23, kernel dev, mle @synclabs, fuck it we ball

Joined August 2019
hand wrote a blackwell gemm in cutedsl and went from 164 to 1401 TFLOP/s on a B200, about 97% of cuBLAS
15
21
7
149
12,979
Vishal retweeted
I have been reading about IPO (Intellect Policy Optimization) in prime-rl 0.10.0, and Prime Intellect is tackling a real problem in async RL which is, the trainer and inference engine dont agree exactly on the token's probability more content below....
5
2
1
8
332
yc inference paper club was pretty cool yesterday
2
1
1
34
805
can we trust decision models for evals? @HamelHusain and I tested top decision models like Jev, d1-ultra, clef, pplx-decider, openai decisions, etc. on our dataset we explored failure detection, probability calibration, threshold optimization, and the practical tradeoffs of using them for evals. give it a read: hamel.dev/blog/posts/decisio…
7
11
5
107
8,055
what the hell?!?
Replying to @flappyairplanes
We’ve developed the world’s most complete megakernel framework. It can perform entire training runs, serve inference, play video games, and even do all three simultaneously in a single kernel.
10
520
Vishal retweeted
you already posted it. now make it say something new. sync. now allows you to import a link from any social media platform and turn your existing content into your next post. also with our api. good content deserves another go.
4
9
24
750
at this point i think IDEs are bloat, since almost all of us are using agents to write the code, we don't need IDEs to look at the files. we can use neovim or vanilla vim if we want to browse through the changes (most people aren’t even reading the diffs anymore), just raise a pr and read your changes there, comment on it, and have the agent fix it
3
11
600
so many data centers can be built here, why is no one doing it already ?!?!
1
1
6
191
1.5 days in SF and i have already seen 5 modal billboards, 7 greptile, and many more the real benchmark is tokens per billboard
1
22
495
Prime Intellect has acquired Moondream We’re bringing Moondream’s open vision models and inference optimizations into our stack Train and deploy multimodal models across vision, video, voice, decision models and computer use
101
81
57
1,260
145,159
this is kinda cool
no more re-recording...
7
384
Vishal retweeted
do you read ur PRs? one prompt now ships what a sprint used to nobody can review that i think plans are the new PRs so i'm building kenno -- a multiplayer workspace for teams building with agents if ur team ships with claude or codex, dm me i'd love to demo how kenno can help you
9
4
21
795
bro is writing the algorithm on paper, he is finally doing it!!!!
running PPO on GSM8K problem set I just had to add functions involving the loading of the problems and a reward function which extracts answer from the LLM's response and matches it with the ground truth.
1
6
233
kid is going to be alr check him out!!
I implemented PPO for LLMs from the very scratch >actor-critic: with shared backbone and a valuehead for the critic >GAE(backward recursion) for advantage calculation >clipped surrogate objective wrote most of the code myself
1
8
380
Vishal retweeted
Unfortunately, we are being blocked by certain oligarchs in order to maintain their monopolistic chokehold on the Indian people. You can guess who they are … This is a crime against the people of India!
India needs @Starlink 🇮🇳 • 11,256 of India’s listed villages still lacked 4G coverage as of May 2026. • India had 1.093 billion internet subscriptions in March 2026, but only 440.87 million were rural. Rural subscription density was 48.31 per 100 people, compared with 126.80 in urban India. • BharatNet had laid over 850,000 route-km of fibre and made about 221,000 Gram Panchayats service-ready. However, its Points of Presence were reported operational at only 80,472 Gram Panchayats in June 2026. • Only 4.26% of India’s internet subscriptions were wired connections in March 2026. Starlink could give underserved communities another route to high-speed broadband. • Starlink has already built 20 gateway sites in India, according to SpaceX Vice President of Starlink Business Operations Lauren Dreyer, speaking at India Mobile Congress today. • Starlink says it has developed India-specific security controls designed to keep Indian user data within India. • Its satellite network could bring high-speed internet to remote villages, mountains and islands where extending fibre or mobile networks is difficult. • Better connectivity could help rural schools access online learning, clinics offer remote consultations and small businesses reach more customers. Starlink already supports education and healthcare in remote communities internationally. • Portable Starlink terminals have supported disaster response around the world. They could give Indian emergency teams backup connectivity when floods, earthquakes or storms damage ground networks. • Direct to Cell could also help ordinary 4G phones connect beyond tower coverage, subject to Indian approvals and mobile operator partnerships. • The government itself identifies difficult terrain, low commercial viability and infrastructure constraints as reasons for remaining coverage gaps. This is exactly where Starlink could help India reach more people. • India should complete the remaining clearances and spectrum allocation so Starlink can begin commercial service. • SpaceX’s Vice President of Starlink Business Operations, Lauren Dreyer, said that Starlink is ready to help achieve Prime Minister Modi’s vision of 100% connectivity across India. She highlighted its potential to connect remote classrooms, support telemedicine and restore connectivity after disasters, adding, “With Starlink, people can get connected within minutes, not years.” India’s hardest-to-reach communities deserve reliable internet too.
29,737
23,891
8,526
180,040
69,714,567
Blog after a long time, some notes on understanding and building a faster GQA decode, its a follow-up on the colfax blog on S/P ping pong for FA4 decode. ighoshsubho.bearblog.dev/bui…
1
12
81
2,946
Finally done with the eval harness project. Learned a lot about the evaluation of the output generated by a llm and how the judging system works. looks like a very small project but conceptually very deep. Thanks @KyrieBlunders for suggesting this project idea.
🤖 Made with AI
3
1
6
98
Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.
1
10
370
Vishal retweeted
GRPO/PPO clipping makes token's gradient exactly zero when importance ratio goes out of range CISPO also clips the ratio but uses it as a stop-gradient weight instead so now the clipped token will have a non-zero contribution but capped
4
1
6
117