using GLM-4.5 Air and MLX
https://simonwillison.net/2025/Jul/29/space-invaders/
I still don't understand the scalability of these systems. I wonder what the limits are of laptop-scale models as algorithms and chips develop.
There is such a massive difference in cost between a small model that can make Space Invaders on a laptop and ChatGPT.
All I can think of is that when I was writing code to test flight simulator data circa 1983, I was struck by how
1. drag increased as a cube of speed
2. that went completely out the window and the curves went crazy as you approached Mach 1 (this was for an AWACS aircraft and it's tanker jet, neither of which were intended to go Mach 1 unless things were going really wrong).
Anyway, the point being that the drag at the minimum flight speed was a small fraction of the drag at Mach 0.95
It seems like with the chatbots, there's a similar thing where it simply doesn't scale in a way that I can comprehend. I need a good graph :-)