Gemini Diffusion is my fav GoogleIO announcement vibe coding at 1000tok/s hits different multi-turn looks good so far (no video speedup or anything) insanely bullish on diffusionLM

probably hands down the best AI conference in recent years not to mention, u can already access most of these apps RIGHT AWAY Google is on an insane streak lately (throwback when we have to wait a year for a sota video gen👀)

they are just aura farming at this point

the power of arrow..?

CLI AI tools just redirected my pain point from spending time to figure out venv to looking at how much $ it just spent to figure out venv (wdym it took 20 cents just to setup my venv)

writing style I've observed: gemini 2.5 pro loves nested bulletpoints claude 4 writes in paragraphs, occasional short bullets o3 loves tables and bulletpoints, not as nested like gemini

AI companies saying that AI will replace SWEs but it still doesn't one shot my use case

LLM routers just make so much sense

🚨This week's top AI/ML research papers: - Claude 4 System Card - Emerging Properties in Unified Multimodal Pretraining - dKV-Cache - Thinkless - GUI-G1 - Scaling Diffusion Transformers Efficiently via μP - Faster Video Diffusion with Trainable Sparse Attention - Model Merging

Multi-Token Prediction: The Fascinating LLM Performance Boost No One Talks About youtu.be/4BhZZYg2_J4

> be me > omf i miss conda > *looks inside* > maybe i dont

damn its literally faster than any other (priv/os) models API except gemini 2.5 flash reasoning (332tok/s via artificial analysis) I was literally complaining a few months back how dsv3 is only like 25tok/s which made it kinda unusable lol it's basically #1 for speed vs perf

since waiting is too boring and not cool, we added some visuals to preview the papers that the search function found findmypapers.ai

o3 is actually the best at writing english stuff, follow instructions, and summarizing

is it normal to feel pessimistic after reading all the RLVR stuff came out this week

p1: i have no idea how he found me p2: iykyk 😎