GPT-6 Astra points to looped transformers as a new scaling direction
The author argues GPT-6 Astra suggests looping the same weights, 4 times for simple questions and 64 for hard ones, separating model size from how much thinking a problem gets.
THE NEXT BIG AI SCALING LAW MAY NOT BE MORE PARAMETERS. IT MAY BE KNOWING WHEN TO THINK TWICE. For years, making a model “deeper” basically meant building a bigger stack. More layers. More parameters. More memory. GPT-6 Astra points at a different direction. Instead of passing through a computational block once and moving on, a looped transformer can send its internal state through the same weights again. Same model. More computation. A simple question might need 4 loops. A hard proof might need 32. A difficult coding problem might deserve 64. And none of that necessarily requires the model to write a longer chain of thought before answering. That's the part I find much more interesting than another benchmark record. We may be separating two things that used to scale together: HOW BIG THE MODEL IS and HOW MUCH THINKING THE PROBLEM GETS. Eventually the important question may not be: “How many parameters does this model have?” But: “Did it know the answer was good enough, or was this token worth another trip around the loop?”
This post quotes another post: x.com/mikadzyki_nft/status/2097336663429046689