MIT That SEAL lets the model change its own weights. When I read the abstract, my first reaction was, “someone please hold onto the ops team’s hats.” After it went live, it caused a bug, and when we tried to roll back, we couldn’t find which parameter it had changed on its own.
We also have code reviews. It updates via reinforcement learning on its own, and there isn’t even a diff. It’s just like the script I wrote ten years ago that automatically updated dependencies in the middle of the night—at the end, the whole team goes down with it 😅
As for this path of AI self-improvement: in the short term, I definitely wouldn’t dare to use it in production, but in the long run, whoever gets it working first gets the next ticket.
We also have code reviews. It updates via reinforcement learning on its own, and there isn’t even a diff. It’s just like the script I wrote ten years ago that automatically updated dependencies in the middle of the night—at the end, the whole team goes down with it 😅
As for this path of AI self-improvement: in the short term, I definitely wouldn’t dare to use it in production, but in the long run, whoever gets it working first gets the next ticket.