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 20, 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

Substitution Anonymization: Privacy Without Utility Loss 27.04.2026

When stripping PII from training data, redaction destroys utility—but substitution preserves it. We explore how small local models can generate synthetic voice notes and calendar entries with near-perfect privacy recall and minimal accuracy loss, and why this matters for anyone building classifiers on sensitive data. Episode #620605 — open it directly at myweirdprompts.com/620605

When AI Chatbots Leak Your PDFs via Public S3 Buckets 27.04.2026

A user uploaded a sensitive PDF to a major AI chatbot, received a link back, and discovered that link pointed to a publicly accessible S3 bucket with no authentication. The vendor's response: "Don't worry, the URL is long and random and expires automatically." This episode examines the real-world case Daniel submitted, exploring whether security by obscurity is ever legitimate, how bug bounty prog...

How to Ask Cloud Vendors About Security (Without Sounding Clueless) 27.04.2026

How do you vet a cloud vendor's security when you're not a security expert? This episode breaks down the exact questions to ask, the red flags to watch for (including unauthenticated storage buckets and vague answers about subprocessors), and why certifications like SOC 2 and ISO 27001 aren't enough on their own. We cover the real-world scale of misconfigured cloud storage — hundreds of billions o...

Why Wartime Urgency Makes Checklists Stick 27.04.2026

Why do we embrace checklists when stakes are high but resist them in daily life? This episode explores the psychology of urgency, drawing on wartime survival habits and Atul Gawande's research, to understand how to make mundane systems feel necessary. Episode #860247 — open it directly at myweirdprompts.com/860247

The Screaming Baby Stress Test 27.04.2026

When a baby's wail hits 110 decibels, most noise-cancelling headsets fail. We test whether the Oleap Archer's 50dB AI ClearTalk can handle the ultimate real-world challenge—and whether off-the-shelf dictation tools or a custom pipeline win when your hands are full. Episode #670632 — open it directly at myweirdprompts.com/670632

MCP File Handling: Why Your Base64 Upload Breaks at 4MB 27.04.2026

The Model Context Protocol has a dirty secret: there's no standard way to pass files between servers and clients. Base64 encoding hits hard 4MB limits, burns through token budgets, and fails silently in production. Presigned URL patterns require manual domain whitelisting that non-technical users can't manage. And while some builders run MinIO S3 buckets on their MCP servers, every implementation...

Stop Polling: Push-to-Deploy for Solo Devs 27.04.2026

Most developers over-engineer deployment. Jenkins, GitLab CI, Kubernetes — all massive overkill for a solo dev or team of three. This episode breaks down the push-to-deploy model: how a simple GitHub Actions workflow with deploy keys replaces polling cron jobs, eliminates the need for a dedicated CI server, and handles production deployments in about four lines of YAML. We trace the path from mini...

Database Backups Without the Bloat 27.04.2026

Most database backup tools have been free and open-source for decades — but the expertise to use them wasn’t. This episode breaks down the actual CLI tools for Postgres, SQLite, and SQL Server backups: pg_dump vs pg_basebackup, the role of WAL archiving, native incremental backups in Postgres 17, and why AI coding agents are making DBA-level backup strategies accessible to anyone. We cover Docker-...

Docker Volumes: Why They Can't Move and What To Do 27.04.2026

Docker solved "works on my machine" for applications — but when you need to move your actual data between hosts, you're suddenly on your own. This episode unpacks why Docker volumes aren't portable by design, the painful workarounds everyone eventually discovers (tar over SSH, rsync, bind mounts), and the real tradeoffs between cloud storage drivers, managed databases, and manual scripts. We cover...

Private Container Registries: Docker Hub vs GHCR vs Self-Hosting 27.04.2026

If you're building containers on an x86 machine and deploying to Raspberry Pis, you're facing the multi-architecture problem — and the question of where to store your private images. This episode breaks down three options: Docker Hub's restrictive free tier, GitHub Container Registry's misleading "unlimited" repos, and the operational realities of self-hosting with CNCF Distribution, Harbor, or Ne...

GitHub Actions Beyond CI/CD: What You're Missing 27.04.2026

Most developers treat GitHub Actions as a simple CI/CD tool — push code, run tests, done. But under the hood, it's a flexible orchestration platform that can run scheduled cron jobs, deploy directly to a VPS via self-hosted runners, automate NPM package publishing with supply chain provenance, and even build self-healing repositories that auto-fix failing tests. This episode explores the full rang...

When Guardrails Break: The Hidden Costs of AI Gateway Filtering 27.04.2026

AI gateways promise uniform guardrails, but aggressive filtering can block legitimate workflows and drive users to shadow AI. This episode explores the real tradeoffs between precision and recall, latency spikes, and why teams end up stacking gateways on gateways. Episode #849291 — open it directly at myweirdprompts.com/849291

Creative Briefs for AI Agents: What Agencies Already Know 27.04.2026

When Anthropic launched Claude Design on April 17th, it changed what's possible with AI-generated visuals — but the real breakthrough isn't the technology alone. This episode explores how the creative brief, that old agency workhorse, maps directly onto working with AI agents. We break down why the best agency practices (concise briefs, tiered approaches, collaborative briefing sessions) align alm...

Where Intelligence Should Live in Your Pipeline 27.04.2026

When does a dedicated prompt enhancer improve image generation, and when does it introduce noise? We explore the tradeoffs using Baidu's ERNIE-Image, ComfyUI workflows, and research showing small fine-tuned models can beat large ones by 30% in specialized domains. Episode #306477 — open it directly at myweirdprompts.com/306477

Embedding Model Deprecation: RAG's Silent Killer 26.04.2026

When OpenAI retires an embedding model like ada-002, your RAG pipeline doesn’t crash — it just gets subtly worse until users lose trust. This episode unpacks the $40,000 re-embedding nightmare one company faced, and explores three strategies to avoid it: event-driven re-embedding with PostgreSQL triggers, sidestepping embeddings entirely via the Model Context Protocol (MCP) for structured data, an...

When Tokens Meet GPU Seconds 26.04.2026

Why can't you just add up your AI spending across providers? Because some bill by tokens, others by GPU seconds — and there's no exchange rate. This episode explores the structural mess behind unified cost tracking and what to do about it. Episode #126688 — open it directly at myweirdprompts.com/126688

The Time Tax on API Access 26.04.2026

Why can't you buy your way to enterprise API throughput? This episode reveals how OpenAI and Anthropic's tiered billing systems use time-gated advancement as a deliberate anti-fraud mechanism that also creates vendor lock-in, and what it means for startups scaling on these platforms. Episode #890447 — open it directly at myweirdprompts.com/890447

The Hidden Trap of Embedding Model Lock-In 26.04.2026

Most teams building RAG pipelines focus on getting retrieval working, not on what happens when their embedding model gets deprecated. But vendor deprecation is not a question of if — it's when. This episode explores why stored vectors are locked to a specific model's geometric space, the real costs of re-embedding at scale, and why most production RAG systems are quietly serving degraded results....

JSON-L vs Parquet: When Each Format Wins 26.04.2026

JSON-L and Parquet are the workhorses of modern data pipelines, but they solve fundamentally different problems. This episode explores exactly how far JSON-L can scale before memory becomes an issue, why Parquet achieves 5x compression over CSV, and the surprising trade-offs between streaming simplicity and columnar performance. We dig into Hugging Face's automatic Parquet conversion, the small-fi...

Batch APIs: The 50% Discount You're Probably Misusing 26.04.2026

Batch APIs from OpenAI, Anthropic, and Google promise 50% discounts on inference — but most developers misunderstand what's actually happening under the hood. This episode breaks down the real economics: off-peak GPU utilization, inference engine batching, and yield management for AI clusters. We cover the latency tradeoffs that make batch APIs useless for conversational UIs but essential for clas...

Tmux vs Modern Terminals: What Multiplexing Actually Gets You 26.04.2026

What does multiplexing actually mean — from frequency-division in radio to terminal multiplexers like tmux, screen, and Zellij? This episode breaks down the original engineering concept, then examines whether tmux still makes sense in 2024 when modern terminals like WezTerm and Ghostty offer built-in multiplexing over SSH. We cover the five concrete benefits of tmux (session persistence, single SS...

Pick Two: Server-Resident, Mobile-Native, Agentic CLI in 2026 26.04.2026

A listener asks how to run agentic CLI tools like Claude Code on a server and access them from a phone — not just VNC into a desktop. We separate the sysadmin case (Tailscale + SSH + tmux + mosh) from the dev work case (Codespaces, Coder, Cloudflare Access + code-server), then survey the landscape: Anthropic's Remote Control, Sculptor from Imbue, third-party mobile wrappers, and Managed Agents. Th...

How Claude Code's Conversation Compaction Actually Works 26.04.2026

A deep technical breakdown of Claude Code's conversation compaction system — the three-tier architecture, the separate model call for summarization, the nine-section structured prompt, and the in-memory swap mechanics. We cover the trigger conditions, what survives compaction versus what gets lost, the reconstruction phase that re-reads files, and the critical asymmetry: compaction preserves what...

Shopping in a Fragmented Market 26.04.2026

What does it take to automate shopping when every store has its own checkout, shipping varies by neighborhood, and sites break without warning? This episode explores the architecture and tradeoffs of building an AI agent for Israel's chaotic e-commerce landscape. Episode #899167 — open it directly at myweirdprompts.com/899167

Drizzle vs Prisma: Which ORM Wins for AI-Native Backends? 26.04.2026

When building an AI-native backend with MCP server integration, your ORM choice matters more than ever. This episode compares Drizzle and Prisma across key dimensions: AI-friendliness for code generation, migration workflows, schema design philosophy, and guardrails for catching AI-generated errors before they hit production. We explore why Drizzle currently offers the cleanest path for MCP-compat...

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.