Comment by mark_l_watson

Off topic, but I just bookmarked Tim’s blog, great stuff.

I dismissed the X references to S1 without reading them, big mistake. I have been working generally in AI for 40 hears and neural networks for 35 years and the exponential progress since the hacks that make deep learning possible has been breathtaking.

Reduction in processing and memory requirements for running models is incredible. I have been personally struggling with creating my own LLM-based agents with weaker on-device models (my same experiments usually work with 4o-mini and above models) but either my skills will get better or I can wait for better on device models.

I was experimenting with the iOS/iPadOS/macOS app On-Device AI last night and the person who wrote this app was successful in combining web search tool calling working with a very small model - something that I have been trying to perfect.