saurabhtwq's profile picture. ml research. inference. agents

saurabh

@saurabhtwq

ml research. inference. agents

idc what do you think but f1 is fun cause there is one driver with daddy issues who don’t give a fuck about anything other than winning.


wow gemini 3 is so smart it rewrite my api keys into two random words.


gemini 3

saurabhtwq's tweet image. gemini 3

i would like to see live streaming of podcasts.


wait lemme flex more if i have more work then i can sleep at 10pm and wake up at 4am.

i feel so left out man cause i have habit of waking up early in the morning.



i feel so left out man cause i have habit of waking up early in the morning.


how these double-faced people (some of them) talk about racism when it happens from outsiders (like from whites), but the sad story is they never talk about the racism that happens in India by Indians based on various factors. losers!!! how the fuck can we even lecture other…


got something good stuff to read !!!

publishing the work i did for the MATS program. Unfortunately, I wasn't selected this year. But it was a great learning experience :) I tried to explore why SLMs hallucinates with simple instructions via mech interp. This is also my first work in mech interp :) 🧵1/n

angkul07's tweet image. publishing the work i did for the MATS program. Unfortunately, I wasn't selected this year. But it was a great learning experience :)

I tried to explore why SLMs hallucinates with simple instructions via mech interp.

This is also my first work in mech interp  :)

🧵1/n


twitter, can you stop this? wtf, why am I getting political content on my feed? It’s either rahul or modi. you both are pieces of shit, inhuman, evil, and you do not value human lives. both parties are supposed to make this country a better place to live, and both have fucked up…


make it so you don’t have to fake it.


only way out is through.


saurabh reposteó

new blogpost after a long time! in this series i will talk about how to solve reinforcement learning for long-horizon tasks, incrementally from the most straightforward approaches. (link in replies!) in part I of this series, we throw RL at the cube in its most direct,…

snowclipsed's tweet image. new blogpost after a long time! in this series i will talk about how to solve reinforcement learning for long-horizon tasks, incrementally from the most straightforward approaches. (link in replies!)

in part I of this series, we throw RL at the cube in its most direct,…

startup idea : just create ai solution first then find a problem then raise vc money and burn as much as you can and go bankrupt.


omg

Deal (on the former). Let's just do pretraining, fastest run wins. Who wants to referee this / hold the escrow?



my yt feed is all about ml compilers and chris lattner talking why should i learn mojo.


told you read more code!

our mental model for why codebase understanding is valuable and why vibe coding has limits to scale is expressed by this chart: manual understanding is the bottleneck. Codemaps understanding scales with model intelligence x.com/windsurf/statu…

cognition's tweet image. our mental model for why codebase understanding is valuable and why vibe coding has limits to scale is expressed by this chart: 

manual understanding is the bottleneck. Codemaps understanding scales with model intelligence

x.com/windsurf/statu…


just train a resnet bro!

Stanford just dropped a 1:47 hour masterclass about LLM training on YouTube. this is the 4th lecture in their new Fall 2025 series of transformers and LLM. this is the new go-to course on LLMs imo, and you need to take it seriously. subjects in this video: → Pretraining →…

Hesamation's tweet image. Stanford just dropped a 1:47 hour masterclass about LLM training on YouTube. this is the 4th lecture in their new Fall 2025 series of transformers and LLM.

this is the new go-to course on LLMs imo, and you need to take it seriously. 

subjects in this video:
→ Pretraining 
→…


Loading...

Something went wrong.


Something went wrong.