Riya Patel

@shivanijpatel Red flag

he's a 10 but he only runs 1 claude code session at a time

Pinterest is the new vs code

@OmarHayat0 @OfficialLoganK Yeah looks like the form is only open for US schools

@riyapee @OfficialLoganK They took it away? I got it a few months ago

Can Canadian students get access to a year of free Gemini pro too @OfficialLoganK

@ErikKaum @puffer_ai yes! code is here github.com/riyapatel25/Pu…

@riyapee @puffer_ai This is super cool 👀 are you planning open sourcing/ releasing the model?

A 766K param model with RL outperforms Opus 4.6 on 8 bit games. I put 4 agents into a Pico Park emulation for 30 minutes. 500 million frames later, they’ve mastered cooperation and can consistently win the game. Play alongside my agents in the blog below! Trained with @puffer_ai

@dhruvbhatia0 Can the agents watch these PR videos and create a verifiable loop ?

static testing is useless in 2026 Our log inspector had a bug in the PR with the bugfix, Glance realized its account didnt have logs to test. it organically went to the playground and used it to generate logs -> properly tested the PR

@morphllm 🫡

great work @riyapee and thanks for making good use of our GPUs if you're doing cool work with multi agent RL + inference acceleration and want free compute for it, dm us

@Samhanknr @puffer_ai yup! code here : github.com/riyapatel25/Pu…

@riyapee @puffer_ai Super cool. Have you open sourced the code ?

@Anishfishhh @puffer_ai I made the game, this wasn’t the og game so was easy to expose the game state

@riyapee @puffer_ai ooo, was it easy to get the game data? is it just available in an env or did you have to actually scrape/monitor it yourself?

@Anishfishhh @puffer_ai No not visual I don’t feed the pixel values in. I have access to game data so map all objects in screen which is the input to the cnn

@riyapee @puffer_ai how'd you get the grid input? Is it purely visual -> cnn or do you have access to game data?

@silennai @puffer_ai yup, beat opus on cost to train and final performance

@riyapee @puffer_ai Great blog! I assume it beats Opus on cost and time?

@shivanijpatel @puffer_ai they do, their names are written under them

@riyapee @puffer_ai request: can they each have cute names

@puffer_ai The model is trained with PPO as the core algorithm using actor-critic architecture. The encoder uses both a CNN for the grid input to keep the spatial information and an MLP for the self data vector.

@puffer_ai trained on a single GH200, ty to @morphllm for the compute and ty @jsuarez for Pufferlib

the type of comeback im gonna make

@dhruvbhatia0 biebian support coming soon

@riyapee Could not repro on my OS (Biebian). please help. biebian.sourceforge.net

find out if your ngmi or cracked uvx cursor-wrapped to get your Cursor 2025 Wrapped🎁

@nico_jeannen uvx cursor-wrapped to see your stats x.com/riyapee/status…

Any cheaper alternative to Cursor?? $1400 in a month is absolute madness, no matter the amount of code 💀

@jia_seed lemme see yours

@riyapee nice