Xijin Ge, Ph.D.
Professor
Department of Mathematics and Statistics
South Dakota State University
Abstract. Artificial intelligence tools are transforming research practice, yet claims of fully autonomous "AI scientists" remain largely aspirational. In this talk, I draw on my own experiences—both successes and failures—using AI throughout the bioinformatics research pipeline to argue that the most reliable gains today come not from autonomy but from agentic AI tools that augment brainstorming, literature search, writing, data analysis, and coding. I begin with RTutor.ai, one of the earliest data chatbots built for scientists, and trace its evolution into Datably.ai, an experience that revealed core lessons about how large language models succeed and fail. I then discuss DataMap, a lightweight heatmap tool built with Claude; the drafting of a LaTeX manuscript from a single prompt; and my transition to coding agents such as Claude Code and Codex. Building on this foundation, I describe using coding agents to generate research hypotheses directly from genomics data, and share what I learned while preparing several resulting manuscripts—namely, that ostensibly routine tasks like literature search and accurate citation remain surprisingly difficult for current AI systems. Finally, I recount a failed attempt to build an autonomous bioinformatics research agent: complex, multi-step tasks require long chains of LLM calls, and small errors and hallucinations compound quickly across them. From this failure I distill practical lessons—skills, MCP servers, subagents, and structured workflows—that I believe can already help researchers work more effectively today, even short of full autonomy.