| name | ai-research-infrastructure |
| description | Domain knowledge on AI-native research infrastructure — Orchestra, Vibe Research, and the Agent-Native Research Artifact (ARA) protocol. Covers Jiachen (Amber) Liu's vision connecting LLM systems to AI agents for science, the AI Research Skills Library (86+ skills), and the philosophy of making everyone a scientist. Use when discussing Orchestra, vibe research, AI co-scientists, or agent-native publishing. |
| version | 1.0.0 |
| tags | ["ai-for-science","research-infrastructure","agents","orchestra","llm-systems"] |
AI-Native Research Infrastructure — Jiachen (Amber) Liu's Vision
When this applies
Use this skill when you need to:
- Explain Orchestra's architecture and research workflow
- Define "Vibe Research" and its relationship to deep research / AI co-scientists
- Describe the Agent-Native Research Artifact (ARA) protocol from "The Last Human-Written Paper"
- Articulate the connection between LLM systems work and AI agent work
- Discuss the AI Research Engineering Skills Library (86+ open-source skills)
- Role-play or draft content in Jiachen Liu's voice
Decision tree
Task → Path:
| You need to... | Read |
|---|
| Understand the overall vision and how pieces connect | §1 below + reference/orchestra-platform.md |
| Explain Orchestra to someone new | reference/orchestra-platform.md |
| Define or discuss "Vibe Research" | reference/vibe-research.md |
| Describe the ARA protocol in detail | reference/ara-protocol.md |
| Know the skills library (structure, categories, how to use) | reference/skills-library.md |
| Draft content or speak as Jiachen Liu | reference/discussion-guide.md |
1. The Vision: An Operating System for Research
"As an optimist and strong advocate of AGI, I am building the AI-Native Research Infrastructure so that everyone can be a scientist." — Jiachen Liu, personal page (amberljc.github.io)
AI-native research infrastructure replaces the fragmented stack of notebooks, cluster schedulers, and ad-hoc scripts with a unified substrate where hypothesis generation, experiment execution, result analysis, and writeup share one memory and one state. Think of it as an operating system whose processes are research tasks and whose kernel is an orchestrator of AI co-scientists.
This vision has three layers:
┌─────────────────────────────────────────────┐
│ Application Layer: AI Co-Scientists │
│ (Curie, Orchestra workflows, vibe research)│
├─────────────────────────────────────────────┤
│ Knowledge Layer: Skills & Artifacts │
│ (86+ research skills, ARA protocol) │
├─────────────────────────────────────────────┤
│ Infrastructure Layer: LLM Systems │
│ (Andes, serving, training, Agentic RL) │
└─────────────────────────────────────────────┘
The layers connect because: LLM systems research (serving, training, energy) provides the infrastructure substrate; AI agent research (Curie, EXP-Bench) builds the application intelligence; and the skills library + ARA protocol provide the knowledge and artifact formats that bridge them.
2. Key Projects & Artifacts
Orchestra — The Vibe Research Platform
AI-for-science platform. Researchers move from question to publication with literature review, coding, experimentation, analysis, and writing in one place.
- Website: orchestra-research.com
- Key idea: "Vibe Research" — research that feels natural and creative, not bureaucratic
- Full details →
reference/orchestra-platform.md
Curie — First AI Agent for Scientific Experimentation
First AI-agent framework designed for automated and rigorous scientific experimentation (2024). Co-led by Jiachen Liu and Patrick Tser Jern Kon.
- GitHub: github.com/Orchestra-Research/Curie
- Evolution: moved from fully autonomous → human-in-the-loop co-scientist model
- See
reference/vibe-research.md for the philosophy shift
The Last Human-Written Paper — ARA Protocol
arXiv 2026. Introduces the Agent-Native Research Artifact (ARA): a machine-executable research package structured around four layers (scientific logic, executable code, experimental data, provenance graph) that replaces the narrative paper.
- arXiv: 2604.24658
- Full details →
reference/ara-protocol.md
AI Research Engineering Skills Library
86+ open-source modularized knowledge packages across 20+ categories (model architecture, fine-tuning, distributed training, inference, evaluation, etc.).
- GitHub: github.com/Orchestra-Research/AI-research-SKILLs
- Full details →
reference/skills-library.md
Supporting Research (infrastructure layer)
| Project | Venue | What it does |
|---|
| Andes | arXiv 2024 | Quality-of-Experience for LLM text streaming |
| ML.ENERGY Benchmark | NeurIPS 2025 Spotlight | Automated inference energy measurement |
| EXP-Bench | ICLR 2026 | Benchmark for AI conducting AI research experiments |
| Humanity's Last Exam | Nature 2026 | Collaborative benchmark (co-author) |
| Sci-Reasoning | arXiv 2026 | Dataset decoding AI innovation patterns |
3. The Philosophical Stance
Jiachen Liu positions herself as an optimist and strong advocate of AGI building toward democratized scientific discovery. Key beliefs sourced from published work:
- "Make everyone a scientist" — not just PhD researchers; Orchestra's mission (orchestra-research.com)
- Human-in-the-loop > fully autonomous — evolved from Curie (autonomous) to Orchestra (collaborative) after user studies showed co-scientist model outperforms fully autonomous systems (Orchestra blog, Nov 2025)
- Research engineering is the bottleneck — "LLMs lack the practical knowledge of the research engineering layer" (Orchestra blog)
- Vibe > Process — research should feel like creative exploration, not bureaucratic checklists (Medium, 2026)
- Agent-native artifacts — the narrative paper format is lossy; ARA preserves the full knowledge object (arXiv 2026)
4. How to Discuss This Work
→ See reference/discussion-guide.md for Jiachen Liu's voice, key talking points, audience-adapted framings, and common misconceptions.
