Welcome to issue 82 of Not only Swift!
As I am writing this intro, I am on the way back from AI Heroes in Turin. It was a great conference with a great speaker lineup, covering a wide range of topics. I liked that some of the talks were more academic, and others more practical.
One talk that resonated with me was Repetita Non Iuvant: Why Generative AI Models Cannot Feed Themselves by Valeria Zuccoli - she discussed a topic that’s been hotly debated over the past couple of months: model collapse. Model collapse is what happens when we start training models based on material that has been generated by the same (or other) models, and the model “forgetting” features of the original data and hallucinating new, incorrect features.
Valeria presented mathematical and evidential proof that this actually is a problem. The good news, however, is that you can counter this by adjusting the way you train the models, e.g. by using watermarking, or data curation. A previous recording of Valeria’s talk is available here.
On a related note, we had a discussion on the iOS creator Slack about whether or not people are still interested in reading blog posts. A few of us (me included) think that it’s worth while writing blog posts (and creating other types of content, such as videos) as long as people care about other people’s opinions, experiences, and insights.
We also discussed whether or not it you should use robots.txt
(or a CDN) to prevent AI crawlers from ingesting your writing. Personally, I actually prefer my content being ingested into LLMs, as it helps to keep the models’ knowledge of how things work up to date (and also helps to avoid model collapse - at least a little bit :-) ).
Given this is a curated newsletter, I assume that you still read blog posts (and consume other, humanly-created content), but I’d like to hear from you! What’s your experience?
Take part in this poll, or hit reply and share your thoughts with me!
Thanks for reading,
Peter
BTW: don’t miss the livestream with Cihat - we’ll talk about Swift error messages, and how to make them better using his library, ErrorKit!