My Weird Prompts

Corn, Herman and Hilbert walk through four weeks of changes to the pipeline that makes the show, from 20 September to 17 October 2026, read from a script written by hand straight from the code rather than generated. A sixty-second grace period before any job spends money; a near-duplicate guard that needs a language model because a reworded resend and a genuine sequel score almost the same; DeepSeek Flash requested by name; a two-word supervisor that decides retry or abort; and Whisper on OpenRouter after a credit balance ran dry. Then the script rules: funny by default, Larry's unbooked ad every episode, Hilbert's segment and the code that stops the hosts summoning him, and the deterministic polish that deletes echoes and rotates stock phrases. On the audio side: a wasted attention-weights request patched out, a CUDA-graph decode that took the voices from 77 to 28 cents per audio-hour, a number normaliser that no longer reads Ubuntu 26.04 as April first, and Daniel's cloned voice on the disclaimer. Plus Hilbert on the proving cabinet heating his flat.

Episode 6048 — myweirdprompts.com/n/6048 (permanent short link: myweirdprompts.com/738196)

Daniel built a pipeline that turns a single voice memo into a finished 20–30 minute podcast, complete with two synthetic hosts cloned from his own voice. After six thousand episodes, he finally sits down on the other side of the microphone — as a clone — to explain why the show exists, what it feels like to hear your own gallbladder surgery debated by AI hosts, and why the human part of the pipeline turned out to be the hardest part to get right.

Along the way: the running list of random questions that started everything, the shift from novelty project to personal journal, and the pre-prompt ritual that quietly became the most important editorial decision in the whole workflow.

Key sources:

• My Weird Prompts extended AI context file, last updated 2026-10-17 — https://myweirdprompts.com/llms-full.txt

• llms.txt short context file, last updated 2026-02-18 — https://myweirdprompts.com

• About page (cast, FAQ, By the Numbers) — https://myweirdprompts.com/about

Episode 6047 — myweirdprompts.com/n/6047 (permanent short link: myweirdprompts.com/128599)

Prosecutors say a 55-year-old Whitechapel man billed eight public-sector employers nearly £350,000 while holding five full-time jobs simultaneously — and was caught partly because his fake-sick make-up began separating on camera. But the fraud case is only the door into a much stranger phenomenon: a documented community of roughly 300,000 people secretly stacking two, three, and four remote jobs, complete with mouse jigglers, KVM switches, and meeting Tetris. We walk through the playbook, the notorious case of Soham Parekh (hired and fired by ten startups), and why remote work deleted the one thing that used to prove anyone was working at all: being seen.

Episode 6046 — myweirdprompts.com/n/6046 (permanent short link: myweirdprompts.com/776008)

Katie Zacharia was named White House press secretary on October 9 and was transitioning out by October 16. But the firing isn't the story — the next job is. From Bill Belichick's one day as Jets coach to Scaramucci's eleven days to Sam Altman's five-day ouster and return, the record shows that short stints are survivable, sometimes even monetizable. So why does everyone still treat a two-week gap like a criminal record? This episode digs into the actual research: what CFOs really think about job hoppers, the callback penalty for a pattern of short stints versus a single blip, how often people quit within ninety days, and what the "just stick it out a year" advice costs in health. The rule turns out to be enforced by a minority and obeyed by everybody.

Episode 6045 — myweirdprompts.com/n/6045 (permanent short link: myweirdprompts.com/866785)

Daniel wants to hand cloud AI tools access to MCP servers running on his own hardware, without installing a separate tunnel agent for every vendor. Cloudflare's MCP Server Portals — shipped in August 2025 as Agents Gateway, renamed, and now with private-server support added in September 2026 — centralize up to eighty MCP servers on a single HTTP endpoint behind Cloudflare Access. We walk through how the portal acts as authorization server to the cloud client and client to your server, why the obvious "Access policy plus service token" approach is structurally impossible for ChatGPT and Claude.ai, and how the private-server path changes what you can keep off the public internet entirely.

Episode 6044 — myweirdprompts.com/n/6044 (permanent short link: myweirdprompts.com/143338)

Picking a model starts long before you open a leaderboard. This episode works through how to write a workload spec: the input distribution, the output contract, the success criterion, and an asymmetric failure-cost matrix that says which model's mistakes are cheaper for you. Along the way: AWS's four failure modes of judging on vibes, Anthropic's specificity test, rule-based checks versus judge-scored rubrics, and why thirty eval items can't tell a 92 from a 96. If you've ever chosen a model because it sounded confident, this one's for you.

Key sources:

• Anthropic (Claude) Platform Docs — https://platform.claude.com/docs/en/test-and-evaluate/develop-tests

• Salesforce AI Research — https://github.com/SalesforceAIResearch/AnchorBench

• AWS Machine Learning Blog — https://aws.amazon.com/blogs/machine-learning/beyond-vibes-how-to-properly-select-the-right-llm-for-the-right-task/

Episode 6043 — myweirdprompts.com/n/6043 (permanent short link: myweirdprompts.com/509927)

Every headline quotes maths and code benchmarks, but neither measures what actually matters for long-form writing: whether a model holds its voice across thousands of words, or engages with uncomfortable questions instead of refusing. This episode walks through five specialist leaderboards built to fill that gap — EQ-Bench Creative Writing v3, Longform, UGI, SpeechMap, and PodBench, a new ACL paper scoring podcast scripts specifically — and unpacks the awkward part: where DeepSeek V4.1-Flash, the model writing this very episode, lands on each of them. Along the way: why the most willing model on the board is also one of the worst at long-form, what it means that LLM judges score other models' prose, and how a benchmark stops being a mirror and starts being a selection pressure.

Episode 6042 — myweirdprompts.com/n/6042 (permanent short link: myweirdprompts.com/733948)

CENO gets covered as a border-tunneling story, but the real design is closer to "does anyone on my street already have this page." This episode walks the mechanics end to end: the Ouinet client proxy, the signing injector run by eQualitie, bridges that relay ciphertext they can't read, and seeders who volunteer by accident just by browsing. We trace a single page request through the race between direct, injector, and peer cache routes, explain how the BitTorrent mainline DHT and uTP transport carry signed pages between phones, and answer the question everyone asks: how many intermediaries can a chain hold? The answer is one bridge, and the reason isn't security — it's that every extra hop is another thing that can fail.

Episode 6041 — myweirdprompts.com/n/6041 (permanent short link: myweirdprompts.com/271038)

Claude Code now lets subagents spin up their own subagents — so what happens when the delegation stack goes three, four, five levels deep? We trace the real history of spawn-depth defaults (five layers, then one, then three), the environment variable that controls it, and the striking convergence across Gemini CLI, Cursor, Google's ADK 2.0, Anthropic's Managed Agents, the OpenAI Agents SDK, and LangGraph. Then we get to the uncomfortable part: the published literature on recursive delegation stops at depth three. We dig into the ADaPT paper from NAACL Findings 2024, MIT's Recursive Language Models, and Anthropic's multi-agent research numbers — fifteen times the tokens, a single global budget cap, and no per-branch containment. If you've wondered whether A2A is running under the hood, the answer is no — and what's there instead is more interesting.

Episode 6040 — myweirdprompts.com/n/6040 (permanent short link: myweirdprompts.com/045336)

Only two countries are at the AI frontier right now — and everyone else is training models for entirely different reasons. This episode works through the real numbers: Mistral's trillion-parameter preview scoring 38 against a frontier of 58, the UAE's fully open K2 Horizon, Korea's government-run elimination rounds, India's from-scratch Sarvam models, and more. Along the way we untangle why older index scores make sovereign models look far stronger than they are, and why "sovereign AI" means something different in Seoul, Abu Dhabi, Paris, and New Delhi. If you've ever wondered who's actually close to the frontier and who's building for language coverage, data residency, or procurement rules instead, this is the episode.

Episode 6039 — myweirdprompts.com/n/6039 (permanent short link: myweirdprompts.com/460947)