Despite occasional silicon-valley-venture-capital-frosting that sticks in my craw1, I’ve been really enjoying the Latent Space AI engineering podcasts. There is technical stuff that’s beyond what I can understand at times but the majority of it is understandable, adds interesting depth to my AI knowledge, and has led to me to interesting things outside of AI. I like it particularly because it’s entirely outside anything to do with k12 or highered’s view on AI.
Here are a few episodes I’d recommend listening to. My recollections are likely blurry as I’ve listened to these over a couple weeks and I took no notes.
Emulating Humans with NSFW Chatbots – with Jesse Silver. I don’t think there’s anything remotely NSFW in this but it’s focused on chatbots for Only Fans. I found it insanely fascinating. The whole mechanization of money extraction was crazy to hear broken down like this. It’s sociology. It’s AI. It’s capitalism gone wild.
Yeah. So this is super unique in that people are paying thousands of dollars to interact with the product for an hour. And so no audience economizes like this. I’m not aware of another audience where a chatting system can economize like this or another use case where on a per fan basis, people are just spending so much money. We’re working with one creator and she has 100 fans on her profile. And every day we earn her $3,000 to $5,000 from 100 people. And like, yeah, the 100 people, you know, 80% of them churn. And so it’s new people. But that’s another reason why you can’t do this on OpenAI because then you’re spending $30 on a fan versus doing this in an open source way. And so open source is really the way to go. You have to get your entire pipeline fine tuned. You can’t do more than some percentage of it on OpenAI or anyone else.
How AI is eating Finance — with Mike Conover of Brightwave My memory of this one is a bit hazy but I liked how they talked about AI as a thought partner and said something like “the worst thing possible was to have AI that was believably wrong.” The focus on the LLM filling a particular role and not doing spreadsheet work in finance was also interesting. You’ve also got RAG talk and some other stuff in there.
The End of Finetuning — with Jeremy Howard of Fast.ai This led me to Fast.ai and it explained a bunch of stuff that I had no idea about regarding fine tuning, training, etc. I liked Jeremy Howard’s goals around making all this accessible and how he feels his initial approach (which is the foundation for all this stuff) is wrong.
That’s not much, but if you don’t like these, you won’t like the others. I’ll have to figure out a better way to take notes on podcasts and get that somewhere outside my head in the moment. That’s rough to figure out in a car. I may have to break down and establish a decent voice option for my phone. I turned off all that stuff because my experience has been so painful in the past.
1 Best intro blog post word I’ve managed yet. Craw was also a word I couldn’t imagine being used in a venture capital pitch.