 14.10.2026 26:40
Most speech recognition coverage leads with accuracy — word error rate, character error rate, how close the transcript got. But that's not the number that decides whether a system is usable. This episode digs into the computational side of ASR benchmarking: real-time factor, why an RTF above one means broken rather than merely slow, and why the mean latency on your slide describes a request that n...
When Governments Answer FOI Requests With Silence 14.10.2026 28:03
Freedom of information laws stretch back to Sweden in 1766, and roughly 140 countries now have one. But a 2025 study filed identical requests in 76 countries and found that 38% drew no substantive reply — not a denial, just silence. This episode traces the history from Peter Forsskål's 1766 Freedom of the Press Act through the post-1990 adoption wave, compares how laws differ on who can request, w...
The Scientific Plugins Hiding in Plain Sight 14.10.2026 21:43
ChatGPT's plugin directory now lists 5,723 plugins with 238,000 callable tools, yet only 52% have useful descriptions. The leaderboard rewards commercial SaaS with clean schemas — Ashby, Supabase, Miro, Stripe — while scientific tools like Wolfram and Consensus don't crack the top rankings. Meanwhile, the genuinely capability-expanding drug discovery work — LIDDIA, MADDIE, DrugPilot — lives in arX...
Will ChatGPT Eat Your Grocery App? 14.10.2026 22:07
MCP turned the AI assistant into a universal front end, and OpenAI's Apps SDK is built on it. So why is a grocery list the battleground? This episode works through a prediction that conversational AI becomes the new interface for everything — and that vendors, not your employer, end up selling you the bolt-on. We trace the MCP timeline from Anthropic's open-source release through the ChatGPT App D...
Work Lights: Lumens, Lux, and the Specs That Matter 14.10.2026 23:19
If you've ever bought a work light on its lumen rating and been disappointed, you're not alone — lumens measure light leaving the fixture, not light landing on your work surface. This episode works through the specs that actually predict whether a light will do the job: lux and illuminance at working distance, beam angle, runtime per setting, thermal step-down, CRI, colour temperature, and the IP...
Layered Safeguards and the Math Behind Them 14.10.2026 27:59
Three safeguards that each work ninety percent of the time should fail together only once in a thousand tries. That arithmetic is correct, and it's also where the trouble starts. This episode works through the real mathematics of layered redundancy, why the independence assumption behind it so often breaks, and what reliability engineers call common cause failure. From the beta-factor model and th...
How Militaries Decide What to Buy 14.10.2026 36:30
A billion shekels for twelve Apache helicopters, funded partly by cutting a five-year plan for Arab communities, eleven days before an election. But the real question isn't why twelve Apaches — it's how any military decides what it needs, in what quantity, and how urgently. This episode uses Israel's October 2026 approval as a doorway into the American requirements system: JCIDS, the Joint Require...
Inside Malmab: Israel's Most Secretive Security Body 14.10.2026 25:58
Most people know Mossad, Shin Bet, and Aman — but there's a fourth body inside Israel's Defense Ministry that almost no one can name. Malmab, the Director of Security of the Defense Establishment, protects the country's most sensitive nuclear and biological facilities, and its practices reveal how governments actually organize information security when a breach could be catastrophic. From clearanc...
The Reasonable Person Is a Legal Fiction 14.10.2026 19:29
Reasonableness is the load-bearing wall of the legal system — negligence, self-defence, harassment law, and constitutional equality all run on it. But the "reasonable person" is a fiction, and what counts as reasonable is shaped by class, culture, gender, and professional habit. Drawing on Mayo Moran's work on the reasonable person and Victoria Nourse's argument that the standard was never a perso...
Self-Hosting Your Search Stack: SearXNG, Crawl4AI, and the Real Bill 14.10.2026 28:42
Daniel wants to know if you can bolt together a search-and-grounding stack from open-source parts and never rent a search API again. On the table: SearXNG for discovery, Crawl4AI for extraction, and a TypeScript aggregation library for grounding — plus how you'd actually run the whole thing on Modal. The honest answer is yes, with one large asterisk: the bill you save is not the bill you pay. You...
Inside Deep Research Agents: Claims, Citations, and the 84.7% Orchestrator Problem 14.10.2026 23:36
What actually separates a research agent that grounds its claims from one that piles up web searches and writes a confident summary? This episode breaks down the four-stage deep research pipeline — planning, retrieval, verification, synthesis — and the three failure classes that survive every surface check: unsupported claims, silent coverage gaps, and unbounded drift. The fix isn't more search to...
LangGraph, LangChain, and Deep Agents in 2026 14.10.2026 22:12
LangChain spent 2026 hardening the parts of the agent stack nobody demos. LangGraph 1.2 cut checkpoint storage by 41x with DeltaChannel, added per-node timeouts, error handlers, and graceful shutdown. Deep Agents v0.7 deleted 65% of its base input tokens — including the entire base system prompt — and moved orchestration out of the tool-call loop and into code the model writes itself. Meanwhile La...
Who Actually Trades on Polymarket? 14.10.2026 26:26
Polymarket markets look like a crowd forecasting the future. The data says otherwise: a tiny cohort of professional wallets generates the overwhelming majority of volume, while millions of retail accounts slowly pay for the forecast. This episode digs into who actually trades, what liquidity and volume on the screen really mean, and why the platform's fee revenue comes almost entirely from crypto...
Inside Humanity's Last Exam: 2,500 Questions AI Can't Ace 14.10.2026 32:50
Humanity's Last Exam launched in January 2025 with 2,500 expert-written questions across 100+ subjects — and the best model in the world scored 2.7%. Eighteen months later, the leaderboard leader sits at 54.8%. Is that a triumph of capability or a confession about the test? This episode digs into the origins and mechanics of HLE: Dan Hendrycks and the Center for AI Safety, the crowdsourced pipelin...
How a Reward Model Actually Trains an LLM 14.10.2026 28:58
Everyone draws the reward model as a smiley face between "human feedback" and "better model" — but what is it actually doing? This episode takes the reward model apart: how it's built by cutting the unembedding layer off an SFT checkpoint and bolting on a single linear head, why it's trained on preference trios with a Bradley-Terry loss for exactly one epoch, and why a 6B critic was used to steer...
Keeping the Rollout Engine Fed: Inside RLHF Pipelines 14.10.2026 24:48
Daniel wrote in with a practical question: he understands RLHF concepts, but now he has to stand up a real pipeline. What does it actually take? This episode walks the full three-phase shape — supervised fine-tuning, reward model training, and the policy optimization loop — and lands on a counterintuitive conclusion: the algorithm is the commodity, and the job lives in the rollout engine. We compa...
Tourniquets From Belts and Socks: What Actually Works 14.10.2026 26:07
When a group of students was caught in an explosion on the Ramim Ridge in the Upper Galilee, there were no commercial tourniquets on scene — so they improvised from belts and socks, in one case coached over a video call by a father simultaneously scrambling rescue services. This episode digs into what the major first aid and military organizations actually recommend about improvised tourniquets, w...
Service Workers: The Proxy That Never Sleeps 14.10.2026 27:57
Service workers sit between your page and the network as a programmable proxy, running on their own thread with no DOM access and no window. They can wake up when no tab is open, intercept every request under their scope, and serve content from a cache you control — which is exactly why a deploy can appear to do nothing at all. This episode walks through the full lifecycle: download, install, acti...
Five Minutes Reading Your Dictation Prompt 13.10.2026 27:14
Daniel sent in a general speech-to-text cleanup prompt built to fix the failure patterns he actually hits, and asked for a full read, a critique, and improvements. So the prompt is the artifact: we read it first, argue about it second. It does two genuinely unusual things — executing spoken self-corrections and stripping out speech aimed at someone nearby — but it also assumes every error is a wor...
The Missing Middle: Content APIs and ISR for a 6,000-Episode Catalog 13.10.2026 32:14
Daniel has built a 6,000-episode podcast with no admin panel — just Claude Code and a custom admin MCP talking to Neon. But Neon's docs explicitly warn against connecting MCP agents to production databases, and Vercel's 45-minute build ceiling looms over every deploy. This episode works through the architecture question: does the catalog need a bespoke content API, or do Vercel's ISR with tag-base...
Android Notifications: PWA vs Native, What You Actually Get 13.10.2026 22:10
Daniel asked us to look at Android notifications from a developer's seat, and the honest answer isn't "native wins." It's about where control actually sits in the stack — and it flips depending on whether you mean the developer's control or the user's. Native Android hands you a rich, opinionated channel architecture with importance levels, categories, and vibration patterns, then deliberately fre...
One Regex Rule Worked, One Was Impossible 13.10.2026 27:19
A listener wrote one text-replacement rule in an Android dictation app and it erased every "um" and "uh" from his transcripts. Then he wrote a second rule about capitalization, and it failed — not because it was fiddly, but because it's provably impossible. This episode uses that gap as a case study in what regular expressions actually are: a 1951 notation from Stephen Cole Kleene, first implement...
Similar podcasts
Replaio is not a podcast publisher; show names, artwork and audio belong to their authors and are distributed through public RSS feeds.