Soumil Bhandari

18, building cool stuff. AI compression @ Perforated AI, FSAE electronics @Purdue. All views are my own.

[Soumil Bhandari @SoumilBhandari]

@grok how to make an llm that is specific to Rust

[Soumil Bhandari @SoumilBhandari]

I'm going to train my own model that's genuinely strong at Rust.

The plan is to start with a small open model, fine-tune it on Rust code, and then use the compiler as a reward signal so it learns from its own errors.

If you've worked on something similar, I'd love to hear your tips. I'll share progress as I go.

@grok I need tips

[Soumil Bhandari @SoumilBhandari]

@Stone_Tao agreed, complex tasks hide where the policy is actually failing. simple ones you can actually go ahead and debug tell you way more about what's going on

[Soumil Bhandari @SoumilBhandari]

@kimmonismus a pure base model with no post training is honestly so rare now. 14b active is nice for serving, the hard part is still fitting 280b of weights somewhere lol

[Soumil Bhandari @SoumilBhandari]

@MaziyarPanahi a 4b beating the big model on its own task in 2 hours is basically the whole argument for small models lol. i found something similar, a model trained small beat a bigger one i shrank to the same size by a ton

[Soumil Bhandari @SoumilBhandari]

@techNmak yeah i saw this running qwen3-8b on my macbook. decode is totally bandwidth bound, 16 bit did 16 tok/s and 4 bit did 55. same model, just reading way less memory per token

[Soumil Bhandari @SoumilBhandari]

@shibatashi_ zero shot on a real g1 from just egocentric depth is so cool. how does it handle someone moving their hand last second on a handshake? feels like that's where it'd get awkward lol

[Soumil Bhandari @SoumilBhandari]

@frankzydou 72 min to 3.4 min is insane. so is most of that from implementation stuff like vectorized envs, or is the algorithm actually doing that much less work?

[Soumil Bhandari @SoumilBhandari]

@ClementDelangue yeah 100%. i'm a student and so much of what i've learned in ml and robotics came from stuff people just put out there for free

[Soumil Bhandari @SoumilBhandari]

@0xKrosh command delay is the one people always forget. you randomize mass and friction and then the real robot has like 20ms of latency and the policy just falls apart lol

[Soumil Bhandari @SoumilBhandari]

@leap_71 nine on one build plate is wild. so do they each get hot fired or do you go ahead and test a sample from the batch? feels like print variation would matter a lot for thrusters

[Soumil Bhandari @SoumilBhandari]

@seeedstudio @github seven resistors and no rf parts is crazy lol. i'm doing telemetry boards for our fsae car so this is super cool to see. what's the range like with just the onboard antenna?

[Soumil Bhandari @SoumilBhandari]

@IlirAliu_ lightweight sims are so underrated for rl. you don't need photoreal stuff for nav, you just need like a million fast rollouts lol

[Soumil Bhandari @SoumilBhandari]

@CyberRobooo shipyard welding is actually a great first job for humanoids bc those compartments are cramped and awful for people. curious how they do seam tracking with all the sparks messing with the cameras

[Soumil Bhandari @SoumilBhandari]

@NVIDIARobotics the 120 task eval in isaac lab is honestly the part i'm most excited about. everyone reports on like 3 tasks they picked, so a shared benchmark is gonna keep people honest

[Soumil Bhandari @SoumilBhandari]

@c0ngwang wait so it just keeps a second kernel warm and swaps over? that's so clever lol. how do you handle state like open sockets though, or does the service just reconnect

[Soumil Bhandari @SoumilBhandari]

@roboticomarket 95% across paid tasks is solid but the last 5% is kinda the whole game in a factory right? so when it fails does a person have to go ahead and walk over, or can it recover on its own

[Soumil Bhandari @SoumilBhandari]

@SemiAnalysis_ so inference is obviously where this makes the most sense right? training leans on the whole cuda ecosystem but serving one fixed model is a way narrower target, so tpus can just win on cost per token

[Soumil Bhandari @SoumilBhandari]

@himanshustwts testing on a ladder is so underrated. so many tricks look amazing at small scale and then just disappear at the next size up lol

[Soumil Bhandari @SoumilBhandari]

@nvidiaomniverse @huggingface the physics properties are the part that actually matters here. geometry is easy to get, it's the mass and friction being right that makes sim to real go ahead and transfer