I'm on the latest episode of the Rooftop Ruby podcast with @collin and @joeldrapper talking about Large Language Models
It was a really excellent conversation - we covered a huge amount of ground
I'm trying something new: I put together my own transcript with Whisper, then cleaned that up and added inline links and section headings. Here's the result, complete with an embedded audio player that can jump to each different section: https://simonwillison.net/2023/Sep/29/llms-podcast/
@collin @joeldrapper I used GPT-4 to help build my own custom audio player, with a 3x speed button!
Here's that GPT-4 transcript: https://chat.openai.com/share/4ea13846-6292-4412-97e5-57400279c6c7
I find LLM code gen works best as a way to generate standalone functions with unit tests that can be integrated into existing apps. As such, it is like generating a custom library instantly.
Generating whole apps is much harder, especially if a UI is involved. It can be quite tedious to explain all the UI behaviors of a webapp at this point. Ultimately, eventually it will boil down to specifications-as-programming which has long been a goal.
@baclace @collin @joeldrapper Yeah, UI programming is definitely a lot less well served - I can get bits and pieces out of it, but fundamentally these models don't have a great idea of 2D space yet so they're not the best for interface work
I'm looking forward to seeing if that changes with the new GPT-4 image inputs
Simon Willison
Matt May
Paul Baclace