he knows me soooo well 🙈

One of the hardest parts of being an immigrant in America is not being there for your family back home if something happens Now on an H-1 visa, you literally cannot even fly back home for a family emergency without risking your entire life here This needs to change asap

this is the dream but we are early lots of experiments to run, frontiers to explore and train models to make magic.

the reason why I don’t like Anthropic - this video shows pure arrogance and disdain for anyone else who isn’t big labs their words and actions around safetyism reflect the same - only we can train the models and you peasants can’t open source AI is a wonderful thing and will win

model is the product two best products of this year - openai deep research and claude code are the results of agentic post training

why model and product are the same or at least should be codeveloped prompts and parameters are two of the biggest levers use them both for aha moment

forged not born - such a beautiful and powerful message

curious about GRPO and why it works or how far behind it is ? this tweet is all you need to get started 👇

(maybe) the next algorithmic improvement with GRPO family frontier labs rl training algorithms are different ? feels like they have figured out better credit assignment than treating all tokens as equal

makes me feel bullish about India

fun part is Indians accept colonization was an horrible part of our history and moved on to build better future otoh left is busy apologizing for the sins of their ancestors and right is busy glorifying the horrors of the past. in India, i heard colonization word only in

Kaurav - raised the biggest round ( army) - hired lots of best talent > lost. Pandav - raised less - had fewer superstars - acquired best CEO by leaving money (army) on the table - listened to the advice of the CEO even when it was hard > won

happy to see we and sensei karpathy converged to the mostly same RL algo configuration

the more I learned linear algebra, the more non-linear returns I got training the models

Israel Palestine conflict might need 25 textbooks for AGI or even ASI, reading this one RL book will take us far

RL is just like democracy "Democracy is the worst form of government — except for all those other forms that have been tried from time to time." - Churchill "Reinforcement Learning is terrible. It just so happens that everything we had before it is much worse." - Karpathy