How this was made

Transcripts. Every episode in the show's podcast feed was transcribed by machine: OpenAI's Whisper (large-v2) for the YouTube uploads through January 2025, and for the feed's MP3s since then, and for older episodes whose Whisper text had looped, NVIDIA's Parakeet (a few with Whisper large-v3-turbo). Neither writes down who is speaking, and both mangle names: Friedberg comes out as "Free Berg" or "David Rupert", Chamath as "Jamath", Sacks as "Sachs".

Finding the predictions. The annual predictions specials were read whole. For the weekly shows, a script pulled about three minutes around every line that sounds like a forecast ("I predict", "I'll bet", "mark my words", "by the end of the year", odds), and a Claude model read those passages and kept the calls that are specific enough to check later. Jokes, "could" and "might", and conditionals on things nobody would ever settle were left out. Interview guests are left out too; guest hosts on the panel are in.

Who said it. Each call is tied to a host by what the transcript shows, marked on every card: called on (Jason asks "Sacks, your pick?" and the answer follows), self (the speaker names their own firm or past pick), or context (only the run of the argument says who it was — the weakest kind). Calls nobody could place were dropped: . Every quote was checked word for word against its transcript; the link on each card opens the episode at that moment so you can hear it.

Two readings. Every episode was read twice, independently, by the same kind of model with the same instructions, and its calls were matched against the first reading's: . Firm calls almost always turn up twice; hedged asides and calls placed only by context drop out more often, and on some matched calls the second reading named a different speaker. Each card says which, and the scorecard can be limited to the calls both readings found.

Grading. A Claude model with web search graded each call that has come due as right, wrong, partly right or open, and had to cite a source for every grade but open. A pick in a judgment category ("biggest business winner") is graded by an obvious yardstick — the year's stock return, the headlines — and the card says which. Read the reason and the source before trusting a grade; corrections are welcome.

What it can't see. A forecast phrased without any of the cue words in a weekly show can be missed, and so can one in an episode the feed never carried. The hosts also hedge, and the line between a prediction and a riff is a judgment call made here once, consistently, not the last word.