Chart crime so bad it's basically propaganda lol

they're playing geese at the starbucks and not just any geese, 3D Country geese thats for the true Geese Lovers

🫡

The coffee scene in SF is elite, holy

obligatory dolores park pic feeling thankful to enjoy a nice day :)

wait, the 16th and Mission BART station is unreasonably pretty

I really really like this idea. Great presentation too

a divine sign

wrapped up some interviewing, thinking of treating myself. not sure which one tho

I'm going to neurips! yay :)

Spent the last few days reading all the documentations/proposals for llm d and the K8s Inference Gateway Extension. It's fascinating, really enjoyed reading the docs and the code. Highly recommend.

Kinda want to design an interview question around GRPO. 1. Implement a function that rolls out N random strings 2. Implement a reward that's sum of character-wise ord. 3. Calculate the advantage and choose a winning rollout. Then chat about design e.g. reward, advantage design

Silksong (2025)

Simple prompt trick. Instead of asking a model to choose the best of N options, try: 1. Saying that you've already chosen the best option privately, 2. Only if you both agree will you move on to implementation I see a more reasoning but shorter responses. Less slop overall.

GPT-5 is feeling much smarter right now than at launch. I wonder if during launch demand many more queries got routed to the smaller models; now that demand is down, there's more reasoning capacity? not sure.