Pipeline Changelog

How the show gets made, and how that's changed

Every episode is written, voiced and assembled by an automated pipeline. This page tracks its major updates — new models, new voices, new stages — and the first episode each one produced. Small prompt tweaks happen weekly and aren't listed.

Running now

Script model GPT Sol, latest version (6.1 today)
Other text stages DeepSeek V4.1 Flash (latest-Flash alias)
Titles, descriptions, blog post Gemini Flash (latest)
Voices Chatterbox, self-hosted on Modal L4 GPUs (CUDA-graph fp16 decode)
Research Multi-round agent: web, news, PubMed, arXiv, Hacker News
Transcription Gemini Flash (latest), Whisper fallback
Cover art Krea 2 Turbo Style with style references
Embeddings Gemini Embedding 2

Pipeline V5.22 · prompts 4.29 · since 24 October 2026. Every episode page lists the pipeline version that made it.

2026

Download 2026 changelog (PDF)

Q4 2026In progress

October – December · 8 changes

PDF
Script model Pipeline V5.22 Prompts 4.29 Current

V5.22: the script writer follows GPT Sol's latest version

The script writer now asks OpenRouter for the newest GPT Sol rather than naming GPT-6.1 Sol directly. Today that is still 6.1, so nothing changes yet, but when OpenAI ships a new Sol the show moves to it without a code change. The credit at the end of each episode still names the exact version that wrote the script.

Script model Pipeline V5.21 Prompts 4.29

V5.21: every model brought up to date

I audited every model the pipeline calls apart from the script writer and brought them all up to date. Where a stage only needs a good fast model, it now asks OpenRouter for the latest version of its family, so it moves to a new generation without anyone touching the code. Titles, descriptions and the blog post come from Gemini Flash, as structured output with length limits, and the blog post is much shorter. Covers are drawn by a newer image model that is shown earlier covers as a style guide.

  • Title, description, URL and image prompt: one Gemini Flash call that returns fields rather than free text, with limits on each (title up to 60 characters, description 50 to 90 words). A miss is regenerated. Before this, one DeepSeek call wrote all of it plus the blog post in labelled text.
  • Blog post: a separate call, about 100 words instead of 150 to 220.
  • Voice-memo transcription: Gemini Flash, latest version, instead of a Gemini 3.1 Flash Lite preview. Whisper is still the fallback.
  • Search embeddings: the released Gemini Embedding 2 instead of its preview. The vectors are identical, so nothing had to be re-embedded.
  • Covers: Krea 2 Turbo Style, given three earlier covers as style references, about a cent per cover. Flux Schnell, the previous model, is the fallback.
  • Embeddings stay on a fixed version, because a silent change there would invalidate every stored vector.
  • The experiment and interview episode types were removed: the guest models they relied on had been retired.
Writing Pipeline V5.3 Prompts 4.7

V5.3: better recommendations

Episodes that were being recommended almost every time are now dropped from the related-episode suggestions. One episode had turned up in 12 of 19 recent runs. The recommendation is also placed directly before "This has been My Weird Prompts" rather than at the start of the host's line. And the script writer now treats a word in my prompt that doesn't fit as a likely speech-to-text error, rather than building a claim on it.

  • New episode addresses now end in a random five-digit number, so two episodes on the same topic can't end up with the same URL. Existing addresses are unchanged.
  • Later the same day the site gained an entity index at /indices/: the people, organisations, places, works, events and products the show discusses, each linked to its episodes.
  • Also later that day, the voices briefly moved to cheaper T4 GPUs, then back to L4 once the speech decoder was rebuilt around a CUDA graph with an fp16 backbone. On the L4 it renders nearly three times faster than before, for about $0.28 of GPU time per hour of audio, down from $0.77. It is on trial: if the voices drift, it can be switched back.
Infrastructure Pipeline V5.2 Prompts 4.2

Anyone can send in a prompt

A public Telegram bot, t.me/mwplistenerbot, takes prompts from listeners. It asks four questions (your name, a line about you, where you are, and the question itself) and emails the submission to me. Nothing becomes an episode automatically: I read each one and decide. Episodes now close by reading out the bot's address instead of an email address.

Writing Pipeline V5.2 Prompts 4.1

Hilbert gets ridiculous, and deep topics get thirty minutes

Hilbert's stories are now meant to be absurd, landed through exact detail and a straight face, where before they had to be believable. His turn grew to three or four minutes. Episodes on topics that need it can now run to thirty minutes. The planner had been offering that length all along, but the code quietly cut it back to twenty-five.

  • Later the same day (prompts 4.3), Hilbert started treating Corn and Herman as colleagues he knows. His crankiness is aimed at the world, not at them.
Infrastructure Pipeline V5.2 Prompts 4.1

Episode Books, an MCP server and an AI sandbox

Three ways into the archive that don't involve listening. Every episode now has an Episode Book: a PDF with the episode's record, synopsis, transcript and sources, pre-rendered for the whole archive. AI tools can query the show directly through a public, read-only MCP server at myweirdprompts.com/mcp. And /sandbox/ hands out a condensed version of the show's format prompt and character sheets, so anyone can run their own version of the show in a chatbot.

Q3 2026

July – September · 7 changes

PDF
Voices Pipeline V5.2 Prompts 4.1

Hilbert reads out each episode's short code

Every episode now opened with a short slate after the intro jingle: Hilbert reading the episode's title and its permanent short code, plus the address myweirdprompts.com/e/ followed by that code, where the transcript lives. The /e/ route was added so the address he reads aloud actually resolves.

  • Retired on 9 October 2026 (see below). The /e/ links still work for every episode.
Script model Pipeline V5.2 Prompts 4.1

DeepSeek V4.1 Flash

DeepSeek had started answering V4 Pro requests with V4.1 Flash until V4.1 Pro ships. The pipeline now asks for V4.1 Flash by name and uses it for every stage, because each episode ends by naming the model that wrote it, and that credit should be accurate.

  • Images can now be sent with a prompt. Up to five images go straight to the script writer.
  • The hosts now receive Daniel's prompt exactly as he wrote it, not a tidied rewrite.
  • The model credit at the end of each episode is now read by Hilbert, the show's producer.
Writing Pipeline V5 Prompts 4.0

Hilbert gets a turn in every episode

Hilbert Flumingtop, the show's producer, had been an occasional end-of-show beat that turned up in roughly one episode in seven and was gone after a line or two. He now has a fixed segment of his own in every episode, about four-fifths of the way through, while there is still real episode left to run.

  • He speaks several times in his segment, not once, and the hosts no longer close it with a polite "thanks, Hilbert".
  • The closing beat that asked the hosts for a dated prediction was dropped, and so was the push to end every episode on an actionable takeaway.

Q2 2026

April – June · 6 changes

PDF
Script model Pipeline V5 Prompts 4.0

DeepSeek V4 Pro

Scripts moved to DeepSeek V4 Pro, called directly through DeepSeek's own API, with its lighter Flash sibling handling the surrounding stages. V4 Pro wrote close to 3,000 episodes, more than any other model in the show's history.

Q1 2026

January – March · 1 change

PDF
Voices Pipeline V4

The voices come in-house

Chatterbox now runs on the show's own Modal GPUs instead of a hosted API. Every episode since has been voiced this way.

  • Switching from Chatterbox Turbo to Chatterbox Regular cut TTS hallucinations (garbled or invented audio) by about 95%.
  • Each host's voice is pre-computed and cached, and an episode's lines are spread across parallel GPU workers. Voicing an episode went from about 36 minutes to about 10.

2025

Download 2025 changelog (PDF)

Q4 2025

October – December · 4 changes

PDF
Infrastructure Pipeline V3

Two AI hosts, one voice memo

The earliest episodes in the archive. Daniel records a voice memo, Gemini 2.5 Flash transcribes it and writes a two-host dialogue between Corn and Herman, and cloned voices read it out.

  • Script and transcription: Gemini 2.5 Flash.
  • Voices: Chatterbox voice clones, run on fal.ai.
  • Roughly $2 of compute per episode.