Followed by the transcription of the podcast

[aminaas1576]

Which brings me to the latest, continual tinkering with the tools and environment available to analyze materials with algorithms. I began asking a large language model to look at the generated transcriptions and create a summary of a given podcast episode.

Two-plus years into generalized algorithmic intensity access, it’s still very much a lumpy and oddly spell-casting endeavor. Instead of asking it to “summarize this transcription”, my request reads like a headmaster at a school or a Dungeons and Dragons game text:

You read transcripts of a podcast and carefully write out descriptions, in the form of narrative paragraphs, to accurately describe the content of the podcast. Longer and more complete descriptions are better, and encouraged. You describe the main subjects, conclusions by the participants, and provide helpful context for the subjects. The podcast you listen to is:

…, time-codes and all.

A matrix of calculation, fast beyond my reckoning phone number database but not less mysterious-and-not-mysterious as CPUs and networking itself, begins analyzing the language in the transcriptions, cross-connecting ideas mentioned, occasionally volunteering more information based on matches to terms, and within a few minutes, language comes out.


And oddly enough, it brings up memories.

It reminds me of being in my first neighborhood and all the kids circled around a fun and weird toy, a tape recorder, where you pressed two buttons and the … cassette, it was called? Would turn slowly and you would shout whatever came to the top of your head, press STOP, and then rewind and hear your own voice. It was distorted and weird, but it was my own voice, and I’d not heard it before from outside my own head. The world shifted, a little bit.