Pipeline Changelog

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.

V5.2: primary sources for technical topics

For technical prompts the research agent now reads arXiv papers and Hacker News discussion, not just whatever a web search ranks highest. It also reports what it looked for and couldn't find, so the hosts don't fill the gap with something that only sounds similar.

V5.1: versions that mean something

This was the first version bump in nine months, and from here every change that can affect an episode gets one. Prompts 4.1 targets verbal tics that were measured across the whole episode archive, rather than ones that were guessed at.

A research agent before every script

Before any writing starts, an agent now researches the topic over several rounds, using web search, open news data (GDELT), PubMed for medical topics, and source code repositories. What it finds becomes the script's research brief.

Scripts written in one pass again

For a few weeks, scripts were written section by section. That produced episodes that restarted or repeated themselves, so scripts are written in a single pass again.

Character lorebooks

Each host now has a lorebook of backstory, and the entries relevant to an episode's topic are pulled into the script prompt. The hosts' personal history comes up when it fits instead of being repeated in every episode.

Studio-style mastering

Assembled episodes now get a mastering pass: loudness normalisation, fades, room tone between lines, and EQ tuned to each character's voice.

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.

DeepSeek takes over

Scripts moved from Claude Sonnet to DeepSeek V3.2, at about a thirty-sixth of the price per output token. Research grounding also got a dedicated section and a three-round search loop.

Prompts 4.0, and the hosts start remembering

The script prompt was rewritten as version 4.0. The same day, each host got a written character card and a memory layer that looks up earlier episodes by meaning, so the hosts can refer back to things they've already discussed.

Claude Sonnet 4.6 writes every script

Gemini Flash kept writing episodes that ran short, so long-form episodes went to Anthropic's Claude Sonnet 4.6 first (from 8 April). Three days later Sonnet was writing all of them.

• Smaller stages ran on Claude Haiku 4.5.
• Voice memos are now transcribed by OpenAI Whisper.

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.

Pipeline V5: production moves to Modal

Generation moved onto Modal, a serverless GPU platform, which has run the pipeline ever since. The V5 label then stayed in place for nine months; for a while V4 and V5 were both still running.

Scripts move to Gemini 3 Flash

The script writer was upgraded from Gemini 2.5 Flash to Gemini 3 Flash. From two days later it ran in preview, and wrote nearly 2,000 episodes before April 2026 brought a run of model changes.

Pipeline V4: auditioning new voice engines

Text-to-speech moved to Fish Audio. This began a month of trying different voice engines (Fish Audio, then Inworld, then Chatterbox hosted on Replicate) before the show settled on running its own.

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.