That OpenRouter stat about reasoning models hitting 50% of inference is nuts when you think about it. A year ago we were basically at zero. The shift from "answer questions" to "actually work through problems" happened way faster than most people expected, and I think that's what's really freaking everyone out about the Gemini 3 vs GPT competition.
Hey, great read as always; the insight that OpenAI's new "Garlic" model aims to solve pretraining efficiency issues to allow smaller models to pack more capability is particularly acute, highlighting a critical shift towards more sustainabil and scalable AI development.
This was a really good rundown!
Thank you so much! I read AI Supremacy religiously so means a lot :)
I sent you a direct message with an idea.
excited! just sent a reply back on DM
That OpenRouter stat about reasoning models hitting 50% of inference is nuts when you think about it. A year ago we were basically at zero. The shift from "answer questions" to "actually work through problems" happened way faster than most people expected, and I think that's what's really freaking everyone out about the Gemini 3 vs GPT competition.
Thanks so much for reading and for the thoughts here! I do agree that was the most astounding stat I saw!
Hey, great read as always; the insight that OpenAI's new "Garlic" model aims to solve pretraining efficiency issues to allow smaller models to pack more capability is particularly acute, highlighting a critical shift towards more sustainabil and scalable AI development.
10000% - something I'm paying close attention to