hham

6 ms·

1. 1.▲
   by hham 19d ago

   memory plus memory bandwidth > GPU, this is the gist of it.
2. 2.▲
   by hham 2mo ago

   Totally agree, we wrote an article about ho this is a senseless "red queen race" few days ago, hope is ok to post: https://news.ycombinator.com/item?id=48892559
3. 3.▲
   by hham 2mo ago

   https://news.ycombinator.com/item?id=48892559
4. 4.▲
   by hham 2mo ago

   Smart that she says "not the strongest overall model, open or closed". This is a rare for an AI lab to say out loud. They basically decided to compete on customizability, and not on topping the temporary leaderboard. Also corrobor
5. 5.▲
   by hham 2mo ago

   This is accelerant #3 and #4 from our article converging in one release: a 27B-class model, built on Qwen (already one of our examples of local models "good enough to matter"), now running on an iPhone. The hardware layer and the