Orchestra Reseach

We're living in the BEST era for doing research. 💪 After I graduated from my PhD, the rise of AI-native research gave me a new chance to revisit my research experience. Lately, doing research feels incredibly rewarding to me. I get to experience the pure joy of curiosity-driven science because I no longer have to worry about the lower-level implementations or getting bogged down by infrastructure 🚀 (I'll be sharing some of my own recent research driven by this very soon!) But today, let me introduce the New Orchestra 🎻. We wanted to ship a product that absorbs the friction and brings science back to the curiosity.

@ZechenZhang5 @orch_research I tried this out after the demo at OpenClaw Boston meetup. Its great for finding related research and structure your project. This is auto-research-management vs. auto-research-a-la-karpathy . Great work @ZechenZhang5

Every developer has an IDE. Researchers never had one. We don’t need another research agent, but a full stack workstation powered by AI. Today we're launching Orchestra @orch_research, the world's first Research IDE 🔬🧵

I have been exploring autoresearch (based on @karpathy 's script) recently, via @openclaw , @claudeai Code, and online platforms like @orch_research. But I have gotten some frustrations. 1. 🖥️ Private infra. Instead of paying for extra computation resources, I prefer to use the existing resources that I have paid. 2. ⏲️Long-running experiments with extremely unstable infra. Unfortunately, the connection with my infra is not 'ssh' based. Hard to maintain a stable connection. I hacked it through a WebSocket. But a lot of frustrations in handling the running and monitoring of experiments. 3. 🔒Private data. Yeah, robot, stay away from my privacy-sensitive data. To make AI do research without sending data to a third party, I have to manually make up some fake data. Kind friends. Let me know if there is a system that can help auto my research.

yeah we have the skills ready github.com/Orchestra-Rese…

Introducing #autoresearch skill, the number 0 skill on our AI-research-skills collection. github.com/Orchestra-Rese…

ML Systems researchers, we got you📖 Our open-source AI Research Skills library now ships LaTeX templates for OSDI, NSDI, ASPLOS, and SOSP — complete with venue-specific review criteria, submission checklists in addition to ML conferences github.com/Orchestra-Rese…

If you like our AI research skills and want to learn more about how skills can be evolved via agent collaborations @skill_evolve , register for the skill hackathon at @agihouse_org . There are amazing speakers and prizes lined up! app.agihouse.org/events/Agent-S… Co-hosted by @xdotli @JIACHENLIU8 @belindmo

Our AI Research Skills have crossed 4k stars! Thank you supporters who find it useful. We will work hard to bring it more useful to people to do AI research more smoothly and seemlessly. Please let us know if you have any suggestions/feedbacks. github.com/Orchestra-Rese…

Quietly shipping V2. Waitlist is open. We will progressively admit new users ⏰ Let’s vibe researching… move at the speed of thought 💭 orchestra-research.com

V2 just shipped 🚀 An end-to-end research workflow — designed for deep human × AI collaboration. Join the waitlist. Let’s do Vibe Research together. orchestra-research.com

这个项目好像有意思？ 它把大模型预训练的知识都写成skills了， 这意味着初中辍学生，都可以做自己的预训练。 搞大模型根本不需要博士学位.... github.com/Orchestra-Rese…

thanks for the support!

Last year, we tried to build an research agent that could run AI research experiments, and we realized LLMs lack the practical knowledge of the research engineering layer 🔧 — how to configure Megatron for distributed training, how to run RLHF with TRL, how to quantize models without breaking them.. So we started documenting skills for every ML framework, every tool, and open-sourced all of them. Now it's 83 skills across 20 categories: → Distributed training (@DeepSpeedAI , FSDP, Megatron-LM) → Inference & optimization (@vllm_project , TensorRT-LLM, @sgl_project ) → Post-training & RLHF (VeRL , OpenRLHF, TRL) → Agents & RAG (@LangChain , @llama_index ,FAISS , @qdrant_engine ) → Writing AI research papers (LaTeX templates, citation verification) One command installs them into any coding agent — Claude Code, Codex, Gemini CLI, Cursor: 𝚗̲𝚙̲𝚡̲ ̲@̲𝚘̲𝚛̲𝚌̲𝚑̲𝚎̲𝚜̲𝚝̲𝚛̲𝚊̲–̲𝚛̲𝚎̲𝚜̲𝚎̲𝚊̲𝚛̲𝚌̲𝚑̲/̲𝚊̲𝚒̲–̲𝚛̲𝚎̲𝚜̲𝚎̲𝚊̲𝚛̲𝚌̲𝚑̲–̲𝚜̲𝚔̲𝚒̲𝚕̲𝚕̲𝚜̲ Do AI research by prompting with your hypothesis, not debugging infrastructure. github.com/Orchestra-Rese…

The best part about having AI research skills? turning a random shower thought into a real experiment is stupidly easy... This is a cute demo that answers my random curiosity: when quantizing an LLM, does it matter WHERE you place the lower-precision layers? Early layers? Late layers? Attention vs FFN? So I just ask my agent (equipped with llama.cpp skill) to run it on llama.cpp - 10 quantization strategies on Qwen2-0.5B. With Zero lines of code written by me, I found → Early layers are the most fragile to quantization compared to later → FFN weights benefit more from higher precision than attention weights → ....

We made a one-click installation of all the AI research skills. You can install them all or on-demand. Really excited to hear about your use cases and feedbacks. npx@orchestra-research/ai-research-skills

We have bitsized lessons for you. Try reproducing them by chatting with Orchestra

Update: we have since implemented a more efficient work around for resource allocation, and all agents are back to normal! We are still working on scaling up in the next few days and agents might be pending for sometime dependent on user traffic.

We’ve received more traffic than expected and have temporarily hit our compute limits. We’re actively working on scaling things up — thanks for your patience. so it eventually collapse into a single, ancient problem: the allocation of scarce resources 🤣😂🤣