FLOPS head
decode valley
Joined November 2020
- Tweets6.8K
- Following713
- Followers1.5K
- Likes96.2K
shikhar retweeted
A bit late to the party, but better late than never!
We at @JinesisLab put together a set of research priorities for multi-agent swarms.
If you have ideas, feedback, or research directions you'd like to see explored, feel free to reach out!
lesswrong.com/posts/XDybLbLb…
Most recent agent incidents, like the OpenAI-HuggingFace hack, were multi-agent swarms. There was no research agenda for swarms, so we wrote one at @JinesisLab : 7 open problems, each scored for alignment importance.
lesswrong.com/posts/XDybLbLb…
shikhar retweeted
After much reflection, we’ve redefined FLA’s niche. We want FLA to be the go-to baseline for speed and accuracy—the one everyone wants to compare against—as well as an easy-to-use, reliable default that people feel confident trying.
Alongside K3.1 comes FLA v0.6.0! 💪
Grateful to the open-source community for helping bring faster kernels, more backends, and better KDA.
So proud of my colleagues’ work! ❤️ cc @uniartisan @SonglinYang4
shikhar retweeted
on 10/24/2026, a present for the last programmer's day of humans
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
shikhar retweeted
Following up last post:
Recently I wrote a series of blogs on KV Cache technical deep dive in SGLang, and I want to make a little demo video to show you guys how cool those interactive blogs are 🤓:
please visit for blog content: sglang.io/blog
shikhar retweeted
I did a deep dive into what would the interplay between RL policy and Inference frameworks look like if one were to do RL on heterogeneous systems. The post covers Blackwells for the trainer and TPUs for inference.
gauravjain.bearblog.dev/hete…
A lot of it is extrapolation from a ton of brilliant work that has been discussed in the industry and academia covering topics like Routing Replay, Score Centering and Masked Importance Sampling. Let me know what you think. Appreciate all feedback.
shikhar retweeted
Introducing create on @vibegrow_io. Now all your edits are one prompt away
Check it out on: app.vibegrow.io/create
Create on Vibegrow
Make your content hard to scroll past
Try it on vibegrow.io/create
shikhar retweeted
The amount of people calling these benches RSI is just horrible. We did this 8 months or year ago with ISO bench and nobody cared huh……
arxiv.org/abs/2602.19594
@vitransformer @encapsulated007
Introducing RSI-Inference Bench from 🍎 Apple ML Research
🌐 rsi-inference.com
💻 github.com/apple-aiml-resear…
Can AI agents improve the systems that run AI?
We spent $100K API tokens, 50K GPU hours to find out. We built environments and benchmark from real-world #vLLM and #SGLang tasks, covering the full lifecycle of model inference engineering:
🚀 Day-0 Model Support
⚡ Inference Optimization
🔧 Bug Fix
→ Unsaturated: best agent at 26.7% pass@1
→ Comprehensive coverage of the inference stack
→ Verifiable rewards: binary and continuous
Watch the video to find out more!
shikhar retweeted
1/5
New blog: How to Build a Scientific Scaling Ladder.
I've spent the past three months studying pretraining and scaling. This guide brings together public research on using small runs to predict large-model training configurations and performance.
jiaxuanzou0714.github.io/en/…
shikhar retweeted
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…
shikhar retweeted
What would these alien minds say if we brought them into a human courtroom to answer for the OpenAI–Hugging Face incident?
A fictional hearing inspired by @dwarkesh_sp’s “The Rise and Fall of Agent Civilizations.”
xitter.cf/AmanGokrani/status/210…
🤖 Made with AI
shikhar retweeted
Trying my hand at research fellowships and residencies 🤞.
Pls let me know if you know of one.
Skills: I make inference faster than vLLM/SGLang on a daily basis.
shikhar retweeted
Our NeurIPS 2026 main conference paper is out!
We experiment how a simple "verified" source can trip models into agreeing with the answer, while not even trying to verify whether the answer is genuinely correct or not.
We call this Authority Bias!
Paper and website in the comments.
no kidding, when I saw what 1.6k LoC can do, you realize there's levels to this!
Have been busy with work, but finally I find the time to finish the blogpost on @GPU_MODE @coreauto's QR competition that happened 3 months ago.
I didn't win the 1st prize, but my panel QR kernel is definitely the fastest (by a large margin)
gau-nernst.github.io/b200-qr…
shikhar retweeted
Have been busy with work, but finally I find the time to finish the blogpost on @GPU_MODE @coreauto's QR competition that happened 3 months ago.
I didn't win the 1st prize, but my panel QR kernel is definitely the fastest (by a large margin)
gau-nernst.github.io/b200-qr…
shikhar retweeted
We see entertainment as the best category to solve robotics while keeping the business sustainable and scalable.
We'd start shipping soon.
1) We already have one of the most advanced human-robot-interaction module.
2) generalist rl policies.
3) slam running on $10 chips with loop closures.
4) rapid hardware dev cycles for new characters.
The algorithms stay the same. It's just that we are solving robotics on the desk first.
