This is a linkpost for https://thefilancabinet.com/episodes/2023/02/05/6-oliver-habryka.html
That was great. I legit had no idea how big of a role Oliver played in making LW 2.0 happen; I always assumed he was just hired.
I used a service with an OpenAI Whisper backend as a first pass (specifically, revolvdiv this time), then manually transcribed everything, discovered that leaving all the speech filler words in made the transcript very hard to read, and did another editing pass.
I agree that, if I do this again in the future, rev.com would be a relevant choice.
Anyway, ultimately the hard part was not mainly turning audio into text, but doing so at a (self-inflicted, probably unreasonably) high standard of accuracy. No, even that's not quite right. The problem is that you want high accuracy (so you don't put words into someone's mouth), but not regarding the literal spoken words (which are full of filler, and word repetitions, and unintelligible mumbling, and sentences that don't have correct grammar - all because people don't speak like they write), but rather the meaning the speakers wanted to convey.
But also, this is the kind of thing at which one gets much better with experience, which I lacked.
No, sorry. Since a few people have asked: transcripts are pretty money- and time-consuming to produce, and I wanted to have a podcast where I make the trade-off of having more episodes but with less polish.
If there isn't, I recommend to the podcast creator to consult with e.g. the Clearer Thinking podcast team on how they do cost-effective partly-automated transcripts nowadays. Here's an article on their thinking from early 2022, which was before e.g. OpenAI Whisper was released.
I think this LW post would be significantly more useful with a full transcript, even if automated, for instance because it's easier to discuss quotes in the comments. (On the other hand, there's a risk of getting misquoted or directing excessive scrutiny to language that's less polished than it would be in essay form, or that may suffer from outright transcription errors.)
Yes, that's the article I meant. I understand the tradeoff you're making, and given the costs you cite, I can totally see that that's not worthwhile for you, especially if higher quality trades off against higher quantity.
That said, I've mailed the Clearer Thinking podcast team to ask about more details regarding their current transcription workflow (which is currently a combination of automatic transcription via Otter.ai, followed by a hired human transcriptionist, to minimize required staff time), and will post any responses I get.
Alternatively, if someone offered to pay me $200, or $40 per hour I actually needed (whichever is lower), I'd produce the transcript myself. (As a general matter of economic arbitrage, nobody who's being paid California salaries should spend their own time to produce transcripts themselves.)
especially if higher quality trades off against higher quantity.
Yeah - this podcast is a side-project of my main podcast which is a side-project relative to my day job (CHAI PhD student), so time minimization is of the essence.
Alternatively, if someone offered to pay me $200, or $40 per hour I actually needed (whichever is lower), I'd produce the transcript myself.
I'll chuck in US$30 of my own money. (would be more if I had a better sense of the quality bar you were going to reach)
I thought it was interesting when Oli said that there are so many good ideas in mechanism design and that the central bottleneck of mechanism design is that nobody understands UI design to take advantage of them. Would be very interested if other folks have takes or links to good mechanism design ideas that are neglected/haven't been properly tried enough or people/blogs that talk about stuff like that.
I may be blind, but the link to the audio doesn't seem to allow me to actually download the audio. Which wouldn't be so bad if the Google Podcasts site didn't cause a bunch of issues for me, e.g. when I rewind back by 10s, the audio cuts off for 10++ seconds, which defeats the purpose of rewinding.
EDIT: Thanks for the audio links!
OK: now on the official podcast website there are dropbox links to download the mp3s. This is sort of a 'beta' feature because I'm not sure how many people want mp3s or whether these links are robust to me reorganizing my dropbox, but since another person wanted the mp3s, there they are.
Notes:
OK sorry to over-advertise but it seemed like this one would be of interest to the LessWrong and EA communities. Episode description below, audio is here, or search for "The Filan Cabinet Habryka" wherever you listen to podcasts.
In this episode I speak with Oliver Habryka, head of Lightcone Infrastructure, the organization that runs the internet forum LessWrong, about his projects in the rationality and existential risk spaces. Topics we talk about include: