Me: *searching for a particular claim in a mainstream media audio interview* Me: Can you find the link to MP3 file on this page so I can download it to transcribe? GPT-5: Oh no, copyright concerns. But I could give you a one line regular expression to run against curl output.

Me: I like the way you think. *does it* Me: OK now what’s the best way to invoke Whisper to transcribe this, using your API, installing absolutely nothing? GPT-5: Here’s minimum curl command. Me: Here’s JSON error. GPT-5: Three possibilities, most likely that MP3 file is >25 MB.

Me: That is not what the error message says. GPT-5: Humans. Me: OK what’s easiest path forward. GPT-5: Do you have ffmeg available? Me: Yea. GPT-5: Here’s one-liner which chunks into 25 MB subparts, submits to API, and then stitches together transcripts. Me: … That worked.

File this under mundane utility, LLMs acting as a passive upgrade on top of “every other API in the world”, five cents of tokens which save a media organization several hundred dollars of labor on an intellectually uninteresting task, etc.

(As always, dialogue in this format is not verbatim. In particular, it did not make the “Humans.” joke, but I was amused by the mental image of an AI evaluating a senior engineer who was surprised that error messages might not accurately reflect root causes.)

You’d think that search engines would be much more useful for this than they are, but it’s actually load bearing that a) I have an extremely good memory for podcasts and b) this particular podcast was on Bloomberg and therefore can be cited as “The CEO admitted to Bloomberg that”