drubinstein

Making models go brrr | Engineering @reflection_ai | Occasional PufferLib contributor

[drubinstein @dsrubinstein]

Excited to finally share our progress in developing a reinforcement learning system to beat Pokémon Red. Our system successfully completes the game using a policy under 10M parameters, PPO, and a few novel techniques.

Blog posted below

[drubinstein @dsrubinstein]

@deforestpeg Trust me. It works :)

[drubinstein @dsrubinstein]

Tried to turn it into a skill https://t.co/U1GU5tXTog

Remember to measure first, optimize second.

[drubinstein @dsrubinstein]

Teaching my Claude sessions how to benchmark and profile has been a great help recently.

[drubinstein @dsrubinstein]

My current use case is to profile speedruns for a tool I'm working on. Profiling has helped find inefficient navigation subroutines. It's great.

[drubinstein @dsrubinstein]

@DanAdvantage Nothing to do with Reflection. This work is entirely my own.

[drubinstein @dsrubinstein]

A few years ago, I wanted to find a way to maximize shade during my long runs in the summer. Recently, I decided to experiment with a solution. Presenting Shady Route Finder.

[drubinstein @dsrubinstein]

https://t.co/bFxtrDpTnq . Supports 8 cities, can tell you what side of the street to walk on, mobile and even has a sunny route finder mode on desktop. I've tested some routes, but obviously not all. Any feedback would be great.

[drubinstein @dsrubinstein]

@yacinelearning It's like 30% the way there. Mostly working on a new harness

[drubinstein @dsrubinstein]

Move over nvidia-smi, ibstat is my new best friend

[drubinstein @dsrubinstein]

I dared him to try an ai assisted native rewrite. As expected, 10k sps at best has now become 4M sps. Nice.

[Dan Advantage @DanAdvantage]

i did start with a rudimentary implementation of pokemon stemming from a native rewrite of pokemon firered. the starting point i used gets around 4,000,000 steps per second as an rl env.

here is the entire prompt (caution: long!!!):

[drubinstein @dsrubinstein]

대박! A year ago we announced our series A. Today we’re announcing an amazing partnership with Shinsegae. Who knows what’ll come next?

[Reflection @reflection_ai]

Reflection is partnering with Shinsegae Group to build a 250-megawatt sovereign AI factory for the Republic of Korea.

Open intelligence. Built on trust between allies. Owned by the nations that need it most.

The future of sovereign AI. Read more in the @WSJ.

[drubinstein @dsrubinstein]

Underrated: Letting a coding agent run when you're in meetings.

[drubinstein @dsrubinstein]

@kywch500 @jsuarez Read to learn more
https://t.co/Db8Jwwx1f6

[drubinstein @dsrubinstein]

Had some fun helping out @kywch500 and @jsuarez simplifying Pufferlib's 2048 env the last couple of weeks. 2x better results with fewer observations, rewards and a new model architecture!

[drubinstein @dsrubinstein]

@TheSayakMondal More or less. Getting to 128k is literally that.

[drubinstein @dsrubinstein]

2048 is an interesting RL env. It can take over 20k steps to get to 65536.

[drubinstein @dsrubinstein]

@aturker01 Any idea what a benchmark for backward would look like?

[drubinstein @dsrubinstein]

@brandondamos @reflection_ai Welcome to the team! Super excited to have you here!