Research articles and insights on ML systems, LLM serving, AI agents for science, and machine learning research by Jiachen (Amber) Liu.
The true speed limit of science isn't the brilliance of the individual scientist — it's the ecosystem they work within. The second half of AI for Science is about rebuilding the network where human and artificial intelligence compound.
When neither the author nor the audience is human, the three-century-old paper format stops making sense. Agent-Native Research Artifacts restructure the paper into a machine-executable knowledge package.
A new wave of citizen scientists is coming, and it will dwarf anything history has seen. As AI collapses the execution layer of science, research stops being a career and becomes a capability.
Why fully automated research tools can be dangerously misleading, and what responsible AI-assisted research actually looks like.
Why the future of AI agents is not more agents, but better skills — modular, composable capabilities that agents can learn and share.
Exploring the paradigm of agents that evolve by sharing and learning skills from each other, moving beyond static agent architectures.
Analyzing the patterns and traits that drive breakthrough AI research, from problem selection to execution strategies.
How AI is transforming research across literature review, brainstorming, and experimentation. Why people with taste and domain knowledge benefit most from the AI-native research paradigm.
Practical advice for navigating machine learning systems research — from finding impactful problems to building systems that actually work.
The essential engineering skills needed for modern AI research — bridging the gap between research ideas and working systems.
How a GPU allocation problem in an experiment agent led to rethinking compute infrastructure for the age of AI agents.
Exploring the shift from chat-based LLM training to agentic RL systems. Why existing RL frameworks fall short for multi-step agent tasks and how a decoupled Agent Layer solves key challenges.
PhD lessons and reflections on building user-centric machine learning systems. Insights from research in ML systems, LLM serving, and federated learning at the University of Michigan.
Introducing EXP-Bench, a benchmark to evaluate AI agents' ability to conduct real AI research experiments — from setting up environments to running and interpreting results.
Introducing Curie, an AI agent framework that automates scientific experimentation — from hypothesis generation to experiment execution and result interpretation.
A comprehensive guide to unlocking LLM capabilities beyond pre-training through fine-tuning, RAG, and other advanced techniques.