Yesterday, we launched a generative image model that competes with the top of the field. I figured some HN folks might be interested in this deep dive from the research team.
Watching them figure out how to make the best use of our relatively-limited resources has demonstrated "necessity is the mother of invention" for me. Instead of just "pour more money on it," they came up with a different training strategy / hypothesis (the title's Layout Bet), and it's exciting to see that bet start paying off.
Watching them figure out how to make the best use of our relatively-limited resources has demonstrated "necessity is the mother of invention" for me. Instead of just "pour more money on it," they came up with a different training strategy / hypothesis (the title's Layout Bet), and it's exciting to see that bet start paying off.