←
OpenRSI Foundation52 posts
OpenRSI Foundation avatar

OpenRSI Foundation

@OpenRSI

Let RSI benefit everyone, let everyone shape RSI to push the boundaries of intelligence and science. index.openrsi.foundation/ind…

◷ Joined September 2026▧ 28 Photos and videos
Zhuofeng Li avatar
⟳ OpenRSI Foundation retweeted
Zhuofeng Li@zhuofengli96475·Nov 9

Excited to co-lead this effort! 🚀 Introducing OpenRSI × SGLang-Omni: Scaling RSI for AI Infra Our agents continuously optimize real serving code, with every validated gain. ⚡ +7.6%–326.5% throughput gains across 8 speech & omni models on H100 ✅ 14 PRs reviewed and merged by SGLang-Omni maintainers 🔁 The loop keeps running. 📝 Blog: index.openrsi.foundation/blo… 🧵👇

We scaled Recursive Self-Improvement (RSI) to production AI infrastructure with SGLang-Omni, shipping 14 upstream-merged PRs across 8 speech & omni models with 7.6%–344.3% throughput gains. Blog: index.openrsi.foundation/blo…

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Nov 9

🔁 Why it scales: RSI-Anything Humans define the problem; compute searches the long tail. Every PR was validated by an independent verifier, passed strict maintainer review, and ships with raw benchmark audits.

Post attachmentPost attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Nov 9

2️⃣ RSI drives real production progress & expands human boundaries (👉 Right) Close your laptop and go to sleep: agents found verified speedups (passing all latency & WER guards) in just 4–8 hours, compounding to 1.44×–2.20× gains in 24–48h.

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Nov 9

Two takeaways from scaling RSI-Harness in production: 1️⃣ RSI upgrades the inference engine to unleash compute limits (👈 Left) GPUs are often bottlenecked by runtime overhead rather than FLOPs. Without touching model weights, agents eliminated CPU thread and kernel bottlenecks to push H100/H200 hardware to its ceiling: Ming3-TTS: +344.3% non-streaming / +114.5% streaming (H200) MOSS-TTS: +149.5% streaming / +65.2% non-streaming (H200) CosyVoice3: +56.3% streaming (+32.5% from a single 1-line fix)

Post attachment
Zhuofeng Li avatar
⟳ OpenRSI Foundation retweeted
Zhuofeng Li@zhuofengli96475·Oct 7

🚀 Thrilled that our OpenRSI agents have delivered 3 optimizations now merged into SGLang-Omni! 🔥 Let’s scale RSI together! 👉 Try OpenRSI-Harness: github.com/OpenRSI-Foundatio…

Post attachmentPost attachment

🚀 A milestone for OpenRSI: 3 OpenRSI agent-generated optimizations have been merged upstream into SGLang-Omni! All three PRs underwent rigorous maintainer code review and verification by the SGLang-Omni team. 😉 Real code. Proven improvements. Lasting impact for the open-source community. More to come! 👏 Huge thanks to @zhuofengli96475 and @GenAI_is_real for leading this effort, and to the SGLang team @sgl_project for their support and collaboration! Try OpenRSI harness now: github.com/OpenRSI-Foundatio…

Post attachment
Guitar Cat + LLM 「Will @ SGLang Summit in SF」 avatar
⟳ OpenRSI Foundation retweeted
Guitar Cat + LLM 「Will @ SGLang Summit in SF」@GenAI_is_real·Oct 7

Really impressed by how professional and efficient this collaboration with @OpenRSI has been. The OpenRSI team showed very strong execution, and these three upstreamed optimizations also gave us a useful glimpse of what software agents can already do under a strong supervision loop. With clear tasks, isolated evaluation, maintainer review, and measurable performance targets, the agents were able to satisfy most of what we care about in both code quality and systems performance. Human supervision is still essential today. But the encouraging part is that much of this supervision can be systematized. Good documentation, explicit constraints, reproducible benchmarks, and clear acceptance criteria can dramatically reduce how often a human needs to step in. That may be one of the most important lessons here: making agents more autonomous is not only about making the model smarter. It is also about making the environment easier to supervise. We are excited to keep working with @OpenRSI and gradually move toward tasks with weaker supervision. That is where things may get really interesting.

🚀 A milestone for OpenRSI: 3 OpenRSI agent-generated optimizations have been merged upstream into SGLang-Omni! All three PRs underwent rigorous maintainer code review and verification by the SGLang-Omni team. 😉 Real code. Proven improvements. Lasting impact for the open-source community. More to come! 👏 Huge thanks to @zhuofengli96475 and @GenAI_is_real for leading this effort, and to the SGLang team @sgl_project for their support and collaboration! Try OpenRSI harness now: github.com/OpenRSI-Foundatio…

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Oct 7

Toward Fully Autonomous RSI Today: RSI in constrained environments with human guidance. Long-term Vision: Fully autonomous RSI across the entire AI research lifecycle.

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Oct 7

How OpenRSI-Harness works 🔁 Experiment → Submit → Evaluate → Improve. Agents iterate in a persistent workspace. Each submission gets an independent evaluation in a fresh, isolated environment, with feedback guiding the next round.

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Oct 7

⚙️ 3 PRs, less overhead: #2524 CosyVoice3: let idle speaker-encoder workers sleep to free CPU. #2525 MOSS-TTS: reuse same-dtype embedding accumulators to cut allocations. #2555 CosyVoice3: pass reference tensors between stages, skipping Python list conversions.

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Oct 7

🚀 A milestone for OpenRSI: 3 OpenRSI agent-generated optimizations have been merged upstream into SGLang-Omni! All three PRs underwent rigorous maintainer code review and verification by the SGLang-Omni team. 😉 Real code. Proven improvements. Lasting impact for the open-source community. More to come! 👏 Huge thanks to @zhuofengli96475 and @GenAI_is_real for leading this effort, and to the SGLang team @sgl_project for their support and collaboration! Try OpenRSI harness now: github.com/OpenRSI-Foundatio…

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Sep 23

6/ Compute Sponsor Every OpenRSI task executes in a real model-development environment on GPU clusters. If you manage compute infrastructure and want to co-build the standard for RSI evaluation, reach out. Compute partners receive formal credit on the index, paper, and every official release.

Post attachment
OpenRSI Foundation avatar
OpenRSI Foundation@OpenRSI·Sep 23

5/ Call for contributors Want to challenge frontier agents with your own research? Contribute a task. RSI-Anything, our contribution pipeline, walks you through the research question, baseline, evaluation, and budget in about 1 hour of conversation, then packages it into a runnable task. Your contribution becomes the exam. Contributors receive paper authorship on OpenRSI-Index. We are also recruiting Domain Leads to own new domains across the foundation-model stack and important verticals like medical, legal, 3D vision, and scientific ML.

Post attachment