building brains @AsariAILabs, looking for branes
San Francisco, CA
Joined December 2017
- Tweets331
- Following227
- Followers1.2K
- Likes259
Pinned Tweet
It works! Our co-inventors improve trillion-param LLM inference e2e while preserving model quality. The progress compounds: each verified gain helps find the next one faster.
It’s early: our invention loop is still running. Proud of the team making this happen!
Our self-improving agents optimized the full @vLLM_project inference stack, with up to 16% more throughput and interactivity for @deepseek_ai v4 Pro and @Zai_org GLM 5.2 on B200s (no MTP).
Every change was verified and our agents got better and faster at it with each iteration.
Great work by our team! Infrastructure powered by agents will look quite different in the near future
Our agents built a custom inference engine in just a few days that serves GLM 5.3 FP8 on @AMD MI350X, 37% faster than on NVIDIA B200, B300 (8-GPU nodes). This achieves a p90 interactivity of 352 tokens/sec/user on @SemiAnalysis_ AgentX, even matching the best result for the lower-precision FP4 model. All outputs are consistent with stock implementation. Lots of potential for this approach!
Stephan retweeted
Faster and faster: yesterday @Alibaba_Qwen’s Qwen3.8 weights came out, today our agents got +10% throughput and +11% interactivity for it with vLLM on B200s. Our co-inventors keep going…
Stephan retweeted
A little fusion goes a long way (for generalization)! For Inkling (975B params) by @thinkymachines, our agents worked at c = 32 and made fused kernels that are faster in the whole c = 1 - 128 range. Interactivity and throughput both improve by up to 18% ↗️
Stephan retweeted
Neurosymbolic AI is back. Can’t wait for recursive self-improvement to give us neuro(symbolicneuro)symbolic AI.
Our co-inventors continue to learn and get results faster: now for Kimi K3 in just a few days. We’re seeing the curves compound
Our co-inventor agents are compounding! Over the weekend, they made @Kimi_Moonshot Kimi K3 32% faster, using vLLM on B200s.
Their invention process gets better and better with each cycle: working on DeepSeek v4 Pro and GLM 5.2 helped them get results more efficiently on Kimi K3.
Stephan retweeted
This Asari approach works, imagine the improvement in inference and the reduction of data center serving costs when people adopt this! The power of agents that are incredibly smart.
Our self-improving agents optimized the full @vLLM_project inference stack, with up to 16% more throughput and interactivity for @deepseek_ai v4 Pro and @Zai_org GLM 5.2 on B200s (no MTP).
Every change was verified and our agents got better and faster at it with each iteration.
Stephan retweeted
Ever wished you could rewind your AI agent when it goes off track?
EnCompass lets you rewind any Python program containing LLM calls.
While normal program execution runs a program once from start to finish, 1/
Stephan retweeted
AI agents can be very effective when they use LLMs, but coding agents to work backwards to fix mistakes is time-consuming.
MIT method executes AI agent programs by backtracking & making multiple attempts, helping coders work w/these systems efficiently: bit.ly/45QNta0
Search and you will find... and @zli11010 made it easier for agents
h/t @MIT
news.mit.edu/2026/helping-ai…
Stephan retweeted
System design and verification for large-scale infrastructure is hard.
We chose migration of code from C to Safe Rust as a proving ground for our AI agents: libyaml, gzip, zlib, libmcs. Production code with 40,000+ total lines.
Our agents succeeded in migrating all four libraries with comprehensive test validation. 🚀
Here's what that means → asari.ai/blog/migrating-c-to…
Excited to share a big milestone: the @AsariAILabs team has built agents that migrated multiple critical C libraries to Safe Rust (@rustlang) — thoroughly tested, no shortcuts, no slop.
One example: libmcs, software that literally runs in space 🚀! Our agents met and exceeded ESA (@esa) testing standards, verified in partnership with @GTDsystems, a space-industry testing provider.
This requires thousands of reasoning steps about systems and efficiently collaborating with humans on key design decisions.
The Safe Rust code is open-source (links below) — check yourself and give us feedback!
Blog: asari.ai/blog/migrating-c-to… + more deep dives soon.
Thread 🧵!
[4/n] We’re open-sourcing the Safe Rust code so anyone can audit, verify, and contribute:
- Space math: github.com/Asari-AI/libmcs-s…
- Parsing: github.com/Asari-AI/libyaml-…
- Compression: github.com/Asari-AI/zlib-saf…, github.com/Asari-AI/gzip-saf…
Please let us know what you think — GitHub Issues and comments are very welcome!
[n/n] I’m super proud of the @AsariAILabs team that pushed our agents to this level and are moving towards engineering superintelligence.
This project offers a glimpse of how mission-critical systems will be built and invented together with AI.
We’re continuously scaling and generalizing our agents & testing with partners.
If you care about building AI that helps invent the systems of the future, come build with us — we’re hiring: jobs.ashbyhq.com/asari.ai
Stephan retweeted
Awesome coverage of our work on EnCompass with @AsariAILabs It’s a new way to program AI Agents that allows developers to capture key logic while abstracting away inference/search strategies. Can substantially decrease coding burden when designing complex agent logic.
w/ @zli11010, @StephanZheng, Armando Solar-Lezama
caltech.edu/about/news/helpi…
Stephan retweeted
Honored to have our work featured in Caltech News!
caltech.edu/about/news/helpi…
Thanks to my amazing mentors Armando Solar-Lezama, @yisongyue and @StephanZheng, and I’m grateful for @AsariAILabs for their support in bringing this project to fruition!
Excited to be presenting the EnCompass search framework at NeurIPS today! Stop by to learn how EnCompass makes LLM-calling programs more reliable.
Poster #2410 4:30-7:30pm neurips.cc/virtual/2025/loc/…
Paper: arxiv.org/abs/2512.03571
dms are open if you’d like to chat about LLM agent frameworks, code gen, or AI for math!
Agents search and improve — great work by @zli11010
👋 Our first blog post is up! We’re excited to share some of what we’ve been working on.
At @NeurIPSConf 2025, we’re presenting EnCompass: a flexible framework for AI agents that separates workflow logic from search strategy. Reduces code needed by 3-6×, makes implementing more sophisticated search strategies practical, and potentially unlocks better scaling laws.
Read the blog post here: asari.ai/blog/encompass