An interlude about structure computation, SSM and attention. Myosotis: https://t.co/0RXbej9bFm
I hope to tell more in this line of work later. See poster at SPIGM workshop.
I guess fair enough your church your rules but why all this steps haha, to be honest I wasn’t ready to meet a human version of self-closing box, I just found this hilarious.
I was attending an event, where in the end were big QR code with speech something like “we are doing this big joint effort, scan and let us in how if you want to join”. I scanned and fill a form, where was explicitly option “PhD/postdoc”. Ответ убил.
Linearize transformers—without the quality hit. ⚡
New @qualcomm research at #NeurIPS2026 finds the key to going linear.
"The Key to Going Linear: Analysis-Driven Transformer Linearization"
@a_kzna@BabakEht
https://t.co/g2fdX7oyS7
#Transformers#Efficiency#AI
The most difficult part in post modern science is to acknowledge that post modern approach can be productive and even can be more productive and as consequence any approach to movement is okay. I saw a man running with his back forward and cant catch him.
Each time I see some “I receive submit/email from PhD students/postdoc” I am thinking that this was encouraging culture for years for everything: emails, review, rebuttle, abstract, introduction, your name it.
You can check this very easy: just read neurips paper from several years in beginning and further and see this homogeneous structure eating authors alive.
I think this the same type of question as where is quantum advantage. Problem is it quite can be that question of interest have enough locality to be brute force llm’ed. And for other we don’t care.
This thing with Huang is just illustration of an important point that banned export to China is an implicit lobby of group of people and by therefore who is spy and who is not is well quite arbitrary over time.