crazy that programmers re-invented phones, money & cars

training time compute can be amortized across all future inference calls but inference time compute is only valuable for one user inference time compute will therefore dramatically increase compute demands

Playing with the new Qwern QVQ "visual reasoning" model on Hugging Face. Comes back blank for tank man, Mao Zedong, Xi Jinping, and Jack Ma. It will identify Yao Ming though! huggingface.co/spaces/Qwen/QV…

Deepseek greatly decreased training costs by reducing inference from O(n^2) to O(n) via Latent Attention. Instead of computing each pairwise attention score, they converted the tokens to a latent space vector (length L) thus O(n*L) forward pass.

Comprehensive analysis of the crime scene of Suchir’s apartment. It wasn’t a suicide… #JusticeForSuchir

every day should be an adventure

less work, more joy

accelerationism is the consummation of capital and technology that removes humans from the loop

the reason his posts work is bc there is no irony his new bio goes hard

Grok breaks down the difference between dropping a hash of a photo vs zk proof of photo: K Kulkarni grok.com/share/bGVnYWN5…

thank you Tim Sweeney for defending the open web & empowering developers you freakin’ did it

not sure what i saw/heard, but i love it